text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: |
Consider a stationary renewal point process on the real line and divide each of the segments it defines in a proportion given by [i.i.d. ]{} realisations of a fixed distribution $G$ supported by \[0,1\]. We ask ourselves for which interpoint distribution $F$ and which division distributions $G$, the division points is again a renewal process with the same $F$? An evident case is that of degenerate $F$ and $G$. Interestingly, the only other possibility is when $F$ is Gamma and $G$ is Beta with related parameters. In particular, the division points of a Poisson process is again Poisson, if the division distribution is Beta: ${\mathrm{B}}(r,1-r)$ for some $0<r<1$.
We show a similar behaviour of random exchange models when a countable number of ‘agents’ exchange randomly distributed parts of their ‘masses’ with neighbours. More generally, a Dirichlet distribution arises in these models as a fixed point distribution preserving independence of the masses at each step. We also show that for each $G$ there is a unique attractor, a distribution of the infinite sequence of masses, which is a fixed point of the random exchange and to which iterations of a non-equilibrium configuration of masses converge weakly. In particular, iteratively applying ${\mathrm{B}}(r,1-r)$-divisions to a realisation of any renewal process with finite second moment of $F$ yields a Poisson process of the same intensity in the limit.
**Keywords:** renewal process, neighbour-dependent shifts, adjustment process, random exchange, fixed point, Poisson process, Gamma distribution, Dirichlet distribution, random operator, attractor
**AMS 2010 Subject Classification.** Primary: 60G55, Secondary: 60B12, 60D05
author:
- 'Anton Muratov[^1]'
- Sergei Zuyev
title: 'Neighbour-dependent point shifts and random exchange models: invariance and attractors'
---
Introduction {#sec:introduction}
============
A Poisson point process is one of the fundamental objects in Probability. Despite being one of the simplest to define and one of the most-studied models, recent developments in stochastic calculus, stochastic geometry, differential geometry of configuration spaces, variational analysis on measures, continue to bring new insights and deepen our understanding of this seemingly elementary concept, [see, e.g., ]{}[@PecRei:15] and the references therein. It is hard to underestimate the usefulness of the Poisson process in applications which are due to its appealing properties often enabling to produce mathematically tractable models. One of such fundamental properties of a Poisson process is that whenever we apply independent random shifts to its points, the resulting process is again Poisson. In particular, if the shifts are i.i.d., the result of such a transformation of a homogeneous Poisson process is again a homogeneous Poisson process of the same intensity.
The fact that random independent shifts preserve the Poisson process distribution reflects ‘independence’ of its points. However, transformations which depend on two or more neighbouring points destroy this independence and hence, as one may naturally think, the Poisson process. For instance, the mid-points of the consecutive segments in a homogeneous Poisson process on the line do *not* form a Poisson process, as it can be easily checked. If one considers again the midpoints of this process, this second iteration of the initial process actually corresponds to the transformation by which the points move to the centres of their one-dimensional Voronoi cells (which are also their centres of gravity). Such a transformation, known as the *adjustment process*, can be defined in any dimension with respect to any metric and it forms the basis of a popular Lloyd’s algorithm in computational geometry. The adjustment process is used to model behaviour of repulsing particles or animals, [see, e.g., ]{}[@OBSC:00 Chap. 7.3.2], Lloyd’s algorithm and its variations is widely used in image compression and optimisation, [see, e.g., ]{}[@DFG:99] and the references therein. It can be shown that the variance of the inter-point distances in one-dimensional adjustment model on any compact set vanishes, so iterations of the adjustment procedure converge (in a suitable sense) to a regular array of points, see [@HasTan:76]. A similar phenomenon of convergence to a lattice is observed in a multi-dimensional case, although questions of the uniqueness of the limiting configuration still remain, see [@DEJ:06].
If dividing the consecutive segments in the Poisson process in half (as well as in any other given non-random proportion) produces a non-Poisson process, an interesting question raises: is there a way to divide each of the segments independently of the others in a *random* proportion so that the division points still form a Poisson process?
Somewhat surprisingly, the answer is *yes*: the segments must be divided in Beta-distributed proportions with parameters $(r,1-r)$ for some $r\in(0,1)$ and this is *the only* class of division distributions that preserves the Poisson process!
More generally, in Section \[sec:2\] we consider a stationary renewal point process, the segments between its consecutive points having lengths drawn independently from a distribution $F$. Divide its every segment in a random proportion independently drawn from a given distribution $G$ on $[0,1]$. One may think of the processes of the division points as a transformation of the original point process by which its every point moves to the closest division point on its right, for instance. Since the distribution of the division points depends on the distance to the neighbouring point to the right only, we call such transformation *neighbour-dependent shifts*. We show in Section \[sec:2\] that Beta-distributed shifts are the only non-trivial ones which preserve a renewal process and that a renewal process preserved by neighbour-dependent shifts is necessarily a process with Gamma-distributed segments. Poisson process with exponential segment lengths provides an example.
Since the resulting process has interpoint segments composed each of pieces of two original segments, one may naturally consider general compositions involving more than two of these. To this end, in Section \[sec:exchange\] we establish correspondence of the neighbour-dependent shifts model to the so-called random exchange process which allows for such a composition interpretation. In this process a set of ‘agents’ simultaneously exchange their ‘masses’ with the neighbours in randomly drawn proportions. We show that independent gamma distributed initial masses are preserved by Dirichlet-distributed proportions. Moreover, we prove that the iterations of the exchange process starting from any stationary sequence of masses with a finite second moment weakly converge to a limit which is a sequence of independent Gamma-distributed masses in the case of Dirichlet-distributed proportions. To our knowledge, so far only the exchange models with a finite number of agents were studied in the literature, the main tool here is analysis of convergence of the product of random matrices, see [@McKin:14] and the references therein. The rôle of the agents and masses in our models play the segments and their lengths, respectively, so we have to deal with the product of infinitely-dimensional linear operators instead. Although the results of Section \[sec:exchange\] are somewhat reminiscent of the finite case, the behaviour in this non-compact framework is rather different.
Throughout this paper, $\Gamma(a,\gamma)$ denotes the Gamma distribution with shape parameter $\alpha$ and rate parameter $\gamma$, its density is given by $$f_\Gamma(x)=\frac{\gamma^\alpha}{\Gamma(\alpha)}x^{\alpha-1}e^{-\gamma
x},\ x>0,$$ and ${\mathrm{B}}(\alpha,\beta)$ is the Beta distribution with density $$f_{\mathrm{B}}(x)=\frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)}
x^{\alpha-1}(1-x)^{\beta-1}, x\in(0,1).$$
More generally, a random vector $X=(X_1, \dotsc, X_r)$ with support on a $(r-1)$-dimensional simplex $\{(x_1,x_2,\dotsc,x_r):x_1+x_2+\dotsc+x_r=1, r\geq 2\}$ has a Dirichlet distribution with positive real parameters $(\alpha_1,
\alpha_2,\dotsc,\alpha_r)$ if its density is given by $$f_X(x) = \frac{\Gamma(\sum_{i=1}^r\alpha_i)}{\prod_{i=1}^r
\Gamma(\alpha_i)}\prod\limits_{i=1}^r x_i^{\alpha_i-1}$$ We write $X\sim{\mathrm{Dirichlet}}(\alpha_1, \alpha_2, \dotsc, \alpha_r)$. Note that ${\mathrm{B}}(\alpha,\beta)$ is ${\mathrm{Dirichlet}}(\alpha,\beta)$.
If $Y_i,\ i=1{,\dotsc,}r$ are independent random variables $Y_i\sim
\Gamma(\alpha_i,\gamma)$ with a common $\gamma>0$, then the vector $(Y_1/Y{,\dotsc,}Y_r/Y)$, where $Y=Y_1+\dotsc +Y_r$, has ${\mathrm{Dirichlet}}(\alpha_1{,\dotsc,}\alpha_r)$ distribution.
Invariance under neighbour-dependent shifts {#sec:2}
===========================================
The main object of our study in this section is a stationary point process $T$ on the real line. Each realisation of $T$ can be associated with a countable set of the intervals ${\mathcal{I}}(T)=\{I_k\}_{k\in{\mathbb{Z}}}$ between its consecutive points. Given a realisation, we are going to introduce its transformation which involves dividing the segments between its consecutive points in a random proportion drawn independently from a given distribution and then studying the distribution of the resulting process of the division points. We fix a *division distribution* $G$ supported by $[0,1]$ and define the following (random) operator $\Psi_G$ acting on the set of countable subsets of ${\mathbb{R}}$ without accumulation points: $$\Psi_GT=T'=\cup_{I\in{\mathcal{I}}(T)} c(I),$$ where $c(I)=x+b_I(y-x)$ for an interval $I=(x,y)\in{\mathcal{I}}(T)$ and $b_I$ is a random variable taken from the distribution $G$ independently of anything. Geometrically, one may think that every point $x$ of $T$ is shifted to a new location $c(I)$, where $I$ is the interval to the right from $x$, this is why we call $\Psi_G$ the *neighbour-dependent shift operator*. Obviously, when $G$ is concentrated on 0 or on 1, the corresponding operator $\Phi_G$ preserves the distribution of any stationary point process. We do not specify how the indexing of the intervals is done: the right neighbour to $I_k$ may or may not be $I_{k+1}$. A common way to define $I_0$ as the zero-interval, i.e. the one containing the origin, introduces a size bias. In addition, the zero-interval $I'_0$ of $T'$ may be composed either of pieces of $I_0$ and $I_1$ or of $I_{-1}$ and $I_0$. To avoid unnecessary technicalities involving either re-indexing of $T'$ or a point-stationary indexing of $T$, we choose to work on the Palm space instead.
It is well-known that there is one-to-one correspondence between the distributions of a stationary point process and of a stationary sequence ${\mathcal{T}}=(\tau_k)_{k\in{\mathbb{Z}}}$ of positive random variables with a finite mean. This sequence is related to the Palm version of the point process: under the Palm distribution there is almost surely a point $T_0=0$ of the process at the origin, and if $T_n$ denotes the $n$th closest process point to the origin on the positive semi-axis for $n\geq 1$ and on the negative semi-axis for $n<0$, then $\tau_k=T_k-T_{k-1}$, $k\in{\mathbb{Z}}$, represents the lengths of the $k$th interpoint interval $I_k$. In the case of a renewal process, the sequence ${\mathcal{T}}$ is [i.i.d. ]{}drawn from a distribution $F$ with finite mean. Relation between the distributions of the point process and the corresponding stationary sequence is given in general by the Ryll-Nardzewski exchange formula, [see, e.g., ]{}[@DalVJon:08 Sec. 13.3] for details.
This allows us to define a (stochastic) *exchange operator* $\Phi_G$ acting on sequences of non-negative numbers as follows: take an [i.i.d. ]{}sequence $\{b_k\}_{k\in{\mathbb{Z}}}$ of $G$-distributed random variables and set $$\label{eq:rec}
\Phi_G{\mathcal{T}}={\mathcal{T}}'=(\tau'_k)_{k\in{\mathbb{Z}}},\ \text{where}\
\tau'_k = (1-b_{k})\tau_{k} + b_{k+1}\tau_{k+1}.$$
The main result of this section is the characterisation of the class of fixed points of the exchange operator $\Phi_G$, and thus of the random shift operator $\Psi_G$. The degenerate distribution concentrated at point $x$ is denoted by $\delta_x$.
\[th:fixpt\] Let ${\mathcal{T}}$ be an [i.i.d. ]{} sequence of positive integrable random variables with the distribution $F$ corresponding to a stationary renewal process $T$ and $\Phi_G$ be the exchange operator . Then $\Phi_G({\mathcal{T}}) {\stackrel{D}{=}}{\mathcal{T}}$, and thus $\Psi_G(T) {\stackrel{D}{=}}T$, if and only if one of the following alternatives is true:
(i) $F=\Gamma(\alpha,\gamma)$ and $G={\mathrm{B}}(r\alpha, (1-r)\alpha)$ for some constants $\alpha>0$, $\gamma>0$ and $r\in(0,1)$,
(ii) $F=\delta_s$ for some $s\in (0,\infty)$ and $G=\delta_b$ for some $b\in [0,1]$.
*Necessity.* Considering three consecutive elements $X,Y,Z$ in ${\mathcal{T}}$, we note that $\Phi_G({\mathcal{T}})={\mathcal{T}}' {\stackrel{D}{=}}{\mathcal{T}}$ implies, in particular, that the two consecutive elements in ${\mathcal{T}}'$ they contribute to, should also be independent, identically distributed with $F$, which in turn implies the following condition:
> If $X, Y, Z$ are independent $F$-distributed random variables and $a, b, c$ are independent $G$-distributed, then the random variables $$(1-a)X+bY, (1-b)Y+cZ$$ are also independent and $F$-distributed.
In terms of the Laplace transforms, for $x_1, x_2<0$ we obtain: $$\begin{gathered}
\phi_{(1-a)X+bY, (1-b)Y+cZ}(x_1,x_2)
= \operatorname{{\bf E}}\exp\{x_1((1-a)X+bY)\\
\qquad\qquad\qquad\qquad+x_2((1-b)Y+cZ)\}\\
=\operatorname{{\bf E}}\exp\{x_1(1-a)X\}\operatorname{{\bf E}}\exp\{x_2cZ\}\operatorname{{\bf E}}\exp\{x_1bY+x_2(1-b)Y\}\\
\qquad =\phi_{(1-a)X}(x_1)\phi_{cZ}(x_2)\phi_{bY,(1-b)Y}(x_1,x_2).\end{gathered}$$ On the other hand, the independence of $(1-a)X+bY$ and $(1-b)Y+cZ$ implies $$\begin{gathered}
\phi_{(1-a)X+bY, (1-b)Y+cZ}(x_1,x_2) = \phi_{(1-a)X+bY}(x_1)
\phi_{(1-b)Y+cZ}(x_2)\\
=\phi_{(1-a)X}(x_1)\phi_{bY}(x_1)\phi_{(1-b)Y}(x_2)\phi_{cZ}(x_2),\end{gathered}$$ and hence $$\phi_{(1-b)Y,bY}(x_1,x_2)=\phi_{(1-b)Y}(x_1)\phi_{bY}(x_2),$$ so the random variables $\eta_1=bY$ and $\eta_2=(1-b)Y$ are independent.
Let us first suppose $Y\sim F$ is degenerate. Then the only case when $Y$ is a sum of the two independent random variables $bY$ and $(1-b)Y$ is when both of them are degenerate, too. That means, the random variable $b\sim G$ must be degenerate, leading us to alternative (ii).
Suppose now that $Y\sim F$ is non-degenerate. Then so are $\eta_1$ and $\eta_2$. In that case the random variables $b=\eta_1/(\eta_1+\eta_2)$ and $Y=\eta_1+\eta_2$ can be understood as shape and size variables of the random vector $(\eta_1,\eta_2)$. Note that $b$ and $Y$ are independent by construction, i.e. the shape is independent of the size. Therefore, the only possibility for the joint distributions of $\eta_1, \eta_2$ is for them to be independent, Gamma-distributed with some positive shape parameters $a_1, a_2$ and a common rate $\gamma$, see [@Mos:70 Theorem 4]. Put $r=a_1/(a_1+a_2)$ and $\alpha=a_1+a_2$. Then $b$ becomes ${\mathrm{B}}(r\alpha,
(1-r)\alpha)$-distributed and $Y$ conforms to $\Gamma(\alpha,\gamma)$, proving the alternative (i).
*Sufficiency.* The alternative (ii) trivially leads to the invariance. For the sufficiency in case (i) it is enough to notice that due to [@Mos:70 Theorem 4], for every $k\in{\mathbb{Z}}$ the random variables $\tau_k b_k$ and $\tau_k (1-b_k)$ are independent, distributed as $\Gamma(r\alpha,\gamma)$ and $\Gamma((1-r)\alpha, \gamma)$, respectively. Hence the random variable $\tau'_k = (1-b_k)\tau_k +
b_{k+1}\tau_{k+1}$ is again Gamma-distributed, and moreover, $\tau'_k=(1-b_k)\tau_k + b_{k+1}\tau_{k+1}$ is independent of $\tau'_{k+1}=(1-b_{k+1})\tau_{k+1} + b_{k+2}\tau_{k+2}$, since all their summands are independent. Hence the sequence $\{\tau'_k\}_{k\in{\mathbb{Z}}}$ is again i.i.d., $\tau'_0\sim\Gamma(\alpha,\gamma)$, thus finishing the proof.
Since a homogeneous Poisson process with rate $\gamma$ has exponential $\Gamma(1,\gamma)$ distributed interpoint distances, the Beta division point distribution $G=B(r,1-r)$ for some $0<r<1$ is the only non-degenerate distribution preserving the Poisson process.
Every second point in a homogeneous Poisson process form a renewal process with $\Gamma(2,\gamma)$-distributed interpoint distances. Thus a uniform division distribution which is also $G=B(1,1)$ preserves it. This also follows from a known elementary fact that if $X,Y$ are independent Exponentially-distributed random variables and $U$ is a uniform variable independent of them, then $U(X+Y)$ and $(1-U)(X+Y)$ are independent Exponentially-distributed random variables.
Random exchange model: fixed points and convergence {#sec:exchange}
===================================================
As we have shown in the previous section, Beta division distribution $G$ defines a neighbour-dependent shift operator which preserves a renewal process with gamma-distributed interpoint distances. The two immediate questions arise. Suppose we start from a renewal process realisation ${\mathcal{T}}$ which is *not* Gamma and apply the operator $\Phi_G$ to it iteratively. Will the iterations $\Phi_G^{(n)}({\mathcal{T}})=\Phi_G(\Phi_G^{(n-1)}({\mathcal{T}}))$ converge to a Gamma renewal process? The answer is *yes*, provided the interpoint distances have a finite second moment. Another question: if the ‘new’ interpoint intervals are composed of more than two ‘old’ ones, how much of the previous results still hold? We show below that Gamma-distributed renewal process appears again, but the rôle of the Beta distribution now plays the Dirichlet distribution.
Recall that at each iteration of $\Phi_G$ each interpoint segment, independently of everything else, cuts a $G$-distributed proportion of its length and passes it to the segment to the right of it, while at the same time receiving a portion of length from the interval to the left of it.
If we regard the intervals as ‘agents’, and their lengths as ‘masses’, then the shift procedure defined by can be interpreted as a simultaneous random exchange, where at each application of $\Phi_G$ every agent $i$ splits its current mass $\tau_i$ into two random pieces in proportion $b_i\,:\, 1-b_i$ and shares it between itself and its neighbour to the right, while at the same time receiving a piece of length of its neighbour to the left. More generally, we define a *random mass exchange* model in discrete time as follows.
Consider a countable collection of *agents*, labelled by a sequence of integers $i\in {\mathbb{Z}}$. Each agent is supplied with a non-negative entity called its *mass*. Assume that at the beginning of step $n$ the $i$th agent has mass $\tau^n_i$, starting from some initial mass (row-)vector $\tau^0 = (\tau^0_i)_{i\in {\mathbb{Z}}}$ at $n=0$. Then each agent $i$ samples a new vector of proportions $(\pi_{i, j}(n+1))_{j \in {\mathbb{Z}}},\ \sum_j
\pi_{i,j}(n+1)=1$, and distributes all of its mass between itself and other agents accordingly, so that agent $j$ gets a portion $\tau_i^n
{{\pi_{i,j}(n+1)}}$ of its mass from $i$, or, in a vector form, $$\label{eq:randexmatrix}
\tau^{n+1} = \tau^{n} \Pi(n+1),\ n=0,1,2,\dotsc$$ Here $$\Pi(n) = ({{\pi_{i j}(n)}})_{i,j \in {\mathbb{Z}}}, n=1,2, \dotsc$$ is a mass exchange (two-side infinite) matrix, with proportion vectors $({{\pi_{i,j}(n)}})_{j \in {\mathbb{Z}}}$ as its rows. Obviously for every $n$, $\Pi(n)$ is row-stochastic:
- ${{\pi_{i j}(n)}} \geq 0,\ i,j \in {\mathbb{Z}},\ n =1,2,\dotsc$
- $\sum_{j \in {\mathbb{Z}}} {{\pi_{i j}(n)}} = 1,\ i\in {\mathbb{Z}},\ n=1,2,\dotsc$
The mass exchange model can be regarded as a discrete version of a randomised Potlatch process first defined in [@Hol:81], [@Lig:81], where instead of having a Poisson clock at each site, all of the sites’ transitions are synchronised.
For the sequel we assume two conditions on the initial mass configuration $\tau^0$:
\[cond:A1\] $\tau^0_i$ are non-negative [i.i.d. ]{}random variables for different $i \in {\mathbb{Z}}$ with $\mu=\operatorname{{\bf E}}\tau^0_i < \infty$;
\[cond:A2\] The second moments of initial masses are finite: $\sigma^2={\mathbf{var}\,}\tau^0_i < \infty$.
To stay in the stationary framework, we will also assume that the random exchange model is translation invariant on ${\mathbb{Z}}$:
\[cond:B1\] There exists a random probability *sharing distribution* $(\pi_j)_{j\in {\mathbb{Z}}}$ on ${\mathbb{Z}}$ such that the vectors $({{\pi_{i,i+j}(n)}})_{j\in {\mathbb{Z}}}$ for different $i \in {\mathbb{Z}},$ and $n=1,2,\dotsc$ are [i.i.d. ]{}copies of $(\pi_j)_{j\in {\mathbb{Z}}}$.
Denoting by $p_{ij} = \operatorname{{\bf E}}{{\pi_{i j}(n)}}$ and by $p_i = \operatorname{{\bf E}}\pi_i,\ i,j\in {\mathbb{Z}}, n=1,2,\dotsc$, this implies that the matrix of proportions’ expected values $$P = (p_{ij})_{i,j\in {\mathbb{Z}}}, n=1,2,\dotsc$$ does not depend on $n$ and the following local balance condition is satisfied: $$\label{eq:balance}
\operatorname{{\bf E}}\sum_i \pi_{ij}(n)=\sum_i p_{0,j-i}=\sum_k p_k=1.$$ for all $n\in{\mathbb{N}}$ and $j\in{\mathbb{Z}}$. Here and below the indices in the sums run over all integers, unless specified otherwise.
We will also use the following notation for matrix products: $$\Pi(1:n) = \Pi(1)\dotsc\Pi(n),\ \ \Pi(n:1) = \Pi(n)\dotsc \Pi(1).$$ Note that the neighbour-dependent shifts in the previous section can be regarded as a random exchange model with the two-diagonal exchange matrix $\Pi(n) = ({{\pi_{i,j}(n)}})_{i,j\in{\mathbb{Z}}}$ given by $${{\pi_{i,j}(n)}} = \begin{cases}
1-b_i(n), & i=j,\\
b_i(n), & i=j+1,\\
0, & \text{otherwise,}
\end{cases}$$ where $b_i(n)\sim G,\ i\in{\mathbb{Z}},\ n=1,2,\dotsc$ are [i.i.d. ]{}
#### Random walk in random environment (RWRE) {#paragraph:RWRE}
There is a correspondence between the random exchange and a certain version of RWRE which we define here.
Assume the translation invariance condition . Introduce $\{W^n\}_{n\geq 0}$, a random walk in a random environment (RWRE) on ${\mathbb{Z}}$, governed by the random transition probabilities $\pi_{i,i+j}(n)$, conditional on the environment ${\mathcal{E}}=
\sigma(\Pi(n), n=1,2,\dotsc)$: $${{\bf P}}\left(W^{n+1}=i+j \vert W^n=i, {\mathcal{E}}\right) = \pi_{i,i+j}(n+1) {\stackrel{D}{=}}\pi_{j}$$ Then its $n$-step transition matrix is given by $\Pi(1:n)$.
In this interpretation of RWRE the environment is re-sampled on every step $n=1,2,\dotsc$. Note that if we integrate out the environment, then due to independence of $\Pi(n)$ we can regard a single random walker’s trajectory as a usual time-homogeneous random walk on ${\mathbb{Z}}$ with transition probabilities given by $${{\bf P}}(W^{n+1}=i+j|W^n=i)= \operatorname{{\bf E}}\pi_{i,i+j}(n+1)= p_j,\ i,j \in Z,\ n=1,2,\dots$$
However if there is more than one random walker, the joint dynamics are more involved. Below we will be interested in running several copies of $W^n$ together in the same realisation of the environment $\{\pi_{i,i+j}(n)\}$. In particular, we are going to consider the process $Z_n$ of the difference between two conditionally on ${\mathcal{E}}$ independent copies $W^n, {\widetilde{W}}^n$ of such a random walk: $$Z^n = W^n - \widetilde{W}^n,\ n=1,2,\dots$$ Note that given ${\mathcal{E}}$, $Z^n$ is not a random walk, and not even a Markov chain, since we have to know the positions of both walkers to determine the conditional probabilities for the next step of $Z_n$. However, if we integrate out the environment, $Z^n$ becomes a Markov chain on ${\mathbb{Z}}$ with the transition probabilities $$\label{eq:ztranprob}
{{\bf P}}(Z^{n+1} = i+j | Z^n = i) =\begin{cases}
\sum\limits_{j_1-j_2=j} \operatorname{{\bf E}}\pi_{j_1} \pi_{j_2} , & i=0,\\
\sum\limits_{j_1-j_2=j} p_{j_1} p_{j_2}, & i\neq 0.
\end{cases}$$
Having this machinery at hand, we are ready to prove the sufficient conditions for existence and uniqueness of an equilibrium point of a random exchange model. We use notation $\Rightarrow$ to denote the weak convergence of a sequence of infinite random sequences, meaning the weak convergence of all of their finite-dimensional sub-vectors.
Under condition , there exists a unique (up to a distributional copy and scaling by a constant factor) fixed point for dynamics , i.e. a sequence $\tau^\infty=(\tau^\infty_i)_{i\in{\mathbb{Z}}}$ of (not necessarily independent) random variables such that:
(i) $\tau^\infty \Pi(1) {\stackrel{D}{=}}\tau^\infty,$ \[cond:fixedpoint\]
(ii) \[cond:weakconv\] for any $\tau^0$ satisfying (-), $$\label{eq:wc}
\tau^0 \Pi(1:n) \Rightarrow \tau^\infty\ \text{as $n\to\infty.$}$$
Introduce a dual version of the process: $${\widetilde{\tau}}^n = \tau^0 \Pi(n:1)$$ The proof is based on the distributional equality: $${\widetilde{\tau}}^n {\stackrel{D}{=}}\tau^n, n=0,1,2,\dotsc$$ that holds for any fixed $n=1,2,\dots$
First, prove the theorem for the constant initial conditions $\tau^0 =
{\mathbf{1}}$, where ${\mathbf{1}}$ is the sequence of ones: $\tau^0_j =1$ for all $j\in{\mathbb{Z}}$. Denote by ${\widetilde{\pi}}^n_{ij}$ the elements of matrix $\Pi(n:1)$ and by ${\mathcal{B}}^n$ the $\sigma$-algebra generated by $\tau^0$ and $\{\Pi(k),\ k=1,\dotsc, n\}$. We have that $$\begin{aligned}
\operatorname{{\bf E}}[{\widetilde{\tau}}^{n+1}_j {\hspace*{1ex} \rule[-1ex]{0.15ex}{3ex} \hspace*{1ex}}{\mathcal{B}}^n] = \operatorname{{\bf E}}\bigl[\sum_{i} {\widetilde{\pi}}_{ij}^{n+1} {\hspace*{1ex} \rule[-1ex]{0.15ex}{3ex} \hspace*{1ex}}{\mathcal{B}}^n\bigr]
= \operatorname{{\bf E}}\bigl[\sum_{i} \sum_{k} \pi_{ik}(n+1) {\widetilde{\pi}}_{kj}^{n} {\hspace*{1ex} \rule[-1ex]{0.15ex}{3ex} \hspace*{1ex}}{\mathcal{B}}^n\bigr]\nonumber\\
= \sum_{k} {\widetilde{\pi}}_{kj}^{n} \operatorname{{\bf E}}\sum_{i } \pi_{ik}(n+1). \label{eq:sumreverse}\end{aligned}$$ Due to , $\operatorname{{\bf E}}\sum_{i } \pi_{ik}(n+1)=1$ for any $k\in{\mathbb{Z}}$, so we continue with $$\sum_{k} {\widetilde{\pi}}_{kj}^{n} \operatorname{{\bf E}}\sum_{i } \pi_{ik}(n+1)
= \sum_{k} {\widetilde{\pi}}_{kj}^{n} = {\widetilde{\tau}}^{n}_j.$$ Thus, for every $j$, the sequence $\{{\widetilde{\tau}}^{n}_j\}_{n\geq 0}$ is a non-negative martingale, therefore it has an almost sure limit, call it ${\widetilde{\tau}}^\infty_j$. The sequence ${\widetilde{\tau}}^\infty = ({\widetilde{\tau}}^\infty_j)_{j\in {\mathbb{Z}}}$ obviously satisfies as well as for the initial condition $\tau^0={\mathbf{1}}$. Moreover, $$\begin{aligned}
{\mathbf{var}\,}{\widetilde{\tau}}_j^n &= {\mathbf{var}\,}\sum_{i} {\widetilde{\pi}}_{ij}^n \leq
\operatorname{{\bf E}}(\sum_{i} {\widetilde{\pi}}_{ij}^n)^2
=\operatorname{{\bf E}}\sum_{i}({\widetilde{\pi}}_{ij}^n)^2 + \operatorname{{\bf E}}\sum_{k,l, k\neq l}{\widetilde{\pi}}_{kj}^n{\widetilde{\pi}}_{lj}^n
\nonumber \\
&\leq \sum_{i}\operatorname{{\bf E}}({\widetilde{\pi}}_{ij}^n)^2 + \sum_{k\neq l}p_k p_l
\leq \sum_{i}\operatorname{{\bf E}}({\widetilde{\pi}}_{ij}^n)^2 + 1. $$ Below in we show that the first term vanishes so for any $j$ the martingale $\{{\widetilde{\tau}}_j^n\}_{n\geq 0}$ is bounded in ${\mathcal{L}}^2$ and hence it converges in ${\mathcal{L}}^2$, as well as any finite subvector of ${\widetilde{\tau}}^n$.
Now assume $\tau^0$ is general satisfying , . Define $$\Delta^n = \Delta^0 \Pi(n:1),\ \text{where } \Delta^0 = \tau^0 - \mu{\mathbf{1}}.$$ Now we can rewrite ${\widetilde{\tau}}^n$ as the sum: $$\label{eq:ttaudecomp}
{\widetilde{\tau}}^n = \Delta^n+\mu{\widetilde{\tau}}^n.$$ The second term converges to the limit $\mu{\widetilde{\tau}}^\infty:=\tau^\infty$ coordinate-wise almost surely and in ${\mathcal{L}}^2$. Now show that each coordinate of the first term vanishes in ${\mathcal{L}}^2$, this will imply the desired weak convergence in (ii).
The variance of $\Delta^n_j$ is given by the following expression: $$\begin{aligned}
\label{eq:vardelta}
{\mathbf{var}\,}\Delta^n_j &= {\mathbf{var}\,}\sum_{i} \Delta^0_i {\widetilde{\pi}}^n_{ij} =
\sum_{i} {\mathbf{var}\,}\Delta^0_i {\widetilde{\pi}}^n_{ij}+
\sum_{i\neq k} {\mathbf{cov}}(\Delta^0_i {\widetilde{\pi}}^n_{ij},\Delta^0_k {\widetilde{\pi}}^n_{kj})\nonumber\\
&= \sum_{i} \operatorname{{\bf E}}({\widetilde{\pi}}^n_{ij})^2 \operatorname{{\bf E}}(\Delta^0_i)^2 = \sigma^2 \sum_{i}\operatorname{{\bf E}}(\pi^n_{ij})^2\end{aligned}$$ Here the third identity is based on the following fact that can be checked directly: if $\xi,\eta,X,Y$ are random variables such that $(X,Y)$ is independent of $(\xi,\eta)$ and ${\mathbf{cov}}(X,Y)=0$, then $${\mathbf{cov}}(\xi X, \eta Y) =\operatorname{{\bf E}}X\,\operatorname{{\bf E}}Y\,{\mathbf{cov}}(\xi ,\eta),$$ implying all the covariance terms in are 0.
We now prove that $\operatorname{{\bf E}}\sum_{i}(\pi^n_{ij})^2 \to 0, n\to\infty$ by making use of the RWRE construction introduced earlier in this section. Let $W^n, {\widetilde{W}}^n,
n=0,1,\dotsc$ be the two copies of conditionally independent (given ${\mathcal{E}}$) RWRE with transition probabilities $\Pi(n)$, and let $Z^n=W^n-{\widetilde{W}}^n,
n=0,1,\dotsc$ with $W^0={\widetilde{W}}^0=j$. As already noted, $Z^n$ is a Markov chain, with transition probabilities .
The key observation is the following: $$\label{key-obs}
{{\bf P}}(Z^n=0) = \operatorname{{\bf E}}\sum_{i}{{\bf P}}(W^n={\widetilde{W}}^n=i+j{\hspace*{1ex} \rule[-1ex]{0.15ex}{3ex} \hspace*{1ex}}{\mathcal{E}})
= \operatorname{{\bf E}}\sum_{i}(\pi^n_{j, i+j})^2,$$ where the last equality follows from the conditional independence of $W^n, {\widetilde{W}}^n$. Now, use the translation invariance to continue: $$\operatorname{{\bf E}}\sum_{i}(\pi^n_{j, i+j})^2 = \sum_{i}\operatorname{{\bf E}}(\pi^n_{j-i, j})^2 =
\sum_{i}\operatorname{{\bf E}}({\widetilde{\pi}}^n_{ij})^2.$$
It is easy to see that $0$ is a null state of the Markov chain $Z^n$. Indeed, starting from $0$, $Z^n$ leaves $0$ after a geometrically distributed with parameter $\bigl(1-\sum_{i}\operatorname{{\bf E}}\pi_i^2\bigr)$ number of steps, and while out of $0$, $Z^n$ behaves as a symmetrical random walk on integers, therefore $${{\bf P}}(Z^n=0)=\sum_{i}\operatorname{{\bf E}}({\widetilde{\pi}}^n_{ij})^2\to 0,\ \ n\to\infty,\label{eq:1}$$ hence by , $\Delta_j^n$ tends to $0$ in ${\mathcal{L}}^2$ for any $j$.
We have shown that both terms in the decomposition converge in ${\mathcal{L}}^2$ coordinate-wise: ${\widetilde{\tau}}^n_j{\stackrel{{\mathcal{L}}^2}{\to}}\tau_j^\infty$ for all $j\in{\mathbb{Z}}$, which in turn implies the weak convergence $\tau^n \Rightarrow
\tau^\infty$, as well as the uniqueness of the distribution of $\tau^\infty$, thus finishing the proof.
The finite second moment condition is essential for the convergence even in much more restrictive settings than ours as the following example shows.
Assume that all exchange proportions are non-random $\pi_{i,i+j}(n) = p_{j}$ with $p_{-1}=1-p_0=p\in(0,1)$ and $p_j=0,\
j\in{\mathbb{Z}}\setminus\{0,-1\}$. Then $$\label{eq:eulersum}
\tau_0^n = \sum_{j=0}^n \tau^0_j p^j(1-p)^{n-j}\binom{n}{j}$$ and the (almost sure) limit of the latter expression when $n$ goes to infinity, if it exists, is called the *Euler sum* of the sequence $\tau^0=(\tau^0_j)_{j\in{\mathbb{Z}}}$ with parameter $p$. If $\tau^0$ is a sequence of [i.i.d. ]{}random variables, the almost sure limit of exists if and only if the second moment of $\tau^0_0$ is finite, see [@BinMae:85 Theorem 1] and the references therein. In that case the limit is equal to $\mu=\operatorname{{\bf E}}\tau^0_0$.
\[cor:balance\] If the vector of proportions $\pi$ has a Dirichlet distribution with the vector of non-negative parameters $\alpha =
(\alpha_i)_{i\in{\mathbb{Z}}}$, $\sum_{i}\alpha_i = a < \infty$, then the fixed point is a vector $\tau^\infty$ of independent $\Gamma(a,
\gamma)$-distributed random variables. In particular, if $\tau^0$ satisfies the conditions ,, then $$\tau^0 \Pi(1:n) \Rightarrow \tau^\infty,$$ where the components of $\tau^\infty$ are independent, with distribution $\Gamma(a, \gamma)$ where $\gamma =a/\mu= a/\operatorname{{\bf E}}\tau^0_0$.
When $\alpha_0=(1-r)\alpha$ and $\alpha_1=r\alpha$ for some $\alpha>0$ and $r\in(0,1)$, the statement of the last corollary is the sufficiency part of Theorem \[th:fixpt\](i). It is left open whether the necessity statement is also true for the the cases when the support of the sharing proportion distribution $\pi$ can have more than 2 indices.
A partial answer is provided by the next theorem: in the case when $\pi$ is exchangeable on some finite subset of ${\mathbb{Z}}$, the only scenario for which the masses of different agents in the equilibrium are independent is indeed when $\pi$ has a Dirichlet distribution.
\[th:fixexch\] Assume the conditions , , are satisfied. Assume additionally, that the support of $\pi$ is almost surely in a compact set: $|K| = |\{i\in{\mathbb{Z}}: {{\bf P}}(\pi_i>0)>0\}|<\infty$, and that every subvector of $\pi$ is exchangeable. Then $$\label{eq:stationarity}
\tau'= \tau \Pi {\stackrel{D}{=}}\tau$$ if and only if the components of $\tau$ are Gamma-distributed: $\tau_j
\sim \Gamma(a,\gamma)$ for some $a, \gamma>0$ and $\pi$ is Dirichlet-distributed: $(\pi_i)_{i\in K} \sim
{\mathrm{Dirichlet}}((a/|K|)_{i\in K})$.
For the simplicity of the presentation, take $K=\{0,1,2,\dotsc,m\}$, $m\geq 1$.
The ‘if’ part follows directly from the shape vs. size independence property of Gamma random vectors, as in the proof of Theorem \[th:fixpt\]. Now prove the ‘only if’ part.
First, consider the joint distribution of the two components of $\tau'$ which are at distance $m$: $(\tau'_0, \tau'_m)$. By the invariance assumption they are independent, i.e. in terms of Laplace transforms we have: $$\begin{aligned}
\label{eq:genfactorize1}
\phi_{\tau'_0,\tau'_m}(x_1, x_2) =\ &
\phi_{\tau'_0}(x_1)\phi_{\tau'_m}(x_2)\nonumber\\
=\
&\prod_{i=-m}^{0}\phi_{\tau_{i}\pi_{i0}}(x_1)
\times\prod_{j=0}^m\phi_{\tau_{j}\pi_{jm}}(x_2).
\end{aligned}$$ Alternatively, we can express the Laplace transform of the pair $(\tau'_0, \tau'_m)$ directly, taking into account the independence of $\tau_{i}\pi_{ij}$ for different $i$: $$\begin{gathered}
\label{eq:genfactorize2}
\phi_{\tau'_0,\tau'_m}(x_1, x_2) = \operatorname{{\bf E}}\exp\{x_1\tau'_0 + x_2\tau'_m\} \\
= \operatorname{{\bf E}}\exp\Big\{x_1\sum_{i=-m}^0\tau_{i}\pi_{i0}
+ x_2\sum_{j=0}^m \tau_{j}\pi_{jm} \Big\}\\
= \prod_{i=-m}^{-1}\operatorname{{\bf E}}\exp\{x_1\tau_{i}\pi_{i0}\} \times
\operatorname{{\bf E}}\exp\{x_1\tau_{0}\pi_{00} + x_2\tau_{0}\pi_{0m}\}\times
\prod_{j=1}^m\operatorname{{\bf E}}\exp\{x_2\tau_{j}\pi_{jm}\}\\
= \prod_{i=-m}^{-1}\phi_{\tau_{i}\pi_{i0}}(x_1) \times
\phi_{\tau_0\pi_{00},\tau_{0}\pi_{0m}}(x_1,x_2)\times
\prod_{j=1}^m\phi_{\tau_{j}\pi_{jm}}(x_2).
\end{gathered}$$ Comparing and , we conclude that the two quantities $\tau_0\pi_{00}, \tau_0\pi_{0m}$ are independent. Since the distribution of the vector $$(\pi_{00},\pi_{01},\dotsc,\pi_{0m})$$ is exchangeable, the random variables $\tau_0\pi_{00},
\tau_0\pi_{01},\dotsc, \tau_0\pi_{0m}$ are pairwise independent.
Next, consider the joint Laplace transform of the three components in $\tau'$: say, $\tau'_0$, $\tau'_1$ and $\tau'_m$. We can repeat the previous argument to arrive at the conclusion that the joint Laplace transform of the three quantities $\tau_0\pi_{00},
\tau_0\pi_{01}, \tau_0\pi_{0m}$ factorizes into the product of their marginal Laplace transforms, making them mutually independent. Using the exchangeability assumption, we conclude that the random variables\
$\tau_0\pi_{00}, \tau_0\pi_{01},\dotsc, \tau_0\pi_{0m}$ are 3-independent.
Repeating this argument $m$ times yields the joint independence of all the components of the random vector $(\tau_0\pi_{00},
\tau_0\pi_{01},\dotsc, \tau_0\pi_{0m})$. Notice that $(\pi_{00},
\pi_{01},\dotsc, \pi_{0m})$ is its shape vector and $\tau_0$ is its size variable. Moreover, $\tau_0$ is independent of $(\pi_{00},
\pi_{01},\dotsc, \pi_{0m})$ by construction, so the application of the shape vs. size characterisation of Gamma-distributed random vectors [@Mos:70 Theorem 4] finishes the proof.
Open problems and generalisations {#sec:open-probl-gener}
=================================
The models we have considered here admit a variety of generalisations and raise many intriguing questions. First of all, already mentioned extension of Theorem \[th:fixexch\] to a non-exchangeable sharing distribution would give a generalisation of Theorem \[th:fixpt\]. We conjecture that a non-degenerate [i.i.d. ]{}limiting sequence $\tau^\infty$ is possible only for a Dirichlet sharing distribution $\pi$, but a counterexample may well exist. If, however, the conjecture *is* true, there are arguments that a more general statement may hold without assuming shift-invariance of $\Pi=(\pi_{ij})$. For instance, the agents may be indexed by another countable group such as ${\mathbb{Z}}^d$ with $d\geq2$ rather than by ${\mathbb{Z}}$. Assume there exists a sequence $q=(q_k)_{k\in {\mathbb{Z}}}$ satisfying the balance equation $qP=q$ for the matrix $P$ of the expectations $p_{ij}=\operatorname{{\bf E}}\pi_{ij}$. Using the same relations between Gamma and Dirichlet distributions, one can show that the vector of masses $\tau^0$ with independent components $\tau_i^0$ distributed as $\Gamma(a q_i,\gamma)$ with some $a,\gamma>0$ is left invariant by the following sharing distributions now depending on the node $k$: $(\pi_{ki})_{i\in{\mathbb{Z}}}\sim{\mathrm{Dirichlet}}((aq_kp_{ki})_{i\in{\mathbb{Z}}})$, $k\in{\mathbb{Z}}$. In the shift-invariant case $\pi_{ij}=\pi_{j-i}$ for all $i,j\in{\mathbb{Z}}$, the unit sequence ${\mathbf{1}}$ satisfies the balance equation and we obtain Corollary \[cor:balance\]. It is interesting if this is the only possibility to have non-trivial independent masses as a fixed point.
Finally, an intriguing question is, if there exists a fixed point for multi-dimensional analog of the neighbour-dependent shifts models. For a point process in ${\mathbb{R}}^d,\ d\geq2$ neighbouring relation can be defined in many different ways, for instance, neighbours can be declared the nodes connected by en edge in any stationary graph having the process points as vertices. We already mentioned in the Introduction the adjustment process where the nodes move to the centre of mass of their Voronoi cells. This provides an example of a shift depending on the Delaunay graph neighbours with a hexagon lattice vertices in ${\mathbb{R}}^2$ left intact. Whether there are any non-degenerate point processes preserved by these neighbouring shifts is an open question, as well as if there are neighbour-dependent shifts of any kind preserving a multi-dimensional Poisson process. Note in this respect that even the balance equation is hard to satisfy on stationary graphs in ${\mathbb{R}}^d$, the reason being typically unbounded degree of their vertices.
#### Acknowledgement {#sec:acknowledgement}
The authors are thankful to Stas Volkov for many fruitful discussions and suggestions.
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
G. Peccati, M. Reitzner (Eds.), Stochastic analysis for [P]{}oisson point processes: [M]{}alliavin calculus, [W]{}iener-[I]{}tô chaos expansions and stochastic geometry, Springer, 2015.
A. Okabe, B. Boots, K. Sugihara, S. N. Chiu, Spatial Tessellations — Concepts and Applications of [Voronoi]{} Diagrams, 2nd Edition, Wiley, 2000.
Q. Du, V. Faber, M. Gunzburger, Centroidal [V]{}oronoi tessellations: Applications and algorithms., [SIAM Rev.]{} 41 (4) (1999) 637–676.
M. Hasegawa, M. Tanemura, On the pattern of space division by territories, Ann. Inst. Statist. Math. 28 (1976) 509–519, part B.
Q. Du, M. Emelianenko, L. Ju, Convergence of the [L]{}loyd algorithm for computing centroidal [V]{}oronoi tessellations., [SIAM J. Numer. Anal.]{} 44 (1) (2006) 102–119.
S. McKinlay, [A characterisation of transient random walks on stochastic matrices with dirichlet distributed limits](http://dx.doi.org/10.1239/jap/1402578642), Journal of Applied Probability 51 (2) (2014) 542–555. [](http://dx.doi.org/10.1239/jap/1402578642). <http://dx.doi.org/10.1239/jap/1402578642>
D. J. Daley, D. Vere-Jones, An Introduction to the Theory of Point Processes. Volume II: General Theory and Structure, 2nd Edition, Springer, New York, 2008.
J. E. Mosimann, Size allometry: size and shape variables with characterizations of the lognormal and generalized gamma distributions, Journal of the American Statistical Association 65 (330) (1970) 930–945.
R. Holley, T. M. Liggett, Generalized potlatch and smoothing processes, Zeitschrift f[ü]{}r Wahrscheinlichkeitstheorie und verwandte Gebiete 55 (2) (1981) 165–195.
T. Liggett, F. Spitzer, [Ergodic theorems for coupled random walks and other systems with locally interacting components](http://dx.doi.org/10.1007/BF00531427), Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete 56 (4) (1981) 443–468. [](http://dx.doi.org/10.1007/BF00531427). <http://dx.doi.org/10.1007/BF00531427>
N. Bingham, M. Maejima, Summability methods and almost sure convergence, Z. Wahrsch. Verw. Geb. 68 (3) (1985) 383–392.
[^1]: Chalmers University of Technology, Department of Mathematical Sciences, Gothenburg, Sweden. Email: `[muratov|sergei.zuyev]@chalmers.se`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We construct a geodesic net in the plane with four unbalanced (boundary) vertices that has $16$ balanced vertices and does not contain proper geodesic subnets. This is the first example of an irreducible geodesic net in the Euclidean plane with $4$ boundary vertices that is not a tree.'
address: 'Fabian Parsch, Department of Mathematics, University of Toronto, 40 St. George Street, Toronto, ON M5S 2E4, Canada'
author:
- Fabian Parsch
bibliography:
- '../../bib/main.bib'
title: An example for a nontrivial irreducible geodesic net in the plane
---
Introduction
============
Geodesic nets in Riemannian manifolds are critical points (but not necessarily local minima) of the length functional on spaces of embedded graphs or multigraphs, where some vertices must be mapped to prescribed vertices that are called [*boundary*]{} or [*unbalanced*]{}. As such they can be regarded as generalizations of geodesics (that arise in the case of two boundary points) as well as $1$-dimensional analogs of minimal surfaces spanning a given contour. Surprisingly, very little is known about the classification of geodesic nets even in the case when the ambient Riemannian manifold is just the Euclidean plane. In this case the geodesic nets are simply embedded graphs so that (i) each edge is a straight line segment; (ii) Some of the edges are mapped to prescribed points (“unbalanced vertices”) and (iii) At each of the remaining vertices $v$ (“balanced vertices”) the following balancing condition holds: The sum of all unit vectors directed along all edges from $v$ to the opposite end of the edge is equal to the zero vector. (This characterization holds also in the more general case of ambient Riemannian manifolds with the only distinction that edges are supposed to be geodesics.)
(-1,0) – +(30:1); (-1,0) – +(150:1); (-1,0) – +(270:1); (-1,0) circle \[radius=0.05\];
(1.5,-0.2) – +(10:1); (1.5,-0.2) – +(-20:1); (1.5,-0.2) – +(160:1); (1.5,-0.2) – +(190:1); (1.5,-0.2) circle \[radius=0.05\];
(4,0) – (4.085159600262456,0.9963673230707331); (4,0) – (3.154511096969029,0.5339929913879817); (4,0) – (3.088735454976287,-0.4118214770780241); (4,0) – (4.975942150302623,0.2180296293229263); (4,0) – (3.5248510852651163,-0.8799053976571926); (4,0) – (4.2773364928491535,-0.9607728502274256); (4,0) – (4.870934923071531,0.49139837176611095); (4,0) circle \[radius=0.05\];
Note that in the present paper we are not allowing integer multiplicities of edges (or, equivalently, we assume that each edge has multiplicity $1$). As one can add and remove vertices of degree $2$ inside each edge at will, the role of such vertices in the classification is clear, and we will be assuming that [**all balanced vertices have degree greater than or equal to $3$.**]{} Also, as one can add or remove straight line edges between unbalanced vertices, we will consider only geodesic nets without such edges. Finally, we are going to consider only connected geodesic nets.
As geodesics nets in the plane are obviously contained in the convex hull of the unbalanced point, the classification really starts from the case of $3$ balanced vertices. The first non-trivial example arises when three unbalanced points $A_1$, $A_2$, $A_3$ form a triangle with all angles less than $120{\degree}$. In this case there exists the unique point $F$ called the Fermat point of the triangle $A_1A_2A_3$, such that all angles $A_iFA_j$, $i\not= j$, are equal to $120{\degree}$. This condition implies that $F$ is the balanced vertex in the geodesic net formed by the three straight line segments connecting $F$ with $A_i$ for $i=1,2,3$.
In [@Parsch:2018aa] we prove the theorem asserting that this is the only possible example of a geodesic net with three unbalanced vertices in the Euclidean plane (and, more generally, any Riemannian manifold endowed with a non-positive curved Riemannian metric). This theorem is surprisingly difficult to prove. It can be stated as follows:
\[thm:three\] Each geodesic net with 3 unbalanced vertices (of arbitrary degree) on the plane endowed with a Riemannian metric of non-positive curvature has exactly one balanced vertex.
What happens in the next case, when the number of unbalanced vertices is equal to $4$? Consider first the geodesic net in figure \[fig:fournet\] (that we first constructed in [@Parsch:2018aa]). At first sight, it seems rather complicated. However, the figure shows that it is essentially an “overlay” of several geodesic nets, each of them being a tree: four geodesic nets with $3$ unbalanced vertices and the Fermat point in the middle, and three well-known tree-shaped geodesic nets with $4$ unbalanced vertices and $1$ or $2$ balanced vertices (plus the balanced vertices that appear as points of the intersections of edges of these elementary geodesic nets). With this example in mind, it is of interest to define [*irreducible*]{} geodesic nets as geodesic nets without proper geodesic subnets. (Here $G_1$ is a proper gedesic subnet of $G$ if (i) The set of balanced (resp. unbalanced) vertices of $G_1$ is the set of balanced (resp. unbalanced) vertices of $G$; (ii) The set of edges of $G_1$ is a non-empty subset of the set $E$ of edges of $G$ that does not coincide with $E$.) Then we immediately realize that the only known examples of irreducible geodesic nets with four unbalanced vertices are the trees with $1$ or $2$ balanced vertices that can be seen at the right side of figure 1.2.
This brings us to the questions that we are trying to answer in this paper:
[**Question.**]{} Do there exist irreducible geodesic nets with $4$ unbalanced vertices in the Euclidean plane with at least $3$ balanced vertices? Can they contain cycles of balanced points?
The main result of the paper is that the answer for these questions is yes.
[**Main Theorem.**]{} There exists an irreducible geodesic net in the Euclidean plane that has $16$ balanced vertices and $4$ unbalanced vertices.
It is tempting to conjecture that our example is one of a series of similar examples with arbitrary large number of balanced vertices, but at the moment this is the only new example of an irreducible geodesic net with four unbalanced vertices that we were able to construct.
(-5.5,1.5) rectangle (20,12); at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(A.center) – (B1.center) – (C.center); (A.center) – (B2.center) – (C.center); (A.center) – (B3.center) – (C.center);
(X.center) – (Y1.center) – (Z.center); (X.center) – (Y2.center) – (Z.center); (X.center) – (Y3.center) – (Z.center);
(B1.center) – (X.center); (B3.center) – (Z.center);
(Y1.center) – (A.center); (Y3.center) – (C.center);
(B2.center) – (Y2.center);
(A.center) – (L.center); (X.center) – (L.center); (Z.center) – (N.center); (C.center) – (N.center); (L.center) – (N.center);
(A.center) – (Z.center); (C.center) – (X.center);
at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(A.center) – (B1.center) – (C.center);
(B1.center) – (X.center);
at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(A.center) – (B3.center) – (C.center);
(B3.center) – (Z.center);
at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(A.center) – (B2.center) – (C.center);
(X.center) – (Y2.center) – (Z.center);
(B2.center) – (Y2.center);
at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(X.center) – (Y1.center) – (Z.center);
(Y1.center) – (A.center);
at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(X.center) – (Y3.center) – (Z.center);
(Y3.center) – (C.center);
at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(A.center) – (L.center); (X.center) – (L.center); (Z.center) – (N.center); (C.center) – (N.center); (L.center) – (N.center);
at (0,0) (Q) ; at (150:3) (X) ; at (125:3) (Y1) ; at (90:3) (Y2) ; at (55:3) (Y3) ; at (30:3) (Z) ;
at (0,15) (P) ; (P) – ++(210:6) node (A) ; (P) – ++(255:6) node (B1) ; (P) – ++(270:6) node (B2) ; (P) – ++(285:6) node (B3) ; (P) – ++(330:6) node (C) ;
at (-.84,4.48) (L) ; at (.84,4.48) (N) ;
(A.center) – (Z.center); (C.center) – (X.center);
Construction of the example
===========================
Before we start, it is worth noting that the geodesic net $G$ we are constructing here will be symmetric under a rotation by $90{\degree}$.
In this section, the notation $ABC$ means the angle at $B$ from $A$ to $C$. Furthermore $\mathcal{O}=(0,0)$ will denote the origin. The end result of the construction is given in figure \[fig:fullnet\].
The octagon of $a_i$ and $b_i$
------------------------------
Fix four vertices at $a_1=(1,0)$, $a_2=(0,1)$, $a_3=(-1,0)$ and $a_4=(0,-1)$. Now add another four vertices $b_1$, $b_2$, $b_3$, $b_4$ so that we arrive at an octagon $a_1b_2a_2b_2a_3b_3a_4b_4$ where the interior angle at each $a_i$ is $150{\degree}$ and the interior angle at each $b_i$ is $120{\degree}$.
(0,0) circle (0.05) node \[below left=0.1\] [$\mathcal{O}$]{};
(0:1) coordinate (a1) – (45:1.12) coordinate (b1) – (90:1) coordinate (a2) – (135:1.12) coordinate (b2) – (180:1) coordinate (a3) – (225:1.12) coordinate (b3) – (270:1) coordinate (a4) – (315:1.12) coordinate (b4) – cycle;
(a1) circle (0.05) node \[right=0.2\] [$a_1$]{}; (a2) circle (0.05) node \[above=0.2\] [$a_2$]{}; (a3) circle (0.05) node \[left=0.2\] [$a_3$]{}; (a4) circle (0.05) node \[below=0.2\] [$a_4$]{};
(b1) circle (0.05) node \[above right=0.2\] [$b_1$]{}; (b2) circle (0.05) node \[above left=0.2\] [$b_2$]{}; (b3) circle (0.05) node \[below left=0.2\] [$b_3$]{}; (b4) circle (0.05) node \[below right=0.2\] [$b_4$]{};
(0:[tan(76)]{}) coordinate (c1) circle (0.05) node \[right=0.2\] [$c_1$]{};
(0,0) – (c1) – (a2) – cycle;
The four unbalanced vertices $c_i$
----------------------------------
There is a uniquely defined triangle as follows, see figure \[fig:octagon\]:
- The left side is $a_2\mathcal{O}$.
- The angle at $\mathcal{O}$ is $90{\degree}$.
- The angle at $a_2$ is $\arccos\left(\frac12-\cos 75{\degree}\right)\approx 76.04{\degree}$.
The resulting third vertex of this triangle is denoted by $c_1$. By rotation around $\mathcal{O}$ we get vertices $c_2$, $c_3$ and $c_4$, see figure \[fig:fullnet\].
(0:1) coordinate (a1) – (45:1.12) coordinate (b1) – (90:1) coordinate (a2) – (135:1.12) coordinate (b2) – (180:1) coordinate (a3) – (225:1.12) coordinate (b3) – (270:1) coordinate (a4) – (315:1.12) coordinate (b4) – cycle;
(0:[tan(76)]{}) coordinate (c1) circle (0.05) node \[right=0.2\] [$c_1$]{}; (90:[tan(76)]{}) coordinate (c2) circle (0.05) node \[above=0.2\] [$c_2$]{}; (180:[tan(76)]{}) coordinate (c3) circle (0.05) node \[left=0.2\] [$c_3$]{}; (270:[tan(76)]{}) coordinate (c4) circle (0.05) node \[below=0.2\] [$c_4$]{};
(a1) – (c2) – (a3) – (c4) – (a1); (a2) – (c3) – (a4) – (c1) – (a2);
(a1) – (c1); (a2) – (c2); (a3) – (c3); (a4) – (c4);
(b1) – ++ (45:0.07) coordinate (d1); (c1) – (d1) – (c2);
(b2) – ++ (135:0.07) coordinate (d2); (c2) – (d2) – (c3);
(b3) – ++ (225:0.07) coordinate (d3); (c3) – (d3) – (c4);
(b4) – ++ (315:0.07) coordinate (d4); (c4) – (d4) – (c1);
(b1) circle (0.3); (a1) circle (0.3);
(\[c\]b1) circle (1.7); (\[c\]a1) – (\[c\]b1) – (\[c\]a2); (\[c\]a1) – (\[c\]c2); (\[c\]a2) – (\[c\]c1); (\[c\]b1) – (\[c\]d1); (\[c\]c1) – (\[c\]d1) – (\[c\]c2); (\[c\]b1) circle (0.05) node \[below left\] [$b_1$]{}; (\[c\]d1) circle (0.05) node \[above right\] [$d_1$]{};
(\[c\]b1) circle (1.7);
(\[c\]a1) circle (1.7); (\[c\]a1) – ++ (120:3); (\[c\]a1) – ++ (-120:3); (\[c\]a1) – ++ (105:3); (\[c\]a1) – ++ (-105:3); (\[c\]a1) – (\[c\]c1); (\[c\]a1) circle (0.05) node \[left\] [$a_1$]{}; (\[c\]a1) – ++(325:0.8) node \[text width=2cm,align=center\]
angles exaggerated
; (\[c\]a1) – ++(0:0.5) coordinate (arcstart1); (\[c\]a1) – ++(0:1) coordinate (arcstart2); (arcstart1) arc (0:105:0.5); (arcstart2) arc (0:120:1);
(\[c\]a1) circle (1.7);
The fermat points $d_i$
-----------------------
We define $d_1$ as follows (again, $d_2$, $d_3$ and $d_4$ will be defined by rotational symmetry): It is the Fermat point of the triangle $b_1c_1c_2$. Recall that the Fermat point is the unique point $x$ in a triangle such that the angle at $x$ between any two corners of the triangle is $120{\degree}$. It exists as long as all interior angles of the triangle are less than $120{\degree}$. So we are left to show:
All three interior angles of the triangle $b_1c_1c_2$ are less than $120{\degree}$.
It is obvious that the interior angles at $c_1$ and $c_2$ are less than $120{\degree}$ (in fact they are both significantly smaller than $90{\degree}$). So it remains to show that the angle at $b_2$ is less than $120{\degree}$. This can be shown as follows:
- $c_1Ob_1=45{\degree}$ follows from the symmetries of the octagon.
- The segment $a_2c_1$ is above the segment $b_1c_1$ (see figure \[fig:octagon\]). It follows that $$\begin{aligned}
b_1c_1O< a_1c_1O\approx 180{\degree}-90{\degree}-76.04{\degree}=13.96
\end{aligned}$$
- We now have estimates for two of the angles of the triangle with corners $\mathcal{O}$, $c_1$, $b_1$ and get $$\begin{aligned}
Ob_1c_1&=180{\degree}- c_1Ob_1- b_1c_1O>180{\degree}- c_1Ob_1- a_1c_1\mathcal{O}\\
&\approx 180{\degree}-45{\degree}-13.96{\degree}=121.04{\degree}\end{aligned}$$
- We can us the symmetry of the picture and conclude: $$\begin{aligned}
c_1b_1c_2&=360{\degree}- Ob_1c_1 - c_2b_1O \\
&= 360{\degree}- 2\cdot Ob_1c_1\approx 117.92<120{\degree}\end{aligned}$$
The edges
---------
Finally we add the edges of the geodesic net. We recommend referring again to figure \[fig:fullnet\] for better understanding.
The following definitions have to be read circular, e.g. $a_5$ is the same as $a_1$. With that in mind, the edges of the geodesic net for $i=1,2,3,4$ are:
- The edges of the octagon, given by $a_ib_i$ and $a_{i+1}b_i$.
- The radial edges given by $a_ic_i$.
- The edges given by $a_ic_{i+1}$ and $a_{i+1}c_i$.
- Finally, the edges for each Fermat point, given by $d_ib_i$, $d_ic_i$ and $d_ic_{i+1}$.
The additional vertices $x_i$
-----------------------------
Note that we are getting four additional vertices of degree $6$ which are situated on the edge connecting $b_1$ and $d_1$ (again, see figure \[fig:fullnet\]). We call these vertices $x_1$, $x_2$, $x_3$ and $x_4$.
Checking the edges and the balance
----------------------------------
The net constructed above is a valid geodesic net with four unbalanced vertices, more specifically:
- Different edges can intersect only at their common endpoints (in other words: there never is an “overlay” of edges, so all edges have weight one).
- Each $a_i$, $b_i$, $d_i$, $x_i$ is balanced.
Note again that the picture is rotationally symmetric by design. So we can concentrate on the corner $c_1Oc_2$ (the upper right corner).
As long as we prove that none of the edges are parallel, the result follows. We will go through the edges as defined above, again adding them step by step.
- It is apparent that none of $a_1b_1$, $a_2b_1$ and $a_1c_1$ are.
- Adding $a_1c_2$, note that the angle between $a_1b_2$ and $a_1c_2$ is approximately $1.04{\degree}$, so these two edges are not parallel. It is apparent that $a_1c_2$ is never parallel to any other edge. By symmetry, adding $a_2c_1$ doesn’t create issues either.
- The edge $d_1b_1$ is radial at the angle $45{\degree}$. No other edge is. This finally brings us to the only two interesting edges: adding $d_1c_1$ and $d_1c_2$. We will consider the former. Symmetry will then deal with the latter. The only problem could arise if $d_1c_1$ coincides with the previously added $a_2c_1$ (which would also imply that $d_1=x_1$). Elementary calculations involving the angle sum in triangles, however, show that $d_1c_1O=15{\degree}$ whereas $a_2c_1O\approx 13.96{\degree}$. It follows that the two edges in question are not parallel.
We finish with showing that all vertices except the $c_i$ are balanced. By symmetry, it is again enough to consider $i=1$:
- Each of the $a_1$ is a degree $5$ balanced vertex. Putting the origin of the coordinate system at $a_1$, the sum of the unit vectors parallel to the five edges can be written as follows (refer to the zoom-in in figure \[fig:fullnet\]): $$\begin{aligned}
\langle 1,0\rangle &+\langle \cos(180{\degree}-\arccos(1/2-\cos(75{\degree}))),\sin(180{\degree}-\arccos(1/2-\cos(75{\degree})))\rangle\\
&+ \langle \cos(180{\degree}-75{\degree}),\sin(180{\degree}-75{\degree})\rangle\\
&+\langle \cos(180{\degree}+\arccos(1/2-\cos(75{\degree}))),\sin(180{\degree}+\arccos(1/2-\cos(75{\degree})))\rangle\\
&+ \langle \cos(180{\degree}+75{\degree}),\sin(180{\degree}+75{\degree})\rangle=\langle 0,0\rangle
\end{aligned}$$ In fact, the very reason for choosing the “odd angle” $\arccos(1/2-\cos(75{\degree}))$ early in the construction was to ensure that the $a_i$ are balanced.
- $b_1$ is a degree three balanced vertex. This follows from the fact that two of the incident edges belong to the octagon, so the angle between them at $b_1$ is $120\degree$. The third edge at $b_1$ is the bisector of the larger angle between the other two edges by symmetry. The balancedness of $b_1$ follows.
- $d_1$ is balanced by the definition of a Fermat point.
- Finally $x_1$ is just the point of intersection of several straight edges and is trivially balanced.
Proof that $G$ is irreducible
=============================
While it is obvious that $G$ is not a tree, we need to show that:
$G$ is irreducible.
We are going to give a proof by contradiction. Assume that $G_1$ is a proper geodesic subnet of $G$. First, assume that the set of (balanced) vertices of $G_1$ does not contain any $a_i$. Then $G_1$ does not contain any edges incident to $a_i$. Now it is easy to see that $G_1$ does not contain any vertices $b_i$ as well as edges incident to $b_i$. From here it is easy to see that $G_1$ is empty - a contradiction.
So, we can assume without any loss of generality that $a_1$ is a (balanced) vertex of $G_1$.
A simple check of the $2^5$ subsets of edges incident to $a_1$ (there are many symmetric cases) show that the only way that $a_1$ can be balanced is if *all* incident edges are used. It follows that $G_1$ includes all vertices adjacent to $a_1$.
We therefore know that $a_1,b_1,b_4,c_1,c_2,c_4$ are in the vertex set of $G_1$.
Consider $b_1$ which is a degree three vertex. Obviously one can’t take a proper subset of the set of incident edges to balance $b_1$ (and the same will be true for all degree $3$ vertices ). It follows that $G_1$ includes all vertices adjacent to $b_1$.
We therefore know that $a_1,a_2,b_1,b_4,c_1,c_2,c_4,d_1$ are in the vertex set of $G_1$.
Now that $a_2$ is in the net, we can reuse the argument based on $a_1$ above, adding $b_2$ and $c_3$ to the picture. Again, reuse previous arguments for $b_2$ and it follows that $d_2$ and $a_3$ are part of $G_1$. It should now be apparent how to conclude that $c_4$, $b_3$, $d_3$, $a_4$ and $d_4$ are in $G_1$.
So $G_1$ includes all balanced vertices of $G$, except possibly the $x_i$. However, as previously argued, since all the $b_i$ and $d_i$ are of degree $3$, all their incident edges are in $G_1$. Also since the $a_i$ can only be balanced with all incident edges included, all edges of $G$ are in $G_1$. Since the $x_i$ are just points of intersection of edges, they are also in $G$.
We can conclude that $G_1=G$. So, $G_1$ is not proper, and we obtain the desired contradiction. Hence, $G$ is irreducible.
Acknowledgements
================
The author would like to thank his PhD advisor Alexander Nabutovsky for insightful discussions and for suggestions to improve the exposition of this paper. This research was partially supported by an NSERC Vanier Scholarship.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'A.V. Lipatov, N.P. Zotov'
title: |
Higgs boson production at hadron colliders\
in the $k_T$-factorization approach
---
[*D.V. Skobeltsyn Institute of Nuclear Physics,\
M.V. Lomonosov Moscow State University,\
119992 Moscow, Russia*]{}\
[**Abstract** ]{}
We consider the Higgs boson production at high energy hadron colliders in the framework of the $k_T$-factorization approach. The attention is focused on the dominant gluon-gluon fusion subprocess. We calculate the total cross section and transverse momentum distributions of the inclusive Higgs production using unintegrated gluon distributions in a proton obtained from the full CCFM evolution equation. We show that $k_T$-factorization gives a possibility to investigate the associated Higgs boson and jets production. We calculate the transverse momentum distributions and study the Higgs-jet and jet-jet azimuthal correlations in the Higgs + one or two jet production processes. We demonstrate the importance of the higher-order corrections within the $k_T$-factorization approach. These corrections should be developed and taken into account in the future applications.
Introduction
============
It is well known that the electroweak symmetry breaking in the Standard Model of elementary particle interactions is achieved via the Higgs mechanism. In the minimal model there are a single complex Higgs doublet, where the Higgs boson $H$ is the physical neutral Higgs scalar which is the only remaining part of this doublet after spontaneous symmetry breaking. In non-minimal models there are additional charged and neutral scalar Higgs particles. The search for the Higgs boson takes important part at the Fermilab Tevatron experiments and will be one of the main fields of study at the CERN LHC collider \[1\]. The experimental detection of the $H$ will be great triumph of the Standard Model of electroweak interactions and will mark new stage in high energy physics.
At LHC conditions, the gluon-gluon fusion $gg \to H$ is the dominant inclusive Higgs production mechanism \[2, 3\]. In this process, the Higgs production occurs via triangle heavy (top) quark loop. The gluon fusion and weak boson fusion ($qq \to qqH$ subprocess via $t$-channel exchange of a $W$ or $Z$ bosons) are also expected to be the dominant sources of semi-inclusive Higgs production (in association with one or two hadronic jets) \[4\]. The detailed theoretical studies of such processes are necessary, in particular, to determine an optimal set of cuts on the final state particles.
It is obvious that the gluon-gluon fusion contribution to the Higgs production at LHC is strongly dependend on the gluon density $xG(x,\mu^2)$ in a proton. Usually gluon density are described by the Dokshitzer-Gribov-Lipatov-Altarelli-Parizi (DGLAP) evolution equation \[5\] where large logarithmic terms proportional to $\ln \mu^2$ are taken into account. The cross sections can be rewritten in terms of hard matrix elements convoluted with gluon density functions. In this way the dominant contributions come from diagrams where the parton emissions in the initial state are strongly ordered in virtuality. This is called collinear factorization, as the strong ordering means that the virtuality of the parton entering the hard scattering matrix elements can be neglected compared to the large scale $\mu^2$. However, at the LHC energies, typical values of the incident gluon momentum fractions $x \sim m_H/\sqrt s \sim 0.008$ (for Higgs boson mass $m_H = 120$ GeV) are small, and another large logarithmic terms proportional to $\ln 1/x$ become important. These contributions can be taken into account using Balitsky-Fadin-Kuraev-Lipatov (BFKL) evolution equation \[6\]. Just as for DGLAP, in this way it is possible to factorize an observable into a convolution of process-dependent hard matrix elements with universal gluon distributions. But as the virtualities (and transverse momenta) of the propagating gluons are no longer ordered, the matrix elements have to be taken off-shell and the convolution made also over transverse momentum ${\mathbf k}_T$ with the unintegrated ($k_T$-dependent) gluon distribution ${\cal F}(x,{\mathbf k}_T^2)$. The unintegrated gluon distribution ${\cal F}(x,{\mathbf k}_T^2)$ determines the probability to find a gluon carrying the longitudinal momentum fraction $x$ and the transverse momentum ${\mathbf k}_T$. This generalized factorization is called $k_T$-factorization \[7–10\]. It is expected that BFKL evolution gives the theoretically correct description at assymptotically large energies (i.e. very small $x$). At the same time another approach, valid for both small and large $x$, have been developed by Ciafaloni, Catani, Fiorani and Marchesini, and is known as the CCFM model \[11\]. It introduces angular ordering of emissions to correctly treat gluon coherence effects. In the limit of asymptotic energies it is almost equivalent to BFKL \[12–14\], but also similar to the DGLAP evolution for large $x$ and high $\mu^2$. The resulting unintegrated gluon distribution depends on two scales, the additional scale ${\bar q}^2$ is a variable related to the maximum angle allowed in the emission and plays the role of the evolution scale $\mu^2$ in the collinear parton densities. The following classification scheme \[15\] is used: ${\cal F}(x,{\mathbf k}_T^2)$ denote pure BFKL-type unintegrated gluon distributions and ${\cal A}(x,{\mathbf k}_T^2,\mu^2)$ stands for any other type having two scale involved. In this paper we will apply the CCFM gluon evolution to study of the inclusive and semi-inclusive Higgs production at LHC conditions.
In the collinear factorization, the calculation of such processes is quite complicated even at lowest order because of the heavy quark loops contribution. For example, in Higgs + one jet production, triangle and box loops occur, and in Higgs + two jet production the pentagon loops occur \[16\]. However, the calculations of the Higgs production rates can be simplified in the limit of large top quark mass $m_t \to \infty$ \[17\]. In this approximation the coupling of the gluons to the Higgs via top-quark loop can be replaced by an effective coupling. Thus it reduces the number of loops in a given diagram by one. The large $m_t$ approximation is valid to an accuracy of $\sim 5$% in the intermediate Higgs mass range $m_H < 2 m_t$, as long as transverse momenta of the Higgs or final jets are smaller than of the top quark mass ($p_T < m_t$) \[16\]. Within this approach, the total cross section for $gg \to H + X$ is known to next-to-next-to-leading order (NNLO) accuracy \[18\]. Higher-order QCD corrections to inclusive Higgs production were found to be large: their effect increases the leading order cross section by about $80 - 100$% \[19\] (see also \[20\]).
A particularly interesting quantity is the transverse momentum distribution of the produced Higgs boson. The precise theoretical prediction of the $d\sigma/d p_T$ at the LHC is important for quantitative evaluation of the required measurement accuracies and detector performance. It is well-known that the fixed-order perturbative QCD is applicable when the Higgs transverse momentum is comparable to the $m_H$. Hovewer, the main part of the events is expected in the small-$p_T$ region ($p_T \ll m_H$), where the coefficients of the perturbative series in $\alpha_s$ are enhanced by powers of large logarithmic terms proportional to $\ln m_H^2/p_T^2$. Therefore reliable predictions at small $p_T$ can only be obtained if these terms will be resummed to all orders. Such procedure is called soft-gluon resummation \[21-23\] and has been performed in collinear calculations at leading logarithmic (LL), next-to-leading logarithmic (NLL) \[24\] and next-to-next-to-leading logarithmic (NNLL) \[25\] levels. Recently it was shown \[26\] that in the framework of $k_T$-factorization approach the soft gluon resummation formulas are the result of the approximate treatment of the solutions of the CCFM evolution equation (in the $b$-representation).
There are several additional motivations for our study of the Higgs production in the $k_T$-factorization approach. First of all, in the standard collinear approach, when the transverse momentum of the initial gluons is neglected, the transerse momentum of the final Higgs boson in $gg \to H$ subprocess is zero. Therefore it is necessary to include an initial-state QCD radiation to generate the $p_T$ distributions. It is well known at present that the $k_T$-factorization naturally includes a large part of the high-order perturbative QCD corrections \[27\]. This fact is illustrated more detailed in Figure 1, which is a schematical representation of a typical Higgs + jet production process. Figure 1 (a) shows the fixed-order perturbative QCD picture where the upper part of the diagram (above the dash-dotted line) corresponds to the $gg \to gH$ subrocess, and the lower part describes the gluon evolution in a proton. As the incoming gluons are assumed to have zero transverse momentum, the transverse momentum distributions of the produced Higgs and jet are totally determined by the properties of the ${\cal O}(\alpha_s^3)$ matrix element. In the $k_T$-factorization approach (Figure 1 (b)), the underlying partonic subprocess is $gg \to H$, which is formally of order ${\cal O}(\alpha_s^2)$. Some extra powers of $\alpha_s$ are hidden in the gluon evolution represented by the part of the diagram shown below the dash-dotted line. In contrast with the collinear approximation, the $k_T$-factorization takes into account the gluon transverse motion. Since the upper gluon in the parton ladder is not included in the hard interaction, its transverse momentum is now determined by the properties of the evolution equation only. It means that in the $k_T$-factorization approach the study of transverse momenta distributions in the Higgs production via gluon-gluon fusion will be direct probe of the unintegrated gluon distributions in a proton. In this case the transverse momentum of the produced Higgs should be equal to the sum of the transverse momenta of the initial gluons. Therefore future experimental studies at LHC can be used as further test of the non-collinear parton evolution.
In the previous studies \[26, 28, 29\] the $k_T$-factorization formalism was applied to calculate transverse momentum distribution of the inclusive Higgs production. The simplified solution of the CCFM equation in single loop approximation \[30\] (when small-$x$ effects can be neglected) were used in \[26\]. In such approximation the CCFM evolution is reduced to the DGLAP one with the difference that the single loop evolution takes the gluon transverse momentum $k_T$ into account. Another simplified solution of the CCFM equation was proposed in Ref. \[31\], where the transverse momenta of the incoming gluons are generated in the last evolution step (Kimber-Martin-Ryskin prescription). The calculations \[26, 29\] were done using the on-mass shell (independent from the gluon $k_T$) matrix element of the $gg\to H$ subprocess and rather the similar results have been obtained. In Ref. \[28\] in the framework of MC generator CASCADE \[32\] the off-mass-shell matrix element obtained by F. Hautmann \[33\] has been used with full CCFM evolution.
In present paper we investigate Higgs production at hadron colliders using the full CCFM-evolved unintegrated gluon densities \[28\]. We obtain the obvious expression for the $g^* g^* \to H$ off-mass-shell matrix element in the large $m_t$ limit apart from Ref. \[33\]. After that, we calculate the total cross section and transverse momentum distribution of the inclusive Higgs production at Tevatron and LHC. To illustrate the fact that in the $k_T$-factorization approach the main features of collinear higher-order pQCD corrections are taken into account effectively, we give theoretical predictions for the Higgs + one jet and Higgs + two jet production processes using some physically motivated approximation.
In Section 2 we recall the basic formulas of the $k_T$-factorization formalism with a brief review of calculation steps. In Section 3 we present the numerical results of our calculations and discussion. Finally, in Section 4, we give summary of our results.
Basic formulas
==============
We start from the effective Lagrangian for the Higgs boson coupling to gluons \[16\]: $${\cal L}_{\rm {eff}} = {\alpha_s \over 12 \pi}\left(G_F \sqrt 2\right)^{1/2} G_{\mu \nu}^a G^{a\,\mu \nu} H, \eqno (1)$$
where $G_F$ is the Fermi coupling constant, $G_{\mu \nu}^a$ is the gluon field strength tensor and $H$ is the Higgs field. The triangle vertex $T^{\mu \nu}(k_1,k_2)$ for two off-shell gluons having four-momenta $k_1$ and $k_2$ and color indexes $a$ and $b$ respectively, can be obtained easily from the Lagrangian (1): $$T^{\mu \nu}(k_1,k_2) = i \delta^{a b} {\alpha_s \over 3\pi} \left(G_F \sqrt 2\right)^{1/2} \left[ k_2^{\mu} k_1^{\nu} -
(k_1 \cdot k_2) g^{\mu \nu} \right]. \eqno (2)$$
To calculate the squared off-mass-shell matrix element for the $g^* g^* \to H$ subprocess it is necessary to take into account the non-zero virtualities of the initial gluons $k_1^2 = - {\mathbf k}_{1T}^2 \neq 0$, $k_2^2 = - {\mathbf k}_{2T}^2 \neq 0$. We have obtained[^1] $$|\bar {\cal M}|^2(g^* g^* \to H) = {\alpha_s^2(\mu^2)\over 576 \pi^2} G_F \sqrt 2
\left[ m_H^2 + {\mathbf k}_{1T}^2 + {\mathbf k}_{2T}^2 + 2 |{\mathbf k}_{1T}||{\mathbf k}_{2T}|\cos \phi \right]^2 \cos^2 \phi, \eqno (3)$$
where $\phi$ is the azimuthal angle between transverse momenta ${\mathbf k}_{1T}$ and ${\mathbf k}_{2T}$, the transverse momentum of the produced Higgs boson is ${\mathbf p}_{T} = {\mathbf k}_{1T} + {\mathbf k}_{2T}$ and the virtual gluon polarization tensor has been taken in the form \[7, 8\] $$\sum \epsilon^{\mu} \epsilon^{*\,\nu} = {k_T^{\mu} k_T^{\nu} \over {\mathbf k}_T^2 }. \eqno (4)$$
The cross section of the inclusive Higgs production $p\bar p \to H + X$ in the $k_T$-factorization approach can be written as $$\displaystyle d\sigma(p \bar p \to H + X) = \int {dx_1\over x_1} {\cal A}(x_1,{\mathbf k}_{1T}^2,\mu^2) d{\mathbf k}_{1T}^2 {d\phi_1 \over 2\pi} \times \atop
\displaystyle \times \int {dx_2\over x_2} {\cal A}(x_2,{\mathbf k}_{2T}^2,\mu^2) d{\mathbf k}_{2T}^2 {d\phi_2 \over 2\pi} d\hat \sigma(g^* g^* \to H), \eqno (5)$$
where $\hat \sigma(g^* g^* \to H)$ is the Higgs production cross section with off-mass-shell gluons, $x_1$ and $x_2$ are the longitudinal momentum fractions, and ${\cal A}(x,{\mathbf k}_{T}^2,\mu^2)$ is the unintegrated gluon distributions in a proton. Let $s = (p_1 + p_2)^2$ and $p_1$, $p_2$ are the four-vectors of the incoming protons. Then the differential cross section reads $$\displaystyle {d\sigma(p \bar p \to H + X)\over dy_H} = \int {\alpha_s^2(\mu^2)\over 288 \pi} {G_F \sqrt 2 \over x_1 x_2 m_H^2 s} \left[m_H^2 + {\mathbf p}_T^2\right]^2 \cos^2 \phi_2 \times \atop
\displaystyle \times {\cal A}(x_1,{\mathbf k}_{1T}^2,\mu^2) {\cal A}(x_2,{\mathbf k}_{2T}^2,\mu^2) d{\mathbf k}_{1T}^2 d{\mathbf k}_{2T}^2 {d\phi_2 \over 2\pi}, \eqno (6)$$
where $y_H$ is the Higgs rapidity in the proton-proton c.m. frame. The longitudinal momentum fractions $x_1$ and $x_2$ are given by $$x_1 = \sqrt{m_H^2 + {\mathbf p}_T^2\over s} \exp (y_H),\quad x_2 = \sqrt{m_H^2 + {\mathbf p}_T^2\over s} \exp (-y_H). \eqno (7)$$
If we average the expression (6) over transverse momenta ${\mathbf k}_{1T}$ and ${\mathbf k}_{2T}$ and take the limit ${\mathbf k}_{1T}^2 \to 0$, ${\mathbf k}_{2T}^2 \to 0$, we obtain well-established expression \[2\] for Higgs production cross section in leading-order perturbative QCD: $$d\sigma(p \bar p \to H + X) = {\alpha_s^2(\mu^2)\over 576 \pi} G_F \sqrt 2 {m_H^2 \over x_1 x_2 s} x_1{\cal G}(x_1,\mu^2) x_2{\cal G}(x_2,\mu^2) dy_H, \eqno (8)$$
where $x {\cal G}(x,\mu^2)$ is the usual (collinear) gluon density which is related with the unintegrated gluon distribution ${\cal A}(x,{\mathbf k}_{T}^2,\mu^2)$ by $$x {\cal G}(x,\mu^2) \sim \int {\cal A}(x,{\mathbf k}_{T}^2,\mu^2) d{\mathbf k}_{T}^2. \eqno (9)$$
Here the sign $\sim$ indicates, that there is no strict equality between the left and the right parts of the equation (9)[^2].
The multidimensional integration in the expression (6) has been performed by means of the Monte Carlo technique, using the routine VEGAS \[35\]. The full C$++$ code is available from the authors on request[^3].
Numerical results and discussion
================================
Inclusive Higgs production
--------------------------
We now are in a position to present our numerical results. First we describe our theoretical input and the kinematical conditions. Besides the Higgs mass $m_H$, the cross section (6) depend on the uninterated gluon distribution ${\cal A}(x,{\mathbf k}_{T}^2,\mu^2)$ and the energy scale $\mu$. The new fits of the unintegrated gluon density (J2003 set 1 — 3) have been recently presented \[28\]. The full CCFM equation in a proton was solved numerically using a Monte Carlo method. The input parameters were fitted to describe the proton structure function $F_2(x,Q^2)$. Since these gluon densities reproduce well the forward jet production at HERA, charm and bottom production data at Tevatron \[28\] and charm and $J/\psi$ production at LEP2 energies \[35\], we use it (namely J2003 set 1) in our calculations. As is often done for Higgs production, we choose the renormalization and factorization scales to be $\mu = \xi m_H$, and vary the scale parameter $\xi$ between $1/2$ and $2$ about the default value $\xi = 1$. Also we use LO formula for the strong coupling constant $\alpha_s(\mu^2)$ with $n_f = 4$ active quark flavours and $\Lambda_{\rm QCD} = 200$ MeV, such that $\alpha_s(M_Z^2) = 0.1232$.
In Figure 2 and 3 we display our prediction for the transverse momentum and rapidity distributions of the inclusive Higgs production at the LHC ($\sqrt s = 14$ TeV). The calculations were done for four choices of the Higgs boson mass under interest in the Standard Model with default scale $\mu^2 = m_H^2$. The solid, dashed, dash-dotted and dotted lines correspond $m_H = 125$ GeV, $m_H = 100$ GeV, $m_H = 150$ GeV (where $WW$ decay channel is dominant) and $m_H = 200$ GeV (above $WW$ and $ZZ$ decay tresholds), respectively. One can see that mass effects are present only at low $p_T < m_H$, whereas all curves practically coincide at large transverse momenta. We note that our predictions which correspond to the Higgs mass $m_H = 125$ GeV slightly underestimate results obtained in the combined fixed-order + resummed approach \[37\]. In this approach fixed-order predictions (at LO or NLO level) and resummed ones (at NLL or NNLL level, respectively) have to be consistenly matched at moderate $p_T$. The NNLL + NLO results \[25\] are smaller than NLL + LO ones \[24\] by about $20$% at low transverse momenta. We see that our predictions lie below NNLL + NLO calculations by about $15$% in this kinematical region. Usage the doubly unintegrated gluon distributions results in more flat behaviour of the $p_T$-distribution \[29\] in comparison with both our and NNLL + NLO predictions.
We note also that the peak in the transverse momentum distribution occurs at a smaller value of $p_T$ compared to the NNLL + NLO calculations. The location of this peak as a function of Higgs boson mass is shown in Figure 4. We find that at $m_H = 125$ GeV the peak occurs at $p_T \sim 10$ GeV, whereas NNLL + NLO line peaks at $p_T \sim 15$ GeV \[37\]. The similar effect has been obtained \[29\] when doubly unintegrated gluon distributions were used.
The total cross sections of the inclusive Higgs production at Tevatron ($\sqrt s = 1.96$ TeV) and LHC conditions as function Higgs mass are plotted in Figure 5 and 6 in the mass range $m_H = 100 - 200$ GeV. The solid lines are obtained by fixing both the factorization and renormalization scales at the default value $\mu = m_H$. In order to estimate the theoretical uncertainties in our predictions, we vary the unphysical parameter $\xi$ as indicated above. These uncertainties are presented by upper and lower dashed lines. We find that our default predictions agree very well with recent NNLO results \[18\]. For example, when Higgs boson mass is $m_H =
120$ GeV, our calculations give $\sigma = 0.84$ pb at Tevatron and $\sigma = 35.9$ pb at LHC. However, the scale dependences are rather large. At LHC energy, it changes from about $20$% when $m_H = 100$ GeV, to about $50$% when $m_H = 200$ GeV. At Tevatron, it range from $40$% to $50$%, respectively. This fact indicates the necessarity of high-order corrections inclusion in the $k_T$-factorization formalism. But one should note that in the $k_T$-factorization the role of such correction is very different in comparison with the corrections in the collinear approach, since part of the standard high-order corrections are already included at LO level in $k_T$-factorization[^4]. At the same time the theoretical uncertainties of the collinear QCD calculations, after inclusion of both NNLO corrections and soft-gluon resummation at the NNLL level, are about $10$% in the low mass range $m_H < 200$ GeV \[18\].
Higgs production in association with jets
-----------------------------------------
Now we demonstrate how $k_T$-factorization approach can be used to calculate the semi-inclusive Higgs production rates. The produced Higgs boson is accompanied by a number of gluons radiated in the course of the gluon evolution. As it has been noted in Ref. \[38\], on the average the gluon transverse momentum decreases from the hard interaction block towards the proton. As an approximation, we assume that the gluon $k'$ closest to the Higgs compensates the whole transverse momentum of the virtual gluon participating in the gluon fusion, i.e. ${\mathbf k'}_T \simeq - {\mathbf k}_T$ (see Figure 1). All the other emitted gluons are collected together in the proton remnant, which is assumed to carry only a negligible transverse momentum compared to ${\mathbf k'}_T$. This gluon gives rise to a final hadron jet with ${\mathbf p}_{{\rm jet}\,T} = {\mathbf k'}_T$.
From the two hadron jets represented by the gluons from the upper and lower evolution ladder we choose the one carrying the largest transverse momentum, and then compute Higgs with an associated jet cross sections at the LHC energy. We have applied the usual cut on the final jet transverse momentum $|{\mathbf p}_{{\rm jet}\,T}| > 20$ GeV. Our predictions for the transverse momentum distribution of the Higgs + one jet production are shown in Figure 7. As in the inclusive Higgs production case, we test four different $m_H$ values in the transverse momentum ditributions. All curves here are the same as in Figure 2. One can see the shift of the peak position in the $p_T$ distributions in comparison with inclusive production, which is direct consequence of the $|{\mathbf p}_{{\rm jet}\,T}| > 20$ GeV cut. We note that the rapidity interval between the jet and the Higgs boson is naturally large. It is because there is angular ordering in the CCFM evolution, which is equivalent to an ordering in rapidity of the emitted gluons.
The investigation of the different azimuthal correlations between final particles in semi-inclusive Higgs production provides many interesting insights. In particular, studying of these quantities are important to clean separation of weak-boson fusion and gluon-gluon fusion contributions. To demonstrate the possibilities of the $k_T$-factorization approach, we present here the two azimuthal angle distributions. First, we calculate azimutal angle distribution between the Higgs boson and final jet transverse momenta in the Higgs + one jet production process. Second, we calculate azimuthal angle distributions between the two final jet transverse momenta in the Higgs + two jet production process. In this case the Higgs boson is centrally located in rapidity between the two jets and it is very far from either jet, and the kinematical cut $|{\mathbf
p}_{{\rm jet}\,T}| > 20$ GeV was applied for both final jets. We set no cuts on the jet-jet invariant mass. Our results are shown in Figure 8 and 9, respectively. Figure 8 demonstrated roughly the back-to-back Higgs + one jet production. In Figure 9 we obtained a dip at $90$ degrees in jet-jet azimuthal correlation, which is characteristic for loop-induced Higgs coupling to gluons \[39\]. The fixed-order perturbative QCD calculations of the $gg \to gg H$ subprocess give the similar result \[16\]. However, as it was already mentioned above, such calculations are very cumbersome even at leading order. The evaluation of the radiative corrections at ${\cal O}(\alpha_s)$ to Higgs + two jet production would imply the calculation of up to hexagon quark loops and two-loop pentagon quark loops, which are at present unfeasible \[20\]. We note that contribution from the weak-boson fusion to the Higgs + two jet production has flat behavior of the jet-jet angular distribution \[16, 20\].
To illuminate the sensitivity of the Higgs production rates to the details of the unintegrated gluon distribution, we repeated our calculations for jet-jet angular correlations using J2003 set 2 gluon density \[28\] (dashed line in Figure 9). This density takes into account the singular and non-singular terms in the CCFM splitting function, where the Sudakov and non-Sudakov form factors were modified accordinly. We note that J2003 set 1 takes into account only singular terms. Both these sets describe the proton structure function $F_2(x,Q^2)$ at HERA reasonable well. However, one can see the very large discrepancy (about order of magnitude) between predictions of J2003 set 1 and set 2 unintegrated gluon densities. The similar difference was claimed \[28\] for charm and bottom production at Tevatron also. This fact clearly indicates again that high-order corrections to the leading order $k_T$-factorization are important and should be developed for future applications.
Conclusions
===========
We have considered the Higgs boson production via gluon-gluon fusion at high energy hadron colliders in the framework of the $k_T$-factorization approach. Our interests were focused on the Higgs boson total cross section and transverse momenta distributions at Tevatron and LHC colliders. In our numerical calculations we use the J2003 set 1 unintegrated gluon distribution, which was obtained recently from the full CCFM evolution equation.
We find that $k_T$-factorization gives the very close to NNLO pQCD results for the inclusive Higgs production total cross sections. It is because the main part of the high-order collinear pQCD corrections is already included in the $k_T$-factorization. Also we have demonstrated that $k_T$-factorization gives a possibility to investigate the associated Higgs boson and jets production in much more simple manner, than it can be done in the collinear factorization. Using some approximation, we have calculated transverse momentum distributions and investigated the Higgs-jet and jet-jet azimuthal correlations in the Higgs + one or two jet production processes. However, the scale dependence of our calculations is rather large (of the order of $20 - 50$%), which indicates the importance of the high-order correction within the $k_T$-factorization approach. These corrections should be developed and taken into account in the future applications.
We point out that in this paper we do not try to give a better prediction for Higgs production than the fixed-order pQCD calculations. The main advantage of our approach is that it is possible to obtain in straighforward manner the analytic description which reproduces the main features of the collinear high-order pQCD calculations[^5]. But in any case, the future experimental study of such processes at LHC will give important information about non-collinear gluon evolution dynamics, which will be useful even for leading-order $k_T$-factorization formalism.
Acknowledgements
================
The authors are very grateful to H. Jung for possibility to use the CCFM code for unintegrated gluon distributions in our calculations, for reading of the manuscript and useful discussion. We thank S.P. Baranov for encouraging interest and helpful discussions. N.Z. thanks P.F. Ermolov for support and the DESY directorate for the hospitality and support.
[36]{}
ATLAS Collaboration, Technical Design Report, Vol. 2, CERN/LHCC/99-15, 1999;\
CMS Collaboration, Technical Proposal, CERN/LHCC/94-38, 1994. F. Wilczek, Phys. Rev. Lett. [**39**]{}, 1304 (1977);\
H.M. Georgi, S.L. Glashow, M.E. Machacek and D.V. Nanopoulos, [*ibid.*]{} Phys. Rev. Lett. [**40**]{}, 692 (1978);\
J.R. Ellis, M.K. Gaillard, D.V. Nanopoulos and C.T. Sachrajda, Phys. Lett. [**B83**]{}, 339 (1979);\
T.G. Rizzo, Phys. Rev. [**D22**]{}, 178 (1980); [**D22**]{}, 1824 (1980). D. Graudenz, M. Spira and P.M. Zervas, Phys. Rev. Lett. [**70**]{}, 1372 (1993);\
M. Spira, A. Djouadi, D. Graudenz and P.M. Zervas, Nucl. Phys. [**B453**]{}, 17 (1995). N. Kauer, T. Plehn, D. Rainwater and D. Zeppenfeld, Phys. Lett. [**B503**]{}, 113 (2001);\
T. Plehn, D. Rainwater and D. Zeppenfeld, Phys. Rev. [**D61**]{}, 093005 (2000);\
D. Rainwater and D. Zeppenfeld, JHEP [**9712**]{}, 005 (1997). V.N. Gribov and L.N. Lipatov, Yad. Fiz. [**15**]{}, 781 (1972);\
L.N. Lipatov, Sov. J. Nucl. Phys. [**20**]{}, 94 (1975);\
G. Altarelly and G. Parizi, Nucl. Phys. [**B126**]{}, 298 (1977);\
Y.L. Dokshitzer, Sov. Phys. JETP [**46**]{}, 641 (1977). E.A. Kuraev, L.N. Lipatov and V.S. Fadin, Sov. Phys. JETP [**44**]{}, 443 (1976);\
E.A. Kuraev, L.N. Lipatov and V.S. Fadin, Sov. Phys. JETP [**45**]{}, 199 (1977);\
I.I. Balitsky and L.N. Lipatov, Sov. J. Nucl. Phys. [**28**]{}, 822 (1978). V.N. Gribov, E.M. Levin and M.G. Ryskin, Phys. Rep. [**100**]{}, 1 (1983). E.M. Levin, M.G. Ryskin, Yu.M. Shabelsky and A.G. Shuvaev, Sov. J. Nucl. Phys. [**53**]{}, 657 (1991). S. Catani, M. Ciafoloni and F. Hautmann, Nucl. Phys. [**B366**]{}, 135 (1991). J.C. Collins and R.K. Ellis, Nucl. Phys. [**B360**]{}, 3 (1991). M. Ciafaloni, Nucl. Phys. [**B296**]{}, 49 (1988);\
S. Catani, F. Fiorani and G. Marchesini, Phys. Lett. [**B234**]{}, 339 (1990);\
S. Catani, F. Fiorani and G. Marchesini, Nucl. Phys. [**B336**]{}, 18 (1990);\
G. Marchesini, Nucl. Phys. [**B445**]{}, 49 (1995). J.R. Forshaw and A. Sabio Vera, Phys. Lett. [**B440**]{}, 141 (1998). B.R. Webber, Phys. Lett. [**B444**]{}, 81 (1998). G.P. Salam, JHEP [**03**]{}, 009 (1999). B. Andersson [*et al.*]{} (Small-$x$ Collaboration), Eur. Phys. J. [**C25**]{}, 77 (2002). V. Del Duca, W. Kilgore, C. Olear, C. Schmidt and D. Zeppenfeld, Nucl. Phys. [**B616**]{}, 367 (2001); Phys. Rev. [**D67**]{}, 073003 (2003). J.R. Ellis, M.K. Gaillard and D.V. Nanopoulos, Nucl. Phys. [**B106**]{}, 292 (1976);\
M.A. Shifman, A.I. Vainstein, M.B. Voloshin and V.I. Zakharov, Yad. Fiz. [**30**]{}, 1368 (1979). R.V. Harlander and W.B. Kilgore, Phys. Rev. Lett. [**88**]{}, 201801 (2002);\
C. Anastasiou and K. Melnikov, Nucl. Phys. [**B646**]{}, 220 (2002);\
V. Ravindran, J. Smith and W.L. van Neerven, Nucl. Phys. [**B665**]{}, 325 (2003). S. Dawson, Nucl. Phys. [**B359**]{}, 283 (1991);\
A. Djouadi, M. Spira and P.M. Zervas, Phys. Lett. [**B264**]{}, 440 (1991). V. Del Duca, hep-ph/0312184. J.C. Collins and D.E. Soper, Nucl. Phys. [**B193**]{}, 381 (1981); [*ibid.*]{} [**B213**]{}, 545 (1983); [**B197**]{}, 446 (1982). J.C. Collins, D.E. Soper and G. Sterman, Nucl. Phys. [ **B250**]{}, 199 (1985). R.K. Ellis and S. Veseli, Nucl. Phys. [**B511**]{}, 649 (1998);\
R.K. Ellis, D.A. Ross and S. Veseli, [*ibid.*]{} Nucl. Phys. [**B503**]{}, 309 (1997). S. Catani, E. D’Emilio and L. Trentadue, Phys. Lett. [ **B211**]{}, 335 (1988);\
R.P. Kauffmann, Phys. Rev. [**D45**]{}, 1512 (1992). D. de Florian and M. Grazzini, Phys. Rev. Lett. [**85**]{}, 4678 (2000); Nucl. Phys. [**B616**]{}, 247 (2001). A. Gawron and J. Kwiecinski, Phys. Rev. [**D70**]{}, 014003 (2004). M.G. Ryskin, A.G. Shuvaev and Y.M. Shabelski, Phys. Atom. Nucl. [**64**]{}, 120 (2001). H. Jung, Mod. Phys. Lett. [**A19**]{}, 1 (2004). G. Watt, A.D. Martin and M.G. Ryskin, Phys. Rev. [**D70**]{}, 014012 (2004), Erratum: [*ibid.*]{} [**D70**]{}, 079902 (2004), hep-ph/0309096. B.R. Webber, Nucl. Phys. Proc. Suppl. [**C18**]{}, 38 (1991);\
G. Marchesini and B.R. Webber, Nucl. Phys. [**B386**]{}, 215 (1992);\
A. Gawron and J. Kwiecinski, Acta. Phys. Polon. [**B34**]{}, 133 (2003). M.A. Kimber, A.D. Martin and M.G. Ryskin, Phys. Rev. [ **D63**]{}, 114027 (2001). H. Jung, Comput. Phys. Comm. [**143**]{}, 100 (2002). F. Hautmann, Phys. Lett. [**B535**]{}, 159 (2002). J. Andersen [*et al.*]{} (Small-$x$ Collaboration), Eur. Phys. J. [**C35**]{}, 67 (2004). G.P. Lepage, J. Comput. Phys. [**27**]{}, 192 (1978). A.V. Lipatov and N.P. Zotov, hep-ph/0412275, submitted to Eur. Phys. J. C. S. Catani, D. de Florian and M. Grazzini, Nucl. Phys. [**B596**]{}, 299 (2001);\
JHEP [**0201**]{}, 015 (2002);\
G. Bozzi, S. Catani, D. de Florian and M. Grazzini, Phys. Lett. [ **B564**]{}, 65 (2003);\
S. Catani, D. de Florian, M. Grazzini and P. Nason, JHEP [**0307**]{}, 028 (2003). S.P. Baranov and N.P. Zotov, Phys. Lett. [**B491**]{}, 111 (2000). T. Plehn, D. Rainwater and D. Zeppenfeld, Phys. Rev. Lett. [**88**]{}, 051801 (2002).
[^1]: We would like to remark that the expression (3) differs from the one obtained in Ref. \[33\].
[^2]: See Refs. \[15, 34\] for more details.
[^3]: [email protected]
[^4]: See also \[15, 34\] for more detailed discussion.
[^5]: In this part our conclusions coincide with ones from Ref. \[29\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The possibility of observing formation of hidden-charm pentaquarks as $s$-channel resonances in antiproton - deuteron collisions is discussed. It is pointed out that the masses of the reported by LHCb pentaquark resonances in the $J/\psi \, p$ channel are very close to a special value of the mass at which formation of a pentaquark by antiproton incident on a deuteron at rest requires exactly the same momentum of the $\bar p$ as needed for the formation in the $s$ channel of the charmonium resonance in $\bar p p$ collisions with the proton being at rest. For this reason the former process can be rather completely described within the notion of the deuteron being a shallow bound state of two nucleons without resorting to models describing its short-distance structure. It is argued that a similar kinematical coincidence can be expected for (yet) hypothetical pentaquark resonances in the $\eta_c \, N$ channel, and that these can be sought for once antiproton - deuterium collisions become available for experimentation.'
---
[**William I. Fine Theoretical Physics Institute\
University of Minnesota\
**]{}
FTPI-MINN-19/10\
UMN-TH-3819/19\
March 2019\
[**Hidden-charm pentaquark formation in antiproton - deuterium collisions\
**]{} William I. Fine Theoretical Physics Institute, University of Minnesota,\
Minneapolis, MN 55455, USA\
School of Physics and Astronomy, University of Minnesota, Minneapolis, MN 55455, USA\
and\
Institute of Theoretical and Experimental Physics, Moscow, 117218, Russia\
The rapidly-growing family of exotic multiquark states containing a heavy quark-antiquark pair, $c \bar c$ or $b \bar b$, has recently been expanded to baryons by the observation [@lhcb1; @lhcb2] of the resonances $P_c(4380)$ and $P_c(4450)$ in the hidden-charm pentaquark channel $J/\psi \, p$ produced in the decays $\Lambda_b \to J/\psi \, p \, K^-$. The new states received explanations in a number of theoretical models, many of which can be found in the recent review [@ghmwzz]. Clearly, a further study of these resonances as well as a search for other baryonic states of similar nature with hidden heavy flavor present a great interest for understanding multi-quark systems. Such studies would certainly be facilitated if other sources of the pentaquark states could be found besides the production in the LHC environment. In particular, it has been pointed out [@wlz; @kv; @kr] that the observed pentaquark resonances should be produced in the $s$-channel in photoproduction on hydrogen, i.e. in $\gamma +p$ collisions. This conclusion largely does not depend on specific models of the ‘internal’ dynamics of the pentaquark states and is based on the mere fact of the coupling of the resonances to the $J/\psi \, p$ channel and the known interaction of the charmonium state $J/\psi$ with photon. One can readily notice then that in a similar manner the formation of the hidden-charm pentaquarks can be effected by arranging an interaction of a nucleon with any other state that couples to charmonium. It is the purpose of this paper to point out that a realistic and largely model-independent possibility of producing hidden-charm pentaquarks is offered by the collisions of antiprotons with a deuterium target, which process is possible due to coupling of charmonium states to the $p \bar p$ channel, e.g. $J/\psi \to p \bar p$ [@pdg]. Furthermore, this process is possible with other states of charmonium, e.g. it can be used for a search of resonances in the $\eta_c \, n$ channel, which, as will be discussed, has a certain advantage over the $J/\psi$ due to a significantly larger $p \bar p$ decay width of $\eta_c$. Moreover, the discussed formation of hidden-charm pentaquarks is not suppressed by the ‘softness’ of the deutron due to a kinematical coincidence, thus making it plausible that actual experimental searches can be performed by the PANDA experiment [@panda] at FAIR using a deuterium target.
The process discussed in this paper is schematically shown in Fig. 1. Clearly, the dominant part of the wave function of the nucleons inside the deuteron, treated as a loose bound state, can be effective only if the kinematical constraints in the graph do not require the relative momentum of the neutron and the proton in the trangle to be large in comparison with the inverse nucleon size. Considering the process in the rest frame of the deutron (which frame coincides with the lab frame in a realistic experiment, e.g. in PANDA), one readily finds that both nucleons in the triangle can be on-shell and simultaneously at rest if the mass $M$ of the pentaquark is related to the mass $m$ of the charmonium state and the nucleon mass $\mu$ as $M=M_0(m)$ with M\_0\^2(m) = 2m\^2 + \^2 . \[m0\] (The small binding energy $\epsilon = -2.22\,$MeV in the deuteron is obviously neglected in this expression.) In particular, for the charmonium mass of $J/\psi$ and $\eta_c$ the special value of the pentaquark mass is estimated as respectively $M_0(m_{J/\psi}) = 4.48\,$GeV and $M_0(m_{\eta_c}) = 4.33\,$GeV. It can be readily noted that the former of these values is quite close to the measured mass of $P_c(4450)$, while the latter, as will be argued below, is close to the expected mass of a possible pentaquark resonance in the $\eta_c \, N$ channel. With the pentaquark mass within a certain range around the special value $M_0$ the triangle graph of Fig. 1 is dominated by the region of low momentum $|\vec q|$ of each of the two nucleons, and the wave function of the deuteron can be approximated by that for the free motion (q) = [ \^2 + q\^[ 2]{}]{} , \[ff\] where $\kappa = \sqrt{ \mu |\epsilon|} \approx 46\,$MeV is the virtual momentum corresponding to the binding energy $\epsilon$ in the deuteron. Clearly, this approximation is valid only as long as the momentum is smaller than the hadronic scale $\Lambda$. Choosing, for an estimate $\Lambda \approx 200\,$MeV, one can find that the condition for the on-shell nucleons in the triangle to have momentum $|\vec q| < \Lambda$ corresponds to the range of the pentaquark resonance masses approximately $4.22 - 5.0$GeV for a $J/\psi \, N$ resonance and approximately $4.07 - 4.81$GeV for a resonance in the $\eta_c \, N$ channel, with the amplitude decreasing with $M$ away from $M_0$ toward both ends of each of these ranges. Outside the mass range where Eq.(\[ff\]) can be applied any estimate of the amplitude becomes dependent on the model for the short-distance wave function inside the deuteron, and in either case the amplitude is small at those values of the mass. Thus in what follows only the region of $M$ sufficiently close to $M_0$ is considered.
Furthermore, the spin and orbital partial-wave structure of the coupling of $P_c$ to charmonium and a nucleon is not known. Even the spin-orbital structure of the vertex for the decay $J/\psi \to p \bar p$ is known only partially [@babar]. For this reason it would be quite premature to consider various spin and orbital-momentum waves in the vertices in Fig. 1. Such consideration involving an off-shell behavior of the vertices would be necessary for a full calculation. Here however, for an estimate, only the absorptive part of the amplitude is evaluated, corresponding to the neutron and the charmonium in the loop being on-shell. Then the kinematical factors associated with the orbital-momentum partial waves in the $P_c \to (c \bar c) \, N$ and the $(c \bar c) \to p \bar p$ vertices are the same as the actual on-shell decay processes, and in fact can be absorbed in the definition of the corresponding vertices, once the rate resulting from the mechanism of Fig. 1 is expressed in terms of the widths $\Gamma[P_c \to (c \bar c) \, N]$ and $\Gamma[(c \bar c) \to p \bar p]$. The spin-dependent factors result in an overall factor of order one and are neglected in the estimates discussed here. Thus in effect all involved particles are considered here as spinless.
The cross section for the process $\bar p + d \to P_c \to anything$ can be evaluated using the well known Breit-Wigner expression (see e.g. in Ref. [@pdg], Sec. 48) in terms of the branching fraction $Br(P_c \to \bar p \, d)$. At the maximum of the resonance the expression reads as (|p + d P\_c anything) = [2 J +1 6]{} [4 k\^2]{} Br(P\_c |p d) , \[bw\] where $J$ is the spin of the resonance and $k$ is the c.m. momentum in the decay $P_c \to \bar p \, d$: k= [ 2 M]{} . \[mk\] The factor six in the denominator of the overall spin factor in Eq.(\[bw\]) is the combined total number of spin states for the antiproton and the deuteron. Setting, as discussed above, the spin factor to one (the value of $J$ is not quite known anyway), and considering that for the pentaquark mass in the ballpark of 4.3 - 4.4GeV the value of $k$ is approximately 1.6GeV, one can find a simple relation for the maximum resonance cross section (|p + d P\_c anything) Br(P\_c |p d) 2 10\^[-27]{} [cm]{}\^2 . \[xs\]
In order to evaluate the (absorptive part of) the amplitude given by the graph of Fig. 1, we use nonrelativistic normalization for wave functions of all the particles in the graph and denote as $g$ the vertex for the decay $J/\psi \to p \bar p$ (the solid blob in Fig. 1) and as $h$ the vertex for $P_c \to (c \bar c) \, n$ (the open rectangle in Fig. 1). The rates of the relevant decays are then expressed, in the chosen normalization,[^1] as (J/p |p)= [|g|\^2 8 ]{} m \[psiw\] and (P\_c J/ n) = [|h\^2| 8 M\^4]{} (M\^2 - m\^2+ \^2) (M\^2 +m\^2 - \^2) . \[pcw\]
The absorptive part of the amplitude of $P_c \to \bar p \, d$, given by the triangle graph of Fig. 1, can be calculated using the wave function of the deuteron in Eq.(\[ff\]) by means of ordinary (Heitler) perturbation theory in the rest frame of the deuteron, which readily accommodates the nonrelativistic wave function (\[ff\]) as well as the relativistic motion of some of the involved particles.[^2] The expression for the absorptive part $A \equiv -i {\rm Abs} [A(P_c \to \bar p \, d)]$ reads as A = h g (E\_[(c |c)]{} + E\_[n]{} - E\_[Pc]{}) (q) [ d\^3 q (2 )\^3]{} , \[absa\] where ${\vec q}$ is the momentum of the neutron, so that the relevant energies in the argument of the $\delta$ function are given as: for the neutron, $E_n=\sqrt{\vec q^{\,2} + \mu^2}$, for the $(c \bar c)$ charmonium, $E_{(c \bar c)}= \sqrt{(\vec p - \vec q)^2 + m^2}$ with $\vec p$ being the momentum (in the rest frame of the deuteron) of the pentaquark (and the same as the momentum of the antiproton). Finally, $E_{P_c}$ is the energy of the produced pentaquark, E\_[P\_c]{} = [M\^2 + 3 \^2 4 ]{}, \[epc\] so that $|\vec p | = \sqrt{E_{P_c}^2 - M^2}$. It can be noticed that the momentum $\vec p$ and the energy $E_{P_c}$ of the pentaquark are rather large: at $M \approx 4.4\,$GeV one finds $|\vec p| \approx 3.8\,$GeV and $E_{P_c} \approx 5.8$GeV.
For further calculation it is convenient to choose the $z$ axis of the coordinate system along the momentum $\vec p$ and denote $p_3 = p$, and to consider separately the longitudinal component $q_3$ of the vector $\vec q$ and its transverse part $\vec q_\perp$. The integration in Eq.(\[absa\]) over the transverse part can be used to remove the $\delta$ function: (E\_[(c |c)]{} + E\_[n]{} - E\_[Pc]{}) d\^2 q\_= 2 [ E\_[(c |c)]{} E\_[n]{} E\_[Pc]{} ]{} , \[iperp\] where the value of $q_\perp$ is found from the energy conservation equation + - E\_[P\_c]{}=0 . \[econ\] Clearly, this equation has a solution for $q_\perp^2$ in the domain, where the expression in the l.h.s is negative at $q_\perp=0$, which requires that $q_3 > Q$ with $Q$ being the solution for the boundary of this domain, + - E\_[P\_c]{}=0 . \[qz\] The remaining in Eq.(\[absa\]) integration over $q_3$ runs to the right from $q_3=Q$: A= h g [ E\_[(c |c)]{} E\_[n]{} E\_[Pc]{} ]{} \_[Q]{}\^[\~]{} [dq\_3 q\_\^2 + q\_3\^2]{} , \[an\] where the peripheral wave function (\[ff\]) is used. The integral in the latter formula converges at large $q_3$ and by itself does not require an ultraviolet cutoff $\Lambda$. However, as previously discussed, the expression for the wave function is applicable only for sufficiently small momentum: $|\vec q | < \Lambda$. Thus the discussed calculation can be justified only if the absolute value of $|Q|$ satisfies the same condition, and this criterion is used in the quoted above estimate of the range of applicability in terms of the pentaquark mass. It can be also noted that the condition $Q=0$ defines the special value of the pentaquark mass in Eq.(\[m0\]).
The components of the momentum $\vec q$ are much smaller than all other quantities in Eq.(\[econ\]). Therefore one can simplify the equation by expanding the l.h.s to the second order in these components resulting in the equation (\[econ\]) taking the form q\_\^2 ( + ) + ( q\_3\^2 - Q\^2 ) ( + [m\^2 p\^2+m\^2]{} ) - 2 (q\_3 - Q) p = 0 , \[econ2\] which relation defines $q_\perp^2$ in the integral in Eq.(\[an\]) as a function of $q_3$. Using the solution for the latter equation, one readily finds the integral in the logarithmic approximation: A h g [ E\_[Pc]{} ]{} [ + 2 p ]{} , \[anf\] where, using the approximation of small $Q$, the factors $E_{(c \bar c)}$ and $E_n$ from Eq.(\[iperp\]) are replaced by their values at $Q=0$, i.e. respectively by $\sqrt{p^2+m^2}$ and $\mu$.
The formula (\[anf\]) describes only the absorptive part of the amplitude of the decay $P_c \to \bar p \, d$. However, given the unresolvable at present uncertainty in evaluating the dispersive part, the best that could be done is to make an estimate of the rate including only the absorptive contribution: (P\_c |p + d) [k (M\^4 - 9 \^4) M\^3]{} , \[gpc\] where $k$ is the c.m. momentum in the decay and is given by Eq.(\[mk\]).
Using in the expression (\[gpc\]), the result in Eq.(\[anf\]), and determining the factors $g$ and $h$ from the Eqs. (\[psiw\]) and (\[pcw\]), one thus finds the following somewhat lengthy formula for the branching fraction entering Eq.(\[xs\]) for the cross section && Br (P\_c |p + d)\
&&Br \[P\_c (c |c) + n\] [(M\^2 -3 \^2) (M\^4 - 10 M\^2 \^2 +16 m\^2 \^2 + 9 \^4) m (M\^4 -m\^4 +2 m\^2 \^2 - \^4 ) (M\^2+ 3 \^2)]{}\
&& [\[4 \^2 + (M\^4 - 10 M\^2 \^2 +16 m\^2 \^2 + 9 \^4)\^[1/2]{}\]\^2 (M\^4 - 10 M\^2 \^2 + 9 \^4)\^[1/2]{} (M\^4 + m\^4 +\^4 - 2 M\^2 m\^2 - 2 M\^2 \^2 - 2 m\^2 \^2)\^[1/2]{}]{} L\^2 , \[fres\] with the logarithmic factor $L$ given by L= = , \[lf\] and the momentum $p$ is expressed as $p= \sqrt{E_{(c \bar c)}^2 - M^2}$. The formulas (\[fres\]) and (\[lf\]) take a remarkably simple form for the special value of the pentaquark mass given by Eq.(\[m0\]) and corresponding to $Q=0$: Br (P\_c |p + d) Br \[P\_c (c |c) + n\] \^2 . \[sres\] Given that the full expressions only moderately depend on $M$ near the special value $M_0$, the latter simplified formula can be used instead for approximate estimates of the cross section in Eq.(\[xs\]). Proceeding in this way one finds for a pentaquark resonance in the $J/\psi$ channel the estimate Br (P\_c |p + d) 0.6 Br (P\_c J/+ n) [( J/p |p ) ]{} 10\^[-7]{} Br (P\_c J/+ n) , \[psie\] where the experimental value [@pdg] $\Gamma(J/\psi \to p \bar p \, ) \approx 0.2\,$keV is used. The branching fraction for the pentaquark decay to $J/\psi + N$ is not known. If it is in the ballpark of 10%, the latter estimate translates into the cross section in Eq.(\[xs\]) of the order of $10^{-35}$cm$^2$, which looks quite challenging for the expected in the near future experimental capability.
The situation however may be more encouraging if by analogy[^3] with the pentaquark(s) decaying to $J/\psi + N$ there also exist one or more pentaquarks decaying to $\eta_c + N$, whose mass is by about 100MeV lower, and is also close to the corresponding value $M_0(m_{\eta_c})$. Applying the same estimate (\[sres\]) at the mass $m_{\eta_c}$ one finds Br (P\_c |p + d) 0.7 Br (P\_c \_c + n) [( \_c p |p ) ]{} 3 10\^[-5]{} Br (P\_c \_c + n) , \[ece\] where the enhancement mostly comes from a much larger (in absolute terms) $p \bar p$ decay width of $\eta_c$, $\Gamma(\eta_c \to p \bar p ) \approx 48 \,$keV. The expected cross section in the resonance maximum then amounts to a significant fraction of $10^{-32}$cm$^2$ \[assuming $Br (P_c \to \eta_c + n) \sim 10\%$\] and , if such pentaquark states do indeed exist, their search appears to be quite feasible at PANDA provided that a deuterium target can be used.
In summary. The formation of hidden-charm pentaquark resonances in $\bar p + d$ collisions is possible with the nucleons moving slowly inside the deuteron due the masses of the pentaquark, charmonium and the nucleon being close to a special kinemtic relation \[Eq(\[m0\])\]. The cross section of this process is evaluated by estimating the absorptive part of the amplitude. The numerical value of the expected cross section at the resonance maximum depends on the unknown branching fraction for the decay of pentaquark to charmonium and a nucleon, $Br [P_c \to (c \bar c) + n]$. If this fraction is of order 10% the cross section for formation of the resonances in the channel $J/\psi +N$ (in particular, the ones reported by LHCb) is likely quite small, in the ballpark of $10^{-35}$cm$^2$, and may require a significan effort to be observed with the PANDA detector at FAIR. However if similar pentaquark states decaying to $\eta_c + N$ do exist, their formation cross section is estimated to be much larger, due to the much larger $p \bar p$ decay width of $\eta_c$, so that such resonances can be realistically sought for at a luminosity of order $10^{32}$cm$^{-2}$c$^{-1}$ for antiproton collisions with a deuterium target.
This work is supported in part by U.S. Department of Energy Grant No. DE-SC0011842.
[99]{} R. Aaij [*et al.*]{} \[LHCb Collaboration\], Phys. Rev. Lett. [**115**]{}, 072001 (2015) doi:10.1103/PhysRevLett.115.072001 \[arXiv:1507.03414 \[hep-ex\]\].
R. Aaij [*et al.*]{} \[LHCb Collaboration\], Phys. Rev. Lett. [**117**]{}, no. 8, 082002 (2016) doi:10.1103/PhysRevLett.117.082002 \[arXiv:1604.05708 \[hep-ex\]\].
F. K. Guo, C. Hanhart, U. G. Mei[ß]{}ner, Q. Wang, Q. Zhao and B. S. Zou, Rev. Mod. Phys. [**90**]{}, no. 1, 015004 (2018) doi:10.1103/RevModPhys.90.015004 \[arXiv:1705.00141 \[hep-ph\]\].
Q. Wang, X. H. Liu and Q. Zhao, Phys. Rev. D [**92**]{}, 034022 (2015) doi:10.1103/PhysRevD.92.034022 \[arXiv:1508.00339 \[hep-ph\]\].
V. Kubarovsky and M. B. Voloshin, Phys. Rev. D [**92**]{}, no. 3, 031502 (2015) doi:10.1103/PhysRevD.92.031502 \[arXiv:1508.00888 \[hep-ph\]\].
M. Karliner and J. L. Rosner, Phys. Lett. B [**752**]{}, 329 (2016) doi:10.1016/j.physletb.2015.11.068 \[arXiv:1508.01496 \[hep-ph\]\].
M. Tanabashi [*et al.*]{} \[Particle Data Group\], Phys. Rev. D [**98**]{}, 030001 (2018).
M. F. M. Lutz [*et al.*]{} \[PANDA Collaboration\], arXiv:0903.3905 \[hep-ex\].
J. P. Lees [*et al.*]{} \[BaBar Collaboration\], Phys. Rev. D [**87**]{}, no. 9, 092005 (2013) doi:10.1103/PhysRevD.87.092005 \[arXiv:1302.0055 \[hep-ex\]\].
S. Dubynskiy and M. B. Voloshin, Phys. Rev. D [**74**]{}, 094017 (2006) doi:10.1103/PhysRevD.74.094017 \[hep-ph/0609302\].
A. E. Bondar and M. B. Voloshin, Phys. Rev. D [**93**]{}, no. 9, 094008 (2016) doi:10.1103/PhysRevD.93.094008 \[arXiv:1603.08436 \[hep-ph\]\].
M. B. Voloshin, arXiv:1902.01281 \[hep-ph\].
F. K. Guo, C. Hanhart, U. G. Mei[ß]{}ner, Q. Wang and Q. Zhao, Phys. Lett. B [**725**]{}, 127 (2013) doi:10.1016/j.physletb.2013.06.053 \[arXiv:1306.3096 \[hep-ph\]\].
S. Weinberg, Phys. Rev. [**137**]{}, B672 (1965). doi:10.1103/PhysRev.137.B672
K. Chilikin [*et al.*]{} \[Belle Collaboration\], Phys. Rev. D [**90**]{}, no. 11, 112009 (2014) doi:10.1103/PhysRevD.90.112009 \[arXiv:1408.6457 \[hep-ex\]\].
R. Aaij [*et al.*]{} \[LHCb Collaboration\], Eur. Phys. J. C [**78**]{}, no. 12, 1019 (2018) doi:10.1140/epjc/s10052-018-6447-z \[arXiv:1809.07416 \[hep-ex\]\].
[^1]: In the relevant kinematics some of the particles in the process are in fact relativistic, e.g. the incoming antiproton. This behavior is accounted for by the appropriate extra energy-dependent factors, e.g $2 E_p$ in the equations (\[psiw\]) and (\[pcw\]) in comparison with the standard expressions resulting from the relativistic normalization of the wave functions to $2 E$. Also one can notice, as already discussed, that the factors $g$ and $h$ generally depend on kinematic variables which dependence however is the same in the absorptive part of the graph of Fig. 1 and in the on-shell decays.
[^2]: This calculation is similar to those described in Refs.[@dv06; @bv; @mv19] and is equivalent (for the purpose of present treatment) to a calculation in a Lorentz-invariant form [@ghmwz] based on the Weinberg’s formula [@Weinberg] for the invariant coupling of a shallow bound state to its constituents.
[^3]: Such analogy is known to work in at least one case of exotic charmonium-like mesonic resonances $Z_c(4200)$, decaying to $J/\psi + \pi$ [@bellezc], and $Z_c(4100)$ decaying to $\eta_c + \pi$ [@lhcbzc].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Andreas Kipf$^{\,\star}$ Ryan Marcus$^{\,\star\dagger}$ Alexander van Renen Mihail Stoian'
- 'Alfons Kemper Tim Kraska$^{\,\star}$ Thomas Neumann'
bibliography:
- 'ryan-cites-long.bib'
title: 'RadixSpline: A Single-Pass Learned Index'
---
[^1]
[^1]: Andreas Kipf, Ryan Marcus, and Alexander van Renen contributed equally
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present ten high-precision light curves of four transits in the XO-1 planetary system, obtained using $u$, $g$, $r$, redshifted H$\alpha$, $I$ and $z$ filters. We use these to measure the physical properties, orbital ephemeris, and a transmission spectrum of the planet covering the full optical wavelength range. We augment this with published HST/WFC3 observations to construct a transmission spectrum of the planet covering 0.37 to 1.65$\mu$m. Our best-fitting model to this spectrum is for a H$_2$/He-rich atmosphere containing water (3.05$\sigma$ confidence), nitrogen-bearing molecules NH$_3$ and HCN (1.5$\sigma$) and patchy cloud (1.3$\sigma$). [We find that adding the optical to the near-infrared data does not lead to more precise constraints on the planetary atmosphere in this case.]{} We conduct a detailed investigation into the effect of stellar limb darkening on our results, concluding that the choice of limb darkening law and coefficients is unimportant; such conclusions may not hold for other systems so should be reassessed for all high-quality datasets. The planet radius we measure in the $g$-band is anomalously low and should be investigated with future observations at a higher spectral resolution. From the measured times of transit we determine an improved orbital ephemeris, calculate a lower limit on the modified stellar tidal quality factor of $Q_\star^{\,\prime} > 10^{5.6}$, and rule out a previously postulated sinusoidal variation in the transit times.'
title: 'Physical properties and optical-infrared transmission spectrum of the giant planet XO-1b'
---
stars: planetary systems — planets and satellites: atmospheres — stars: fundamental parameters — stars: individual: XO-1
Introduction {#sec:intro}
============
------------- ------------ ------------ ---------- --------------- --------------- ---------------- --------------- ---------------------------- -------- ------------ ---------------- ---------
Telescope / Date of Start time End time $N_{\rm obs}$ $T_{\rm exp}$ $T_{\rm dead}$ Filter Airmass Moon Aperture $N_{\rm poly}$ Scatter
instrument first obs (UT) (UT) (s) (s) illum. radii (px) (mmag)
INT/WFC 2010 04 27 23:56 03:33 95 100 30 red.H$\alpha$ 1.29 $\to$ 1.00 $\to$ 1.01 0.995 25 40 60 [2]{} 0.63
CAHA/BUSCA 2012 05 07 22:34 03:56 107 120 61 SDSS $u$ 1.19 $\to$ 1.01 $\to$ 1.25 0.945 27 33 60 1 3.35
CAHA/BUSCA 2012 05 07 22:34 03:56 107 120 61 SDSS $g$ 1.19 $\to$ 1.01 $\to$ 1.25 0.945 30 40 70 1 0.76
CAHA/BUSCA 2012 05 07 22:34 03:56 107 120 61 SDSS $r$ 1.19 $\to$ 1.01 $\to$ 1.25 0.945 32 42 80 1 0.55
CAHA/BUSCA 2012 05 07 23:19 03:56 92 120 61 SDSS $z$ 1.10 $\to$ 1.01 $\to$ 1.25 0.945 30 40 70 1 0.75
CAHA/BUSCA 2012 05 11 21:35 03:58 142 100–120 52 SDSS $u$ 1.34 $\to$ 1.01 $\to$ 1.31 0.595 23 33 60 1 1.44
CAHA/BUSCA 2012 05 11 21:35 03:47 138 100–120 52 SDSS $g$ 1.34 $\to$ 1.01 $\to$ 1.27 0.595 30 40 70 1 0.64
CAHA/BUSCA 2012 05 11 21:35 03:58 147 100–120 52 SDSS $r$ 1.34 $\to$ 1.01 $\to$ 1.31 0.595 30 45 80 2 1.78
CAHA/BUSCA 2012 05 11 21:35 03:58 145 100–120 52 SDSS $z$ 1.34 $\to$ 1.01 $\to$ 1.31 0.595 25 40 80 2 0.72
CAHA/1.23m 2014 05 26 21:39 02:58 151 105–125 11 Cousins $I$ 1.15 $\to$ 1.01 $\to$ 1.30 0.036 30 45 80 1 0.61
------------- ------------ ------------ ---------- --------------- --------------- ---------------- --------------- ---------------------------- -------- ------------ ---------------- ---------
The atmospheric properties of giant planets are an important indicator of the formation and evolution of planets and planetary systems [@Madhusudhan++14apj; @Mordasini+16apj]. They are also observationally accessible in a large fraction of hot Jupiters (planets with mass $>$0.3[$\,{\rm M}_{\rm Jup}$]{} and orbital period $<$10d) which transit their host star, via the method of [*transmission spectroscopy*]{}.
Transmission spectroscopy offers a way of measuring the radius of the planet as a function of wavelength, by determining the transit depth at multiple wavelengths. It is sensitive to the amount of absorption and scattering of starlight passing though the outer atmosphere of the planet whilst it is backlit by its host star. Transmission spectroscopy can be used to detect enhanced opacity due to atomic absorption, molecular absorption and Rayleigh scattering [e.g. @Pont+13mn; @Madhusudhan++14apj; @Fischer+16apj]. This can yield constraints on the chemical composition of the atmosphere, its temperature-pressure structure, and the presence of cloud or haze particles. The first detection of the atmosphere of an extrasolar planet was [due to]{} sodium absorption in HD209458b [@Charbonneau+02apj], and extensive results have recently been obtained from both the ground and space [e.g. @Nikolov+16apj; @Sing+16nat].
In the near future the NASA [*James Webb Space Telescope*]{} (JWST) will revolutionise this research area with extensive observations covering wavelengths from 0.6$\mu$m to 28$\mu$m [@Beichman+14pasp; @Greene+16apj]. [It is expected to be used to study a significant sample of planets, and by comparison to HST it will achieve much lower Poisson noise, more extensive wavelength intervals, and uninterrupted coverage of individual transits.]{}
[@Stevenson+16pasp outlined an [*Early Release Science*]{} program intended to occur shortly after JWST enters service, in which extensive observations of a small number of transiting planets will be performed using multiple instruments and observing modes. The aims are to allow an assessment of the relative strengths of the observing modes, and to expedite the development of data reduction pipelines for this work. @Stevenson+16pasp selected 12 transiting planets as promising targets. XO-1 is one of the most suitable targets within this list, with a sky position near the continuous viewing zone of JWST, a host star which is bright ($K_s = 9.53$; @Skrutskie+06aj) and inactive (${\ensuremath{\log R^{\prime}_{\rm HK}}}= -4.958$; @Knutson++10apj), and a planet with an atmospheric scale height (277km) suitable for obtaining transmission spectra with a significant signal to noise ratio.]{} In this work we present a detailed analysis of the XO-1 system, based on new transit light curves in six optical passbands plus published infrared transmission spectroscopy, in order to measure the physical properties of the system, refine the orbital ephemeris, and investigate the atmospheric properties of the planet.
XO-1 was only the eleventh transiting extrasolar planet (TEP) discovered [@Mccullough+06apj], and was found to be a 0.92[$\,{\rm M}_{\rm Jup}$]{} and 1.21[$\,{\rm R}_{\rm Jup}$]{} planet orbiting a 1.04[$\,{\rm M}_\odot$]{} and 0.94[$\,{\rm R}_\odot$]{} G1V star every 3.94d [@Me10mn]. Follow-up light curves from @Holman+06apj were analysed using homogeneous methods by both @Torres++08apj and @Me08mn [@Me09mn; @Me11mn]. Occultations (secondary eclipses) were observed by @Machalek+08apj using the [*Spitzer Space Telescope*]{}, from which flux ratios of the planet to the star were measured in the four IRAC passbands (3.6, 4.5, 5.8 and 8.0 $\mu$m). @Tinetti+10apj and @Burke+10apj presented and studied HST/NICMOS transmission spectroscopy of XO-1b, finding evidence for the presence of the molecules H$_2$O, CH$_4$ and CO$_2$. Their results have been questioned by @Gibson++11mn based on a reanalysis of the same data, and by @Deming+13apj based on new HST/WFC3 transmission spectroscopy. @Deming+13apj found evidence for water in the transmission spectrum of XO-1b, a conclusion also reached by @Tsiaras+18aj.
In addition to the works cited above, transit light curves have been presented by @Vanko+09conf [@Caceres+09aa; @Raetz+09an] and @Sada+12pasp, spectroscopic analyses of the host star have been performed by @Ammler+09aa [@Torres+12apj; @Mortier+13aa] and @Teske+14apj, and the orbital eccentricity has been constrained to be less than 0.29 by @MadhusudhanWinn09apj and @Pont+11mn. Most recently, @Bonomo+17aa presented new radial velocity measurements from which they constrained the eccentricity to be less than 0.019 to $1\sigma$ and 0.043 to $2\sigma$.
High-resolution imaging of TEP host stars is an important part of determining the physical properties of the system [e.g. @Evans++16apj]. Lucky Imaging of the XO-1 system was presented by @Wollert+15aa, who found no nearby stars [less than 3.97, 4.85, 5.79 and 6.46 mag fainter than XO-1A]{} (5$\sigma$ detection limits) in the $z^\prime$ band at distances of 0.25[$^{\prime\prime}$]{}, 0.5[$^{\prime\prime}$]{}, 1.0[$^{\prime\prime}$]{} and 2.0[$^{\prime\prime}$]{}, respectively.
[The outline of this paper is as follows. Section\[sec:obs\] presents our new observations of XO-1, which are analysed in Section\[sec:lc\] alongside published light curves. The results are used to measure the physical properties of the system in Section\[sec:absdim\]. Section\[sec:porb\] presents an improved orbital ephemeris and a search for transit timing variations. The transmission spectrum of XO-1b is obtained and analysed in Section\[sec:transspec\], after which the paper is concluded in Section\[sec:summary\].]{}
Observations and data reduction {#sec:obs}
===============================
{width="\textwidth"}
Two transits of XO-1 were observed using the BUSCA four-band imaging photometer on the 2.2m telescope at Calar Alto, Spain. We selected Sloan $u$, $g$, $r$ and $z$ filters [@Fukugita+96aj] from the Calar Alto filter database, which with BUSCA yield a circular field of view approximately 5.8[$^\prime$]{} in diameter. We were not able to obtain good photometry simultaneously in the bands with the lowest and highest counts ($u$ and $r$ respectively), [because the four arms of BUSCA cannot be operated at different focus levels or exposure times. We therefore optimised for the $r$-band on the first night and the $u$-band on the second night.]{} The $u$-band light curve from 2012/05/07 therefore has a large scatter due to low flux levels, and the $r$- band light curve from 2012/05/11 displays systematic effects due to being near the saturation level of the CCD. An observing log is given in Table\[tab:obslog\] and further details of our approach with BUSCA can be found in @Me+12mn2.
One transit of XO-1 was observed with the Isaac Newton Telescope (INT) and Wide Field Camera (WFC) on La Palma, Spain. We used CCD 4, as this is the one on the optical axis, and a redshifted H$\alpha$ filter (ING filter[^1] \#226, central wavelength 689nm, FWHM 10nm) rather than a wide-band filter in order to limit the amount of defocussing used. We were not able to autoguide as the guide CCD is in the same focal plane as the science instrument.
One transit of XO-1 was obtained with the 1.23m telescope at Calar Alto, using a Cousins $I$ filter. The transit ingress was missed due to technical problems, but the light curve is otherwise excellent.
The data were reduced using the [defot]{} pipeline [@Me+09mn; @Me+14mn], which depends on the NASA [astrolib]{} library[^2] [idl]{}[^3] implementation of the [aper]{} routine from [daophot]{} [@Stetson87pasp]. Software apertures were placed by hand and their radii chosen to minimise the scatter in the final light curve. [The apertures were shifted to account for telescope pointing wander, which was measured by cross-correlating each image with a reference image.]{} We did not perform bias or flat-field calibrations as these had little effect on the final light curves beyond a slight increase in the scatter of the data.
A differential-magnitude light curve of XO-1 was generated for each observing sequence versus an ensemble comparison star containing the weighted flux sum of the good comparison stars. A polynomial was also fitted to the observations outside transit and subtracted to rectify the final light curve to zero differential magnitude. [The order of the polynomial was chosen to be the lowest which gave a good fit to the out-of-transit data.]{} The coefficients of the polynomial and the weights of the comparison stars were simultaneously optimised to minimise the scatter in the datapoints outside eclipse. The observational errors were then scaled so each transit had a reduced $\chi^2$ of ${\ensuremath{\chi_\nu^{\,2}}}= 1.0$ versus a best-fitting model calculated with the [jktebop]{} code (see below). Table\[tab:obslog\] includes the polynomial order and the $rms$ of the residuals versus the best fit for each light curve. The final data are shown in Fig.\[fig:lcall\] and listed in Table\[tab:lcdata\]. The times given refer to the midpoint of the exposure on the BJD/TDB timescale [@Eastman++10pasp].
Light curve analysis {#sec:lc}
====================
Approach
--------
We modelled the available transit light curves of XO-1 using the [jktebop]{}[^4] code [@Me13aa and references therein] and the formalism of the [*Homogeneous Studies*]{} project [see @Me12mn and references therein]. The fitted parameters were as follows.
- The sum and ratio of the fractional radii of the two components, $r_{\rm A} + r_{\rm b}$ and $k = \frac{r_{\rm b}}{r_{\rm A}}$, where the fractional radii are the absolute radii in units of the semimajor axis: $r_{\rm A,b}= \frac{R_{\rm A,b}}{a}$. These combinations of parameters were chosen because they are only weakly correlated.
- The orbital inclination, $i$.
- A time of mid-transit, $T_0$.
- The coefficients of a polynomial of differential magnitude versus time. The polynomial order for each light curve is given in Table\[tab:obslog\]. Whilst the fitted polynomials were removed at the data-reduction stage, their inclusion in the [jktebop]{} fit is necessary to propagate their uncertainties into the measured photometric parameters.
- One or two limb darkening (LD) coefficients, depending on the solution performed.
The orbital period was held fixed in each solution, because the uncertainty in its value was utterly negligible for this analysis. We also enforced orbital circularity in the absence of evidence for an eccentric orbit [see discussion in @Anderson+12mn].
------------- ------------ -------------- -------------- -------------
Telescope / Filter BJD(TDB) Diff. mag. Uncertainty
instrument $-$2400000
INT/WFC rH$\alpha$ 55314.497407 $ $0.0003821 0.0006442
BUSCA $u$ 56055.445999 $-$0.0059505 0.0028984
BUSCA $g$ 56055.445999 $-$0.0006606 0.0007777
BUSCA $r$ 56055.445999 $-$0.0007141 0.0005374
BUSCA $z$ 56055.477499 $ $0.0002283 0.0007324
BUSCA $u$ 56059.405206 $-$0.0010300 0.0015896
BUSCA $g$ 56059.405206 $-$0.0006314 0.0006988
BUSCA $r$ 56059.405206 $-$0.0018004 0.0007984
BUSCA $z$ 56059.405206 $ $0.0008813 0.0008654
CAHA123 $I$ 56804.408021 $ $0.0063808 0.0005926
------------- ------------ -------------- -------------- -------------
: \[tab:lcdata\] The first line of each of the light curves presented in this work. The full dataset will be made available at the CDS.
We performed [jktebop]{} solutions using each of four two-parameter LD ‘laws’: quadratic, square-root, logarithmic and cubic [see @Me08mn]. We furthermore calculated solutions with both LD coefficients fixed at theoretical values, the linear coefficient fitted and the nonlinear coefficient fixed, and both coefficients fitted. The theoretical LD coefficients were obtained by bilinearly interpolating[^5] in tabulated predictions to the host star’s [measured]{} effective temperature ([$T_{\rm eff}$]{}) and surface gravity ([$\log g$]{}). We considered multiple sources of theoretical coefficients [@Vanhamme93aj; @Claret00aa; @ClaretHauschildt03aa; @Claret04aa2] and averaged their predictions when necessary.
Least-squares best fits were obtained using the Levenberg-Marquardt method [@Marquardt63] as implemented in the [mrqmin]{} routine [@Press+92book]. The uncertainties in the fitted parameters were estimated using both Monte Carlo and residual-permutation solutions [see @Me08mn for further details], and the larger errorbar was retained for each measured parameter. Uncertainties were further inflated to account for any scatter in the measured values of a parameter from the solutions using different approaches to the inclusion of LD. Tables of results for each light curve can be found in the Supplementary Information. The measured photometric parameters are given in Table\[tab:lcfit\].
Our new data
------------
Source
---------------------------- ---------------- ---------------- ---------------- ---------------- --------------- -------------- ---------------- ---------------- ----------------- -----------------
BUSCA $u$ 0.1048 0.0063 0.1372 0.0032 87.22 1.17 0.0923 0.0053 0.01253 0.00101
BUSCA $g$ 0.0993 0.0024 0.1289 0.0014 89.07 0.74 0.0880 0.0020 0.01133 0.00038
BUSCA $r$ 0.1002 0.0024 0.1331 0.0012 88.82 0.81 0.0884 0.0021 0.01177 0.00037
BUSCA $z$ 0.1014 0.0027 0.1327 0.0015 88.50 0.81 0.0895 0.0023 0.01188 0.00038
INT/WFC rH$\alpha$ 0.0977 0.0015 0.1339 [0.0027]{} 89.88 0.62 0.0862 0.0013 0.01154 0.00022
1.23m $I_C$ 0.1012 0.0029 0.1334 0.0013 88.52 0.58 0.0893 0.0025 0.01191 0.00042
\[3pt\] @Holman+06apj FLWO 0.1009 0.0023 0.1321 0.0012 88.56 0.66 0.0891 0.0019 0.01177 0.00033
@Holman+06apj Palomar 0.0953 0.0043 0.1265 0.0028 89.99 1.03 0.0846 0.0037 0.01070 0.00063
@Caceres+09aa SofI 0.1018 0.0025 0.1324 0.0021 88.47 0.47 0.0899 0.0022 0.01191 0.00036
@Caceres+09aa ISAAC 0.0978 0.0038 0.1321 0.0018 89.81 0.95 0.0863 0.0033 0.01140 0.00063
@Sada+12pasp 0.1082 0.0110 0.1297 0.0063 87.92 1.86 0.0958 0.0092 0.01242 0.00176
\[3pt\] Final results [**0.0997**]{} [**0.0008**]{} [**0.1325**]{} [**0.0008**]{} [**88.84**]{} [**0.22**]{} [**0.0880**]{} [**0.0007**]{} [**0.01166**]{} [**0.00012**]{}
![\[fig:lcfit\] [jktebop]{} best fits to our phased light curves of XO-1. The data are shown as filled circles colour-coded consistently with Fig.\[fig:lcall\]. The best fits are shown as grey lines. The residuals are offset to appear at the base of the figure. Labels give the passband and source for each dataset. The polynomial baseline functions have been subtracted from the data before plotting.](plotLCfit.eps){width="\columnwidth"}
The data from the two transits of XO-1 observed with BUSCA were collected into one light curve for each filter, and each was modelled with [jktebop]{} (see Fig.\[fig:lcfit\]). We made two exceptions to this approach: the $u$-band data from 2012/05/07 were ignored because the low flux levels caused a large scatter, and the $r$-band data from 2012/05/11 were rejected because they suffer from saturation effects. We found that the $g$-band light curves are in excellent agreement with each other (${\ensuremath{\chi_\nu^{\,2}}}= 1.02$ when the individual light curves have ${\ensuremath{\chi_\nu^{\,2}}}= 1.0$). However, the $z$-band light curves are not (${\ensuremath{\chi_\nu^{\,2}}}= 1.56$), as can be seen in Fig.\[fig:lcall\]. Our resulting parameters for the $z$-band are therefore roughly the average for the two datasets, and are in fact in good agreement with the results from other light curves.
In all cases except BUSCA $u$, we adopt the results from [jktebop]{} models with the linear LD coefficient fitted and the nonlinear LD coefficient fixed, as these agree very well both between different LD laws and between different datasets. For BUSCA $u$ we found that the data were unable to support fitting for even one LD coefficient, so we adopt the results obtained with both coefficients fixed. LD coefficients are not available for the redshifted H$\alpha$ filter. We therefore used those for the Johnson $R$ filter, which has a similar central wavelength (0.67$\mu$m; @Johnson66). The effect of the difference in passband on the LD coefficients is expected to be smaller than the intrinsic uncertainty of the coefficients, as judged from the variation in predictions for the same filters from different sources.
Reanalysis of published data
----------------------------
![\[fig:lcfit:pub\] [jktebop]{} best fits to published light curves of XO-1. The data are shown as filled circles and the best fits as grey lines. The residuals are offset to appear at the base of the figure. Labels give the passband and source for each dataset. The polynomial baseline functions have been subtracted from the data before plotting.](plotLCfitpub.eps){width="\columnwidth"}
The literature for XO-1 includes several light curves of a quality sufficient for inclusion in the current work. We have obtained these data and modelled them using the same methods as for our own observations. The results are included in Table\[tab:lcfit\] and are discussed below.
@Holman+06apj presented light curves of two transits of XO-1 obtained with the FLWO 1.2m telescope and KeplerCam in the $z$-band, and one transit observed using the Palomar 1.5m in the $R$ band. According to the webpage for this facility[^6] this corresponds to a Kron-Cousins $R$ band. Both datasets have been analysed in the past by the first author [@Me08mn] but were reanalysed with the modification that a first-order polynomial was applied to each transit, an option added to [jktebop]{} since the previous analysis [see @Me+14mn]. The best fits are shown in Fig.\[fig:lcfit:pub\] and were each obtained with one fitted and one fixed LD coefficient.
@Caceres+09aa published observations of four transits of XO-1, all obtained at near-infrared wavelengths. We ignored their Run A due to the large systematic errors visible in the data, and their Run C due to the patches of very high scatter during transit. We therefore analysed their Run B, which was obtained using NTT/SofI in the $J$-band, and their Run D, observed using VLT/ISAAC in the $J$-band but with a blocking filter to remove flux from a red leak in the $J$ filter. Both runs were obtained at high cadence, with integration times of 0.8s and 0.08s respectively and very low dead times. We therefore binned the light curves by factors of 100 and 1000, respectively, to yield a sampling rate of approximately 80s in both cases. Whereas the SofI data could be satisfactorily modelled with one fitted LD coefficient, we had to fix both to obtain an acceptable solution of the ISAAC observations. In both cases we included a second-order polynomial to model the baseline brightness of the system.
@Sada+12pasp observed one transit in the $z$-band using a 0.5m telescope at Kitt Peak National Observatory. The ingress was missed and the data have an expectedly large scatter of 2.9mmag, but we ran the usual [jktebop]{} solutions in order to determine whether this dataset can provide results worth including on our analysis. We allowed for a second-order polynomial baseline function.
Combined results
----------------
![\[fig:lcfit:compare\] Plots of the measured values of the sum of the fractional radii versus orbital inclination (top) and of the two fractional radii (bottom). In each case the coloured points represent the light curves presented in this work, with colour-coding the same as in Fig.\[fig:lcall\], grey points show results for literature light curves, and the bold black lines indicate the weighted mean value.](plotphotcompare.eps){width="\columnwidth"}
The measured photometric parameters are given in Table\[tab:lcfit\] and show a good agreement between light curves. We calculated the weighted mean value for each measured parameter for use in the next section. The [$\chi_\nu^{\,2}$]{} value of the individual values versus the weighted mean is good for $r_{\rm A}$, $r_{\rm b}$ and $i$ (0.6, 0.6 and 0.9, respectively) but less so for $k$ (1.8). This could be caused by residual systematic errors and/or by a true astrophysical signal (i.e. a non-flat transmission spectrum; see @MeEvans16mn). We have multiplied the uncertainty in the weighted mean of $k$ by $\sqrt{1.8}$ to account for this.
Casual inspection of Table\[tab:lcfit\] suggests that correlations exist between several of the photometric parameters. Such correlations are widely known [e.g. @Me08mn; @Pal08mn; @Carter+08apj] and must be accounted for in the uncertainties of the parameter measurements. In Fig.\[fig:lcfit:compare\] we illustrate two of these correlations: between $r_{\rm A}+r_{\rm b}$ and $i$, and between $r_{\rm A}$ and $r_{\rm b}$. The former arises because $r_{\rm A}+r_{\rm b}$ and $i$ together determine the duration of the transit [@SeagerMallen03apj], and the latter occurs as $k = \frac{r_{\rm b}}{r_{\rm A}}$ is much better-determined than [either]{} $r_{\rm A}$ or $r_{\rm b}$. It is clear from Fig.\[fig:lcfit:compare\] that the correlations are greatly attenuated using the high-quality light curves presented here, and that the errorbars in Table\[tab:lcfit\] are not underestimated. For reference, the linear Pearson correlation coefficients are $-0.89$ and $+0.90$, respectively.
Physical properties {#sec:absdim}
===================
[r@[$\pm$]{}l r@[$\pm$]{}l r@[$\pm$]{}l l]{} & & & Ref\
5750 & 13$^*$ & 0.015 & 0.040$^*$ & 4.53 & 0.065$^*$ & 1\
5754 & 42 & $-$0.01 & 0.05 & 4.61 & 0.05 & 2\
5738 & 65 & $-$0.06 & 0.05 & 4.50 & 0.01 & 3\
5754 & 42 & $-$0.01 & 0.05 & 4.61 & 0.05 & 4\
5695 & 26 & $-$0.11 & 0.06 & 4.42 & 0.12 & 5\
5729 & 25$^*$ & $-$0.07 & 0.010$^*$ & 4.49 & 0.028$^*$ & 6\
\
5740 & 50 & $-$0.03 & 0.05 & &\
Parameter
------------------------------------------------- --------- ----------- ---------- ----------------- -----------
$M_{\rm A}$ ([$\,{\rm M}_\odot$]{}) 1.018 0.028 0.034 [1.027]{} [0.06]{}
$R_{\rm A}$ ([$\,{\rm R}_\odot$]{}) 0.930 0.011 0.010 [ 0.94]{} [0.02]{}
$\log g_{\rm A}$ (cgs) 4.509 0.009 0.005 [ 4.50]{} [0.01]{}
$\rho_{\rm A}$ ([$\,\rho_\odot$]{}) [ 1.23]{} [0.03]{}
\[2pt\] $M_{\rm b}$ ([$\,{\rm M}_{\rm Jup}$]{}) 0.907 0.022 0.020 [ 0.92]{} [0.08]{}
$R_{\rm b}$ ([$\,{\rm R}_{\rm Jup}$]{}) 1.199 0.017 0.013 [ 1.21]{} [0.03]{}
$g_{\rm b}$ ([ms$^{-2}$]{}) [ 15.5]{} [1.1 ]{}
$\rho_{\rm b}$ ([$\,\rho_{\rm Jup}$]{}) 0.492 0.018 0.005 [ 0.48]{} [0.04]{}
\[2pt\] [$T_{\rm eq}^{\,\prime}$]{} (K)
$a$ (AU) 0.04914 0.00045 0.00054 [0.049]{} [0.001]{}
Age (Gyr)
: \[tab:absdim\] Derived physical properties of the XO-1 system [from this work compared to those from @Burke+10apj]{}. When measurements are accompanied by two errorbars, the first refers to the statistical uncertainties and the second to the systematic uncertainties.
We used the results of the photometric analysis from the previous section to obtain the full physical properties of the system. This process also required knowledge of the spectroscopic properties of the host star (effective temperature [$T_{\rm eff}$]{} and metallicity [$\left[\frac{\rm Fe}{\rm H}\right]$]{}) which are summarised in Table\[tab:spec\], and the stellar orbital velocity amplitude, $K_{\rm A} = 115.3 \pm 1.8$[ms$^{-1}$]{} [@Bonomo+17aa]. As the necessary additional constraint, we used tabulated predictions from each one of five sets of theoretical stellar models [@Claret04aa; @Demarque+04apjs; @Pietrinferni+04apj; @Vandenberg++06apjs; @Dotter+08apjs].
We then estimated the value of the velocity amplitude of the planet, $K_{\rm b}$ and calculated the physical properties of the system using this and the measured quantities. We iteratively adjusted $K_{\rm b}$ to optimise the agreement between the calculated $\frac{R_{\rm A}}{a}$ and the measured $r_{\rm A}$, and between the [$T_{\rm eff}$]{} and that predicted by the stellar models for the observed [$\left[\frac{\rm Fe}{\rm H}\right]$]{} and calculated stellar mass ($M_{\rm A}$). We did this for ages from 0.1Gyr to 20Gyr in steps of 0.1Gyr, from which we identified the overall best fit and age of the system [see @Me09mn]. This process was undertaken for each of the five sets of tabulated theoretical model predictions, and the final parameters were taken to be the median of the five different possibilities arising from this repeated analysis.
We propagated the statistical errors in all input parameters using a perturbation analysis, and added all contributions in quadrature for each output parameter. We estimated the systematic uncertainties, which are incurred by the use of theoretical stellar models, by taking the maximum deviation between the final parameter value and the individual values obtained using the different sets of tabulated predictions.
The measured physical properties of the XO-1 system are given in Table\[tab:absdim\]. The mass, radius, gravity and density of the star are denoted by: $M_{\rm A}$, $R_{\rm A}$, $\log g_{\rm A}$ and $\rho_{\rm A}$; and of the planet by $M_{\rm b}$, $R_{\rm b}$, $g_{\rm b}$ and $\rho_{\rm b}$, respectively. [Our results are in good agreement with all previously published measurements]{}. Our measured $r_{\rm A}$ is equivalent to a relatively large stellar density, which means that the best-fitting theoretical star is near the zero-age main sequence. We therefore see a significant systematic uncertainty in our results caused by edge effects in the model tabulations, and by the intrinsic variation in how different stellar evolution codes initialise their stellar models. [Table\[tab:absdim\] also includes a comparison between our measurements and those of @Burke+10apj, which are in very good agreement.]{}
This young age is surprising because it is not supported by other age indicators such as chromospheric activity and rotational velocity. @Knutson++10apj observed the cores of the CaII H & K lines, finding a small core emission due to stellar activity. They measured an activity index of $\log R^\prime_{\rm HK} = -4.958$, which indicates that it is a relatively inactive star. The calibration of @MamajekHillenbrand08apj points to an age of roughly 6Gyr with an uncertainty of perhaps 0.05dex due to astrophysical scatter, and unknown uncertainties due to the $\log R^\prime_{\rm HK}$ value (which is not supplied with an errorbar) and activity cycles on XO-1A (because we only have one measurement of $\log R^\prime_{\rm HK}$). One possible solution to this conflict is inaccuracies in theoretical models [e.g. @Maxted++15aa2], with perhaps a small contribution from an orbital eccentricity which is large enough to affect the measured $r_{\rm A}$ but small enough to hide in the available radial velocity measurements.
Transit timing analysis {#sec:porb}
=======================
----------------- ------------- ---------- ------------ ---------------------------------------------
Time of minimum Uncertainty Cycle Residual Reference
(BJD/TDB) (d) number (d)
2453127.03924 0.00580 $-$555.0 $ $0.00167 @Wilson+06pasp
2453150.68624 0.01060 $-$549.0 $-$0.00036 @Wilson+06pasp
2453154.62574 0.00260 $-$548.0 $-$0.00236 @Wilson+06pasp
2453158.56704 0.00340 $-$547.0 $-$0.00257 @Wilson+06pasp
2453162.51444 0.00250 $-$546.0 $ $0.00333 @Wilson+06pasp
2453166.45124 0.00250 $-$545.0 $-$0.00138 @Wilson+06pasp
2453170.39244 0.00370 $-$544.0 $-$0.00168 @Wilson+06pasp
2453229.51504 0.00450 $-$529.0 $-$0.00165 @Wilson+06pasp
2453237.40504 0.00320 $-$527.0 $ $0.00534 @Wilson+06pasp
2453241.34174 0.00670 $-$526.0 $ $0.00054 @Wilson+06pasp
2453808.91774 0.00110 $-$382.0 $-$0.00012 @Mccullough+06apj
2453875.92321 0.00047 $-$365.0 $-$0.00023 This work (Palomar data from @Holman+06apj)
2453879.86474 0.00110 $-$364.0 $-$0.00021 @Holman+06apj
2453883.80638 0.00018 $-$363.0 $-$0.00007 This work (FLWO data from @Holman+06apj)
2453887.74746 0.00015 $-$362.0 $-$0.00049 This work (FLWO data from @Holman+06apj)
2453887.74774 0.00060 $-$362.0 $-$0.00021 B. Gary (AXA)
2453911.39781 0.00049 $-$356.0 $ $0.00083 J. Ohlert (TRESCA)
2454171.53332 0.00170 $-$290.0 $-$0.00298 @Raetz+09an
2454214.89274 0.00090 $-$279.0 $-$0.00011 B. Gary (AXA)
2454218.83405 0.00114 $-$278.0 $-$0.00030 @Caceres+09aa
2454222.77623 0.00023 $-$277.0 $ $0.00037 This work (SofI data from @Caceres+09aa)
2454222.77671 0.00039 $-$277.0 $ $0.00085 @Caceres+09aa
2454226.71808 0.00033 $-$276.0 $ $0.00072 This work (ISAAC data from @Caceres+09aa)
2454285.84036 0.00097 $-$261.0 $ $0.00043 C. Foote (AXA)
2454506.56417 0.00010 $-$205.0 $-$0.00003 @Burke+10apj
2454518.38906 0.00017 $-$202.0 $ $0.00034 @Burke+10apj
2454553.86244 0.00100 $-$193.0 $ $0.00018 B. Gary (AXA)
2454620.86554 0.00080 $-$176.0 $-$0.00230 B. Gary (AXA)
2454620.86784 0.00080 $-$176.0 $-$0.00000 C. Foote (AXA)
2454624.81004 0.00140 $-$175.0 $ $0.00069 C. Foote (AXA)
2454624.81214 0.00130 $-$175.0 $ $0.00279 C. Foote (AXA)
2454628.75154 0.00040 $-$174.0 $ $0.00069 Healy (AXA)
2454888.89006 0.00070 $-$108.0 $-$0.00012 B. Gary (AXA)
2454959.83746 0.00060 $-$90.0 $ $0.00019 B. Gary (AXA)
2454959.83783 0.00150 $-$90.0 $ $0.00056 This work (data from @Sada+12pasp)
2454967.71916 0.00070 $-$88.0 $-$0.00112 B. Gary (AXA)
2454983.48656 0.00080 $-$84.0 $ $0.00026 J. Gregorio (AXA)
2454987.42836 0.00080 $-$83.0 $ $0.00055 Ayoinemas (AXA)
2455058.37686 0.00100 $-$65.0 $ $0.00196 Srdoc (AXA)
2455290.92347 0.00060 $-$6.0 $-$0.00023 B. Gary (AXA)
2455298.80597 0.00060 $-$4.0 $-$0.00074 B. Gary (AXA)
2455314.57290 0.00014 0.0 $ $0.00017 This work (INT/WFC light curve)
2455365.81217 0.00050 13.0 $-$0.00013 B. Gary (AXA)
2455369.75357 0.00070 14.0 $-$0.00023 B. Gary (AXA)
2455369.75517 0.00070 14.0 $ $0.00137 B. Gary (AXA)
2455629.89263 0.00041 80.0 $-$0.00052 S. Shadic (TRESCA)
2455653.54256 0.00056 86.0 $ $0.00038 R. Naves (TRESCA)
2455700.84090 0.00053 98.0 $ $0.00066 S. Shadic
2455712.66431 0.00077 101.0 $-$0.00045 S. Dvorak (TRESCA)
2455834.85186 0.00017 132.0 $ $0.00044 @Deming+13apj
2455984.62762 0.00061 170.0 $-$0.00100 J. Trnka (TRESCA)
2456055.57528 0.00019 188.0 $-$0.00044 This work ($g$-band light curve from BUSCA)
2456055.57529 0.00013 188.0 $-$0.00043 This work ($r$-band light curve from BUSCA)
2456055.57614 0.00017 188.0 $ $0.00042 This work ($z$-band light curve from BUSCA)
2456059.51669 0.00030 189.0 $-$0.00053 This work ($u$-band light curve from BUSCA)
2456059.51659 0.00014 189.0 $-$0.00063 This work ($g$-band light curve from BUSCA)
2456059.51756 0.00016 189.0 $ $0.00034 This work ($z$-band light curve from BUSCA)
2456059.51881 0.00054 189.0 $ $0.00159 R. Naves (TRESCA)
2456063.45989 0.00063 190.0 $ $0.00116 A. Carreño (TRESCA)
2456067.40080 0.00061 191.0 $ $0.00057 S. Poddaný (TRESCA)
2456106.81198 0.00105 201.0 $-$0.00331 S. Curry (TRESCA)
2456106.81544 0.00127 201.0 $ $0.00015 D. Mitchell (TRESCA)
----------------- ------------- ---------- ------------ ---------------------------------------------
----------------- ------------- -------- ------------ ------------------------------------
Time of minimum Uncertainty Cycle Residual Reference
(BJD/TDB)) (d) number (d)
2456130.46286 0.00185 207.0 $-$0.00146 F. Emering (TRESCA)
2456725.63020 0.00067 358.0 $-$0.00145 M. Zibar (TRESCA)
2456729.57470 0.00039 359.0 $ $0.00155 M. Zibar (TRESCA)
2456737.45381 0.00086 361.0 $-$0.00235 J. Trnka (TRESCA)
2456800.52272 0.00073 377.0 $ $0.00246 CAAT (TRESCA)
2456804.46170 0.00015 378.0 $-$0.00006 This work (CAHA 1.23m light curve)
2457210.43605 0.00097 481.0 $-$0.00079 J. Trnka (TRESCA)
2457210.43733 0.00032 481.0 $ $0.00049 M. Bretton (TRESCA)
2457257.73677 0.00130 493.0 $ $0.00186 O. Mazurenko (TRESCA)
2457454.80941 0.00125 543.0 $-$0.00079 K. Menzies (TRESCA)
2457478.45983 0.00081 549.0 $ $0.00060 A. Marchini (TRESCA)
2457545.46405 0.00101 566.0 $-$0.00078 F. Lomoz (TRESCA)
2457545.46612 0.00083 566.0 $ $0.00129 F. Lomoz (TRESCA)
----------------- ------------- -------- ------------ ------------------------------------
{width="\textwidth"}
A crucial part of obtaining observations of XO-1 with JWST is the availability of a high-precision orbital ephemeris for the scheduling of observations. The most recent detailed study of the ephemeris of XO-1 is as long ago as that of @Burke+10apj. We have therefore redetermined times of minimum [from]{} all available transit light curves in order to [obtain]{} an ephemeris with the highest possible precision.
We first measured the times of mid-transit for each of our own light curves by fitting the data from each passband and each night with $T_0$, $r_{\rm A}+r_{\rm b}$, $k$, $i$, the linear LD coefficient of the quadratic law, and the relevant coefficients of the baseline polynomials as fitted parameters. All times of mid-transit are collected in Table\[tab:tmin\]. The uncertainty in each measured $T_0$ was calculated using 1000 Monte Carlo simulations and residual-permutation simulations and the larger of the two errorbars kept.
We performed the same steps for the published light curves which we included in our analysis above. The photometry from some of these sources [@Holman+06apj; @Caceres+09aa] is given on the “HJD” timescale, which we assumed to mean HJD/UTC and therefore converted into BJD/TDB for consistency with modern analysis methods. The data from @Sada+12pasp are already expressed as a function of BJD/TDB; however we found a large offset between our and their results which is probably due to the differing treatments of the out-of-transit baseline. Our measured $T_0$ has a significantly larger errorbar and also a better agreement with the final linear ephemeris.
@Mccullough+06apj quoted one time of mid-transit based on their follow-up photometry. @Wilson+06pasp presented two timings from the original XO survey data [@Mccullough+06apj] as well as nine times of mid-transit from SuperWASP data [@Pollacco+06pasp]. We ignored one timing with a quoted uncertainty of 31min. One more timing was obtained from @Raetz+09an. The timings discussed in this paragraph so far were quoted as being on the “HJD” system: we have assumed this to represent HJD/UTC and converted them all to BJD/TDB for consistency. Finally, we obtained two timings from @Burke+10apj and one from @Deming+13apj, all three being on the BJD/TDB timescale.
XO-1 was one of the earliest-discovered transiting planetary systems and has a deep transit well suited for observation with small telescopes. It therefore has a rich history of timings obtained by amateur observers. These have been systematically accumulated and fitted by contributors to the Exoplanet Transit Database (ETD[^7]; @Poddany++10newa). We have included all timings based on observations of a complete transit with a scatter sufficiently low to clearly identify the transit shape by eye (sometimes by recourse to the AXA[^8] website), resulting in 43 $T_0$ values. These were all assumed to be on the HJD/UTC system and converted to BJD/TDB.
We fitted all times of mid-transit with a straight line to give the linear ephemeris: $$T_0 = {\rm BJD(TDB)} \,\, 2\,455\,314.572766 (49) \, + \, 3.94150514 (20) \times E$$ where the bracketed numbers show the uncertainty in the final digit of the preceding number and $E$ gives the cycle count versus the reference epoch. We chose the transit observed with the INT as the reference transit because it is close to the weighted mean of the $T_0$ values so the two terms in the ephemeris have a negligible correlation. The [$\chi_\nu^{\,2}$]{} of the fit is 1.66, a typical value for this kind of analysis [e.g. @Me+16mn]. We interpret this as an indication that the errorbars of the individual measurements are modestly underestimated, and not as evidence of transit timing variations. We have multiplied the errorbars for the ephemeris by $\sqrt{1.66}$ to account for this – the orbital period of the XO-1 system is now known to a precision of 0.017s. The residuals versus the linear ephemeris are shown in Fig.\[fig:tmin\].
Constraints on orbital decay
----------------------------
Tidal effects dominate the orbital evolution of short-period giant planets [e.g. @Ogilvie14araa]. Tidally-induced orbital decay is expected to shorten the orbital period of XO-1 and shift its transits earlier in time [in the usual case that the stellar rotation period exceeds the planet orbital period]{} [@Levrard++09apj; @Jackson++09apj]. Tidal evolution timescales depend on the stellar tidal quality factor, $Q_\star$, which has a canonical value of $10^6$ but is uncertain by several orders of magnitude [@OgilvieLin07apj; @Jackson++08apj2; @PenevSasselov11apj; @Penev+12apj].
The relatively long observational history of XO-1 means that it is reasonable to check if transit times are useful in constraining the strength of $Q_\star$. Orbital decay would give rise to a progressive advance of the time of transit, imprinting a quadratic term in its orbital ephemeris. We fitted a quadratic ephemeris to the transit times collected in Table\[tab:tmin\], finding that the quadratic term was consistent with zero ($6.2 \times 10^{-10} \pm 9.0 \times 10^{-10}$dd$^{-1}$, or $9.7 \pm 14.2$msyr$^{-1}$). The Bayesian Information Criterion [@Schwarz78] is higher for this ephemeris (219.6) than for the linear ephemeris (216.1). So is the Akaike Information Criterion [@Akaike81] with 212.7 versus 211.5, respectively. We conclude that there is no observational support for a quadratic ephemeris, and thus no detection of orbital decay in this planetary system.
To derive an upper limit on orbital decay, and thus a lower limit on $Q_\star$, we followed the procedure outlined by @Birkby+14mn and rediscussed by @Wilkins+17apj. In this method, the quadratic term in the orbital ephemeris, $q$, constrains the modified tidal quality factor $$Q_\star^{\,\prime} = \frac{3}{2} \, \frac{Q_\star}{k_2}$$ where $k_2$ is the Love number [@Love11book]. The relevant equation is[^9] $$Q_\star^{\,\prime} = \frac{-27}{~~8} \left(\frac{M_{\rm b}}{M_{\rm A}}\right) \left(\frac{R_{\rm A}}{a}\right)^5 \left(\frac{{\ensuremath{P_{\rm orb}}}}{2\pi}\right) \frac{1}{q}$$ The quantity $(R_{\rm A}/a)$ is of course the fractional radius of the star, $r_{\rm A}$, measured directly from the transit light curves in Section\[sec:lc\].
As the quadratic term is formally greater than zero – which equates to an increasing orbital period – we set the 3$\sigma$ limit on orbital decay to be $(q - 3\sigma_q) = -2.1 \times 10^{-9}$dd$^{-1}$ (i.e. $-33$msyr$^{-1}$). Using the quantities in Table\[tab:absdim\] and this constraint on $q$, we find a lower limit on the tidal quality factor to be $Q_\star^{\,\prime} > (4.0 \pm 0.3) \times 10^5$. The uncertainty was calculated by propagating the errors on $M_{\rm A}$, $M_{\rm b}$ and $r_{\rm A}$ with a Monte Carlo approach. For ease of comparison, this limit can also be expressed as $Q_\star^{\,\prime} > 10^{5.60\pm0.03}$.
Constraints on periodic transit timing variations
-------------------------------------------------
{width="\textwidth"}
@Burke+10apj investigated a possible sinusoidal variation in the transit timing values with a period 118.3 orbital cycles, following a suggestion by B. Gary. They found that this more complex ephemeris provided a better fit to the data but by an amount which fell far short of statistical significance. To check this out we calculated a periodogram of the residuals of the best-fitting linear ephemeris with the [period04]{} code [@LenzBreger04iaus] covering the frequency range from 0.0 to the Nyquist frequency of 0.13 cycles per day (i.e. equivalent to twice the orbital period).
Fig.\[fig:pgram\] shows the resulting frequency spectrum. The red dotted line indicates the possible period at 118.3 orbital cycles (466.3d) mentioned by @Burke+10apj: the periodogram shows no significant power at this period. The two strongest peaks are at much higher frequencies of 0.0807 and 0.0802 cycles per day, and both have a signal to noise ratio of 3.04. This is well below the value of 4.0 typically considered to be the level at which a frequency is significant [e.g. @Breger+93aa]. We therefore conclude that there is no evidence for a periodic variation in the orbital ephemeris of XO-1.
The optical-infrared transmission spectrum of XO-1b {#sec:transspec}
===================================================
[We now]{} study how the transit depth varies as a function of wavelength. This effect is caused by changes in the apparent radius of the planet, which in turn arise from variations in opacity and scattering processes in its extended atmosphere. Its transmission spectrum therefore potentially holds information about the abundances of atoms and molecules, and the temperature structure of the atmosphere.
Following the approach of @Me+12mn2, we modelled all available transit light curves of XO-1 in order to measure the planet radius (in the form of $r_{\rm b}$) as a function of wavelength. It is important to fix the geometric parameters to representative values in order to maximise the consistency between different light curve fits and to avoid sources of uncertainty which are common to all light curves. The choice of these parameters is not simple because of conflicting results from published transmission spectroscopic studies of XO-1b.
Consideration of published results
----------------------------------
{width="\textwidth"} {width="\textwidth"}
@Tinetti+10apj presented HST/NICMOS observations of a transit of XO-1 which yielded a transmission spectrum covering 1.2–1.8$\mu$m. They claimed the detection of H$_2$O, CH$_4$ and CO$_2$ molecules in the planetary atmosphere. @Burke+10apj extended this analysis to the geometric parameters of the system, and included a second (or should that be first?) NICMOS observation of XO-1 obtained 12days (three planetary orbits) prior to the observations utilised by @Tinetti+10apj.
@Gibson++11mn presented a reanalysis of the NICMOS data used by @Tinetti+10apj, with differences of approach concerning the use of decorrelation parameters to remove systematic errors in the data which arise from both HST and NICMOS. @Gibson++11mn obtained a more scattered and much more uncertain transmission spectrum, and concluded that the detection of molecules claimed by @Tinetti+10apj was not supported by the data. @Gibson++11mn concluded that NICMOS is not a suitable instrument for transmission spectroscopy as it displays unremovable systematics of similar size to the astrophysical signal being sought.
@Crouzet+12apj also presented a reanalysis of the NICMOS observations from @Tinetti+10apj, but also included the second transit of XO-1 observed 12days earlier. They performed a similar reduction of the data as @Tinetti+10apj and @Gibson++11mn, but with some different choices of instrumental parameters against which the light curves were decorrelated. They found results which were much closer to those of @Tinetti+10apj than @Gibson++11mn, but with important differences remaining at the level of the expected astrophysical signal in the transmission spectrum.
@Deming+13apj used the improved capabilities of HST/WFC3 to obtain a transmission spectrum of XO-1b over the 1.12–1.65$\mu$m wavelength interval. This was used to claim a detection of water absorption in the planetary atmosphere, as well as to rule out spectral features at the level claimed by @Tinetti+10apj. As the work by @Deming+13apj is based on a more modern analysis of data obtained using a better instrument than previous transmission spectroscopy, we have chosen to anchor our new results on the geometric parameters used in this work. They are, in turn, those found by @Burke+10apj: $r_{\rm A} = 0.0890 \pm 0.0007$ (the inverse of the quoted quantity $\frac{a}{R_\star} = 11.24 \pm 0.09$) and $i = 88.8 \pm 0.2^\circ$.
Analysis method
---------------
For each light curve we calculated the best-fitting model with [jktebop]{}. We fixed $r_{\rm A}$ at 0.0890, $i$ at 88.8$^\circ$ and the orbital period at a representative value. We fitted for $r_{\rm b}$, the time of mid-transit (to guard against possible orbital period variations) and the coefficients of the baseline polynomial [(see Table\[tab:obslog\])]{}. Uncertainties in $r_{\rm b}$ were calculated using both Monte Carlo and residual-permutation simulations, and the larger errorbar for $r_{\rm b}$ was retained in each case. We found that the uncertainties for the BUSCA $z$-band light curve were relatively large, especially for the residual-permutation simulations: this is a result of the moderate differences between the two light curves and therefore is expected.
The phenomenon of LD deserves special consideration. In a recent work on GJ1132 [@Me+17aj], and in provisional analyses for the current work, we found that the transmission spectrum was significantly affected by way in which LD was treated. We therefore modelled the light curves with a range of ways of dealing with LD. The quadratic LD law is the most widely used in the literature, but recent theoretical studies [@EspinozaJordan16mn; @Morello+17aj] have found that other laws, such as logarithmic and square-root [see @Me08mn for the equations], are capable of matching theoretical LD predictions more precisely. Logarithmic should be better than square-root in the current case, particularly for the redder optical passbands under consideration [@Vanhamme93aj].
We therefore obtained solutions to the light curves using the quadratic, logarithmic and square-root LD laws, in each case with both coefficients fixed and with the linear coefficient fitted but the nonlinear coefficient fixed. For consistency we adopted theoretical LD coefficients obtained by @Claret00aa [@Claret04aa2] using the [atlas9]{} atmosphere models [@Kurucz93], for all light curves, with the exception of the redshifted H$\alpha$ filter for which we used LD coefficients from the Johnson $R$ filter tabulated by @Vanhamme93aj.
For a comparison with the results above, and in order to capture the effect of differences in the LD coefficients used, we also fit each light curve using LD coefficients from @Claret00aa [@Claret04aa2] calculated using the [phoenix]{} model atmospheres. Fig.\[fig:ts:r2ld\] shows the results for all alternatives investigated. It can be seen that the measured value of $r_{\rm b}$ is [*not*]{} significantly affected by either the choice of LD law, whether or not one of the LD coefficients is fitted, or whether the LD coefficients come from the [atlas9]{} or [phoenix]{} model atmospheres. We also notice – perhaps counterintuitively – that fixing both LD coefficients can yield larger errorbars despite the loss of one dimension from the area of parameter space in which the solution can be located. This occurs because fixing the LD coefficients can cause a poorer fit to the data, leading to larger errorbars from the residual-permutation algorithm.
From Fig.\[fig:ts:r2ld\] we conclude that the treatment of LD does not have a significant effect on the results for individual light curves, and that it is safe to proceed with a representative set of $r_{\rm b}$ measurements. One possible exception to this rule is the $g$-band, for which the effect of LD treatment on the measured planet radius is significantly above the (very small) errorbars. Notwithstanding this, we [chose]{} as the representative set of $r_{\rm b}$ values those measured using the quadratic LD law with the linear coefficient fitted at values from the [atlas9]{} model atmospheres. Table\[tab:rb\] contains these values, and also for reference contains those from the quadratic LD law with both LD coefficients fixed. Table\[tab:rb\] also includes values for the central wavelength and full width at half maximum of the filters used to obtain our observations with BUSCA[^10] and the INT/WFC[^11], and for published data obtained using the Palomar 50in[^12], SofI[^13] and ISAAC[^14] instruments.
Results
-------
---------------- ---------------------- ----------------- ------------ ---------- -------------- ---------- --------------
Data Filter Central Band full
source wavelength (nm) width (nm)
INT/WFC redshifted H$\alpha$ 689 10 0.011983 [0.000210]{} 0.011911 [0.000150]{}
BUSCA SDSS $u$ 366 38 0.011844 0.000129 0.011690 0.000129
BUSCA SDSS $g$ 478 150 0.011483 0.000048 0.011450 0.000035
BUSCA SDSS $r$ 663 105 0.011840 0.000074 0.011698 0.000123
BUSCA SDSS $z$ 910 90 0.011906 0.000126 0.011837 0.000112
CAHA 1.23m Cousins $I$ 810 110 0.011804 0.000074 0.011741 0.000071
Holman FLWO SDSS $z$ 0.011716 0.000075 0.011750 0.000069
Holman Palomar Cousins $R$ 647 152 0.011297 0.000196 0.011231 0.000164
Cáceres SofI $J$ 1247 290 0.011725 0.000170 0.011719 0.000175
Cáceres ISAAC $J$ + block 1250 290 0.011907 0.000208 0.011822 0.000265
Sada $z^\prime$ 0.011170 0.000355 0.011880 0.000437
---------------- ---------------------- ----------------- ------------ ---------- -------------- ---------- --------------
![\[fig:rvary\] [Measured planetary radius ($R_{\rm b}$) as a function of the central wavelength of the passbands used. The passband names are give at the top of the plot. The horizontal lines indicate the FWHM of the passband used and the vertical lines show the errorbars in the $R_{\rm b}$ measurements. [*The errorbars exclude all common sources of uncertainty*]{}. Results obtained when fitting the linear LD coefficient are shown as filled circles with errorbars. Results from fixing both LD coefficients are shown as open circles without a horizontal line indicating the passband. The colour coding is consistent with Figs. \[fig:lcall\] and \[fig:lcfit\].]{}](plotvarywave.eps){width="\columnwidth"}
In Fig.\[fig:rvary\] we show the transmission spectrum of XO-1b determined from the light curves studied in this work, both new and previously published. Our preferred approach is to fit for the linear LD coefficient, and these results are shown as filled circles. The alternative approach of fixing both LD coefficients yields the results shown using open circles. Fig.\[fig:rvary\] shows the values of $R_{\rm b}$ obtained by multiplying the $r_{\rm b}$ values in Table\[tab:rb\] by the semimajor axis (0.04914AU) and a conversion factor (1AU $=$ 2092.5[$\,{\rm R}_{\rm Jup}$]{}).
It is immediately apparent from Fig.\[fig:rvary\] that different light curves in the same or similar passbands show significant [variations in $r_{\rm b}$]{}. On closer inspection the two worst offenders are the Palomar $R$-band data from @Holman+06apj and the $z$-band light curve from @Sada+12pasp. Both have a high scatter and include no observations on one side of the transit, so it is not surprising that they give $r_{\rm b}$ values which are very uncertain. This issue can be dealt with either by combining results from multiple light curves in the same or similar passbands or by ignoring the problematic results. In the current case, both options give a similar outcome.
![\[fig:rvary2\] [As Fig.\[fig:rvary\] but similar passbands have been combined into weighted mean values (the $z$-bands and the $R$ band and $r$ band), and the $J$-band results are ignored in favour of the HST/WFC3 transmission spectrum of XO-1b obtained by @Deming+13apj and shown using black filled circles.]{}](plotvarywave2.eps){width="\columnwidth"}
In Fig.\[fig:rvary2\] we show the transmission spectrum of XO-1b after some consolidation of the results. The three $z$-band $r_{\rm b}$ values have been reduced into their weighted mean, as have the Palomar $R$ and BUSCA $r$ bands, in order to stop their large errorbars obfuscating such plots. We have not combined the redshifted H$\alpha$ result with any other as the value of $r_{\rm b}$ from this light curve has [much greater wavelength resolution (resolving power $R \approx 70$) than the $R$ and $r$ bands]{}. We have furthermore ignored the $J$-band results from now on because they add nothing to our analysis: they are consistent with and are completely overlapped by published transmission spectra, but are of [lower precision and much lower wavelength resolution]{}.
In Fig.\[fig:rvary2\] we have also plotted the HST/WFC3 transmission spectrum of XO-1b obtained by @Deming+13apj, after converting it from the values of $k^2$ [@Deming+13apj their table3] to $R_{\rm b}$ consistently with our values of $r_{\rm b}$. The treatment of LD by @Deming+13apj is relevant: they used the linear LD law with coefficients fixed to values interpolated from the $J$- and $H$-band coefficients tabulated by @ClaretBloemen11aa. They account for minor variations between different sources of theoretical LD coefficients, but do not allow for any imperfections in the description of real stars by current theoretical model atmospheres. They also neglect the spectral variation of LD coefficients over wavelength intervals smaller than those of the broad-band $J$ and $H$ filters. This approach is quite simplistic, but has less impact in the infrared than at visual wavelengths, because stellar LD is weaker in the infrared.
Interpretation
--------------
![\[fig:madhu\] Best-fitting model transmission spectrum of XO-1b (dark red line). The observed transmission spectrum is shown using coloured points for the optical data and black points for the HST/WFC near-infrared data.](figure_xo1b_v5.eps){width="\columnwidth"}
We used a forward transmission spectrum model to fit the optical and near infrared data of XO-1b. For the pressure-temperature profile, we use the parameterisation of @MadhusudhanSeager09apj which consists of six free parameters. We partitioned our model atmosphere into 100 layers spaced equally in log-pressure between $10^{-6}$ bar and $10^{2}$ bar. For the atmospheric composition, we considered several chemical species with prevalent signatures in the the spectral range of the optical and near-infrared observations [@Madhusudhan12apj; @Moses+13apj; @Venot+13aa]. These include Na, K, H$_2$O, NH$_3$, HCN, and CH$_4$. The mixing ratio of each species was assumed to be uniform in the observable atmosphere and we assumed an atmosphere rich in H$_2$ and He with a He/H$_2$ ratio of 0.17. We considered line absorption from each molecular species and collision-induced opacity from H$_2$-H$_2$ and H$_2$-He. The sources of opacity for the chemical species are described in @GandhiMadhusudhan17mn [@GandhiMadhusudhan18mn]. In addition, we accounted for cloud effects due to small and large modal particle sizes. Large cloud particles were represented by a grey opacity throughout the whole spectrum and small cloud particles and/or hazes modified the H$_2$ scattering Rayleigh slope in the optical.
The full set of observations were best fitted (Fig.\[fig:madhu\]) with a patchy cloud model having a terminator cloud and haze fraction of 0.54. The patchy cloud model is generally preferred to a clear-atmosphere model at the 1.3$\sigma$ confidence level. H$_2$O is present at 3.05$\sigma$ confidence to fit the HST/WFC3 data, signifying water vapour is present with a certainty of 99.87%. Nitrogen chemistry (NH$_3$ and HCN) is hinted at 1.5$\sigma$. The data do not provide evidence for the presence of either Na or K in the planetary atmosphere. Our model fits the optical transmission spectrum in the $u$, $r$/$R$, $i$ and $z$ bands to within 0.5$\sigma$.
The best-fit model is unable to explain the measured planet radius in the $g$ band, which lies 8$\sigma$ below the model transmission spectrum and well below all other planet radius measurements. The reason for this discrepancy is not clear but is very difficult to explain theoretically, as none of our model transmission spectra exhibit a planet radius at any point in the optical which is below the radii in the infrared. It is also hard to understand observationally, as the two light curves in this passband are of high precision and very good mutual agreement, and such an effect has not been seen in this band in previous observations by our team[^15]. Temporal variability of the planet [or stellar [e.g. @Oshagh+14aa; @Rackham+17apj] atmosphere]{} cannot be culpable because both $g$-band light curves were obtained simultaneously with $z$-band and either $u$-band or $r$-band observations.
We conclude that the transmission spectrum is best reproduced by a H$_2$/He-rich planetary atmosphere containing H$_2$O with low confidence levels of patchy clouds and nitrogen-bearing molecules (NH$_3$ and HCN). An anomalously small planet radius in the $g$-band is difficult to explain either observationally or theoretically and should be investigated by obtaining new observations in this wavelength region, preferably with a significantly higher resolution.
Discrepant transit depths
-------------------------
[The referee expressed concern over the discrepant transit depth obtained from the $g$-band light curves. It is clear that there is something affecting the $g$-band data which is not accounted for in our data reduction and analysis procedures. These datasets were processed through the same data reduction and analysis programs as used by our group in many previous studies, which implies that the problem lies with the data themselves rather than with the reduction and analysis. Based on this, we rejected the $g$-band data from the analysis of the transmission spectrum. This implicitly assumes that the problem is isolated to the $g$-band alone; our results could be affected if the problem exists in other light curves or is an artefact of our data reduction pipeline.]{}
[We chose not to reject the $g$-band data when determining the physical properties of the system, and have assessed the impact of this choice by rerunning the analysis without the $g$-band data. We find that the final photometric results (Table\[tab:lcfit\]) differ by 0.3$\sigma$ for $r_{\rm b}$ and less than 0.1$\sigma$ for $i$ and $r_{\rm A}$. The physical properties of the system in Table\[tab:absdim\] are unchanged except that $R_{\rm b}$ increases by 0.2$\sigma$ and $\rho_{\rm b}$ decreases by 0.3$\sigma$. The inclusion of the $g$-band data therefore does not have a significant effect on the measured physical properties of the XO-1 system.]{}
Impact of the optical data
--------------------------
{width="\textwidth"}
[One purpose of the current work was to see what improvement in our understanding of the properties of the atmosphere of XO-1b could be obtained by adding optical transit data to the HST near-infrared transmission spectrum. We investigated this by modelling both the full transmission spectrum and the HST results only.]{}
[We find that the addition of the optical data to the near-infrared observations introduces an alternative water abundance estimate. Fig.\[fig:OPTICAL+WFC3\] shows the retrieved water abundances for the case of our optical observations plus the HST data, and for the HST data alone. In the latter case the modal H$_2$O abundance is approximately $-1$dex with a median and $1\sigma$ errorbars of $-1.45^{+0.50}_{-2.19}$. The slight tail of the posterior distribution arises from a weak degeneracy with HCN. The adjoined observations in the visible offer a complementary interpretation of XO-1b’s atmosphere, adding a second mode to the H$_2$O mixing ratio at $-4$dex and thus altering the median abundance by approximately $-2$dex.]{}
[The two interpretations of XO-1b’s atmospheric H$_2$O concentration emerge from two possible cloud condensate configurations. The [water abundance mode at approximately $-4$dex that is introduced by the optical data suggests]{} an atmosphere with condensate clouds composed of particle sizes $\sim$1 $\mu$m whose cloud-top pressures are 0.01 to 0.1 mbar. The formation efficiency of condensate particles decreases with atmospheric height [@Parmentier++13aa], and therefore clouds extending to low pressures of 0.01–0.1mbar require vertical mixing processes such as convection which could advect material upward. [On the other hand, the second mode constituting a high water abundance of approximately $-1$dex proposes cloud-top pressures greater than 1mbar]{}. Ultimately, elucidating the atmosphere of XO-1b from these two distinct possibilities (low water abundance/high-extending clouds, and high water abundance/low-extending clouds) will have to await more precise observations in the optical.]{}
Summary and discussion {#sec:summary}
======================
[XO-1 has been identified as a good candidate for the JWST Early Release Science program [@Stevenson+16pasp]. A near-infrared transmission spectrum for XO-1b has previously been obtained using HST/WFC3, resulting in the detection of water in the planetary atmosphere.]{} We have obtained a total of ten high-precision transit light curves covering the full optical wavelength range (366nm to 910nm) in order to extend this transmission spectrum to optical wavelengths.
We use our data, alongside published transit light curves and spectroscopic quantities of the host star, to measure the physical properties of the system. Our results are in good agreement with, and more precise than, previous studies. We also assemble all available transit timing measurements and derive a high-precision orbital ephemeris useful for scheduling future observations. We find no evidence for periodic deviations from this ephemeris, contrary to previous suggestions. The non-detection of any quadratic deviation from the linear ephemeris allows us to constrain the tidal quality factor for the host star to be $Q_\star^{\,\prime} > 10^{5.60}$.
We fitted the transit light curves using the same system geometry as for the HST/WFC3 observations, in order to measure the radius of the planet as a function of wavelength. This optical-infrared transmission spectrum is well fitted by a model spectrum for a planet with a H$_2$/He-rich atmosphere and patchy cloud. H$_2$O is detected to 3.05$\sigma$ while suggestions of patchy clouds (1.3$\sigma$) and nitrogen chemistry (1.5$\sigma$) are weak given the present observations. [We find that adding the optical to the near-infrared data leads to [*less*]{} precise constraints on the planetary atmosphere. This indicates that optical observations of a higher precision and spectral resolution would be needed to improve our understanding of the atmosphere of XO-1b, and also that there is some tension between the best-fitting atmospheric properties in the optical and in the near-infrared.]{} The planet radius we measure in the $g$-band is anomalously low, a finding difficult to explain either observationally or theoretically. We advocate further observations in this wavelength region, with a higher spectral resolution. Throughout this work we have paid careful attention to the treatment of LD when fitting transit light curves. When measuring the physical properties of the system we used four different LD laws and two different approaches to fitting the coefficients of these. We find that the range of solutions produced by these different fits is very small when fitting high-quality data, so the treatment of LD is thankfully not a significant hindrance to measuring the system properties. From a similarly detailed investigation concerning the transmission spectrum, we find that the choice of LD law, and whether or not to fit for one of the coefficients, is unimportant, giving rise to a scatter in the planet radius measurements which is small compared to the variation between light curves. The only exception to this rule is for the $g$-band, where the very small uncertainties in the planet radius do not fully cover the scatter between solutions with a different treatment of LD. Whilst the situation for XO-1 is encouraging, we urge that similar analysis should be performed as standard procedure when obtaining transmission spectra. This is particularly true for planets transiting low-mass stars, whose LD may not be well captured by parametric laws and for which LD coefficients are more difficult to derive theoretically.
We confirm that XO-1 is an excellent target for future observations with JWST. Its physical properties are well-understood, the planet’s transmission spectrum has features comparatively easy to measure using existing instrumentation, its solar-type host star shows no sign of chromospheric activity, and our new orbital ephemeris is precise enough to predict transits to within $\pm$5s up to the year 2266.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the referee for comments which improved the paper and encouraged us to test the reliability of our results. The reduced light curves presented in this work will be made available at the CDS ([http://vizier.u-strasbg.fr/]{}) and at [http://www.astro.keele.ac.uk/jkt/]{}. JS acknowledges financial support from the Leverhulme Trust in the form of a Philip Leverhulme Prize. AP is grateful for research funding from the Gates Cambridge Trust. LM acknowledges support from the Italian Minister of Instruction, University and Research (MIUR) through FFABR 2017 fund, and from the Department of Physics of the University of Rome Tor Vergata, through Mission Sustainability 2016 funds. The following internet-based resources were used in research for this paper: the ESO Digitized Sky Survey; the NASA Astrophysics Data System; the SIMBAD database and VizieR catalogue access tool operated at CDS, Strasbourg, France; and the ar$\chi$iv scientific paper preprint service operated by Cornell University. Based on observations collected at the Centro Astronómico Hispano Alemán (CAHA) at Calar Alto, Spain, operated jointly by the Max-Planck Institut für Astronomie and the Instituto de Astrofísica de Andalucía (CSIC), and on observations made with the Isaac Newton Telescope 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
[103]{} natexlab\#1[\#1]{}
, H., 1974, IEEE Transactions on Automatic Control, 19, 716
, M., [Santos]{}, N. C., [Sousa]{}, S. G., [Fernandes]{}, J., [Guillot]{}, T., [Israelian]{}, G., [Mayor]{}, M., [Melo]{}, C., 2009, A&A, 507, 523
, D. R., et al., 2012, MNRAS, 422, 1988
, C., et al., 2014, PASP, 126, 1134
, J. L., et al., 2014, MNRAS, 440, 1470
, A. S., et al., 2017, A&A, 602, A107
, M., et al., 1993, A&A, 271, 482
, J. M., [Fischer]{}, D. A., [Valenti]{}, J. A., [Piskunov]{}, N., 2016, ApJS, 225, 32
, C. J., et al., 2010, ApJ, 719, 1796
, C., [Ivanov]{}, V. D., [Minniti]{}, D., [Naef]{}, D., [Melo]{}, C., [Mason]{}, E., [Selman]{}, F., [Pietrzynski]{}, G., 2009, A&A, 507, 481
, J. A., [Yee]{}, J. C., [Eastman]{}, J., [Gaudi]{}, B. S., [Winn]{}, J. N., 2008, ApJ, 689, 499
, D., [Brown]{}, T. M., [Noyes]{}, R. W., [Gilliland]{}, R. L., 2002, ApJ, 568, 377
, S., et al., 2015, A&A, 577, A54
, A., 2000, A&A, 363, 1081
, A., 2004, A&A, 428, 1001
, A., 2004, A&A, 424, 919
, A., [Bloemen]{}, S., 2011, A&A, 529, A75
, A., [Hauschildt]{}, P. H., 2003, A&A, 412, 241
, N., [McCullough]{}, P. R., [Burke]{}, C., [Long]{}, D., 2012, ApJ, 761, 7
, P., [Woo]{}, J.-H., [Kim]{}, Y.-C., [Yi]{}, S. K., 2004, ApJS, 155, 667
, D., et al., 2013, ApJ, 774, 95
, A., [Chaboyer]{}, B., [Jevremovi[ć]{}]{}, D., [Kostov]{}, V., [Baron]{}, E., [Ferguson]{}, J. W., 2008, ApJS, 178, 89
, J., [Siverd]{}, R., [Gaudi]{}, B. S., 2010, PASP, 122, 935
, N., [Jord[á]{}n]{}, A., 2016, MNRAS, 457, 3573
, D. F., [Southworth]{}, J., [Smalley]{}, B., 2016, ApJ, 833, L19
, P. D., et al., 2016, ApJ, 827, 19
, M., [Ichikawa]{}, T., [Gunn]{}, J. E., [Doi]{}, M., [Shimasaku]{}, K., [Schneider]{}, D. P., 1996, AJ, 111, 1748
, S., [Madhusudhan]{}, N., 2017, MNRAS, 472, 2334
, S., [Madhusudhan]{}, N., 2018, MNRAS, 474, 271
, N. P., [Pont]{}, F., [Aigrain]{}, S., 2011, MNRAS, 411, 2199
, T. P., [Line]{}, M. R., [Montero]{}, C., [Fortney]{}, J. J., [Lustig-Yaeger]{}, J., [Luther]{}, K., 2016, ApJ, 817, 17
, M. J., et al., 2006, ApJ, 652, 1715
, B., [Greenberg]{}, R., [Barnes]{}, R., 2008, ApJ, 678, 1396
, B., [Barnes]{}, R., [Greenberg]{}, R., 2009, ApJ, 698, 1357
, H. L., 1964, Boletin de los Observatorios Tonantzintla y Tacubaya, 3, 305
, H. A., [Howard]{}, A. W., [Isaacson]{}, H., 2010, ApJ, 720, 1569
, R., 1993, ATLAS9 stellar atmosphere programs and 2 km/s grid. Kurucz CD-ROM No. 13
, P., [Breger]{}, M., 2004, in [J. Zverko, J. [Ž]{}i[ž]{}novsky, S. J. Adelman, & W. W. Weiss]{}, ed., The A-Star Puzzle, Cambridge University Press, Cambridge, UK., vol. 224 of *IAU Symposium*, p. 786
, B., [Winisdoerffer]{}, C., [Chabrier]{}, G., 2009, ApJ, 692, L9
, A. E. H., 1911, [Some Problems of Geodynamics]{}, [Cambridge University Press, Cambridge, 1911]{}
, P., [McCullough]{}, P. R., [Burke]{}, C. J., [Valenti]{}, J. A., [Burrows]{}, A., [Hora]{}, J. L., 2008, ApJ, 684, 1427
, N., 2012, ApJ, 758, 36
, N., [Seager]{}, S., 2009, ApJ, 707, 24
, N., [Winn]{}, J. N., 2009, ApJ, 693, 784
, N., [Amin]{}, M. A., [Kennedy]{}, G. M., 2014, ApJ, 794, L12
, E. E., [Hillenbrand]{}, L. A., 2008, ApJ, 687, 1264
, L., et al., 2014, MNRAS, 443, 2391
, D. W., 1963, J. Soc. Ind. Appl. Math., 11, 431
, P. F. L., [Serenelli]{}, A. M., [Southworth]{}, J., 2015, A&A, 577, A90
, P. R., et al., 2006, ApJ, 648, 1228
, C., [van Boekel]{}, R., [Molli[è]{}re]{}, P., [Henning]{}, T., [Benneke]{}, B., 2016, ApJ, 832, 41
, G., [Tsiaras]{}, A., [Howarth]{}, I. D., [Homeier]{}, D., 2017, AJ, 154, 111
, A., [Santos]{}, N. C., [Sousa]{}, S. G., [Fernandes]{}, J. M., [Adibekyan]{}, V. Z., [Delgado Mena]{}, E., [Montalto]{}, M., [Israelian]{}, G., 2013, A&A, 558, A106
, J. I., [Madhusudhan]{}, N., [Visscher]{}, C., [Freedman]{}, R. S., 2013, ApJ, 763, 25
, N., [Sing]{}, D. K., [Gibson]{}, N. P., [Fortney]{}, J. J., [Evans]{}, T. M., [Barstow]{}, J. K., [Kataria]{}, T., [Wilson]{}, P. A., 2016, ApJ, 832, 191
, G. I., 2014, ARA&A, 52, 171
, G. I., [Lin]{}, D. N. C., 2007, ApJ, 661, 1180
, M., [Santos]{}, N. C., [Ehrenreich]{}, D., [Haghighipour]{}, N., [Figueira]{}, P., [Santerne]{}, A., [Montalto]{}, M., 2014, A&A, 568, A99
, A., 2008, MNRAS, 390, 281
, V., [Showman]{}, A. P., [Lian]{}, Y., 2013, A&A, 558, A91
, K., [Sasselov]{}, D., 2011, ApJ, 731, 67
, K., [Jackson]{}, B., [Spada]{}, F., [Thom]{}, N., 2012, ApJ, 751, 96
, A., [Cassisi]{}, S., [Salaris]{}, M., [Castelli]{}, F., 2004, ApJ, 612, 168
, S., [Br[á]{}t]{}, L., [Pejcha]{}, O., 2010, New Astronomy, 15, 297
, D. L., et al., 2006, PASP, 118, 1407
, F., [Husnoo]{}, N., [Mazeh]{}, T., [Fabrycky]{}, D., 2011, MNRAS, 414, 1278
, F., [Sing]{}, D. K., [Gibson]{}, N. P., [Aigrain]{}, S., [Henry]{}, G., [Husnoo]{}, N., 2013, MNRAS, 432, 2917
, W. H., [Teukolsky]{}, S. A., [Vetterling]{}, W. T., [Flannery]{}, B. P., 1992, [Numerical Recipes in FORTRAN 77. The art of scientific computing]{}, Cambridge: University Press, 2nd ed.
, B., et al., 2017, ApJ, 834, 151
, S., et al., 2009, AN, 330, 475
, P. V., et al., 2012, PASP, 124, 212
, G., 1978, Annals of Statistics, 5, 461
, S., [Mall[é]{}n-Ornelas]{}, G., 2003, ApJ, 585, 1038
, D. K., et al., 2016, Nature, 529, 59
, M. F., et al., 2006, AJ, 131, 1163
, J., 2008, MNRAS, 386, 1644
, J., 2009, MNRAS, 394, 272
, J., 2010, MNRAS, 408, 1689
, J., 2011, MNRAS, 417, 2166
, J., 2012, MNRAS, 426, 1291
, J., 2013, A&A, 557, A119
, J., [Evans]{}, D. F., 2016, MNRAS, 463, 37
, J., [Mancini]{}, L., [Maxted]{}, P. F. L., [Bruni]{}, I., [Tregloan-Reed]{}, J., [Barbieri]{}, M., [Ruocco]{}, N., [Wheatley]{}, P. J., 2012, MNRAS, 422, 3099
, J., [Mancini]{}, L., [Madhusudhan]{}, N., [Molli[è]{}re]{}, P., [Ciceri]{}, S., [Henning]{}, T., 2017, AJ, 153, 191
, J., et al., 2009, MNRAS, 396, 1023
, J., et al., 2014, MNRAS, 444, 776
, J., et al., 2015, MNRAS, 454, 3094
, J., et al., 2016, MNRAS, 457, 4205
, P. B., 1987, PASP, 99, 191
, K. B., et al., 2016, PASP, 128, 094401
, J. K., [Cunha]{}, K., [Smith]{}, V. V., [Schuler]{}, S. C., [Griffith]{}, C. A., 2014, ApJ, 788, 39
, G., [Deroo]{}, P., [Swain]{}, M. R., [Griffith]{}, C. A., [Vasisht]{}, G., [Brown]{}, L. R., [Burke]{}, C., [McCullough]{}, P., 2010, ApJ, 712, L139
, G., [Winn]{}, J. N., [Holman]{}, M. J., 2008, ApJ, 677, 1324
, G., [Fischer]{}, D. A., [Sozzetti]{}, A., [Buchhave]{}, L. A., [Winn]{}, J. N., [Holman]{}, M. J., [Carter]{}, J. A., 2012, ApJ, 757, 161
, J., et al., 2018, MNRAS, 474, 5485
, A., et al., 2018, AJ, 155, 156
, W., 1993, AJ, 106, 2096
, D. A., [Bergbusch]{}, P. A., [Dowler]{}, P. D., 2006, ApJS, 162, 375
, M., et al., 2009, in [Pont]{}, F., [Sasselov]{}, D., [Holman]{}, M. J., eds., Transiting Planets, vol. 253 of *IAU Symposium*, p. 440
, O., [H[é]{}brard]{}, E., [Ag[ú]{}ndez]{}, M., [Decin]{}, L., [Bounaceur]{}, R., 2015, A&A, 577, A33
, A. N., [Delrez]{}, L., [Barker]{}, A. J., [Deming]{}, D., [Hamilton]{}, D., [Gillon]{}, M., [Jehin]{}, E., 2017, ApJ, 836, L24
, D. M., et al., 2006, PASP, 118, 1245
, M., [Brandner]{}, W., [Bergfors]{}, C., [Henning]{}, T., 2015, A&A, 575, A23
[^1]: [http://catserver.ing.iac.es/filter/list.php?\
instrument=WFC]{}
[^2]: [http://idlastro.gsfc.nasa.gov/]{}
[^3]: [http://www.harrisgeospatial.com/SoftwareTechnology/ IDL.aspx]{}
[^4]: [jktebop]{} is written in [fortran77]{} and the source code is available at [http://www.astro.keele.ac.uk/jkt/codes/jktebop.html]{}
[^5]: Bilinear interpolation was performed using the [jktld]{} code at: [http://www.astro.keele.ac.uk/jkt/codes/jktld.html]{}
[^6]: [http://www.astro.caltech.edu/palomar/observer/\
P60observers.html]{}
[^7]: [http://var2.astro.cz/ETD/credit.php]{}
[^8]: [http://brucegary.net/AXA/x.htm]{}
[^9]: Note that the term $({\ensuremath{P_{\rm orb}}}/2\pi)$ is inverted in equations 3 and 5 of @Wilkins+17apj.
[^10]: [https://www.caha.es/CAHA/Instruments/filterlist.html]{}
[^11]: [http://catserver.ing.iac.es/filter/list.php?\
instrument=WFC]{}
[^12]: [http://www.astro.caltech.edu/palomar/observer/\
60inchResources/p60filters.html]{}
[^13]: [http://www.eso.org/sci/facilities/lasilla/\
instruments/sofi/inst/Imaging.html]{}
[^14]: [http://www.eso.org/sci/facilities/paranal/\
decommissioned/isaac/doc/VLT-MAN-ESO-14100-0841\_v90.pdf]{}
[^15]: For example WASP-57 [@Me+15mn2], HAT-P-23 and WASP-48 [@Ciceri+15aa2], Qatar-2 [@Mancini+14mn] and HAT-P-32 [@Tregloan+18mn].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quantization is still a central problem of modern physics. One example of an unsolved problem is the quantization of Nambu mechanics. After a brief comment on the role of Harrison cohomology, this review concentrates on the central problem of quantization of QCD and, more generally, quark confinement seen as a problem of quantization. Several suggestions are made, some of them rather extravagant.'
author:
- Christian Frønsdal
date: 'Received: December 5, 2000'
title: Retrospective on Quantization
---
Quantization
============
A well known text book on Quantum Mechanics offers, in an early chapter, a Universal Quantization Paradigm. After the usual familiar applications one finds, towards the end of the book, a chapter dealing with the rigid rotator. Needless to say, the “universal" paradigm has been forgotten.
I remember trying to explain to a colleague, more than 25 years ago, what we were trying to understand about quantization. He would not agree that there was a problem. Of course, we were all aware of the connection between quantization and the theory of group representations; this was explained, in principle, already by von Neumann, in his proof of the uniqueness of the unitary representation of the Heisenberg group. But few people anticipated the greatly expanded role that has come to be played by the Heisenberg group in modern mathematics, or the much more sophisticated applications of group representation theory found in modern physics.
Moshé had come to realize that the essence of quantization is deformation theory, almost any deformation of a classical structure amounts to quantization [@Fl1]. Perhaps the most radical example turned up during his last attack on the problem of quantization of Nambu mechanics [@DFST]. This problem leads to a search for an *abelian* deformation of the ordinary algebra of functions. It turns out that such deformations are governed by Harrison cohomology, just as non-abelian deformations are described within de Rham cohomology. But Harrison cohomology is trivial on smooth manifolds [@Pi]. This is an interesting difficulty, but a greater obstacle is the fact that few mathematicians have taken up the study of this subject, so that no instructive examples of Harrison cohomology are available [@H; @Ba; @GS]. The following example may therefore be of some interest.
Consider the algebraic variety $A$ (a cone, or degenerate conic) associated with conformal field theory in $n-1$ real, Euclidean dimensions, $$A = \Rit^{n+1}/\sum_{i = 1,...,n}x_i^2-y^2=0\,.$$ Let $W$ be the commutative $\Cit$-algebra of polynomials in the coordinates $x_1,...,x_n,y$, $$W = \Cit[x_1,...,x_n,y],$$ and let $W(A)$ be the quotient algebra $$W(A) = W/ \sum_{i = 1,...,n}x_i^2-y^2 = 0$$ of functions on $A$. This algebra admits non trivial Harrison cohomology. We exhibit a particular, non trivial cocycle and use it to construct an abelian deformation of the algebra of functions on $A$.
Every $f\in W(A)$ has a unique decomposition, $$f(x,y) = f_1(x) + yf_2(x),$$ with $f_1,f_2$ being polynomials in $x_1,...,x_n$. Define a deformed product, denoted $*$, by $$f*g = fg + \lambda C(f,g),~~ C(f,g) := f_2 \,g_2,~~ f,g \in W(A).$$ That $C$ is a cocycle is readily verified; that is, $$dC(f,g,h) := fC(g,h) - C(fg,h) + C(f,gh) - C(f,g)h = 0.$$ It follows that, to first order in $\lambda$ (regarded as a formal variable) the deformed product is associative. (Actually, it is associative to all orders in $\lambda$.) That it is not trivial can be seen by the observation that, after fixing $\lambda$ in $\Rit$, it is isomorphic to $$\Cit [x_1,...,x_2,y]/\sum_{i = 1,...,n}x_i^2 -y^2 + \lambda
= 0,$$ an algebra of functions on the nondegenerate conic $\Rit^{n+1}/\sum_{i = 1,...,n}x_i^2 -y^2 + \lambda = 0$.
I suggest that Harrison cohomology may have interesting applications in quantum field theory, especially in connection with certain anomalies, and perhaps in the context of operator product expansions.
Color
=====
The success of the concept of color in strong interaction physics goes far beyond solving the problem for which it was invented, that of allowing quarks to evade the constraint that connects spin and statistics [@T; @G-M]. It is a pity, nevertheless, that this rather primitive idea of color has displaced the far more interesting concept of parastatistics. Wigner’s point of view, that led him to invent parastatistics, is curiously close to the more modern concept of current algebra [@Wig]. Evidently, the observables are the currents. In conventional local quantum field theory they are bilinears in the enveloping algebra of a Heisenberg algebra; their algebraic structure is that of $Sp(\infty)$, and if the linears and the unit are included we obtain the super Lie algebra $OSp(\infty)$. Naturally, one is interested in representations of this algebra, more particularly in those of highest weight, realized in Fock spaces. The familiar ones are induced from one-dimensional representations of the subalgebra that is generated by $1, a_i$ and $a_ia^*_j +
a^*_ja_i,~i,j = 1,2,...~$, of the form $$a_i|0\rangle = 0~~(a_ia^*_j +a^*_ja_i)|0\rangle =
\lambda\,\delta_{ij}|0\rangle.$$ The representation, and with it the statistics, is completely fixed by the number $\lambda$. This is a simple but fundamental fact that deserves to be mentioned in any text that aims to explain Bose-Einstein quantization. The simplest case is $\lambda = 1$; in this case one easily verifies that the vector $$(a^*_ia^*_j -a^*_ja^*_i)|0\rangle$$ has zero norm. To get a Hilbert space, one must project out the subspace that contains all null vectors, hence the postulate that creation operators commute is actually redundant, as is the statement that the many-particle states are symmetric. Higher, integer values of $\lambda$ lead to parastatistics [@GM; @FF1].
Interest in parastatistics had another aim in the context of strong interactions, the problem of confinement [@GM]. Taken in a very general sense, the problem is to construct a theory that contains particles that do not appear as asymptotic states. This is very easy to do in a nonrelativistic context, but very difficult to reconcile with relativity and the precepts of local field theory. Let us talk about this extremely important subject.
The hypothetical theory that I want to discuss contains field operators that have some properties in common with the quark fields of QCD. Namely, they have the same formal transformation properties under various groups, and they appear in perturbation theory in much the same way. But the crucial properties are these:
\(1) The theory contains no asymptotic quarks and
\(2) the asymptotic states are (in some sense) systems consisting of several quarks.
Here are some of the ways that one can fantasize about these objects.
A. The quarks are much like ordinary particles, but they are prevented from separating from each other because they are tied together with rubber bands or flux tubes. This is the most “physical" model. The most basic properties of the theory are very different from the theory of free quarks, and the true nature of the theory is not revealed by perturbation theory.
B. The quarks are singletons; quantum fields that are not locally observable except when produced in pairs [@FF2]. Ordinary massless fields are singleton currents, quantization relies on a generalized version of the Heisenberg algebra. The quark fields are permitted to be non-local since they are not locally observable. Indeed, it seems a pity, when one tries to deal with unobservable fields, not to take advantage of this freedom to step outside the narrow framework of local field theory[^1]. Singleton field theory provides an example of topological gauge fields that can be transformed away locally; possibly there are other ways to introduce quark fields that are artifacts of a gauge symmetry. The problem is to relate them to hadrons.
C. Perturbative local field theory is somewhat at a loss when it comes to incorporating unstable particles. It is possible, however, to concoct a real Lagrangian that describes particles with complex masses, that I shall call quirks, interacting with more conventional fields, and such that the Fourier transforms of the propagators are analytic near the real axis. If $\psi(p)$ is the Fourier transform of such a field operator, then the norm of the asymptotic state associated with it is given by the discontinuity of the generalized function $G(p^2)$ defined by $$\langle 0|\psi(p)\bar\psi(q)|0\rangle = \delta(p-q)G(p^2)$$ across the real axis, and this vanishes for quirks. Hence states that contain an isolated quirk is null. States that contain two or more quirks do not have zero norm; or rather, whether they do or do not depends on the Feynman rules that one invents for them. More about this below.
D. Recall that the scattering amplitudes of string theory, invented by Veneziano [@V], describe the scattering of spinless particles that turned out to be tachyons. String theory came of age when, with the aid of supersymmmetry, the means were found to decouple this unphysical mode from the physical states. Yet all string states are constructed as states of two or more of these objects. The similarity to the philosophy of QCD is striking. The “constituents" of string theory are bosons, those of QCD are fermions. What is wanted is a QCD without physical quarks, or a string theory with constituents that have some of the properties of quarks. This topic will also be developed below.
Quirks
======
The dominant modern attitude to elementary particle physics is based on the idea of scattering matrix, a mapping of a Hilbert space of initial on an identical Hilbert space of final states. The salient point about unstable particles is that they do not appear among these states. Consider a fictitious unstable particle – a “quirk" – and let us try to associate a fundamental field operator to it. If all interactions are switched off, then the unstable particle presumably becomes stable, and in that limit it must appear in the Hilbert space of asymptotic states; at least, that is the conventional point of view. But this would imply an important discontinuity in the structure of the space of asymptotic states, and consequently a discontinuity of the values of physical observables, as functions of the coupling strengths; the very concept of analytic perturbations becomes meaningless.
The suggestion that we offer here is as follows. Let us incorporate the unstable nature of quirks into the point of departure of perturbation theory. In the free theory there are stable particles represented by free fields, but the free Lagrangian also contains fundamental noninteracting quirk fields. Once we understand *this* free field theory, then interactions can be introduced, and now there is a hope of constructing an analytic perturbation theory. Procedures such as this are familiar in the context of quantum mechanics, where the first step in setting up a perturbation scheme consists of splitting the Hamiltonian into two parts. Two considerations govern the splitting; that is, the choice of the unperturbed Hamiltonian and [*ipso facto*]{} the nature of the unperturbed system: (1) The free theory should be exactly solvable, and (2) the perturbation should be analytic. (Example: a Schrödinger atom with an additional potential proportional to $1/r^2$; it is necessary to include this term in the free Hamiltonian. Another example: Dollard’s treatment of the infrared problem [@Do].) The use of counterterms in renormalization theory can be described in the same way. And here the analogy is a close one, for the width of the resonance may well be renormalized in perturbation theory. The shift of the vacuum in the Standard Model, to the minimum of the Higgs potential, offers another close analogy, for the non-vanishing vacuum expectation value redefines the free theory by incorporating into it a feature of the interaction, to render the perturbation analytic (or at least, more likely to be analytic).
Let $ |0\rangle$ be the unique vacuum state, and let $\psi$ denote a free Dirac-type quirk field operator. We anticipate a perturbation theory in which the calculation of S-matrix elements reduces to the evaluation of vacuum expectation values of products of field operators. Suppose that $$\psi(f)|0\rangle \neq 0,~~~\psi(f) = \int f(x)\psi(x),$$ then there is no alternative to accepting the fact that this state must have a place in the theory. But there is no need to jump to the conclusion that it belong to the Hilbert space of asymptotic states. Taking a cue from gauge theories, we suppose that fundamental field operators act in a larger space, possibly indefinite, and that the physical Hilbert space is a quotient space (or a quotient of a semi-definite subspace) by a subspace of null states (states of vanishing norm). In particular, to avoid the appearance of quirks in the asymptotic states, we must suppose that the asymptotic state associated with $\psi(f)|0\rangle$ have zero norm.
Let us further assume the conventional construction of the inner product, in terms of $\psi(f)$ and a hermitian conjugate field operator $\bar \psi(f)$, namely, the norm of this asymptotic state is determined by a discontinuity of the matrix element $\langle 0|~ \bar\psi(f)
\psi(f)~|0\rangle.$ Unitarity of the theory demands that this discontinuity be represented as $$\sum_n \langle 0|~ \bar\psi(f) |n\rangle\langle n|
\psi(f)~|0\rangle,$$ where the sum runs over a set of asymptotic states. Since the quirk field operator does not create such states, the norm must be zero: $${\mathrm{Discontinuity~}} \langle 0|~ \psi(f) \bar\psi(f)~|0\rangle = 0.$$
This is our main premise: that a state created from the vacuum by the quantum field operator associated with an unstable particle have zero norm. Let us emphasize that this refers to the action of [**one**]{} fundamental, unstable field operator; we shall see that some states containing two unstable particles do not have zero norm. This is a minor miracle that reflects the fact that, strictly speaking, quantum field operators are not operators (as we know, for example, by the existence of anomalies).
The quirk Lagrangian
====================
We construct an example of a Lagrangian field theory of noninteracting quirks. Since the Lagrangian must be real we need two spinor field, $\psi_1$ and $\psi_2$. The following vacuum expectation values for quirk fields incorporates all expected properties, $$\begin{aligned}
\langle 0|\psi_1(x)\bar \psi_2(x')|0\rangle &=& \int dp~ e^{ip(x-x')}
\frac{1}{p-m +i\lambda},\\
\langle 0|\psi_2(x)\bar \psi_1(x')|0\rangle &=& \int dp~
e^{ip(x-x')} \frac{1}{p-m -i\lambda},\end{aligned}$$ where $m$ and $\lambda$ are real parameters that characterize the quirk. (The integration runs over $\mathbb{R}^4$.) Namely, (1) it has the usual hermiticity property for Dirac fields, (2) the Fourier transform can be continued to a function analytic near the real axis, with no discontinuity, and consequently the norm of the asymptotic state associated with $\psi(x)|0\rangle$ is zero, (3) the matrix elements fall off exponentially for large values of $|(x-x')^2|$.
This function will take the place, for quirk fields, of the usual vacuum expectation value of the time ordered product. To complete the Wick rules we need to define the vacuum expectation values of products of the fundamental fields; we may assume that this will involve symmetrization, and it is immediately clear that there may be an opportunity for getting around the usual connection between spin and statistics.
A real Lagrangian that is consistent with these rules is, $$L = \bar\psi_1(p-m+i\lambda)\psi_2 - \bar\psi_1J_1 -
\bar\psi_2J_2 + {\rm h.c.}~.$$ A change of basis leads to $$L = \bar\psi_+(p-m)\psi_+ - \bar\psi_-(p -m)\psi_- +
\lambda(\bar\psi_-\psi_+ + \bar\psi_+\psi_-) -(\psi_+J_+ +
\psi_-J_- + h.c.).$$ The $\lambda$-term can be introduced via a Yukawa interaction with a scalar field that is given a non-vanishing vacuum expectation value. Of course, the $m$-term can be interpreted in the same way, by the usual Higgs mechanism. It is perhaps significant that both sectors are unstable without the vacuum expectation value.
The idea of associating certain exotic particles with null states has been advanced in other contexts as well, as we have mentioned already. The most obvious apparent difficulty would seem to be the argument that “if $\psi|0\rangle$ has zero norm, then surely $\psi\psi|0\rangle$ must have zero norm as well". This would be true if the operator products were not singular. The norm of a 2-quirk state; for example, of the asymptotic state associated with $\psi_1(f)\psi_2(f')|0\rangle$, is determined by the discontinuity of $$\int dk \frac{1}{k -m +i\lambda } ~\frac{1}{p-k -m -i\lambda }
f(k)f'(p-k).\eqno(4.1)$$ Ordinary Feynman rules, applicable to stable particles, are evaluated by conversion to complex contour integrals, and deformation of the contours of integration over internal energies, here $k_0$. The location of the singularities, in the second and fourth quadrants, authorizes the rotation of the contour from the real axis to the imaginary axis. (More precisely, this is true for real $\vec k$ and for $p$ real and spacelike.) In (4.1) we have an integrand with poles in all four quadrants of the complex $k_0$ plane, and rotation of the contour is not possible. When $\lambda$ is not zero, and if the $k_0$ contour follows the real axix, then this function has no discontinuity for any real value of $p$. Instead we shall define the two-point quirk correlation function by (4.1), but with the $k_0$ integration following the imaginary axis. The usual treatment of this integral leads to an expression that is analytic in $p$ except for singularities of a parameter-integral of the form $$\int_0^1d\alpha \big[ \alpha^2p^2 - \alpha(p^2 - 4i\lambda m) +
(m-i\lambda)^2\big]^{-2}.$$ Vary $p^2$ along the real axis, from left to right. The two poles start at $\alpha = 0,1$, move into the upper half plane until they meet at the point $(1 + im/\lambda)/2$ when $p^2 = -4\lambda^2$. One of them goes to infinity as $p^2$ passes the origin and returns in the lower half plane, while the other crosses the real axis at $1/2$ when $p^2 = 2m^2-2\lambda^2$. The contour of integration finally gets pinched at at $(1 -i\lambda/m)/2$ when $p^2$ reaches the value $4m^2$. Result: there is branch point at $p^2 = (2m)^2$, independently of the value of $\lambda$.
Thus we learn that the asymptotic state associated with $\bar\psi_1\bar\psi_2 |0\rangle$ does not have zero norm, in spite of the fact that the states $\bar\psi_1 |0\rangle$ and $\bar\psi_2|0\rangle$ are null states. There are physical, propagating states with masses above $2m$, independently of the value of $\lambda$.
Dynamics of string theory
=========================
Now I want to tell you a very ancient story, as briefly as I can. The story begins with the invention of the Dirac equation [@Di], $$(p\cdot \gamma - m)\psi(x) = 0.$$ Recall that $\psi$ is a field on space time, taking values in a finite dimensional Lorentz module. Solutions exist only for timelike momenta, this is of primary importance to the interpretation. Dirac’s equation applies to electrons; for a while it was thought to apply to protons as well. Of course it does, in a very approximate sense. But to keep to the historical order let me next recall the very interesting work of Majorana [@M].
Majorana may have had the same motivation as Dirac, but he was less ready to accept the presence of solutions with negative energy. He asked if it is possible for the operator $\gamma_0$ to have a strictly positive spectrum. Of course, relativistic invariance demands that the field take values in a Lorentz module, and the answer found by Majorana was that yes, it is possible if this module is unitary and hence infinite dimensional. Majorana’s work was ignored.
I come now to the applicability of Dirac’s equation to the proton. As the experimental probing of proton structure advanced, it became clear that it was not a simple Dirac particle. The structure was described in terms of elastic form factors. These form factors decrease with increasing momentum transfer, and this is not consonant with Dirac’s equation. More generally, any particle described by a Dirac-type equation, where the field lives in a finite dimensional Lorentz module, has polynomial form factors. In contrast, the form factor of a Majorana particle does decrease with increasing momentum transfer; in fact, it looks very much like the proton form factor found experimentally.
For this reason, theories of the Majorana type became, briefly, quite popular. One knows what is wrong with these infinite component field theories, but let me first go over some of their good points. Perhaps the most surprising fact is that some theories of this type have a very sensible physical interpretation, as two-particle systems [@Fr1]. In fact, one of the useful applications is a marked improvement of the Bethe-Salpeter method for dealing with bound states in quantum field theory [@FH]. The non-relativistic limit is instructive, it is a formulation of the theory of the Schrödinger hydrogen atom. Of course, here we shall not see a Lorentz module, but a Galilei module. What is important is that we are describing a system with a complicated internal structure. I shall limit myself to discussing just one aspect of the theory, the structure of perturbation theory.
Consider the scattering of photons from an atom [@Fr2]. The Feynman diagram represents a sequence of events taking place: incoming photon and atom, absorption of the photon, propagation of the atom in an excited state, emission of a photon, outgoing photon and atom. The amplitude for the process is a parallell sequence: a vertex factor $\exp(i\vec k\cdot \vec r)$, a propagator $(E-H_0)^{-1}$, a vertex factor $\exp (-i\vec k'\cdot \vec r)$. The propagator depends on the energy $E$ of the intermediate state, but it is not sensitive to the momentum. This is because the momentum of a very heavy object is not observable. But we may, if we wish, define the hamiltonian of an atom with momentum $\vec k$ by setting $$H_{\vec k}:= \exp(i\vec k\cdot \vec r)H_0\exp(-i\vec k\cdot \vec
r).$$ The Schrödinger equation now involves $\vec k$ and so does the wave function. In this entirely equivalent formulation of the problem ... and this is the nonrelativistic approximation to Majorana’s theory ... the amplitude is constructed in terms that are those of relativistic local field theory, namely: Incoming photon with momentum $\vec k$ and atom with momentum $\vec p$, (no exponential factor), propagator for an atom with momentum $\vec k+\vec p$, (no exponential factor), outgoing photon with momentum $\vec k'$ and atom with momentun $\vec p+\vec k-\vec k'$.
String theory took over the main ideas of infinite component field theories. The spark of the string theory revolution came from the realization that pole dominated scattering amplitudes behave better when there is an infinite sequence of poles. But the known infinite component field theories suffered from the ubiquitous space like solutions, about which more below; that is, bad behaviour with respect to crossing symmetry. The additional ingredient of string theory is duality: one contrives to construct an amplitude that is at the same time a sum over pole contributions in the direct channel and a sum over pole terms in the crossed channel. This was accomplished by Veneziano [@V].
In the beginning there was just Veneziano’s amplitude, and the multiparticle generalization. Then it was shown that these amplitudes could be factorized, just like the amplitude that we have been talking about, into a sequence of propagators separated by exponential momentum factors [@FGV]. But what followed created a huge gulf between string theory and Majorana theory. Fubini and Veneziano [@FV], in a very beautiful paper, succeeded in doing exactly the opposite of what I just did for the hydrogen atom. Instead of getting rid if the vertex factors, they managed to elimate all the propagators, thereby creating vertex algebras and two dimensional quantum field theory!
This result is truly wonderful, but something was lost. The well known Lagrangian string theories in high dimensions describe the dynamics of elementary boson and fermion fields. But the dynamical fields of string theory are the vertex operators, exponentials of these elementary fields; scattering amplitudes are constructed entirely out of vertex operators. There may be something like a string field operator, directly related to vertex operators, and we do have some hints about the dynamics, as we shall see.
So far I have stressed only the attractive features of Majorana’s infinite dimensional version of Dirac’s equation. The bad news is that it possesses solutions with space like momenta. Similar equations, studied intensively during the late sixties, including relativistic versions of the hydrogen atom, all shared this devastating feature. There was a time when theories were said to be implausible to the degree that they predict the existence of new particles. The Dirac equation predicted only the positron, and that was not too much, but the Majorana equation and its generalization predicted infinite numbers of excited particle states. Actually, this feature should be counted an advantage, especially after the advent of dual models and the realization that it was responsible for the good properties of form factors. The irony is that the predicted excited states turned out to be too few!
Most models have energy spectra labelled by an integer, $n$ say. The degeneracy of the $n$’th level, in theories of the Majorana type, are polynomial in $n$. In dual models the degeneracy grows exponentially. This may be an essential characteristic of theories that avoid spacelike solutions. The degree of degeneracy associated with the Veneziano amplitude is known from the work of Fubini and Veneziano [@FV] on factorization of this amplitude. In this work there appears a propagator, namely $$(L_0- p^2)^{-1},~~ L_0 = -2(\sum na^*_na_n-1),$$ where $L_0$ is the operator that was later interpreted as the most important operator of the Virasoro algebra, $\{a_n\}$ is an infinite set of oscillator operators and $p$ is the momentum of the state. Let $|0\rangle$ be the vacuum associated with all these oscillators, then the domain of $L_0$ is the space $[[a*_1,a^*_2,... ]] \otimes |0\rangle$, and the mass$^2$ spectrum of the model is the spectrum of $L_0$. In other words, the states satisfy the equation $$( L_0-p^2)\Psi = 0,~~ \Psi \in [[a_1^*,a^*_2,...]] \otimes
|0\rangle.$$ Though second order, this equation is very much of the Majorana type, though the internal space is bigger than anything that Majorana or his emulators had dreamed of.
This equation is a big part of a dynamical theory of strings. It certainly does not contain everything, for there is one (only one!) spacelike solution, and plenty of states with negative norm. Some of the states of negative norm are eliminated from the Veneziano amplitude by the exponential factors associated with the vertices. Fubini and Veneziano showed that the amplitude could be understood entirely in terms of vertex operators, after absorbing the propagators into the vertex operators. My inclination is to try the opposite tack, to improve the propagator by incorporating into it the role of the vertex factors. This should facilitate the task of constructing the interaction.
A promising attempt to formulate an action principle for quantum string dynamics was made by Witten 15 years ago [@Wit]. This work seems to be applicable only to open strings. I believe that this work should be taken up again, but not with the too modest aim of incorporating closed strings. If the result is to have any bearing on the outstanding open problem of our time, then it is not enough just to look for an extension of Witten’s treatment of the Veneziano model; what is needed is a very nontrivial generalization.
Summary
=======
Quantization still offers challenging mathematical problems, but the biggest problem of all is to make sense of QCD. There are hints that string theory, more precisely the Veneziano model, may point in the right direction. The suppression of the tachyon in super string theory suggests that quarks (and gluons) may be the suppressed constituents of a more elaborate version of the Veneziano model. Steps towards such a generalization could include the following small steps:
\(1) Incorporate the Fadde’ev-Popov ghosts into the vertex operators. (2) Construct the vertex operators for the superstring. (3) Ditto on an AdS background. (4) Replace the outer, scalar fields in the amplitude by scalar superfields. (Note that there are no supersymmetric tachyons!) (5) Combine (3) and (4). Possibly, the first two may have been accomplished already [@BV; @DW]. The third, without ghosts and without supersymmetry, is easy. For the 4-point function the result is $A(s,t)\Psi(x_1,x_2,x_3,x_4)$, evaluated at $x_1 = x_2, x_3 =
x_4$, where $A$ is Veneziano’s function, $s = (\bigtriangledown_1
+ \bigtriangledown_2)^2,~ t = (\bigtriangledown_2 +
\bigtriangledown_4)^2$ are covariant d’Alembertians and $\Psi$ is a product of tachyon wave functions. The fourth suggestion seems to lead to a spectrum $m \propto n$ (instead of $m^2
\propto n$). Finally the fifth one may include an interesting possibility involving singletons and massless particles with all spins.
I thank Murray Gerstenhaber for introducing me to Harrison cohomology, Georges Pinczon for the suggestion to study functions on cones, and Dirk Kreimer, Daniel Sternheimer and Ivan Todorov for useful criticism.
Barr, M., “A cohomology theory for commutative algebras. I, II", Proc. Amer. Math. Soc. [**16**]{} (1965), 1379–1384.\
“Harrison Cohomology, Hochschild Cohomology, and Triples", J. Algebra [**8**]{} (1968) 314–323.
Berkovits, Nathan; Vafa, Cumrun. “$N=4$ topological strings". Nuclear Phys. [**B 433**]{} (1995) 123–180 (`hep-th/9407190`).
Dirac, P.A.M., Proc. Roy. Soc. [**A 167**]{} (1938), 148.
Dolan, Louise; Witten, Edward. “Vertex operators for AdS$_3$ background with Ramond-Ramond flux". J. High Energy Phys. 1999, no. 11, Paper 3, 21pp. (`hep-th 9910205`).
Dito, G.; Flato, M.; Sternheimer, D. and Takhtajan, L. “Deformation quantization and Nambu mechanics", Comm. Math. Phys. [**183**]{} (1997), 1–22 (`hep-th/9602016`).
Dollard, John D. “On the definition of scattering subspaces in nonrelativistic quantum mechanics. J. Mathematical Phys. [**18**]{} (1977), 229–232.\
“Asymptotic convergence and the Coulomb interaction". J. Mathematical Phys. [**5**]{} (1964) 729–738.
Faddeev, L.D., “Operator anomaly for the Gauss law", Phys.Lett. [**B 145**]{} (1984) 81–84.
Flato, M.,“Deformation view of physical theories”, Czech.J.Phys. [**B 32**]{} (1982), 472–475.
Flato, M. and Fronsdal, C., “Parastatistics, highest weight $\mathfrak{osp}(N,\infty)$ modules, singleton statistics and confinement", J.Phys.Geom. (1989), 293–309.
Flato, M. and Fronsdal, C., “Quarks or Singletons", Phys.Lett. [**B 172**]{} (1986) 412–416.
Fronsdal, C., “Infinite multiplets and the hydrogen atom", Phys. Rev. [**156**]{} (1967), 1665–1677.
Fronsdal, C., “Compton scattering from bound electrons", Phys. Rev. [**179**]{} (1969), 1513–1517.
Fronsdal, C. and Huff, R.W., “Two-Body Problem in Quantum Field Theory", Phys.Rev. [**D3**]{} (1971), 1689.
Fubini, S., Gordon, D.and Veneziano, G., “A General Treatment of Factorization in Dual Resonance Models", Phys.Lett. [**29B**]{} (1969) 679–82.
Fubini, S. and Veneziano, G., “Duality in operator formalism", Nuovo Cim. [**67A**]{} (1970), 29-47.
Gell-Mann, M., “Current topics in particle physics", in [*Proceedings of the thirteenth International Conference on High-Energy Physics*]{}, Berkeley 1966.
Gerstenhaber, M. and Schack, S. D., “A Hodge type decomposition for commutative algebra cohomology", J. Pure and Appl. Alg. [**48**]{} (1987) 229-247. “Algebraic cohomology and deformation theory. Deformation theory of algebras and structures and applications," (Il Ciocco, 1986), 11–264, NATO Adv. Sci. Inst. Ser. C: Math. Phys. Sci., 247, Kluwer Acad. Publ., Dordrecht, 1988.
Greenberg, O. W., “Spin and unitary-spin independence in a paraquark model of baryons and mesons", Phys.Rev.Lett. [**13**]{} (1964), 1100-1102. Harrison, D. K., “Commutative Algebras and Cohomology", Trans. Amer. Math. Soc. [**104**]{} (1962), 191–204.
Majorana, H., Nuovo Cimento [**9**]{} (1932), 335.
Ohnuki, Y, and Kamefuchi, S., “Wave functions of identical particles", Ann.Phys. [**57**]{} (1969) 337–358.
Pinczon, Georges. “On the equivalence between continuous and differential deformation theories. Lett. Math. Phys. [**39**]{} (1997), 143–156.
Tavkhelidze, N., “Higher Symmetries and Composite Models of Elementary Particles", Proceedings of the Conference on High-Energy Physics and Elementary Particles, International Centre for Theoreticle Physics, Trieste 1965.
Veneziano, G., “Construction of a Crossing-symmetric, Regge-Behaved Amplitude for Linearly Rising Trajectories", Nuovo Cimento [**57A**]{} (1968) 190.
Wigner, E. P.,“Do the equations of motion determine the quantum mechanical commutation relations?", Phys. Rev. [**77**]{}, (1950), 711–712.
Witten, Edward. “Noncommutative geometry and string theory", Nuclear Phys. [**B268**]{} (1986), 253–294.
[^1]: Take note, however, that a resolution of the ills of QCD will be worth a million dollars US only if found strictly within the conventional framework, unlikely as this is.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study a method for mass-selective removal of ions from a Paul trap by parametric excitation. This can be achieved by applying an oscillating electric quadrupole field at twice the secular frequency $\omega_{\text{sec}}$ using pairs of opposing electrodes. While excitation near the resonance with the frequency $\omega_{\text{sec}}$ only leads to a linear increase of the amplitude with excitation duration, parametric excitation near $2\, \omega_{\text{sec}}$ results in an exponential increase of the amplitude. This enables efficient removal of ions from the trap with modest excitation voltages and narrow bandwidth, therefore substantially reducing the disturbance of ions with other charge-to-mass ratios. We numerically study and compare the mass selectivity of the two methods. In addition, we experimentally show that the barium isotopes with 136 and 137 nucleons can be removed from small ion crystals and ejected out of the trap while keeping $^{138}\text{Ba}^{+}$ ions Doppler cooled, corresponding to a mass selectivity of better than $\Delta m / m = 1/138$. This method can be widely applied to ion trapping experiments without major modifications, since it only requires modulating the potential of the ion trap.'
address:
- 'Albert-Ludwigs-Universität Freiburg, Physikalisches Institut, Hermann-Herder-Stra[ß]{}e 3, 79104 Freiburg, Germany'
- 'Laboratoire Kastler Brossel, Sorbonne Université, CNRS, ENS-PSL Research University, Collège de France, 4 place Jussieu, Paris, France'
- 'National Institute of Standards and Technology, Boulder, CO, USA'
- 'Albert-Ludwigs-Universität Freiburg, Physikalisches Institut, Hermann-Herder-Stra[ß]{}e 3, 79104 Freiburg, Germany'
- 'Albert-Ludwigs-Universität Freiburg, Physikalisches Institut, Hermann-Herder-Stra[ß]{}e 3, 79104 Freiburg, Germany'
- 'Albert-Ludwigs-Universität Freiburg, Physikalisches Institut, Hermann-Herder-Stra[ß]{}e 3, 79104 Freiburg, Germany'
- 'Albert-Ludwigs-Universität Freiburg, Physikalisches Institut, Hermann-Herder-Stra[ß]{}e 3, 79104 Freiburg, Germany'
- 'Albert-Ludwigs-Universität Freiburg, Physikalisches Institut, Hermann-Herder-Stra[ß]{}e 3, 79104 Freiburg, Germany'
author:
- Julian Schmidt
- Daniel Hönig
- Pascal Weckesser
- Fabian Thielemann
- Tobias Schaetz
- Leon Karpa
title: 'Mass-selective removal of ions from Paul traps using parametric excitation'
---
=1
Introduction
============
Radiofrequency (rf) Paul traps [@Paul1990; @Dehmelt1990; @Wineland2013] can store ions for a wide range of charge-to-mass ratios $Q/m$ and thus represent a versatile and powerful tool with applications such as quantum information processing [@Wineland2011; @Wineland2009] or cold chemistry [@Haerter2014; @Tomza2019]. Many experiments involve trapping one or two atomic or molecular ion species. Over the course of the experiment, other (parasitic) isotopes or ion species may appear in the trap, e.g. due to chemical reactions of the ion of interest with background gas particles, and create a disturbance [@Schmidt2020]. The parasitic ions can be embedded into the ion crystal after sympathetic cooling [@Haerter2013b] or remain on large orbits [@Guggemos2015].
In principle, it is possible to operate any quadrupole mass filter (including rf ion traps) in a parameter regime for which the trajectories of the desired species are stable and those of other species are unstable due to their specific charge-to-mass ratios. The dynamics are given by the solutions of the Mathieu equation and regions of stability can be described with the $a$ and $q$ parameters (see below) which are related to the electrostatic and radiofrequency voltages [@Paul1990]. Due to technical limitations, it can be challenging to reach such regions of instability in an experiment. Other means of removing ions may then be advantageous, such as resonant excitation [@Fulford1980; @Londry2003].
By applying an oscillating voltage to one of the trap electrodes, the near-harmonic motion of the ions in the effective potential can be resonantly excited by displacing the ion from its equilibrium position [@McCormick2019]. This method, also known as “tickle”, will be referred to as *displacement driving* in the following. Since the secular frequency $\omega_{\text{sec}}$ in the effective radial potential is proportional to $Q/m$, this method is also suitable to selectively remove ions from the trap [@Fulford1980] with a $Q/m$ resolution depending on the driving duration [@Goeringer1992].
Parametric resonances offer another possibility to efficiently drive a mechanical oscillator [@Landau1976]. Ion traps are ideally suited to study parametric excitation due to the high degree of harmonicity of the typical potentials [@Yu1993; @Alheit1997; @Zhao2002]. Applications range from ion removal with higher $Q/m$ resolution [@Vedel1990; @Williams1991; @Haerter2013b] to excess micromotion compensation [@Ibaraki2011; @Tanaka2012]. Very recently, experiments in ion traps have exploited parametric excitation to create squeezed motional states for quantum control in both quantum sensing and experimental quantum simulations [@Heinzen1990; @Burd2019; @Wittemer2019; @Kiefer2019; @Wittemer2020].
Here, we review and discuss the dynamics of a single ion in a Paul trap under the influence of parametric excitation and experimentally use this method for ion removal. We first perform numerical simulations to compare the ion removal efficiency achieved by either parametric excitation or resonant displacement driving. We find that parametric excitation can remove ions within milliseconds (or a few hundred oscillation cycles) with a $Q/m$ resolution of better than $ 10^{-3} $ and vanishing off-resonant excitation. In addition, we study the case in which an ion of mass $m_2$ should be removed while it is sympathetically cooled by a co-trapped, laser cooled ion of different mass $m_1$ which should remain in the trap. Finally, we discuss a first characterization of the experimental performance of the method and compare the results with our numerical calculations.
The method is flexible, as it only requires modulating the potential of the trap (in our case via the rf driving amplitude), representing a minor modification for typical ion trapping experiments. It can be used for any set of $a$ and $q$ parameters and does not require tuning the trap to a higher-order (e.g. octupolar) instability [@Franzen1994; @Drakoudis2006]. This method may be applicable to chemistry experiments which study reactions of ions and a (neutral) atomic ensemble [@Schmidt2020; @Haerter2012; @Chen2014; @Greenberg2018; @Gingell2010], but also to precision measurements where the spectroscopy ions are sympathetically cooled by an ion species amenable to laser cooling [@Schmoeger2015; @Biesheuvel2017; @Alighanbari2018; @Chou2017]. It could be extended to ions in Penning traps or storage rings.
Theoretical model {#sec:theory}
=================
We consider a linear rf ion trap oriented along the $z$-axis. We discuss the experimental realization later and refer to section \[sec:exp\] and Fig. \[fig:setup\]. Two rf electrodes provide dynamical radial confinement while two (segmented) dc electrodes control the electrostatic axial confinement along $z$ as well as radial electric fields. The motion of the ion in the Paul trap in the radial $x-y$-plane, here exemplarily shown for $x$, is described by the Mathieu equation [@Leibfried2003]: $$\frac{d^2x}{dt^2} + \frac{Ze}{m} \left[ U_{\text{dc}} \alpha_x + U_{\text{rf}} \alpha_x^{\prime} \cos(\Omega t)\right] x = 0 \label{eq:Mathieu}\;$$
where $Ze=Q$ is the charge of the ion, $m$ its mass, $U_\text{dc}$ and $U_\text{rf}$ are the voltages applied to the dc and rf electrodes and $\Omega$ is the angular frequency of the rf driving field. $\alpha$ and $\alpha^{\prime}$ are geometry factors with unit m$^{-2}$ determined by the trap dimensions. The motion along the $y$ and $z$ directions can be written analogously. The Mathieu equation can be rewritten in terms of the $a$ and $q$ parameters, which we will use in the following: $$a_x = \frac{4Ze U_{\text{dc}} \alpha}{ m \Omega^2}\; , \quad q_x =\frac{2Ze U_{\text{rf}} \alpha^\prime}{m\Omega^2} \;.$$
In the limit $a_x \ll 1$, $q_x \ll 1$, the solution of the Mathieu equation can be approximated as a secular motion at frequency $\omega_{x,\text{sec}}$ (along $x$) superimposed with a fast micromotion at $\Omega$ [@Leibfried2003]. In this so-called pseudopotential approximation, $\omega_{x,\text{sec}} \approx (\Omega/2) \sqrt{a_x + q_x^2/2}$ [@Paul1990].
Within the pseudopotential approximation, in which the micromotion is neglected, a modulation of $U_{\textrm{rf}}$ at the frequency $\omega_{\textrm{mod}}$ can again be modelled by the Mathieu equation. Similarly to the undriven ion trap, the parametrically driven harmonic oscillator exhibits regions of stability and instability, as has been previously shown and discussed in detail [@Zhao2002].
In order to study the dynamics of the problem and to study the experimental relevance of the method for various scenarios, we numerically solve the Mathieu equation, Eq. \[eq:Mathieu\], after adding a parametric drive. The equation of motion is $$\frac{d^2x}{dt^2} + \frac{1}{2} \Omega^2 x \left[\frac{a_x}{2} + q_x \left(1 + A_{p} \sin\left({\omega_{\textrm{mod}} t}\right) \right) \cos\left({\Omega t}\right) \right] = 0 \; \label{eq:secondmathieu}.$$
We solve this equation in one dimension for our experimental parameters, $\Omega = 2\pi \times 1.4\,\textrm{MHz}$, $a = 0.001$, $q=0.25$, $m=m_{\text{Ba}} = 138~\text{amu}$ (throughout this paper, $m_{\text{Ba}}$ refers to the mass of $^{138}\text{Ba}^{+}$). We tune $\omega_{\textrm{mod}} \approx 2 \omega_{x,\text{sec}}$ with an amplitude (modulation depth) up to $A_{p} = 6 \times 10^{-3}$. On resonance, the parametrically driven oscillator becomes metastable and starts a coherent oscillation for an initial, nonzero displacement from the origin [@Landau1976]. As initial conditions, we therefore choose the position $x_{\text{ini}}=1 \,\mu\text{m}$ and momentum $p_{\textrm{ini}} = 0$. Our parameters correspond to a secular frequency of $\omega_{\text{sec}} \approx 2\pi \times 127 \, \textrm{kHz}$ and an initial potential energy of about $k_{\textrm{B}} \times 5 \,\textrm{mK}$ (and zero kinetic energy).
Typical trajectories are shown in Fig. \[fig:typical\_trap\]. Since the parametric modulation term $q_x A_p \sin \left({\omega_{\text{mod}}t} \right) $ in Eq. \[eq:secondmathieu\] is proportional to the oscillation amplitude $x(t)$ at any time, the energy in the system increases exponentially [@Landau1976; @Heinzen1990; @Yu1993], see Fig. \[fig:typical\_trap\] (c). The rate of energy gain is proportional to $x_{\text{ini}}$, $A_{p}$ and $q_x$. We compare this result with resonant displacement driving in Fig. \[fig:typical\_trap\] (a) and (c), where the amplitude increases linearly and the energy increases quadratically. In the displacement driving case, the motion in $x$ ($y$ analogous) is described by $$\frac{d^2x}{dt^2} + \frac{1}{2} \Omega^2 x \left[\frac{a_x}{2} + q_x \cos\left({\Omega t} \right) \right] + \frac{Ze}{m} A_{d} \sin\left(\omega_{mod} t \right) = 0\; .$$ In our experiment, a typical value for the amplitude is given by $A_{d} = 1\, \textrm{V/m}$, corresponding a voltage amplitude of $10 \, \textrm{mV}$ (all amplitudes given as zero to peak) on a given dc electrode, in agreement with our segmented linear trap geometry with an ion-electrode distance of 9 mm, see section \[sec:exp\].
The mass resolution is determined by the response of a single ion to the excitation as a function of the ratio $m/(Ze)$, with $Z = 1$ in our case. To compare the mass resolution in both the parametric and the displacement driving cases, we solve the equations of motion for the parameters given above. The modulation frequency $\omega_{\textrm{mod}}$ (chosen to excite $^{138}\text{Ba}^{+}$) is constant while $m$ is varied in discrete steps within the range $m_{\text{Ba}} \pm 3\%$. We choose the duration $t_{\text{exc}}$ and amplitude of the excitation for both parametric and displacement driving such that on resonance, the same oscillation amplitude of the ion is reached after $t_{\text{exc}}$ (intersection of the two curves in Fig. \[fig:typical\_trap\] (c)).
The results are shown in Fig. \[fig:mass\_res\] and clearly emphasize the improved mass resolution with parametric excitation. While the full width at half maximum (FWHM) of the response to the parametric driving is only smaller by a factor of 1.6, the striking difference is the response when further detuned from the resonance. The response to displacement driving has a Lorentzian shape and falls off with a power law, while the response to parametric driving is Gaussian and falls off exponentially for $m\neq {138}\, \textrm{amu}$ and allows for an improved mass resolution. This is equivalent to an improvement of the frequency resolution. For a relative mass difference $|m - m_{\text{Ba}}| / m_{\text{Ba}} > \pm 0.2\%$, parametric driving already suppresses the excitation by 40 dB compared to displacement driving.
![Mass dependence of the energy gain using parametric excitation compared to displacement driving. We simulate the dynamics of an ion for different $m/m_{\text{Ba}}$ in both cases while keeping the excitation frequency constant. At $t_{\text{exc}}$, we calculate the energy of the ion, which is plotted on a logarithmic scale. The driving amplitudes have been chosen as in Fig. \[fig:typical\_trap\], and in order to ensure comparability, $t_{\text{exc}}=2.54 \,\text{ms}$ corresponds to the point in Fig. \[fig:typical\_trap\](c) where the energy curves for parametric and displacement driving intersect. Here, the parametric resonance is narrower than the displacement driving resonance by a factor of about 1.6 (full width at half maximum). More importantly, the tail of the parametric driving case falls off as a Gaussian, while that of the displacement driving resonance behaves as a Lorentzian function. Parametric driving thus acts as a knife for the ions close to resonance, allowing us to anticipate the separation of two ions with a mass ratio of $137/138$ within milliseconds, as depicted by the dashed vertical line. \[fig:mass\_res\]](Fig2.pdf){width="45.00000%"}
Separation of ions with different masses
========================================
We now study the case of two co-trapped ions 1 and 2 which form a crystal and are strongly coupled. Ion 1 has mass $m_1$ and is to remain in the trap while ion 2 with mass $m_2$ is a parasitic ion which is to be removed from the trap. We numerically show that this can be achieved by cooling the ion 1 while parametrically driving ion 2. In this case, the crystal formed by the two ions has to melt in order to decouple the motion of the two ions.
We tested that performing the calculation with the full dynamics including $\Omega$ or within the pseudopotential approximation does not significantly change the dynamics of the ions in the simulation. In this section, we therefore use the pseudopotential approximation to speed up the calculation while extending it to three dimensions.
Along $x$, each of the two ions would individually oscillate at a secular frequency $\omega_{x,\text{sec} 1}$ and $\omega_{x,\text{sec} 2}$, where the frequencies depend on $m_1$ and $m_2$ (analogous for $y$, $z$). Due to the strong coupling of the ions, their motion can be described in the normal mode picture with $3 N$ ($N=2$) oscillation frequencies. By exposing ions located near their equilibrium positions to homogeneous electric fields, only the common modes can be excited. However, as the oscillation amplitudes grow, the coupling between the ions becomes weaker until the normal mode description breaks down.
For a two-ion crystal with $m_1 = m_2$, the motion in $x$ can be decomposed into the center-of-mass mode $\omega_{x,\text{com}} = \omega_{x,\text{sec}}$ and the so-called rocking mode $$\omega_{x,\text{rock}} = \sqrt{\omega_{x,\text{com}}^2-\omega_{z,\text{com}}^2} \, , \label{eq:rocking}$$ where $\omega_{z,\text{com}} = \omega_{z,\text{sec}}$ is the center-of-mass frequency along the trap axis $z$. Generally, if $m_1 \neq m_2$, the radial modes have a different structure and the center-of-mass frequency is not equal to either $\omega_{x,\text{sec} 1}$ or $\omega_{x,\text{sec} 2}$. Parametric driving at the frequency of the unperturbed ion 1 ($\omega_{\text{mod}} = 2 \omega_{x,1}$), is then not resonant with the common motion of the ions. The individual detuning from resonance, which is related to the excitation efficiency, depends on $m_1/m_2$. In addition, the sympathetic cooling within the Coulomb crystal is also affected by the mass ratio (more efficient for mass ratios closer to 1), leading to a non-trivial behaviour of the excitation as long as the ions remain close to their coupled equilibrium positions. For our experimental parameters, the residual off-resonant excitation is sufficient to amplify the relative motion of the ions. Once the relative motion of the ions is sufficiently large, the ions are effectively decoupled and ion 1 is further excited parametrically as in the single ion case. The shift of the frequencies of the radial normal modes also depends on the axial confinement, as is apparent from Eq. \[eq:rocking\]. In our case, this shift is small ($\approx 1\%$) as $\omega_{x,\text{com}} \gg \omega_{z,\text{com}}$, which is beneficial for effectively decoupling the ions in the parametric excitation scheme.
The simulation results for two ions with an additional damping term for ion 1 are shown in Fig. \[fig:ion\_sep\]. The Coulomb crystal is initially radially displaced by $1\,\mu\text{m}$. Here, we use a higher modulation amplitude for parametric driving than before to reduce the computation time followed by a qualitative analysis only. When comparing the trajectories of the two ions, we observe that the motion of the two ions is initially driven in common, that is, they still form a Coulomb crystal. Ion 1 (blue) is cooled and damps the motion of ion 2 (red). At a critical amplitude, ion 2 “breaks free” and the motion is decoupled from that of ion 1. The amplitude of ion 2 then increases exponentially as in the single ion case while ion 1 is again efficiently Doppler cooled.
![Calculated trajectories of two ions with different masses in the rf trap. Here, the ions are considered to be $^{137}\text{Ba}^{+}$ (red curve) and $^{138}\text{Ba}^{+}$ (blue curve). The motion of the $^{138}\text{Ba}^{+}$ ion is damped by laser cooling and sympathetically cools the $^{137}\text{Ba}^{+}$ ion through the mutual Coulomb interaction. We set the common initial displacement in the $x$ direction to $1\,\mu\text{m}$. At $t=0$, a parametric excitation at twice the resonance frequency of a single $^{137}\text{Ba}^{+}$ ion is turned on while the $^{138}\text{Ba}^{+}$ ion is still directly laser cooled. The $x$ (radial) and $z$ (axial) positions of the ions are plotted as a function of time. Initially, the ions are aligned along the $z$ axis within a common ion crystal, with their Coulomb interaction in equilibrium with the axial electrostatic potential. The ions then begin oscillating in the $x$ direction in common, until the motion of the $^{138}\text{Ba}^{+}$ ion is sufficiently decoupled and only the amplitude of the $^{137}\text{Ba}^{+}$ ion keeps growing. Here, this critical point is reached after about 2ms and the $^{138}\text{Ba}^{+}$ is cooled back to the trap center. The amplitude of the $^{137}\text{Ba}^{+}$ reaches $1\,\text{mm}$ at $t \approx 2.8\,\text{ms}$, where it is assumed to leave the trap. \[fig:ion\_sep\]](Fig3.pdf){width="50.00000%"}
Comparison with experimental results {#sec:exp}
====================================
![Sketch of the experimental setup. $^{138}\text{Ba}^{+}$ ions (blue circle) are trapped and laser cooled near the center of a segmented linear rf trap. Other isotopes of barium (red circle) are also produced during photoionization and trapped. In addition, we can produce magneto-optically and all-optically trapped clouds of $^{87}$Rb, which sometimes leads to the production of Rb$^{+}$ and Rb$_2^{+}$ ions [@Haerter2013b; @Schmidt2020]. These ions can be crystallized (strongly coupled to the laser cooled ions via Coulomb interaction) or in a plasma (weakly coupled). In order to remove all ions but $^{138}\text{Ba}^{+}$, we modulate the rf drive of our Paul trap at twice the radial secular frequency of these ions through the helical resonator. We can also apply an oscillating displacement via the dc electrodes for determining the trapping frequencies. \[fig:setup\]](exp_sketch.pdf){width="40.00000%"}
![Measurement of the timescale for removing a single ion via parametric excitation. After Doppler cooling, at $t=0$, the cooling and repumping lasers are turned off and the rf trap drive is modulated to parametrically excite the ion at $2\omega_{\text{sec}}$. After $t_{\text{exc}}$, the parametric excitation is turned off and the cooling and repumping lasers are turned on again. Then, the CCD camera signal is monitored to observe whether the ion is recooled into the trap. If, after 1 s, no fluorescence signal can be observed, we consider the removal attempt from the trap as successful. After loading a new ion if necessary, the experiment is repeated until each $t_{\text{exc}}$ has been repeated about 30 times. Finally, the removal probability is shown in dependence on the excitation duration $t_{\text{exc}}$. Due to the finite temperature of the ion, the initial conditions for position and momentum vary according to the Boltzmann distribution and lead to a smoothed transition from 0% to 100% removal probability. The shaded area is an estimate of the removal probability of the ion after $t_{\text{exc}}$ based on a numerical calculation of its trajectory under the influence of parametric excitation and assuming an initial temperature of $T \approx 360\, \mu$K corresponding to the Doppler limit. The ion is considered to be removed when it reaches a distance of 1 mm from the trap center. \[fig:exp\_res\]](Fig4.pdf){width="50.00000%"}
We next perform an experimental investigation of parametric driving as a tool for removing ions from a Paul trap. Our experimental setup [@Lambrecht2017; @Schmidt2018; @Schmidt2020] shown in Fig. \[fig:setup\] consists of a linear segmented Paul trap (with ion-electrode distance of 9 mm) under ultra-high vacuum (UHV). The rf source is a signal synthesizer operated at a frequency $\Omega = 2\pi \times 1.4\,\text{MHz}$. A helical resonator with a quality factor of 40 enhances $U_{\text{rf}}$ up to 2000 V and is connected to two diagonally opposed electrodes of the Paul trap via a standard vacuum high-frequency feedthrough. The other two segmented electrodes are grounded with respect to the rf electrodes and can be offset by small dc fields to compensate stray fields [@Huber2014] and to provide axial confinement of typically $\omega_{z,\text{sec}} = 2\pi \times 12\,\text{kHz}$. $^{138}\text{Ba}$ atoms (and other isotopes) are emitted from an oven, photoionized with a resonant 553 nm and an additional 405 nm laser [@Leschhorn2012], and subsequently Doppler cooled to below one millikelvin using lasers at 493 nm and 650 nm for cooling and repumping, respectively. The fluorescence emitted by $^{138}\text{Ba}^+$ during Doppler cooling is imaged by a charge-coupled device (CCD) camera used for detection.
We first experimentally confirm the efficiency of the parametric excitation method, here with a single $^{138}\text{Ba}^+$ ion as a showcase, see Fig. \[fig:exp\_res\]. In order to investigate the undamped dynamics as in Fig. \[fig:typical\_trap\], we first prepare a single Doppler-cooled $^{138}\text{Ba}^{+}$ ion in the trap. We then turn off the cooling and repumping lasers and turn on parametric excitation for a variable duration $t_{\text{exc}}$. Here, $\omega_{x,\text{sec}} = 2\pi \times(120\pm 1) \, \text{kHz}$ and the corresponding modulation frequency is $\omega_{\text{mod}} = 2\pi \times (240\pm 2) \, \text{kHz}$ with a modulation depth of $8\times 10^{-3}$. We then turn on the cooling and repumping lasers and detect via fluorescence imaging whether the ion has remained in the trap. The probabilistic distribution of the initial position and the velocity after laser cooling translates to a distribution of the duration after which the ion is removed from the trap. This leads to the smooth transition from 0% removal probability to near-unity removal probability after about 5 ms. The experimental results are in reasonable agreement with the results of our simulation depicted by the red curve. Given that the growth of the motional amplitude is exponential, an accurate description of timescales is challenging.
In practice, using a linear frequency chirp of about 1 kHz and sweep time of 1 ms across the resonance improves the repeatability in our experiment which we attribute to two effects. First, due to the high frequency resolution, small drifts of the secular frequency on the order of 300 Hz require frequent tuning of the modulation frequency. Second, we suspect that anharmonicities of the trap lead to a change in $\omega_{x,\text{sec}}$ at higher oscillation amplitudes. After reaching a critical amplitude, $\omega_{\text{mod}}$ becomes detuned from $\omega_{x,\text{sec}}$ and the modulation no longer efficiently excites the motion. The frequency chirp mitigates both possible issues.
Application to mass separation of mixed-species or mixed-isotope ion crystals and clouds
========================================================================================
As discussed above, various processes can involve loading of contaminating ion species into the trap besides the laser-cooled ion species of interest. For two of such processes, we apply parametric excitation to remove the contaminating ion species.
The photoionization process is not perfectly isotope selective in our experiment due to an angle of about $110^{\circ}$ between the photoionization laser axis and the direction of the hot atomic beam from the oven. This leads to loading of barium isotopes with 136 and 137 nucleons into the trap with a probability of about 10%. Reactions of electronically excited $^{138}\text{Ba}^{+}$ with background $\text{H}_2$ molecules can also lead to the formation of $\text{BaH}^{+}$ hydride ions [@Molhave2000; @Kahra2012; @Aymar2012]. After sympathetic cooling by $^{138}\text{Ba}^{+}$, the ensemble consisting of $^{138}\text{Ba}^{+}$ and another ion species forms a crystal. Only the $^{138}\text{Ba}^{+}$ ion emits fluorescence from the near-resonant cooling laser and appears as a bright spot on the CCD camera image. The presence of another (dark) ion shifts the $^{138}\text{Ba}^{+}$ ion from its equilibrium position at the center of the trap.
If the fluorescence image indicates that such a mixed crystal is present in the trap, we apply parametric driving on the trap electrodes. We tune the modulation frequency to the resonance of $^{136}\text{Ba}^{+}$ ($^{137}\text{Ba}^{+}$) while laser cooling $^{138}\text{Ba}^{+}$. We observe that the dark ions can be removed from the trap by applying the parametric driving resonant with the secular frequency of $^{136}\text{Ba}^{+}$ ($^{137}\text{Ba}^{+}$). This procedure now takes a few seconds to successfully remove the parasitic ions. We were not able to remove the parasitic ions using displacement driving while the laser-cooled $^{138}\text{Ba}^{+}$ remain in the center of the trap.
Compared to the numerical simulations presented in Fig. \[fig:ion\_sep\], where it was possible to remove ion 2 within milliseconds while ion 1 remained in the trap, we experimentally observe longer timescales. In our simulations, we found that the behaviour depends on the damping rate, $\omega_{x,\text{sec}}/\omega_{z,\text{sec}}$ and $A_p$. Reaching the critical point at which the parasitic ion is sufficiently decoupled from the laser-cooled ion strongly depends on the initial conditions [@Landau1976; @Yu1993] and can be challenging. In our experiment, this may be assisted by instantaneous heating effects. Given the observed timescales and our background pressure on the order of $5\times 10^{-10} \, \text{mbar}$, it is possible that background gas collisions initiate the process. The effective continuous sampling of the Boltzmann distribution of the ion while it is being laser cooled may also assist the process.
We have also used this method to remove $\text{Rb}^{+}$ and $\text{Rb}_2^{+}$ ions from our trap [@Schmidt2020]. In this case, the ions are typically not embedded into the crystal structure at the center of the trap. Ions on high orbits with large kinetic energy can be removed reliably within tens of milliseconds while only minimally disturbing the $^{138}\text{Ba}^{+}$ in the center of the trap.
Summary and outlook
===================
We have discussed parametric excitation as a method for mass-selectively removing ions from an rf ion trap. We find that parametric excitation at twice the ion’s resonance frequency is a highly efficient and fast method for this purpose. For the high oscillation amplitudes (10% to 50% of the ion-electrode distance) required to remove ions from the trap, the mass resolution is significantly enhanced compared to displacement driving at the resonance frequency. Efficiently removing ions with parametric excitation requires precise tuning of the excitation frequency. Small drifts of the trapping frequency can thus limit the performance of the method. While the secular frequencies can be stabilized via rf and dc voltages [@Johnson2016], we used a sweep of the excitation frequency in a 1 kHz window near the resonance. Further improvement could be achieved by initially transferring a controlled amount of energy to the ions by applying a resonant displacement driving pulse at $\omega_{x,\text{sec}}$ [@Yu1993; @McCormick2019] or giving the ions a kick with a voltage pulse.
The amplitude modulation index is limited due to the bandpass filter constituted by the helical resonator. Other schemes to enhance the rf-amplitude which lead to less damping of the modulation sidebands may be advantageous, especially at low trap drive frequencies where higher voltages can be achieved in the pre-amplification. Using even higher harmonics $n\,\omega_{\text{sec}}$ (with even integer $n$) could further improve the frequency resolution [@Zhao2002]. Parametric excitation may be useful when initializing a large ion trap with a single ion, e.g. in the context of shallow optical trapping potentials [@Lambrecht2017]. Due to the recent interest in experiments relying on a multitude of ion species in a single trap, including ions for sympathetic cooling and e.g. molecular ions, we believe that removal of ions by parametric excitation may be a useful tool for many ion trapping groups.
**Acknowledgements**
This project has received funding from the European Research Council (ERC) under the European Unions Horizon 2020 research and innovation program (Grant No. 648330), and was supported by the Georg H. Endress foundation. J. S., F. T., and P. W. acknowledge support from the DFG within the GRK 2079/1 program. J. S. acknowledges financial support from the Region Île-de-France within the framework of DIM SIRTEQ and the Alexander-von-Humboldt foundation. P. W. gratefully acknowledges financial support from the Studienstiftung des deutschen Volkes. We are indebted to D. Leibfried for stimulating discussions.
[35]{}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 10.1103/revmodphys.62.531) [****, ()](\doibase 10.1103/RevModPhys.62.525) [****, ()](\doibase 10.1103/revmodphys.85.1103) [****, ()](\doibase 10.1002/lapl.201010125) [****, ()](\doibase 10.1088/0031-8949/2009/t137/014007) [****, ()](\doibase 10.1080/00107514.2013.854618) [****, ()](\doibase
10.1103/RevModPhys.91.035001) [****, ()](\doibase
10.1103/PhysRevLett.124.053402) [****, ()](http://dx.doi.org/10.1038/nphys2661), [****, ()](\doibase
10.1088/1367-2630/17/10/103001) [****, ()](\doibase
10.1116/1.570570), [****, ()](\doibase 10.1021/jasms.8b01877), , [****, ()](\doibase
10.1088/2058-9565/ab0513) [****, ()](\doibase 10.1021/ac00037a023) @noop [*,*]{} () [****, ()](\doibase 10.1063/1.353399), [****, ()](\doibase
10.1103/PhysRevA.56.4023) [****, ()](\doibase 10.1103/PhysRevA.66.063414) [****, ()](\doibase 10.1016/0168-1176(90)85025-W) [****, ()](\doibase 10.1016/0168-1176(91)85034-J) [****, ()](\doibase 10.1007/s00340-011-4463-x) [****, ()](\doibase
10.1007/s00340-011-4762-2) [****, ()](\doibase 10.1103/PhysRevA.42.2977) [****, ()](\doibase
10.1126/science.aaw2884), [****, ()](\doibase
10.1103/PhysRevLett.123.180502) [****, ()](\doibase
10.1103/PhysRevLett.123.213605) [ (), 10.1103/PhysRevLett.123.180502](\doibase
10.1103/PhysRevLett.123.180502) [“,” ](http://www.freepatentsonline.com/5331157.html) () [****, ()](\doibase https://doi.org/10.1016/j.ijms.2006.02.006) [****, ()](\doibase
10.1103/PhysRevLett.109.123201) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.98.032702) [****, ()](\doibase
10.1063/1.3505142), [****, ()](\doibase 10.1126/science.aaa2960), [**** (), 10.1007/s00340-016-6576-8](\doibase
10.1007/s00340-016-6576-8) [****, ()](\doibase 10.1038/s41567-018-0074-3) [****, ()](http://dx.doi.org/10.1038/nature22338) [****, ()](\doibase
10.1103/RevModPhys.75.281) [****, ()](\doibase
10.1038/s41566-017-0030-2) [****, ()](\doibase
10.1103/PhysRevX.8.021028) @noop [**** ()]{} [****, ()](\doibase 10.1007/s00340-012-5101-y) [****, ()](\doibase 10.1103/PhysRevA.62.011401) [****, ()](\doibase 10.1038/nphys2214) [****, ()](\doibase 10.1088/0953-4075/45/21/215103) [**** (), 10.1063/1.4948734](\doibase 10.1063/1.4948734)
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'AdS2-CPW.bib'
---
=10000
[**Mass-deformed ABJM and Black Holes in AdS$_{\bfs 4}$**]{}
[**Nikolay Bobev,${}^{(1)}$ Vincent S. Min,${}^{(1)}$ and Krzysztof Pilch${}^{(2)}$\
**]{} ${}^{(1)}$ Instituut voor Theoretische Fysica, KU Leuven,\
Celestijnenlaan 200D, B-3001 Leuven, Belgium ${}^{(2)}$ Department of Physics and Astronomy\
University of Southern California\
Los Angeles, CA 90089, USA\
\
We find a class of new supersymmetric dyonic black holes in four-dimensional maximal gauged supergravity which are asymptotic to the ${\rm SU(3)}\times {\rm U(1)}$ invariant AdS$_4$ Warner vacuum. These black holes can be embedded in eleven-dimensional supergravity where they describe the backreaction of M2-branes wrapped on a Riemann surface. The holographic dual description of these supergravity backgrounds is given by a partial topological twist on a Riemann surface of a three-dimensional $\cN=2$ SCFT that is obtained by a mass-deformation of the ABJM theory. We compute explicitly the topologically twisted index of this SCFT and show that it accounts for the entropy of the black holes.
Introduction
============
Holography has evolved into an indispensable tool to study the dynamics of strongly coupled quantum field theories. In addition, this duality can be used to learn new lessons about the structure of black holes. For a long time, an important outstanding question in black hole physics has been to account microscopically for the entropy of asymptotically AdS black holes in more than three dimensions.[^1] While this problem still remains open for black holes in five or more dimensions, recently there has been a rapid progress in understanding the microstate counting for supersymmetric black holes in AdS$_4$ [@Benini:2015eyy; @Benini:2016rke; @Hosseini:2016tor; @Hosseini:2016ume; @ABCMZ]. These developments were triggered by employing the tools of supersymmetric localization (see [@Pestun:2016zxk] for a recent review) to define and compute a suitable partition function, called “topologically twisted index” [@Benini:2015noa; @Benini:2016hjo; @Closset:2016arn], which can be used to count the microstates of these black holes.
The basic idea of the recent work is to engineer a black hole in M-theory[^2] which is asymptotic to an AdS$_4\times M_7$ solution, where $M_7$ is a Sasaki-Einstein manifold. The horizon of such four-dimensional black holes is a compact Riemann surface, $\Sigma_{\fg}$. This gravitational background in turn is holographically dual to a three-dimensional $\mathcal{N}=2$ SCFT of the ABJM type [@Aharony:2008ug; @Bagger:2007jr; @Bagger:2007vi] placed on $\mathbb{R}\times \Sigma_{\fg}$ with a partial topological twist. For such twisted three-dimensional SCFTs the supersymmetric partition function was studied in [@Benini:2015noa; @Benini:2016hjo; @Closset:2016arn] and it reduces to a matrix model due to supersymmetric localization. In the planar limit of a large number, $N$, of coincident M2-branes, one can solve this matrix model and obtain the free energy of the twisted SCFT to leading order in $N$.[^3] This in turn reproduces the entropy of the black hole. This procedure is best studied for black holes in eleven-dimensional supergravity compactified on $S^7$ [@Benini:2015eyy; @Benini:2016rke], i.e. for the ABJM theory at $k=1,2$, but it can also be generalized to other manifolds $M_7$ [@Hosseini:2016tor; @Hosseini:2016ume; @ABCMZ]. The black holes in AdS$_4$ can also be viewed as holographic duals of RG flows across dimensions in the spirit of Maldacena-Nuñez [@Maldacena:2000mw; @Gauntlett:2001qs; @BC].
In this work we will follow a slightly different approach. Our starting point is the well-known observation that the ABJM theory admits a mass deformation that preserves $\mathcal{N}=2$ supersymmetry and leads to an interacting SCFT in the IR [@Benna:2008zy] (see also [@Klebanov:2008vq]). We refer to this SCFT as mABJM. Although this theory is strongly coupled, some information about its physics can be obtained using symmetries and supersymmetric localization. For example, the partition function of the theory on $S^3$ was computed in [@Jafferis:2011zi] (see, in particular, Section 5). In addition, mABJM has a holographic dual which was constructed in four-dimensional gauged-supergravity by Warner (W) in [@Warner:1983vz; @Warner:1983du] (see also [@Ahn:2000aq; @Ahn:2000mf]) and uplifted to eleven dimensions in [@Corrado:2001nv]. The situation here is akin to the well-known $\mathcal{N}=1$ Leigh-Strassler fixed point arising from a supersymmetric mass-deformation of four-dimensional $\mathcal{N}=4$ SYM [@Leigh:1995ep]. The gravity dual of this four-dimensional $\mathcal{N}=1$ SCFT was studied in [@Freedman:1999gp; @Pilch:2000ej]
There are two main objectives that we have in mind. On one hand, we are interested in studying the topologically twisted index of [@Benini:2015noa; @Benini:2016hjo; @Closset:2016arn] for the mABJM $\mathcal{N}=2$ SCFT. On the other hand we want to construct new four-dimensional supersymmetric black holes that are asymptotic to the AdS$_4$ Warner vacuum [@Warner:1983vz; @Warner:1983du] (or alternatively the CPW solution of eleven-dimensional supergravity [@Corrado:2001nv]) and have a near-horizon AdS$_2$ region. The large $N$ limit of the topologically twisted index should then reproduce the Bekenstein-Hawking entropy of these black holes. It is worth emphasizing that the CPW AdS$_4$ solution in eleven-dimensional supergravity is not of the usual Freund-Rubin type and thus the class of black holes that we study is different from the ones explored recently in the literature [@Benini:2015eyy; @Benini:2016rke; @Hosseini:2016tor; @Hosseini:2016ume; @ABCMZ].
The calculation of the topologically twisted index in the planar limit proceeds similarly as in [@Benini:2015eyy; @Benini:2016rke; @Hosseini:2016tor; @Hosseini:2016ume; @ABCMZ]. However, there are several subtle points related to the electric charge parameters of the index, which we emphasize and clarify along the way.
The construction of the new black hole solutions is more involved. We start with the maximal $\SO(8)$ gauged supergravity in four-dimensions [@deWit:1982bul], which is a consistent truncation to the lowest-lying KK modes of the eleven-dimensional supergravity on $S^7$ [@deWit:1986oxb; @Nicolai:2011cy]. The three-dimensional mABJM SCFT of interest is dual to the $\mathcal{N}=2$ AdS$_4$ vacuum discovered by Warner [@Warner:1983vz; @Warner:1983du]. It has the usual ${\rm U(1)}_R$ R-symmetry along with an ${\rm SU(3)}_F$ flavor symmetry which is manifested on the supergravity side by the presence of a massless ${\rm SU(3)}\times {\rm U(1)}$ gauge field in the AdS$_4$ Warner vacuum. The supersymmetric black hole solutions of interest are similar to the ones found in [@Romans:1991nq; @Caldarelli:1998hg; @Cacciatori:2009iz]. In particular, they have non-vanishing gauge fields lying in the Cartan subalgebra of ${\rm SU(3)}\times {\rm U(1)}$. This allows us to simplify the construction by focusing on an ${\rm U(1)}^3$-invariant consistent truncation of the maximal supergravity. In addition to the metric and three Abelian gauge fields, the bosonic sector of that truncation contains also eight real scalars. By analyzing the BPS equations and the equations of motion, we construct a plethora of magnetic and dyonic supersymmetric black holes in this truncated theory.
The ${\rm U(1)}^3$-invariant truncation can be embedded into a larger ${\rm U(1)}^2$-invariant truncation of the four-dimensional $\mathcal{N}=8$ supergravity. The advantage of doing that is that the resulting theory is a fully-fledged four-dimensional $\mathcal{N}=2$ gauged supergravity coupled to three vector multiplets and one hypermultiplet. The ten real scalars in this truncation parametrize the coset $$\label{MVMHintro}
\cals M_{V}\times \cals M_{H} \eql \left[{\rm SU(1,1)\over U(1)}\right]^3\times {\rm SU(2,1)\over SU(2)\times U(1)}\,.$$ Recasting our black hole solutions in the $\cals N=2$ language offers some additional insights and allows us to use the existing results on black holes in four-dimensional $\mathcal{N}=2$ gauged supergravity, see [@DallAgata:2010ejj; @Hristov:2010ri; @Gnecchi:2013mta; @Halmagyi:2013sla; @Halmagyi:2013qoa; @Halmagyi:2014qza] for a non-exhaustive list of references.
We note that our current set-up is very similar to the one in [@Bobev:2014jva], where a partial topological twist of the Leigh-Strassler theory [@Leigh:1995ep] placed on $\mathbb{R}^2\times \Sigma_{\fg}$ led to a holographic RG flow from the four-dimensional $\mathcal{N}=1$ SCFT to a two-dimensional $(0,2)$ SCFT. The holographic dual to this setup is a family of black string solutions with an AdS$_3$ near-horizon region which are asymptotic to the AdS$_5$ fixed point of the five-dimensional $\mathcal{N}=8$ gauged supergravity found in [@Khavaev:1998fb].
Synopsis
--------
& &&\
& &&\
& && &\
& S\_\^(\_) && &
&&&\
&&&\
&& I\_D\^(\_i,\_|) & &\
&& &
Since the paper is rather long and technical, let us highlight some of the main results first. We begin in Section \[sec:CFT\] with a discussion of the field theory side of the duality, specifically the mABJM theory that is obtained from the ABJM theory by a mass deformation (MD), formally captured by a constraint on the R-charge,[^4] $\Delta_1$. The main object of interest is the topologically twisted index, which is a partition function on $\mathbb{R}\times \Sigma_{\fg}$ that depends on electric charges and magnetic fluxes as well as complex fugacities for the continuous global symmetries of the theory. The calculation of this observable in mABJM proceeds in several steps and has been schematically summarized in Figures \[Figone\] and \[Figtwo\]. It parallels a similar calculation in ABJM [@Benini:2015eyy; @Benini:2016rke], see the right and left columns in the figures, respectively.
First we compute the “magnetic index,” $\mathcal{I}_M(\fn;\Delta)$, which depends on the magnetic topological twist parameters $\fn_{\alpha/i}$ and the real fugacities $\Delta_{\alpha/i}$. We use the observation in [@Hosseini:2016tor] that the index, $\cals I _M(\fn;\Delta)$, is directly related by a topological twist (TT) to the supersymmetric partition function, $F_{S^3}$, of the CFT on $S^3$, where the fugacities, $\Delta_{\alpha/i}$, are identified with the R-charges on $S^3$. The topologically twisted index, $\cals I_M(\fn)$, which is a function of the magnetic fluxes, $\fn_{\alpha/i}$, only, is then obtained from $\mathcal{I}_M(\fn;\Delta)$ by extremization (CE) with respect to the fugacities, $\Delta_{\alpha/i}$, subject to an algebraic constraint (with the corresponding real Lagrange multiplier, $\lambda$) that is imposed by supersymmetry [@Benini:2015eyy]. We show that the end result for the magnetic index, $\cals I^\text{mABJM}_M(\fn_i)$, in the mABJM theory is the same irrespective of whether one first applies the mass deformation to the ABJM twisted index, $\cals I^\text{ABJM}_M(\fn_\alpha;\Delta_\alpha)$, to obtain the corresponding twisted index, $\cals I_M^\text{mABJM}(\fn_i;\Delta_i)$, which is then extremized with respect to its fugacities, or, equivalently, one extremizes $\cals I^\text{ABJM}_M(\fn_\alpha;\Delta_\alpha)$ while imposing simultaneously two constraints on the fugacities: the one for the mass deformation and the one for the topological twist. The resulting extremized index, $\mathcal{I}_M^{\rm mABJM}(\fn_i)$, is shown in Section \[subsec:genBHmag\] to match the entropy of the new family of magnetic black holes that we construct in Section \[sec:solspace\].
In general, the topologically twisted index is dyonic, it depends on both electric charges, $\fq_{\alpha/i}$, and magnetic fluxes, $\fn_{\alpha/i}$, as well as complex fugacities, $u_{\alpha/i}$. To include these extra parameters we follow the approach in [@Benini:2016rke] which is summarized in Figure \[Figtwo\]. We start with $\mathcal{I}_{M}(\fn;\Delta)$, analytically continue it from real fugacities, $\Delta_{\alpha/i}$, to complex fugacities, $u_{\alpha/i}$, and introduce the electric charges, $\fq_{\alpha/i}$, by a Legendre transformation (LT). This yields the dyonic index, $\mathcal{I}_D(\fn,\fq;u)$, which is then extremized (CER) with respect to constrained fugacities, $u_{\alpha/i}$.
This calculation is more subtle than for the purely magnetic index in Figure \[Figone\]. The reason is that there should be a linear relation between the electric charges to ensure supersymmetry, however, it is not entirely clear how to find it. It was proposed in [@Benini:2016rke] that to reproduce correctly the entropy of a macroscopic black hole, the imaginary part of the dyonic index, $\mathcal{I}_{D}(\fn,\fq)$, should vanish. This provides exactly one additional constraint that serves as the expected relation between the electric charges.
It is straightforward to implement this procedure in the ABJM theory, see [@Benini:2016rke] and Section \[sec:dabjm\], as summarized by the left column in Figure \[Figtwo\]. In mABJM there is a further subtlety at which step of the calculation one should impose the massive deformation that eliminates one of the global ${\rm U}(1)$ symmetries. One possibility, see Section \[sec:mabjmdind\], is to perform the mass deformation first. This leaves three ${\rm U}(1)$ global symmetries the corresponding electric charges, $\fq_i$, magnetic fluxes, $\fn_i$, and fugacities, $u_i$. Then the extremization with respect to those fugacities together with the reality constraint yields an unambiguous result for the dyonic twisted index, $\cals I_D^\text{mABJM}(\fn_i,\fq_i)$, that matches the entropy of the new dyonic black holes constructed in Section \[subsec:dyonicBPS\]. The other possibility, suggested by the corresponding calculation of the magnetic index, is to extremize the ABJM dyonic index, $\cals I_D^\text{ABJM}(\fn_\alpha,\fq_\alpha;u_\alpha)$, while imposing two constraints on the fugacities using two complex Lagrange multipliers, $\lambda_1$ and $\lambda_2$, see the diagonal arrow in Figure \[Figtwo\]. Indeed, in Section \[sec:manjmditc\] we find the extremized dyonic index and the magnetic fluxes are the same as above. However, unlike before, this extremization does not yield a unique result for the electric charges because of a shift symmetry that involves the imaginary parts of the Lagrange multipliers. In Sections \[sec:manjmditc\] and \[sec:eleccharg\] we show that this symmetry can be fixed consistently in two ways: (i) one can set the electric charge $\fq_1$ to zero, thus reducing the calculation to the one in the mABJM theory above, and (ii) set the imaginary parts of both Lagrange multipliers to zero, which gives a consistent match with the dyonic black holes in the dual supergravity with four vector fields.
Throughout the paper we work with a consistent truncation of the maximal $\SO(8)$ gauged supergravity, which is discussed in Section \[sec:Sugra\] and Appendices \[appendixA\] and \[appendixB\]. We use the truncation to construct supersymmetric AdS$_2\times \Sigma_{\fg}$ solutions which should be thought of as the near-horizon geometry of a class of dyonic black holes asymptotic to the AdS$_4$ Warner vacuum. In Section \[sec:DyonicComp\], we show explicitly that the Bekenstein-Hawking entropy of these black holes is the same as the topologically twisted index. In Section \[sec:magneticBH\], we study the magnetically charged black holes in more detail. We conclude in Section \[sec:Conclusions\] with a short discussion and some open questions for future work. In Appendix \[appconv\] we summarize our notation and conventions. In Appendix \[appendixA\] we also show how to formulate our truncation in the canonical language of four-dimensional $\mathcal{N}=2$ gauged supergravity. In Appendix \[appendixB\] we present some details on the derivation of the near-horizon BPS equations used in Section \[sec:Sugra\]. Finally, in Appendix \[appattractor\] we show how these BPS equations can be written in a form similar to the “attractor mechanism” equations discussed in [@DallAgata:2010ejj; @Hristov:2010ri].
Field theory {#sec:CFT}
============
ABJM and a mass deformation {#subsec:ABJMmass}
---------------------------
Here we offer a short summary on the ABJM SCFT [@Aharony:2008ug] and a particular supersymmetric mass deformation studied in [@Benna:2008zy] (see also [@Klebanov:2008vq; @Jafferis:2011zi]). The ABJM theory is a double Chern-Simons theory with gauge group ${\rm U(N)}\times {\rm U(N)}$ and equal and opposite levels for the two gauge groups $(k,-k)$. This theory describes the low-energy dynamics of $N$ coincident M2-branes probing a $\mathbb{C}^4/\mathbb{Z}_k$ singularity in M-theory. The dual holographic description at large $N$ is in terms of an AdS$_4\times S^7/\mathbb{Z}_k$ solution in M-theory. For $k>2$ the theory has only $\mathcal{N}=6$ supersymmetry which gets enhanced to $\mathcal{N}=8$ for $k=1,2$. In the following we will focus on $k=1$ where there is no orbifold singularity and the gravitational solution is well-described by eleven-dimensional supergravity.[^5] For $k=1$ the theory has an ${\rm SO(8)}$ R-symmetry which, however, is not manifest at the level of the ABJM Lagrangian.
The ABJM theory can be succinctly described using the $\mathcal{N}=2$ superspace formalism. In addition to the two vector multiplets, there are four chiral multiplets denoted by $A_a$ and $B_c$, $a,c=1,2$, transforming in the $(\overline{\textbf{N}},\textbf{N})$ and $(\textbf{N},\overline{\textbf{N}})$ representation of ${\rm U(N)}_k\times {\rm U(N)}_{-k}$, respectively, with the superpotential $$\label{ABJMsuperp}
W \sim \text{Tr}\left(\epsilon^{ab}\epsilon^{cd}A_aB_cA_bB_d\right)\,.$$ The R-charges of these chiral superfields,[^6] $R[A_a] \equiv \Delta_{A_a}$ and $R[B_c] \equiv \Delta_{B_c}$, must satisfy the constraint $$\label{DeltaconstrABJM}
\Delta_{A_1}+\Delta_{A_2}+\Delta_{B_1}+\Delta_{B_2}=2\,,$$ so that the total R-charge of the superpotential is equal to 2.
In this formulation only an ${\rm U(1)}_R\times {\rm SU(2)}\times {\rm SU(2)}\times {\rm U(1)}_b$ subgroup of the global symmetry is manifest. It is enhanced to ${\rm SU(4)}_R\times {\rm U(1)}_b$ when the Lagrangian is written in components, see for example [@Jafferis:2011zi]. The ${\rm U(1)}_b$ global symmetry has a topological nature characteristic of three-dimensional QFTs and is generated by the current $*_3\text{Tr}(F+\tilde{F})$, where $F$ and $\tilde{F}$ are the field strengths of the two ${\rm U(N)}$ gauge fields and $*_3$ is the Hodge star in three dimensions. Due to this topological current there are gauge invariant monopole operators, $T^{(q)}$, in the theory, which turn on $q$ units of flux for the topological current through an $S^2$ surrounding the insertion point. When $k=1$ the operator $T^{(1)}$ transforms in the $(\textbf{N},\overline{\textbf{N}})$ and the operator $T^{(-1)}$ transforms in the $(\overline{\textbf{N}},\textbf{N})$ representation of the gauge group. This is ultimately responsible for the enhancement of the supersymmetry to $\mathcal{N}=8$ and of the R-symmetry to ${\rm SO(8)}_R$. In the dual holographic description for $k=1$, given by the AdS$_4\times S^7$ solution of M-theory, the ${\rm SO(8)}_R$ is realized as the isometry group of $S^7$. The metric of $S^7$ can be written as a circle fibration over $\mathbb{CP}^3$, then the ${\rm SU(4)}_R$ is the isometry group of $\mathbb{CP}^3$ and ${\rm U(1)}_b$ is realized as the isometry of the fibre.
The $S^3$ free energy of the ABJM SCFT can be computed using supersymmetric localization and is given by the following function of the R-charges:[^7] $$\label{FS3}
F_{S^3}=\frac{4\sqrt{2}\pi}{3}N^{3/2}\sqrt{\Delta_{A_1}\Delta_{A_2}\Delta_{B_1}\Delta_{B_2}} \;.$$ Using $F$-maximization [@Jafferis:2010un; @Closset:2012vg] while satisfying the second relation in , one finds the values of the R-charges at the superconformal point, $$\label{DeltaABJM}
\Delta_{A_1}=\Delta_{A_2}=\Delta_{B_1}=\Delta_{B_2}=\frac{1}{2}\;,$$ so that the free energy on $S^3$ for ABJM reads $$\label{FS3ABJM}
F_{S^3}^{\text{ABJM}} =\frac{\sqrt{2}\pi}{3}N^{3/2}\;.$$ Note that the values of $\Delta_{A_{1,2}}$ and $\Delta_{B_{1,2}}$ in can also be obtained as a condition for enhanced supersymmetry of the SCFT. When the ABJM theory is placed on $S^3$, for values of $\Delta_{A_{1,2}}$ and $\Delta_{B_{1,2}}$ that obey , but not , the theory preserves $\mathcal{N}=2$ supersymmetry but is not conformal, see [@Freedman:2013ryh] for a discussion.
The ABJM superpotential can be deformed by a mass term that preserves $\mathcal{N}=2$ supersymmetry $$\label{CPWsuperp}
\Delta W \sim \text{Tr}(T^{(1)}A_1)^2\;.$$ Adding this deformation triggers an RG flow from the ABJM theory in the UV to an interacting $\mathcal{N}=2$ SCFT in the IR. This was studied in [@Benna:2008zy] (see also [@Klebanov:2008vq; @Jafferis:2011zi]) from a field theory perspective. The holographic description of this mABJM SCFT is given by the Warner vacuum of four-dimensional maximal ${\rm SO(8)}$ gauged supergravity [@Warner:1983vz; @Warner:1983du] (see also [@Nicolai:1985hs]) which was uplifted to eleven-dimensional supergravity in [@Corrado:2001nv]. There have been several consistency checks of this proposed duality including a match between the spectrum of protected operators [@Klebanov:2008vq] (see also [@Nicolai:1985hs] for earlier work) as well as the free-energy to leading order in $N$ [@Jafferis:2011zi]. It is worth pointing out that, while the ABJM theory is parity invariant, the $\mathcal{N}=2$ SCFT obtained by the mass deformation in breaks parity. In the supergravity description this breaking of parity is manifested by the fact that one of the four-dimensional $\cN=8$ supergravity pseudoscalars has a non-vanishing value at the Warner vacuum.
The mABJM SCFT has the following symmetries: The four chiral superfields of the ABJM theory, ordered as $(A_1,A_2,B_1,B_2)$, transform in the ${\bf 4}_{1}$ of ${\rm SU(4)}_R\times {\rm U(1)}_b$ (see, e.g., [@Freedman:2013ryh] for a summary). The superpotential breaks ${\rm SU(4)}_R$ to ${\rm SU(3)}_F$ and only a linear combination of the ${\rm U(1)}_c$ commutant of ${\rm SU(3)}_F$ inside of ${\rm SU(4)}$ and the ${\rm U(1)}_b$ is preserved. We will call that linear combination ${\rm U(1)}_{R}^{\text{W}}$ since it is the superconformal R-symmetry of the mABJM conformal fixed point, which in turn is dual to the Warner vacuum in supergravity. The ${\rm SU(3)}_F$ symmetry does not act on the supercharges and thus deserves the name flavor symmetry. The linear combination of ${\rm U(1)}_b$ and ${\rm U(1)}_c$ orthogonal to ${\rm U(1)}_{R}^{\text{W}}$ is broken by the quadratic superpotential and corresponds to the massive ${\rm U(1)}_m$ vector field in the supergravity discussion below.
The superpotential deformation modifies the R-charge assignments in the theory.[^8] In particular, the value of the R-charge for $A_1$ is set to unity. Combining this with one finds $$\label{DeltaconstrCPW}
\Delta_{A_1}=1 \, , \qquad \Delta_{A_2}+\Delta_{B_1}+\Delta_{B_2}=1\;.$$ The $S^3$ free energy for general values of the three R-charges can be computed by localization [@Jafferis:2011zi]. The final result can be obtained by formally setting $\Delta_{A_1}=1$ in and reads $$\label{FS3CPWgen}
F_{S^3} =\frac{4\sqrt{2}\pi}{3}N^{3/2}\sqrt{\Delta_{A_2}\Delta_{B_1}\Delta_{B_2}} \;.$$ Applying $F$-extremization to and enforcing the second constraint in , we find that at the mABJM fixed point $$\label{DeltaCPW}
\Delta_{A_2}=\Delta_{B_1}=\Delta_{B_2}=\frac{1}{3}\; .$$ This is compatible with the ${\rm SU(3)}_F$ flavor symmetry of the model and leads to the following $S^3$ free energy of the mABJM SCFT: $$\label{FS3CPW}
F_{S^3}^{\text{mABJM}} =\frac{4\sqrt{2}\pi}{9\sqrt{3}}N^{3/2}\;.$$ Thus one finds that the UV (ABJM) and IR (mABJM) SCFTs have the following relation between their $S^3$ partition functions $$F_{S^3}^{\text{mABJM}} =\frac{4}{3\sqrt{3}} F_{S^3}^{\text{ABJM}}\; .$$ As expected this is compatible with the $F$-theorem, namely $F_{S^3}^{\text{ABJM}}>F_{S^3}^{\text{mABJM}}$ [@Jafferis:2011zi; @Casini:2012ei].
The topologically twisted index {#ssec:twistedindex}
-------------------------------
A three-dimensional $\mathcal{N}=2$ SCFT can be placed on the manifold $\mathbb{R} \times \Sigma_{\mathfrak{g}}$, where $\Sigma_{\mathfrak{g}}$ is a closed Riemann surface of genus $\mathfrak{g}$,[^9] while preserving at least two supercharges by employing the topological twist of Witten [@Witten:1988ze]. The procedure amounts to turning on a background gauge field for the ${\rm U(1)}$ R-symmetry of the SCFT with a finely tuned magnitude so as to cancel the curvature of the Riemann surface. In addition, one is free to turn on any appropriately quantized flux for the background gauge fields that couple to the continuous flavor symmetry currents in the CFT. This procedure can be applied to both the ABJM and the mABJM theories discussed above.
An interesting supersymmetric observable [which]{} captures non-trivial information about a topologically twisted three-dimensional $\mathcal{N}=2$ SCFT on $\mathbb{R} \times \Sigma_{\mathfrak{g}}$ is the topologically twisted index $$\label{}
\cals I(\fn_i;\Delta_i)\equiv \log Z_{\mathbb{R}\times \Sigma_{\fg}} (\Delta_i,\fn_i),$$ defined in [@Benini:2015noa; @Benini:2016hjo; @Closset:2016arn]. This is a supersymmetric partition function that depends on the theory at hand, the genus, $\mathfrak{g}$, of the Riemann surface, the magnetic fluxes, $\mathfrak{n}_i$, for the background magnetic fields as well as the fugacities, $\Delta_i$, for those global symmetries. We use the same notation for the fugacities and the R-charges in since they obey formally the same constraint [@Hosseini:2016tor]. The general form of this partition function is quite complicated, but it simplifies dramatically in an appropriate large $N$ limit which will be the focus of our discussion.
The large $N$ limit of the twisted index was first studied in [@Benini:2015eyy] for $\mathfrak{g}=0$. Here we use mostly the results of [@Hosseini:2016tor], combined with the ones in Section 6 of [@Benini:2016hjo], which are applicable for the large $N$ limit of the so called non-chiral quiver gauge theories that include both the ABJM and mABJM theories. The resulting formula for the twisted index can be expressed in terms of the partition function on $S^3$ and the background magnetic fluxes $\fn_i$,[^10] $$\label{index}
\cals I(\fn_i;\Delta_i) = (\fg-1) \bigg ( F_{S^3} (\Delta_i) + \sum_{i=1}^{r_{G}} \left [ \left( \frac{\fn_i}{\fg-1} - \Delta_i \right) \frac{1}{2} \frac{\partial F_{S^3} (\Delta_i)}{\partial\Delta_i} \right ] \bigg ) \,,$$ where $r_G$ is the rank of the continuous global symmetry group. For the ABJM theory, $r_G=4$ since the global symmetry is ${\rm SO(8)}$. For the mABJM theory, the global symmetry is ${\rm SU(3)}_F\times {\rm U(1)}_R^{\text{W}}$ and thus $r_G=3$. As argued in [@Benini:2015eyy], the topologically twisted index can be found by extremizing with respect to $\Delta_i$, subject to the constraint for ABJM and for mABJM. This means that to use the formula in one should first fix the background fields (the genus, $\fg$, and the magnetic fluxes, $\fn_i$), then solve the constrained extremization problem to find the extremal values $\bDelta _i(\fn)$ that are finally plugged back into to obtain the topologically twisted index as a function of the background fields.
### The ABJM twisted index
Let us illustrate this procedure in some detail for the ABJM theory. We begin by turning on background magnetic fields along the four Cartan generators, $T_\alpha$, of the ${\rm SO(8)}$ global symmetry, $$\label{so8mag}
F\eql F^{(\alpha)}T_\alpha\,,\qquad F^{(\alpha)} = \fn_\alpha {\rm vol}_{\Sigma_{\fg}}\;, \qquad \alpha=1,\ldots,4\,.$$ To preserve supersymmetry we have to impose the following relation between the magnetic fluxes: $$\label{mfabjm}
\fn_1+\fn_2+\fn_3+\fn_4\eql 2(\fg-1)\,,$$ which implements the topological twist. In addition, we must ensure the proper flux quantization for the magnetic fields piercing the Riemann surfaces. In our conventions this amounts to $\fn_\alpha \in \mathbb{Z}$.
The general formula for the topologically twisted index in , after using the explicit expression for the free energy on $S^3$ in , is $$\label{IABJM}
\cals I_M(\fn_\alpha;\Delta_\alpha)\eql {\sqrt 2\pi\over 3}\,N^{3/2}\sqrt{\Delta_1\Delta_2\Delta_3\Delta_4}\,\Big({\fn_1\over\Delta_1} +{\fn_2\over \Delta_2}+{\fn_3\over\Delta_3}+{\fn_4\over\Delta_4}\Big)\,.$$ One should extremize it as a function of $\Delta_\alpha$ subject to the constraint in . To this end we introduce the Lagrange multiplier, $\lambda$, and extremize $$\label{}
\cals I(\fn_\alpha;\Delta_\alpha|\lambda)\eql \cals I_M(\fn_\alpha;\Delta_\alpha)+\pi\lambda (\Delta_1+\Delta_2+\Delta_3+\Delta_4-2) \,.$$ This yields the system of equations $$\label{extreqmag}
{\partial\cals I_M\over\partial\Delta_\alpha}+\pi \lambda\eql 0\,,\qquad \alpha\eql 1,\ldots,4\,,$$ which can be solved for the magnetic charges $$\label{magabjmnns}
\begin{split}
\fn_\alpha &= -\frac{3\lambda}{2\sqrt{2} N^{3/2}} \frac{\sum_\beta \sigma_{\alpha\beta} \Delta_\alpha \Delta_\beta}{\sqrt{\Delta_1\Delta_2\Delta_3\Delta_4}} \,,
\end{split}$$ where $$\label{defsigab}
\sigma_{\alpha\beta}\eql \begin{cases}-1 & \text{for~~ $\alpha=\beta$} \,,\\ \phantom{-} 1 & \text{for~~ $\alpha\not =\beta$}\,. \end{cases}$$ Plugging this back into and using , we find $$\mathcal{I}(\fn_\alpha;\Delta_\alpha|\lambda) = -2\pi \lambda \,.$$ Imposing the topological twist condition on the magnetic fluxes , we solve for the Lagrange multiplier, $$\begin{split}
\lambda = -2(\fg-1)\frac{2\sqrt{2}N^{3/2}}{3} \frac{\sqrt{\Delta_1\Delta_2\Delta_3\Delta_4}}{\sum_{\alpha,\beta} \sigma_{\alpha\beta}\Delta_\alpha \Delta_\beta}\,.
\end{split}$$ Then can be rewritten as $$\label{nofDelta}
\begin{split}
\fn_\alpha &= 2(\fg-1) \frac{\sum_\beta\sigma_{\alpha\beta} \Delta_\alpha \Delta_\beta}{\sum_{\gamma,\delta} \sigma_{\gamma\delta} \Delta_\gamma \Delta_\delta} \,,
\end{split}$$ and the topologically twisted index as a function of extremal fugacities, $\Delta_\alpha=\Delta_\alpha^\text{ext}$, is $$\label{topindmagABJM}
\mathcal{I}(\fn_\alpha(\Delta);\Delta_\alpha)\eql \frac{8 \sqrt{2}\pi}{3}(\fg-1) N^{3/2} \frac{\sqrt{\Delta_1\Delta_2\Delta_3\Delta_4}}{\sum_{\alpha,\beta} \sigma_{\alpha\beta}\Delta_\alpha \Delta_\beta}\,.$$
To find the final expression for the topologically twisted index as a function of the magnetic fluxes, $\fn_\alpha$, one has to solve the algebraic equations in for $ \Delta _\alpha(\fn)$ and plug the result in to obtain, $\cals I_M(\fn_\alpha)$. Clearly, given the nonlinearity of , this is in general a complicated algebraic problem that one would rather avoid. So, instead we will work with the implicit formulae above for the magnetic fluxes and the twisted topological index as functions of the extremal fugacities.
There is a special topological twist, the so-called universal twist [@ABCMZ; @BC], for which one can perform the algebraic calculations above explicitly. This twist is characterized by having background magnetic fields that extend only along the unique $\mathcal{N}=2$ superconformal R-symmetry of the ABJM theory. In our conventions this amounts to setting[^11] $\fn_1=\fn_2=\fn_3=\fn_4=(\fg-1)/2$. Solving for these values of the background fluxes one finds $\bDelta _{\alpha}=1/2$ which in turn leads to the following simple expression for the topologically twisted index $$\mathcal{I}_{\text{univ}}^{\text{ABJM}} = (\fg-1) \frac{\sqrt{2}\pi}{3} N^{3/2} = (\fg-1) F_{S^3}^{\text{ABJM}}\;,$$ where we have used . Note that only for $\fg>1$ one finds a positive topologically twisted index at leading order in $N$.
### The mABJM twisted index {#ssec:mABJMmagI}
Now we apply the same procedure to the mABJM theory obtained by a mass deformation of the ABJM superpotential in . As we discussed in Section \[subsec:ABJMmass\], this breaks the global $\SO(8)$ symmetry to ${\rm SU(3)}_F\times {\rm U(1)}_{R}^{\text{W}}$. In terms of the $\SO(8)$ Cartan generators, $T_{\alpha}$, $\alpha=1,\ldots,4$, the Cartan subalgebra of the new global symmetry group is spanned by $$\label{newCart}
T^{(1)}\eql {1\over 2}(T_2-T_3)\,,\qquad T^{(2)}{1\over 2\sqrt 3}(T_2+T_3-2 T_4)\,, \qquad T^{(R)}\eql {1\over 3}(3T_1+T_2+T_3+T_4)\,,$$ where the first two are Cartan generators of $\SU(3)_F$ and the third one is the generator of the new R-symmetry, ${\rm U(1)}_{R}^{\text{W}}$. If we start with a general $\SO(8)$ magnetic field , the symmetry breaking along the RG-flow restricts it to the Cartan subalgebra of the new global symmetry, which is enforced by the condition $$\label{massconstr}
\fn^{(m)}\equiv \fn_1-\fn_2-\fn_3-\fn_4\eql 0\,,$$ while the topological twist along the new R-symmetry generators gives $$\label{toptwist}
\fn^{(R)}\equiv \frac{1}{2}(3 \fn_1 + \fn_2 + \fn_3 + \fn_4)\eql 2(\fg-1)\,.$$ It is illuminating to rewrite the two constraints as $$\label{magconst}
\fn_1=\fg-1\,,\qquad \fn_2+\fn_3+\fn_4\eql \fg-1\,,$$ which is analogous to the condition on the R-charges. Using and we then find the following expression for the topologically twisted index $$\label{IgenCPW}
\cals I(\fn_i;\Delta_i) \eql
{\sqrt 2\pi\over 3}\,N^{3/2}\sqrt{\Delta_2\Delta_3\Delta_4}\,\Big(\fg-1 +{\fn_2\over \Delta_2}+{\fn_3\over\Delta_3}+{\fn_4\over\Delta_4}\Big)\,,$$ which must be extremized as a function of $\Delta_{2,3,4}$ satisfying the constraint . Introducing a Lagrange multiplier and extremizing as above, we obtain the following relations between the extremal values $\bDelta _i=\Delta_i^\text{ext}$ and the magnetic fluxes: $$\label{n234CPW}
\begin{split}
\fn_2 & \eql (\fg-1)\bDelta _2\bigg[{\bDelta _3+\bDelta _4\over \bDelta _2\bDelta _3+\bDelta _3\bDelta _4+\bDelta _4\bDelta _2}-1\bigg]\,,\\[6 pt]
\fn_3 & \eql (\fg-1)\bDelta _3\bigg[{\bDelta _2+\bDelta _4\over \bDelta _2\bDelta _3+\bDelta _3\bDelta _4+\bDelta _4\bDelta _2}-1\bigg]\,,\\[6 pt]
\fn_4 & \eql (\fg-1)\bDelta _4\bigg[{\bDelta _2+\bDelta _3\over\bDelta _2\bDelta _3+\bDelta _3\bDelta _4+\bDelta _4\bDelta _2}-1\bigg]\,.\\
\end{split}$$ Finding the twisted index as a function of $\fn_{2,3,4}$ again amounts to solving the algebraic equations in for $\bDelta _i$ and plugging the result in which is difficult. Hence, we proceed as previously and express the final result in terms of the extremal values $\bDelta _i=\Delta_i^\text{ext}$, $$\label{twistedCPW}
\cals I(\fn_i(\bDelta );\bDelta_i )\eql {2\sqrt{2}\pi\over 3}(\fg-1)N^{3/2}\sqrt{\bDelta _2\bDelta _3\bDelta _4}\,\bigg[{1\over \bDelta _2\bDelta _3+\bDelta _3\bDelta _4+\bDelta _4\bDelta _2}-1\bigg]\,.$$ We have obtained for mABJM from for ABJM by imposing the constraints and on $\Delta_1$ and $\fn_1$, respectively. However, implementing those constraints does not commute with the extremization of the topologically twisted index. Indeed, differs from the result one would have obtained by evaluating the topologically twisted index for ABJM in with $\fn_1$ and $\Delta_1$ set to their mABJM values. If one wants to start with the ABJM index , the correct procedure is to extremize it with both constraints in .
Explicit examples {#subsec:magnexpl}
-----------------
It is instructive to discuss two examples in which we can solve the algebraic equations in and obtain the twisted index in a compact form as an explicit function of the magnetic fluxes.
Our first example is the universal twist which amounts to turning on the magnetic flux only along the R-symmetry generator in . This leads to the following values for $\fn_{2,3,4}$[^12] $$\fn_{2}=\fn_{3}=\fn_{4}= {1 \over 3}\,(\fg-1)\,.$$ Plugging this in one finds the solution $\bDelta _2=\bDelta _3=\bDelta _4=1/3$. As expected on general grounds, see [@ABCMZ; @BC], the topologically twisted index is then $$\label{IunivCPW}
\mathcal{I}_{\text{univ}}^{\text{mABJM}} = (\fg-1) \frac{4\sqrt{2}\pi}{9\sqrt{3}}N^{3/2} = (\fg-1) F_{S^3}^{\text{mABJM}}\;,$$ where for the second equality we used .
The second example is more involved. We impose the following relation between the magnetic fluxes: $$\fn_2= \fn_3 \equiv (\fg-1)\,\fn \, .$$ The remaining magnetic fluxes are then fixed by . Solving the equations in with these restrictions leads to the following four branches of solutions for $\bDelta _{2,3}$: $$\label{branchesCPW}
\begin{split}
\text{Branch 1$_{\pm}$:} \quad & \bDelta _2 = \frac{1 - \fn \pm \sqrt{(1+\fn)(1-3\fn)} }{2} \, , \quad \bDelta _3 = \frac{1 - \fn \mp \sqrt{(1+\fn)(1-3\fn)} }{2} \, , \\
\text{Branch 2$_{\pm}$:} \quad & \bDelta _2 = \bDelta _3 = \frac{1 - \fn \pm \sqrt{(1+\fn)(\fn-1/3)}}{2} \, .
\end{split}$$ Note that $\bDelta _4$ is fixed uniquely by the linear relation in once a choice of a branch of solutions in has been made. The corresponding twisted index reads: $$\label{twistSU2}
\begin{split}
\text{Branch 1$_{\pm}$:} \quad & \mathcal{I}(\fn) = \frac{3\sqrt{3\fn} (1-\fn)}{2} \; (\fg-1) F_{S^3}^{\text{mABJM}} \, , \\[6 pt]
\text{Branch 2$_{\pm}$:} \quad & \mathcal{I}(\fn) = \frac{3}{2\sqrt{2}} \frac{(1 - 2\fn)\left(1+3 \fn^2 \pm(1-3 \fn) \sqrt{(1+\fn) \left(\fn-\frac{1}{3}\right)}\right)}{\sqrt{(1 - 2\fn)\left(1-3 \fn^2 \mp(1-3 \fn) \sqrt{(1+\fn) \left(\fn-\frac{1}{3}\right)}\right)}} (\fg-1) F_{S^3}^{\text{mABJM}} \, ,
\end{split}$$ where once again we have expressed the result in terms of the mABJM free energy on $S^3$, see . Interestingly, we find that the two branches of solutions $1_{\pm}$ have the same twisted index.
The extremized values $\bDelta _{i}$ play the role of R-charges in the one-dimensional quantum mechanical system arising at low energies after the twisted compactification on $\Sigma_{\fg}$. We thus have to impose that $\bDelta _{i}$ are real. In addition the twisted index is expected to reproduce the entropy of the black hole that describes this twisted compactification holographically. For that reason we also have to find $\mathcal{I}>0$. Imposing these two constraints in the expression restricts the value of the magnetic flux for Branch $1_\pm$ to the range $0<\fn<1/3$ and for Branch $2_\pm$ to $1/3<\fn<1/2$. At the special value $\fn=1/3$, the R-charges reduce to $\bDelta _i=1/3$ and one should recover the universal twist. Indeed, when evaluated at $\fn=1/3$, the twisted index reduces to the universal relation for all branches. In addition we find that the Riemann surface has to be hyperbolic, i.e. $\fg>1$.
Dyonic generalization {#ssec:dyonicindex}
---------------------
So far we have limited ourselves to turning on a background metric and magnetic fluxes on the Riemann surface $\Sigma_{\fg}$. There are, however, more background parameters that can be turned on while preserving the supersymmetry of the topologically twisted index [@Benini:2015noa; @Benini:2016hjo; @Benini:2016rke]. In the context of holography, these additional parameters correspond to electric charges that can be in general non-vanishing in the dual supersymmetric AdS$_4$ black holes, see [@Benini:2016rke] and references therein.
A generalization of the topologically twisted index to include electric charges has been proposed in [@Benini:2016rke]. The new “dyonic” index is defined as a Legendre transform of the “magnetic” index discussed in the previous sections and is explicitly given by $$\label{Idyonic}
\mathcal{I}_D(\fn_i,\fq_i;u_i) \equiv \cals I_M(\fn_i;u_i) - \i\, \pi \sum_{i=1}^{r_G}u_i \mathfrak{q}_i\,,$$ where $\fn_i$ and $\fq_i$ are the background magnetic fluxes and electric charges, respectively, while $u_i$ are complex fugacities replacing the real fugacities $\Delta_i$. The magnetic index, $ \cals I_M(\fn_i;u_i)$, as a function of the complex $u_i$ is defined by an analytic continuation. As usual, the magnetic fluxes, $\fn_i$, satisfy the topological twist condition that preserves supersymmetry, such as or . However, it is a priori not known how to impose the corresponding supersymmetry constraint on the electric charges $\fq_i$.
It was argued in [@Benini:2016rke] that in order to obtain the leading saddle point approximation to the dyonic topologically twisted index in the limit of large $N$, one must first fix the values of the electric and magnetic charges $(\mathfrak{q_i},\mathfrak{n_i})$ and then extremize $\mathcal{I}_D(\fn_i,\fq_i;u_i)$ with respect to the complex variables $u_i$, subject to the same constraints as the corresponding $\Delta_i$. The entropy of the dual dyonic black hole, $S_{\rm BH}(\mathfrak{q},\mathfrak{n})$, should then be identified with the real part of the dyonic index at this extremum. Furthermore, it was conjectured in [@Benini:2016rke] that when the index scales with $N$ such that there is a classical dual AdS$_4$ black hole with a regular horizon, i.e. $N^{3/2}$ for the [ABJM]{} and mABJM SCFTs, the supersymmetry constraint on the electric charges is equivalent to $\mathcal{I}_D(\fn_i,\fq_i;u_i)$ being real after the extremization.
In the following subsections we will illustrate this procedure in detail for the two theories of interest and obtain explicit formulae for the twisted index that can be compared directly with the entropy of the dual black holes.
### The ABJM dyonic twisted index {#sec:dabjm}
We start by specializing to the ABJM theory. Using , $$\label{IdABJM}
\cals I_D(\frak n_{\alpha},\frak q_{\alpha};u_{\alpha})\eql {\sqrt 2\pi\over 3}\,N^{3/2}\sqrt{u_1u_2u_3u_4}\,\Big({\fn_1\over u_1} +{\fn_2\over u_2}+{\fn_3\over u_3}+{\fn_4\over u_4}\Big)-\i\,\pi\sum_{\alpha=1}^4 u_{\alpha}\frak q_{\alpha}\,,$$ where the magnetic fluxes satisfy the supersymmetric twist condition while the complex fugacities are constrained by, cf. , $$\label{uabjmconstr}
u_1+u_2+u_3+u_4\eql 2\,.$$ The extremization equations now read $$\label{exdyabjm}
{\partial \cals I_M\over\partial u_{\alpha}}-\i\,\pi\,\fq_{\alpha}+\pi \lambda \eql 0\,,\qquad {\alpha}\eql1,\ldots,4\,,$$ where $\lambda=\mu+\i\,\nu$ is a complex Lagrange multiplier. Solving for the electric charges, $\fq_{\alpha}$, and substituting the result back into , one finds that the extremized index is simply given by $$\label{LagIndabjm}
\cals I_D(\frak n_{\alpha},\frak q_{\alpha};u_{\alpha})\Big|_{u_{\alpha}\eql u_{\alpha}^\text{ext}}\eql -2\pi\lambda(\frak n_{\alpha},\frak q_{\alpha};u_{\alpha}^\text{ext})\,.$$ Hence, by imposing the reality condition on the extremal index, we conclude that $\lambda$ must be real.
Next, we go back to and decompose the equations into their real and imaginary parts. This yields eight real equations that are linear in the magnetic fluxes, $\fn_{\alpha}$, and the electric charges, $\fq_{\alpha}$, but highly nonlinear with respect to the complex fugacities, $u_{\alpha}$. Hence, just as before, determining the extremal fugacities, $u_{\alpha}^\text{ext}$, as functions of $\fn_{\alpha}$ and $\fq_{\alpha}$ is a daunting task. Instead, we solve the linear system for the magnetic fluxes and electric charges.
To present the result in a compact form, it is convenient to set $$\label{upolar}
u_{\alpha}\eql \r_{\alpha}\, e^{\i\,\theta_{\alpha}}\,,\qquad \theta_\alpha\in (-\pi,\pi)\,,$$ and $$\label{analcont}
\sqrt{u_1u_2u_3u_4}\eql \sqrt{\Delta_1\Delta_2\Delta_3\Delta_4}\, \,e^{{\i\over 2}(\theta_1+\theta_2+\theta_3+\theta_4)}\,,$$ which fixes the analytic continuation we are working with.[^13] Let us define the following linear combinations of the phases: $$\label{}
\begin{split}
\theta_{\alpha\beta}\eql \theta_\alpha-\theta_\beta\,,\qquad \theta^*_{\alpha\beta}\eql {1\over 2}\epsilon_{\alpha\beta\gamma\delta}\theta_{\gamma\delta}\,,
\end{split}$$ and $$\label{}
\vartheta_\alpha\eql {1\over 2}(4\theta_\alpha-\theta_1-\theta_2-\theta_3-\theta_4)\,.$$ The solution to the linear system can be simplified using the constraint . This yields $$\label{thenabjm}
\fn_\alpha\eql -{3\sqrt 2\over N^{3/2}}{\mu\over C(\vartheta)}{\r_\alpha\over \sqrt{\r_1\r_2\r_3\r_4}}\sum_{\beta=1}^4 \sigma_{\alpha\beta}\r_\beta\cos\theta^*_{\alpha\beta}\,,\qquad \alpha=1,\ldots,4\,,$$ and $$\label{theqabjm}
\fq_\alpha\eql -{\mu\over C(\vartheta)}\,\Big[S(\vartheta)+{2\over \r_\alpha}\sum_{\beta=1}^4\sigma_{\alpha\beta}\r_\beta\sin\coeff 1 2(\vartheta_\alpha+\vartheta_\beta)\Big]\,,\qquad \alpha=1,\ldots,4\,,$$ where $$\label{}
\begin{split}
C(\vartheta) & \eql \cos \vartheta_1+\cos\vartheta_2+\cos\vartheta_3+\cos\vartheta_4\,,\\[6 pt]
S(\vartheta) & \eql \sin \vartheta_1+\sin\vartheta_2+\sin\vartheta_3+\sin\vartheta_4\,,
\end{split}$$ and $\sigma_{\alpha\beta}$ is defined in . We can now use the topological twist condition to determine the Lagrange multiplier $\lambda=\mu$ and the twisted index as a function of extremal fugacities, $$\label{extdinabjm}
\cals I_D^\text{ABJM}(\fn_\alpha(u),\fq_\alpha(u);u_\alpha)\eql {2\sqrt 2\pi\over 3}N^{3/2}(\frak g-1)\,C(\vartheta)\,{\sqrt{\r_1\r_2\r_3\r_4}\over \sum_{\alpha,\beta}\sigma_{\alpha\beta}\r_\alpha\r_\beta\cos\theta^*_{\alpha\beta}}\,.$$ Finally, one can use to eliminate $\mu$ from and , to obtain a complete solution to the extremization problem.
It should be clear that the reality of the index provided the “missing equation” needed to determine the Lagrange multiplier and hence the electric charges. Somewhere within the solution there is a hidden supersymmetric twist condition one should impose ab initio on the electric charges. Identifying this condition more clearly within the field theory remains a puzzle. We will return to this issue in Section \[sec:abjmstu\] when we discuss the corresponding supergravity calculation.
It is straightforward to check that in the pure magnetic limit, $\theta_\alpha\to0$, the electric charges vanish, while the magnetic fluxes and the extremized index reduce to and , respectively.
### The mABJM dyonic twisted index {#sec:mabjmdind}
The extremization of the twisted dyonic index for mABJM proceeds similarly as for the ABJM index above. We start with $$\label{mabjmdind}
\cals I_D(\fn_i,\frak q_i;u_i) \eql {\sqrt 2\pi\over 3}\,N^{3/2}\sqrt{u_2u_3u_4}\,\Big(\fg-1+\sum_{i=2}^4 {\fn_i\over u_i}\Big)-\i\,\pi\sum_{i=2}^4 u_i\frak q_i\,,$$ that follows from and , the constraint $$\label{uconsmabjm}
u_2+u_3+u_4\eql 1\,,$$ and the corresponding Lagrange multiplier $\lambda=\mu+\i\,\nu$. The extremization equations have the same form as in . Using them to simplify yields the following relation between the extremized index and the Lagrange multiplier, $$\label{extindm}
\cals I_D(\frak n_i,\frak q_i;u_i)\eql
{2\over 3}\sqrt 2 \pi N^{3/2}(1-\fg)\sqrt{u_2u_3u_4}-2\pi\lambda+\i\,\pi (u_2\fq_2+u_3\fq_3+u_4\fq_4)\,,$$ where ${u_i\eql u_i^\text{ext}}$. The more complicated form of this equation in comparison with is due to the fact that unlike the ABJM index , the mABJM index is not a homogenous function of $u_i$. Still, the reality of the extremal index provides an additional equation that leads to a unique solution for the magnetic fluxes, $\fn_i$, the electric charges, $\fq_i$, and the index as functions of the fugacities.
It is convenient to use the polar parametrization and the following linear combinations of the phases:[^14] $$\label{thetastar}
\theta_{ij}\eql\theta_i-\theta_j\,,\qquad \theta^*_i\eql {1\over 2}\epsilon_{ijk}\theta_{jk}\,,\qquad \theta^*_{ij}\eql \epsilon_{ijk}\theta_k\,,$$ and $$\label{varthetam}
\tau_i\eql 3\theta_i-\theta_2-\theta_3-\theta_4\,.$$ Let[^15] $$\label{}
\begin{split}
C(\tau) & \equiv \sum_{i=2}^4\cos\coeff 1 2(\tau_i+\theta_i)+\cos\coeff 1 2(\theta_2+\theta_3+\theta_4)\,,\\
S(\tau) & \equiv \sum_{i=2}^4\sin\coeff 1 2(\tau_i+\theta_i)-\sin\coeff 1 2(\theta_2+\theta_3+\theta_4)\,.
\end{split}$$ Then $$\label{nncpw}
\begin{split}
\fn_i \eql & (1-\fg)\r_i\cos\theta^*_i+ \mu {3\sqrt 2\over N^{3/2}}{1\over C(\tau)}{\r_i\over \sqrt{\r_2\r_3\r_4}}\Big[
2\r_i-\sum_{j=2}^4\r_j(\cos\theta^*_{ij}+\cos\theta^*_i\cos\theta^*_j)\Big]\,,
\end{split}$$ and $$\label{qqcpw}
\begin{split}
\fq_i\eql {\sqrt 2\over 3} N^{3/2} (1-\fg) & {\sqrt{\r_2\r_3\r_4}\over \r_i} \sin\coeff 1 2(\tau_i-\theta_i) \\[6 pt]
& +{\mu\, }{\r_i(\sin\theta_i+\sin\tau_i)-\sum_j \r_j\sin(\tau_i-\theta_{ij})
\over\r_i(\cos\theta_i+\cos\theta^*_i)}+\nu\,.
\end{split}$$ Substituting into the topological twist condition , we can evaluate the real part of the Lagrange multiplier, $\mu$, and from the reality of the extremized index , the imaginary part $\nu$. A tedious algebra yields the following result for the extremized index: $$\label{mABJMttind}
\begin{split}
\cals I^\text{mABJM}_D (\fn_i(u),& \fq_i(u);u_i)\eql -{\sqrt 2\pi\over 3} (\fg-1)N^{3/2} C(\tau) \sqrt{\r_2\r_3\r_4}\\[6 pt]
& \times
{1+ \sum_{i} \r_i\cos\theta^*_i-\sum_{i<j}\r_i\r_j[\cos\theta^*_{ij}+\cos(\theta_i-\theta_j)]\over
\sum_{i} \r_i^2\sin^2\theta^*_i- \sum_{i<j} \r_i\r_j[2\cos\theta^*_{ij}+\cos(\theta^*_i-\theta^*_j)+\cos(\theta_i-\theta_j)]
}\,,
\end{split}$$ which has been further simplified using the constraint .
### The mass deformed twisted dyonic ABJM index {#sec:manjmditc}
At the end of Section \[ssec:mABJMmagI\], we have observed that the extremized mABJM twisted index could be obtained by starting with the ABJM index and extremizing it under two constraints . This is equivalent to using the ABJM constraint together with the first constraint in , where the latter formally imposes the mass deformation from ABJM to mABJM. In this section we discuss this extremization in more detail for the dyonic index, which is also more subtle.
We start with $$\label{IdABJM2c}
\begin{split}
\cals I_D(\frak n_\alpha,\frak q_\alpha;u_\alpha|\lambda_r) & \eql {\sqrt 2\pi\over 3}\,N^{3/2}\sqrt{u_1u_2u_3u_4}\,\Big({\fn_1\over u_1} +{\fn_2\over u_2}+{\fn_3\over u_3}+{\fn_4\over u_4}\Big)-\i\,\pi\sum_{\alpha=1}^4 u_\alpha\frak q_\alpha\\
&\qquad + \pi \lambda_1 (u_1-1)+\pi \lambda_2(u_2+u_3+u_4-1) \,,
\end{split}$$ where the first line is the ABJM index and the second line are the constraints with the corresponding Lagrange multipliers, $\lambda_r=\mu_r+\i\,\nu_r$, $r=1,2$. In addition we impose two conditions on the magnetic fluxes.
The same calculation as previously shows that the extremized index is $$\label{exIhg}
\cals I_D(\frak n_\alpha,\frak q_\alpha;u_\alpha|\lambda_r)\eql -\pi (\lambda_1+\lambda_2)\,,$$ and hence the reality condition sets, $$\label{nu12s}
\nu_1+\nu_2\eql 0 \,.$$ The subtlety, which does not arise in any of the previous examples, is that the extremization does not lead to a unique solution for the electric charges. This comes about from the flat direction in . If we shift the electric charges by $\delta\frak q_\alpha$ and the imaginary parts of the Lagrange multipliers by $\delta\nu_1$ and $\delta\nu_2$, respectively, then and remain invariant provided $$\label{shiftsymm}
\delta\frak q_1\eql -\delta\frak q_2\eql -\delta\frak q_3\eql -\delta\frak q_4\eql\delta\nu_1\eql -\delta\nu_2\,.$$ Differentiating with respect to $u_\alpha$ and solving the imaginary parts of the resulting equations for the electric charges we find $$\begin{split}\label{mdefqabjm}
\fq_1 = \nu_1 & - {2\over 3}\mu _1{1\over C(\tau)}\sum_{j=2}^4 \Big[\sin\theta_j \cos \coeff 1 2\left(\tau_j-\theta_j\right)-\sin\coeff{1}{2}\left(\tau_j-\theta_j\right) \cos \theta_j^*\Big]\\
& -2 \mu _2 {1\over C(\tau)}\sum_{i=2}^4 \Delta _i \sin \coeff{1}{2} (\tau_i-\theta_i)\,,\\[6 pt]
\fq_i = \nu_2 & -2\mu_1{1\over C(\tau)}{1\over \r_i}\sin\coeff 1 2 (\tau_i-\theta_i)\\
&
-\mu_2{1\over C(\tau)}\,\Big[S(\tau) -4\sin\coeff 1 2(\tau_i+\theta_i)+{2\over \r_i}\sum_{j=2}^4 \r_j\sin \coeff12(\theta_i+\tau_j+\theta_{ij})\Big]\,,
\end{split}$$ where the various angles are the same as in Section \[sec:mabjmdind\], see and . Substituting into the real part of the extremization equations we solve for the magnetic fluxes, $$\label{thevns}
\begin{split}
\fn_1 &= \frac{3 \sqrt{2}}{N^{3/2}}{1\over C(\theta)} {1\over\sqrt{\Delta _2 \Delta _3 \Delta _4}} \Big(\mu_1-\mu _2 \sum_{i=2}^{4}\Delta _i \cos \theta_i^*\Big)
\,,\\[6 pt]
\fn_i &= \frac{3 \sqrt{2}}{N^{3/2}}\frac{1}{C(\theta)}\frac{\Delta _i}{\sqrt{\Delta_2 \Delta _3 \Delta _4}}\Big[-\mu _1 \cos \theta _{i}^*+\mu _2 \big(2\Delta_i-\sum_{j=2}^{4}\Delta _j \cos \theta _{ij}^*\big)\Big] \,.
\end{split}$$ Those depend only on the real parts of the Lagrange multipliers, $\mu_1$ and $\mu_2$, which in turn are determined using , $$\begin{split}
\mu_1 &= {\sqrt 2\over 6 }(\fg-1) N^{3/2} {C(\theta)\over \cals D(\Delta,\theta)} \sqrt{\Delta_2\Delta_3\Delta_4}\Big(\sum_{i=2}^4\Delta_i\left(\Delta_i+ \cos \theta^*_i \right)-2\sum_{i<j}\Delta_i\Delta_j\cos\theta_{ij}^*\Big)\,,\\
\mu_2 &= {\sqrt 2\over 6 }(\fg-1) N^{3/2} {C(\theta)\over \cals D(\Delta,\theta)} \sqrt{\Delta_2\Delta_3\Delta_4} \Big(1+\sum_{i=2}^4 \cos \theta^*_i \Delta_i\Big)\,,
\end{split}$$ where $$\label{}
\cals D(\Delta,\theta)\eql \sum_{j=2}^4 \sin^2\theta_j^* \Delta_j^2-\sum_{i<j}\Delta_i \Delta_j \left[2\cos\theta_{ij}^*+\cos\left(\theta^*_i-\theta_j^*\right)+\cos\theta_{ij}\right]\,.$$ In the formulae above, we have implemented explicitly the constraint $u_1^\text{ext}=1$. One can check that as functions of the extremal fugacities, $u_i^\text{ext}$, $i=2,3,4$, subject to the constraint , the magnetic fluxes, $\fn_2,\fn_3,\fn_4$, in reproduce exactly the magnetic fluxes in mABJM in Section \[sec:mabjmdind\]. Similarly, the extremized twisted index , that depends only on the real parts of the Lagrange multipliers, is the same as the dyonic twisted index . In fact, the present calculation yields without using the constraint to simplify intermediate expressions, which is much simpler.
The electric charges remain undetermined due the shift symmetry . One way to fix it, is to compare the four electric charges, $\fq_\alpha$, in with the electric charges, $\tilde \fq_i$, in , where we have introduced the “tilde” to avoid any confusion. By direct calculation, one can check that $\fq_i=\tilde \fq_i$ for that value of $\nu_1=-\nu_2$ for which $\fq_1\eql 0$, precisely the result one would expect.
To summarize, we have shown that the extremization of the ABJM index with two constraints reproduces exactly the mABJM extremized dyonic twisted index and the corresponding magnetic fluxes and electric charges provided, in addition to , we also impose the condition $$\label{fq1cond}
\fq_1\eql 0\,,$$ on the electric charges. However, one also has the option to fix the shift symmetry differently, which then results in four, typically non-vanishing, electric charges. As we will see in Section \[sselecchar\], this freedom will be crucial for matching our field theory results with supergravity calculations.
Supergravity {#sec:Sugra}
============
We expect that the holographic dual description of the twisted compactification of the mABJM SCFT on $\Sigma_{\fg}$ discussed in the previous section is provided by asymptotically AdS$_4$ supersymmetric black holes. In this section we study these black hole solutions within the maximal ${\rm SO(8)}$ gauged supergravity theory of de Wit and Nicolai [@deWit:1982bul], which is a consistent truncation of eleven-dimensional supergravity compactified on $S^7$ [@deWit:1986oxb; @Nicolai:2011cy]. In particular, this means that our solutions can be uplifted to M-theory.
The truncation {#sec:trunc}
--------------
The four-dimensional $\cals N=8$ supergravity has many bosonic fields, but to construct the black hole solutions of interest it is sufficient to work within a subsector of the theory that is invariant under the symmetry of the dual topologically twisted mABJM theory.
The topological twist breaks the ${\rm SU(3)}_F$ symmetry of the mABJM SCFT, and the corresponding AdS$_4$ Warner vacuum, to the Cartan subgroup ${\rm U}(1)^2$. It is natural to impose this ${\rm U(1)}^2$-invariance on the $\cals N=8$ supergravity, which then yields a consistent truncation to a four-dimensional $\mathcal{N}=2$ supergravity coupled to three vector multiplets and one hypermultiplet. The bosonic fields of the resulting theory are the metric, the graviphoton gauge field along with three ${\rm U(1)}$ gauge fields in the vector multiplets. The ten real scalars in the truncation parametrize the manifold and combine into three complex scalars, $z_i$, $i=1,2,3$, in the vector multiplets and two complex scalars, $\zeta_{1}$ and $\zeta_2$, in the hypermultiplet. The details of the truncation and the geometric data of the resulting $\cals N=2$ supergravity are presented in Appendix \[appendixA\].[^16]
The recasting of our truncation into the canonical formalism of $\mathcal{N}=2$ gauged supergravity lets us draw on some standard identities (see, e.g., [@Andrianopoli:1996vr]) and may prove useful for a general analysis of black hole solutions using the results of [@Cacciatori:2009iz; @DallAgata:2010ejj; @Hristov:2010ri; @Halmagyi:2013qoa]. However, given the simplicity of the truncation, we also opt for a more direct approach whenever possible.
In particular, we observe that the topological twists in Section \[ssec:twistedindex\] have additional invariance, namely ${\rm U(1)}_R^{\text{W}}$. Imposing this symmetry on our supergravity theory at the level of the bosonic fields amounts to truncating half of the hypermultiplet by setting $\zeta_1=0$. The remaining complex scalar, $\zeta_2$, in the hypermultiplet will be denoted by $z\equiv \zeta_2$.
The four Abelian gauge fields, $A_{\mu}^{\alpha}$, $\alpha=0,1,2,3$, are related to the standard Cartan gauge fields in $\SO(8)$ by $$\label{Avfieldsmain}
\begin{split}
A^{12} & \eql {1\over 2}(A^0+A^1-A^2-A^3)\,,\qquad
A^{34} \eql {1\over 2}(A^0-A^1+A^2-A^3)\,,\\
A^{56} & \eql {1\over 2}(A^0-A^1-A^2+A^3)\,,\qquad
A^{78} \eql {1\over 2}(A^0+A^1+A^2+A^3)\,.
\end{split}$$ The bosonic Lagrangian for the truncated fields comprises of the usual Einstein-Hilbert term, kinetic terms for the scalars, a Maxwell term and a scalar potential[^17] $$e^{-1}\cals L= {1\over 2}\,R+\cals L_\text{kin}+\cals L_{\text{Max}}-g^2\cals P \,.$$ The details of the derivation can be found in Appendix \[appendixA\] and here we present only the final result.
The scalar kinetic term is given by $$\label{Lkindef}
\begin{split}
\cals L_\text{kin} & \eql -\sum_{i=1}^3 {\partial_\mu z_i\partial^\mu\bar z_i\over (1-|z_i|^2)^2} -{\big[\partial_\mu z-{\rm i}gA^{(m)}_\mu\,z\big]\big[\partial^\mu \bar z+
{\rm i}gA^{(m)}_\mu\,\bar z\big]\over (1-|z|^2)^2}\,,
\end{split}$$ where $$\label{massA}
A^{(m)}_\mu\equiv A^0_\mu-A^1_\mu-A^2_\mu-A^3_\mu\,.$$ The scalars parametrize the coset manifold $$\label{}
\cals M\eql \left[\rm {\SU(1,1)\over U(1)}\right]^3\times {\SU(1,1)\over {\rm U}(1)}\,.$$ From we read-off the diagonal metrics, $$\label{metrics}
g_{z_i\bar z_j}\eql {\delta_{ij}\over (1-|z_i|^2)^2}\,,\qquad g_{z\bar z}\eql {1\over (1-|z|^2)^2}\,,$$ that come from the Kähler potentials, $$\label{}
K_V\eql -\log\big[ (1-|z_1|^2)(1-|z_2|^2)(1-|z_3|^2)\big]\,,\qquad K_H\eql -\log(1-|z|^2)\,,$$ respectively.
The contribution from the complex scalar fields to the Lagrangian, $\cals L_{\text{Max}}$, for the gauge fields is quite complicated. To write it in a compact form it is convenient to use the standard scalar tensors from the $\cals N=2$ formalism [@Andrianopoli:1996vr] as summarized in Appendix \[appendixA\]. To this end we introduce the holomorphic sections, $X^\alpha$, $$\label{theXsmain}
\begin{split}
X^0 & \equiv {1\over 2\sqrt 2}\,(1-z_1)(1-z_2)(1-z_3)\,,\qquad
X^1 \equiv {1\over 2\sqrt 2}\,(1-z_1)(1+z_2)(1+z_3)\,,\\
X^2 & \equiv {1\over 2\sqrt 2}\,(1+z_1)(1-z_2)(1+z_3)\,,\qquad
X^3 \equiv {1\over 2\sqrt 2}\,(1+z_1)(1+z_2)(1-z_3)\,,\\
\end{split}$$ and the prepotential $$\label{prepmain}
F\equiv -2\,\i\, \sqrt{X^0X^1X^2X^3}\,.$$ The Maxwell Lagrangian is then[^18] $$\label{Lmaxdef}
\cals L_{\text{Max}} = {1\over 4}\Big(\cals I_{\alpha\beta}F_{\mu\nu}^\alpha F^{\beta\,\mu\nu}-\cals R_{\alpha\beta}F_{\mu\nu}^\alpha \widetilde F^{\beta\,\mu\nu}\Big)\,,$$ where $\cals R_{\alpha\beta}$ and $\cals I_{\alpha\beta}$ are the real and imaginary parts of $\cals N_{\alpha\beta}=\cals R_{\alpha\beta}+\i\,\cals I_{\alpha\beta}$, $$\label{Nalphabetadef}
\cals N_{\alpha\beta}\equiv \overline F_{\alpha\beta}+2 \,\i\,{(\Im F_{\alpha\gamma})(\Im F_{\beta\delta})X^\gamma X^\delta\over (\Im F_{\gamma\delta})X^\gamma X^\delta}\,,\qquad F_{\alpha\beta} \equiv {\partial^2 F\over\partial X^\alpha\partial X^\beta}\,.$$
The potential for the scalars is $$\label{potential}
\begin{split}
\cals P & \eql {2\over (1-|z|^2)^2}\bigg(3-\sum_{i=1}^3 {2\over 1-|z_i|^2}\bigg)\\
& \quad +{2\,|z|^2\over (1-|z|^2)^2}\bigg(\prod_{i=1}^3 {1\over 1-|z_i|^2}\bigg)\,\Big[4+4|z_1|^2|z_2|^2|z_3|^2 -(z_1+\bar z_1)(z_2+\bar z_2)(z_3+\bar z_3)
\\
&\hspace{100 pt} -(1+|z_1|^2)(z_2-\bar z_2)(z_3-\bar z_3)-(1+|z_2|^2)(z_1-\bar z_1)(z_3-\bar z_3)\\[6 pt]
&\hspace{100 pt} -(1+|z_3|^2)(z_1-\bar z_1)(z_2-\bar z_2)\,\Big]\;.
\end{split}$$ Let us define the $\cals N=1$ “holomorphic” superpotential, $$\label{holV}
\mathcal{V} = \frac{|z|^2}{1-|z|^2}(1-z_1)(1-z_2)(1-z_3)+\frac{2}{1-|z|^2}(z_1z_2z_3-1)\;.$$ Then $$\mathcal{P} = \frac{1}{2}e^{K_V}\big[\,g^{z_i\bar{z}_j}\,\nabla_{z_i}\mathcal{V}\nabla_{\bar{z}_j}\overline{\mathcal{V}}+4g^{z\bar z}\,\partial_{z}\mathcal{V}\partial_{\bar{z}}\overline{\mathcal{V}}-3\mathcal{V}\overline{\mathcal{V}}\,\big]\;,$$ where $$\label{defnabla}
\nabla_{z_i}\mathcal{V} = \partial_{z_i}\mathcal{V}+(\partial_{z_i}K_V)\mathcal{V}\;,$$ is a covariant derivative.
There are two supersymmetric AdS$_4$ solutions in this truncation corresponding to the critical points of the potential and the superpotential .[^19] The first one is the ${\rm SO(8)}$-invariant vacuum at $$\label{AdS4SO8}
z_i=0\;, \qquad z=0\;, \qquad \mathcal{P}_* = -6\;,$$ where $\cals P_*$ is the value of the potential at the critical point, which uplifts to the AdS$_4\times S^7$ solutions of the eleven dimensional supergravity. This solution is dual to the conformal vacuum of the ABJM theory.
The second supersymmetric AdS$_4$ solution was found by Warner [@Warner:1983vz] and is dual to the mABJM theory. In our parametrization of the potential, it is at $$\label{cpwz}
z\eql \pm {{\rm i}\over \sqrt 3}\,,\qquad z_1\eql z_2\eql z_3\eql \sqrt 3-2\;,\qquad \mathcal{P}_* = -\frac{9\sqrt{3}}{2}\,.$$ It has an $\mathcal{N}=2$ supersymmetry and is invariant under the ${\rm SU(3)}\times {\rm U(1)}_R^{W}$ subgroup of ${\rm SO(8)}$.
The scale of AdS$_4$ is set by $\cals P_*$. Hence we have $$\label{L4gCPW}
L_{\text{AdS}_4}^2 = - \frac{3}{g^2\mathcal P_*} = \begin{cases}
\displaystyle {1\over 2g^2} & \quad \text{for SO(8)}\,,\\[6 pt]
\displaystyle \frac{2}{3\sqrt{3}g^2} & \quad \text{for\, W}\,.
\end{cases}$$ This four-dimensional background uplifts to the CPW solution [@Corrado:2001nv] of the eleven-dimensional supergravity. A more detailed discussion of these (and other) AdS$_4$ vacua in this truncation as well as the spectrum of scalar excitations around them can be found in [@Bobev:2010ib].
A crucial fact that motivates much of the discussion in this paper is that there exists a supersymmetric gravitational domain wall solution which connects the two AdS$_4$ vacua described above [@Ahn:2000aq; @Bobev:2009ms]. This domain wall is the holographic dual realization of the RG flow described in Section \[subsec:ABJMmass\], which connects the ABJM SCFT to the mABJM SCFT.
There are two further consistent truncations of the supergravity model described above that are of interest for our discussion. The first one is the STU-model obtained by setting the hyperscalar, $z$, to zero and retaining the three complex scalars, $z_i$, and the four Abelian gauge fields, $A^\alpha$. For a discussion of this model in the present context, see for example [@Benini:2015eyy].
The second truncation is to the $\rm SU(3)\times U(1)_R$-invariant sector originally studied in [@Warner:1983vz] and recently discussed in [@Bobev:2010ib]. It is obtained by setting $$\label{}
z_1\eql z_2\eql z_3\eql -\bar{z}^\text{BHPW}\,,\qquad z\eql \zeta_2^\text{BHPW}\,.$$ The superscript BHPW refers to the scalars in [@Bobev:2010ib], where one also has to set $\zeta_1^\text{BHPW}=0$. In addition, one must impose $A^1=A^2=A^3$, which leaves only two Abelian fields in the truncation.
The BH Ansatz {#sec:ansatz}
-------------
Our goal is to study supersymmetric black hole solutions in the supergravity model presented above that are dual descriptions of the partial topological twists of the mABJM SCFT discussed in Section \[sec:CFT\]. These solutions should interpolate between one of the two supersymmetric AdS$_4$ vacua, the $\SO(8)$-invariant vacuum in or the $\SU(3)\times {\rm U}(1)$-invariant vacuum and a near horizon region with the metric of the form AdS$_2\times \Sigma_{\fg}$, where $\Sigma_{\fg}$ is a Riemann surface. As in other known examples of black holes solutions in $\rm AdS_4$ (see, e.g., [@Cacciatori:2009iz; @DallAgata:2010ejj; @Hristov:2010ri]), we need to turn on both scalar fields with nontrivial profiles, as well as non-vanishing gauge fields carrying the dyonic charges of the black hole at asymptotic infinity. In the presence of both the magnetic and electric charges, this turns out to be a difficult problem in general.
Fortunately, the entropy of these black holes can be determined by a much simpler set-up, namely by studying the solutions in the near horizon region only. This is what we will do in the remainder of this section. We will return to the more difficult problem of constructing full black hole solutions in Section \[sec:magneticBH\], where we present both analytic and numerical solutions for magnetically charged black holes, but with vanishing electric charges.
To construct the near horizon AdS$_2\times \Sigma_{\fg}$ solutions of interest, we take the scalar fields, $z_i$ and $z$, to be constant and the metric of the form, $$\label{AnsatzAdS2}
ds^2 = e^{2f_0} ds_{{\rm AdS}_2}^2 + e^{2h_0} ds^2_{\Sigma_{\fg}}\;,$$ where the unit radius metric on ${\rm AdS}_2$ is $$\label{}
ds_{{\rm AdS}_2}^2\eql {1\over r^2}(-dt^2+dr^2)\,,$$ and $f_0$ and $h_0$ are real constants. Given the results of the analysis in [@Anderson:2011cz], we expect that without a loss of generality we can use a constant curvature metric on $\Sigma_\fg$ given in .
The gauge field fluxes, $F^\alpha=dA^\alpha$, and their (local) potentials, $A^\alpha$, are $$\label{FdAnsatz}
\qquad F^{\alpha} = e_{\alpha} {\rm vol}_{{\rm AdS}_2} + m_{\alpha} {\rm vol}_{\Sigma_{\fg}}\;,
\qquad
A^\alpha\eql e_{\alpha} \,\omega_{\rm AdS_2}+m_\alpha \,\omega_{\Sigma_\fg}\,,\qquad \alpha=0,\ldots,3\,.$$ One can now plug this Ansatz into the supersymmetry variations and the equations of motion of maximal $\cals N=8$ gauged supergravity and derive a system of algebraic equations between the metric constants, the scalar fields and the magnetic and electric charges. This is a straightforward but tedious calculation summarized in Appendix \[appendixB\]. There we also show that the black holes we construct preserve 2 real supercharges which are enhanced to 4 in the near horizon AdS$_2$ region.
Dyonic BH near horizon BPS equations {#subsec:dyonicBPS}
------------------------------------
The truncation of the equations of motion and the supersymmetry variations of $\cals N=8$ $d=4$ gauged supergravity discussed in Appendix \[appendixB\] yields four types of algebraic equations for the supersymmetric near horizon dyonic black holes:[^20]
- Four real equations for the electric and magnetic parameters, $e_\alpha$ and $m_\alpha$: $$\begin{aligned}
e_0 & \eql 0\,,\label{e0eqs}\\
e_0-e_1-e_2-e_3 & \eql 0\,,\label{e123eqs}\\
m_0 & \eql -{\kappa\over 2g}\,,\label{m0eqs}\\
m_0-m_1-m_2-m_3 & \eql 0\,,\label{m123eqs}\end{aligned}$$ where $\kappa=1$, 0 or $-1$ is the normalized curvature of the Riemann surface.
- Four complex equations for the scalar dressed components, $\Phi_\alpha$, of the fluxes: $$\begin{aligned}
\cfb_0 & \eql -2g\,\overline\fW\,,\label{Ph0eqs}\\[6 pt]
\cfb_i & \eql -2g(1-|z_i|^2)\,D_{z_i}\fW\,,\qquad i=1,\ldots,3\,,\label{Phieqs}\end{aligned}$$ where, cf. , $$\label{defofW}
\fW\eql e^{K_V/2}\,\cals V\,,$$ The fluxes $\Phi_\alpha$ are defined by $$\label{Smatactmn}
S_{\alpha\beta}\cfb_\beta \eql e^{-2h_0}m_\alpha+\i\,e^{-2f_0}e_\alpha\,,$$ where $$\label{Smatrixmn}
\begin{split}
S_{\alpha0} & \eql {1\over \sqrt 2}\, L^\alpha\,,\\
S_{\alpha i} & \eql -{1\over \sqrt 2} (1-|z_i|^2) \overline{D_{z_i}L^\alpha}\,.
\end{split}$$ Here $L^\alpha=e^{K_V/2}X^\alpha$ are the symplectic sections, cf. . The Kähler covariant derivative in and is defined as[^21] $$\label{}
D_{z_i}\eql \partial_{z_i}+{1\over 2}\partial_{z_i}K_V\,.$$ Note that $D_{z_i}L^\alpha=f_{z_i}{}^\alpha$, see .
- One complex equation for the metric constant, $f_0$, and the phase $\Lambda$, $$\label{eqf0La}
e^{-f_0-\i\,\Lambda}\eql \sqrt 2\,\i\,g\,\fW\,.$$
- A complex cubic constraint for the scalars, $z_i$, $$\label{thezconstr}
\cals C\equiv z_1z_2 z_3 +z_1z_2 +z_2 z_3+z_3 z_1-z_1-z_2-z_3-1\eql 0\,.$$ Note that $$\label{cubeconstr}
\cals C\eql \sqrt2\,\left( X^0-X^1-X^2-X^3 \right)\,,$$ where $X^\alpha$ are the holomorphic sections .
An indirect check of the consistency of these equations with the ones obtained for general dyonic black holes using the formalism of $\cals N=2$ gauged supergravity [@DallAgata:2010ejj; @Hristov:2010ri; @Halmagyi:2013sla] is to rewrite them as “attractor equations.” This is briefly summarized in Appendix \[appattractor\].
Our task here is to solve the equations (i)-(iv) so that we can compare directly the black hole entropy $$\label{}
S_{\text{BH}} = \frac{\text{Area}}{4G_N^{(4)}} = {\pi |\fg-1|\over G_N^{(4)}}\, e^{2h_0}\,,$$ with the twisted topological index . The strategy is to solve for the metric parameters, the magnetic and electric parameters, and the hyperscalar in terms of the three scalars, $z_i$, which then will be mapped onto the fugacities, $u_i$, of the mABJM theory.
We start by acting with the matrix $S_{\alpha\beta}$ on and . Using , this yields $$\label{abseqs}
e^{-2h_0}m_\alpha+\i\,e^{-2f_0}e_\alpha \eql -\sqrt 2g\Big[L^\alpha\,\overline\fW-\sum_{i=1}^3 g^{z_i\bar z_i}\,\overline{D_{z_i}L}{}^\alpha D_{z_i}\fW\Big]\,.$$ The next step is to project these equations onto the real and imaginary part and then use -.
To this end, first note that $\fW$ given in can be rewritten entirely in terms of the symplectic sections, $L^\alpha$, $$\label{calWL}
\begin{split}
\fW & \eql -2\sqrt 2\,L^0 +{\sqrt 2\over 1-|z|^2}\,(L^0-L^1-L^2-L^3)\,,
\end{split}$$ where the second term is proportional to the cubic constraint , $$\label{Wwconstr}
\fW \eql -2\sqrt 2\,L^0 +{1\over 1-|z|^2}\,\fG\,,\qquad \fG\equiv e^{K_V/2}\,\cals C\,.$$ We can use this to simplify the first term in the square bracket in .
Next, we have the “useful relation” [@Ceresole:1995ca] $$\label{usefulrel}
\sum_{i=1}^3 g^{z_i\bar z_i} D_{z_i}L^\alpha \overline{D_{z_i}L}{}^\beta \eql -{1\over 2}\,(\cals I^{-1})^{\alpha\beta}-{\overline L}{}^\alpha L^\beta\,,$$ where $\cals I_{\alpha\beta}$ is real, the imaginary part of $\cals N_{\alpha\beta}$ in . Using those identities in , we find $$\label{ealsol}
e^{-2f_0}e_\alpha\eql -4\,\i\,g\left (\,\overline{L}{}^0 L{}^\alpha-L^0\overline{L}{}^\alpha \,\right)\,,$$ and $$\label{malsol}
\begin{split}
e^{-2h_0}m_\alpha & \eql 4\,g\,\left (\,\overline{L}{}^0 L{}^\alpha+L^0\overline{L}{}^\alpha \,\right)\\[6 pt]
&\qquad +2g (\cals I^{-1})^{0\alpha}-{g\over 1-|z|^2}\left[(\cals I^{-1})^{0\alpha}-(\cals I^{-1})^{1\alpha}-(\cals I^{-1})^{2\alpha}-(\cals I^{-1})^{3\alpha}\right]\,.
\end{split}$$ The solution is consistent with the equations and for the electric parameters. Indeed, the first one is satisfied manifestly, while the left hand side in the second one is proportional to the cubic constraint . Substituting for $\alpha=0$ in , we solve for $e^{-2h_0}$ and then similarly for $|z|$. Finally, from and we obtain an explicit solution for all electric parameters and magnetic fluxes.
This shows that the equations - have a unique solution for all the dyonic black hole parameters in terms of the scalars, $z_i$, that are constrained by . The problem is that the expressions such as are quite difficult to use because of the complicated form of the inverse matrix, $\cals I^{-1}$.
To obtain simpler explicit expressions for $e^{-2h_0}$ and $|z|$, we go back to . Taking a linear combination of these equations such that the left hand side vanishes using and and observing that the first terms on the right hand side sum up to the constraint, we are left with $$\label{}
2\sqrt 2\,\sum_{i=1}^3 g^{z_i\bar z_i}\, \overline{D_{z_i}\fG} D_{z_i}L^0-{1\over 1-|z|^2}\,\sum_{i=1}^3 g^{z_i\bar z_i} \,\overline{D_{z_i}\fG} D_{z_i}\fG\eql 0\,.$$ Note that by the “useful relation,” the first term above is real. Let us introduce the shorthand notation $\langle\,\cdot\,,\,\cdot\,\rangle$ for the scalar product defined by the sums. Then we have $$\label{zsols}
{1\over 1-|z|^2}\eql 2\sqrt 2\,{\langle D\fG, DL^0\rangle \over \langle{D\fG}, D\fG\rangle}\,.$$ One can verify explicitly that $$\label{}
\langle DL^0,DL^0\rangle \eql 3|L^0|^2\,.$$ This allows to further simplify the solution for $e^{-2h_0}$ that follow from with $\alpha=0$ after using . The result is $$\label{}
e^{-2h_0}\eql 8 g^2\kappa\left(2|L^0|^2-{|\langle D\fG, DL^0\rangle|^2\over \langle{D\fG}, D\fG\rangle}\right)\,.$$ Finally, one can evaluate the scalar products above explicitly and use $$\label{}
D_{z_i}\fG \big|_{\cals C=0}\eql e^{K_V/2}\,\partial_{z_i}\cals C\,.$$ Let us define $$\label{}
\Xi \eql \sum_{i=1}^3 (1-|z_i|^2){1-z_i\over 1-\bar z_i}{\partial_{z_i} \cals C}\,,\qquad
\Gamma\eql \sum_{i=1}^3 (1-|z_i|^2)^2\left|{\partial_{z_i} \cals C}\right|^2\,,\qquad \Pi\eql \prod_{i=1}^3(1-z_i)\,.$$ Then we find $$\label{solem2h0}
e^{-2h_0}\eql 8 g^2\,\kappa\,|L^0|^2 \left(2-{|\Xi|^2\over \Gamma}\right)\,,$$ and[^22] $$\label{oneoverzzb}
{1\over 1-|z|^2}\eql -{\overline \Pi\,\Xi\over \Gamma}\,,$$ and, from and , $$\label{solef0}
e^{-f_0}\eql -4 \,\i\,g\, e^{\i\,\Lambda}\,L^0\,.$$ This completes the solution for the AdS$_2\times\Sigma_\fg$ near horizon black holes in our model.
### Comments
The result of our analysis above is an explicit solution for the metric parameters and the hypermultiplet given in - as functions of the constrained vector multiplets’ scalars. The solutions for the magnetic fluxes and the electric parameters can then be read-off from or more directly from and . It should be noted that in this near horizon solution the hypermultiplet scalar appears only through its absolute value $|z|$.
It is well known that the STU black holes with electric charges can exist only for nontrivial axions, that is complex scalar fields, $z_i$. The same is true in our model. Indeed, if we set $z_i=\bar z_i$ to be real, the electric parameters given by automatically vanish.
The regularity of a solution requires that the left hand sides in and be strictly positive and, since $|z|<1$, the left hand side in be greater than 1. This fixes the phase $\Lambda$ and the supersymmetric projectors , and excludes the possibility of black hole solutions with toroidal ($\kappa=0$) horizons. Then we are left with two conditions $$\begin{aligned}
\label{h0cond}
\kappa\left(2-{|\Xi|^2\over \Gamma}\right) & >0\,,\qquad \kappa\eql \pm 1\,,\\
\label{zcond}
-{\overline \Pi\,\Xi\over \Gamma}-1 & >0\,.\end{aligned}$$ It is easy to check numerically that both for spherical ($\kappa=1$) and hyperbolic ($\kappa=-1$) horizons, there exist constrained scalars, $z_i$, for which both inequalities are satisfied. Once more something interesting happens in the purely magnetic limit. One can check that for real $z_i$’s the ratio of the left hand sides in and is constant and equal to $-2$. Hence $$\label{highgenus}
-2\kappa>0\qquad \text{for}\qquad z_i\eql \bar z_i\,,$$ which excludes spherical horizons for purely magnetic black holes. We show in Section \[sec:magneticBH\] that the hyperbolic near horizon solutions with only magnetic fluxes indeed give rise to bona fide black holes with AdS$_4$ asymptotics.
Ideally one would like to know for which values of the electric and magnetic parameters, $e_\alpha$ and $m_\alpha$, there are regular near-horizon black hole solutions. This entails finding an explicit solution for the scalars $z_i$, $z$, and the metric parameters, $f_0$ and $h_0$, in terms of $e_\alpha$ and $m_\alpha$. We were not able to analyze explicitly this complicated algebraic problem for the general dyonic solutions above. However, in Section \[sec:solspace\] we show how to answer this question for the purely magnetic black holes.
### STU black holes {#sssec:STUBH}
It is straightforward to extract from the supersymmetry variation in Appendix \[appendixB\] the STU-limit of our model and reproduce the black holes studied in [@Cacciatori:2009iz; @Benini:2015eyy; @Benini:2016rke]. The resulting BPS equations can be summarized as follows:
- The four equations - are replaced by two equations defining the topological twist $$\begin{aligned}
\label{eRstu}
e_0+e_1+e_2+e_3 & \eql 0\\[6 pt]
m_0+m_1+m_2+m_3 & \eql -{\kappa\over g}\,.\label{mRstu}\end{aligned}$$
- Equations - remain the same, except that $$\label{}
\fW^\text{STU}\equiv \fW\big|_{z=\bar z=0}\eql -\sqrt 2(L^0+L^1+L^2+L^3)\,.$$
- Equation is the same, but with $\fW^\text{STU}$.
- There is no constraint, which was due to the hypermultiplet scalar.
The equations in (ii) give rise to the analogue of . By taking the sum of those four equations one finds that is identically satisfied. Then from one finds $$\label{solh0stu}
e^{-2h_0}\eql -g^2\kappa (|\fW^\text{STU}|^2-\langle D\fW^\text{STU},D\fW^\text{STU}\rangle)\,.$$ Substituting and in one finds $e_0,\ldots,m_3$.
The duality {#sec:DyonicComp}
===========
Our goal now is to test directly the conjecture that the topologically twisted dyonic/magnetic indices in Section \[ssec:twistedindex\] match the entropy of the near horizon black hole solutions above. We also compare directly the magnetic fluxes and electric charges on both sides of the duality.
The translation between the gravitational and field theory quantities of interest is provided by the free energy, $F_{S^3}$, on both sides of the duality. On the gravity side, the free energy of pure AdS$_4$ with $S^3$ as an asymptotic boundary can be computed from an on-shell action which diverges unless properly regulated. With the correct counterterms described in [@Emparan:1999pm], one finds $$\label{Fdef}
F_{S^3} = \frac{ \pi L_{\text{AdS}_4}^2}{2G_N^{(4)}} \;,$$ where $G_N^{(4)}$ is the four-dimensional Newton constant and $L_{\text{AdS}_4}$ is the radius of AdS$_4$. This supergravity result agrees with the free energy of the ABJM SCFT and the mABJM SCFT to leading order in $N$ given in and , respectively. Indeed, it was shown in [@Jafferis:2011zi] that the ratios of these free energies and the radii of the corresponding AdS$_4$ vacua given in are universal, $$\label{}
{F_{S^3}\over L_{\text{AdS}_4}^2}\eql {2\sqrt 2 \pi\over 3} g^2 N^{3/2}\,.$$ Using these results and , we arrive at the following string of equalities: $$\label{Sdeff}
S_{\text{BH}} \equiv \frac{\text{Area}}{4G_N^{(4)}} \eql \frac{\pi |\fg-1| }{G_N^{(4)}}\,e^{2h_0} \eql 2 \, |\fg-1| {F_{S^3}\over L_{\text{AdS}_4}^2}\,e^{2h_0}\eql
{4\sqrt 2 \pi\over 3}\,{|\fg-1|}\,g^2\, N^{3/2}\,e^{2h_0}\,.$$ where “Area” is the area of the black hole horizon.
We have parametrized the field strengths in terms of “bare” magnetic fluxes, $m_\alpha$, and the electric parameters, $e_\alpha$. Those are related to the actual magnetic and electric charges of the AdS$_4$ black holes by (see, e.g., [@Benini:2016rke]) $$\begin{split}\label{nqcharges}
n_\alpha \eql {1\over 4\pi |\fg-1|} \int_\Sigma F^\alpha\,,\qquad
q_\alpha &= \frac{1}{4\pi |\fg-1|} \int_\Sigma \frac{\delta \mathcal{L}_\text{Max}}{\delta F^\alpha} \,,
\end{split}$$ where $ \mathcal{L}_\text{Max}$ is the Maxwell action . Starting with the Ansatz and evaluating the integrals using , we find (cf. [@Halmagyi:2013sla]), $$\label{msandqssugra}
n_\alpha\eql m_\alpha\,,\qquad q_\alpha\eql - e^{2(h_0-f_0)} \mathcal{I}_{\alpha \beta} e_\beta + \mathcal{R}_{\alpha \beta} m_\beta \,.$$ These are those magnetic fluxes and electric charges that should be matched with their field theory counterparts in Section \[sec:CFT\].
The ABJM SCFT and STU supergravity {#sec:abjmstu}
----------------------------------
The equality between the twisted dyonic index in ABJM SCFT and the entropy of the corresponding AdS$_4$ black holes in STU supergravity was first shown in [@Benini:2016rke] by mapping the extremization problem for the index onto the BPS equations rewritten in the form of the “attractor equations” in [@DallAgata:2010ejj; @Hristov:2010ri]. Given the explicit form for the extremized index derived in Section \[sec:dabjm\] and the entropy given by and , we can now verify that equality directly.
Following [@Benini:2016rke], let us consider the map between the complex fugacities, $u_\alpha$, in Section \[sec:dabjm\] and the scalar fields, $z_i$, at the black hole horizon given by $$\label{bhzus}
u_{\alpha}(z) \eql {2 X^{\alpha-1}\over X^0+X^1+X^2+X^3}\,,\qquad \alpha\eql 1,\ldots,4\,,$$ which automatically solves the constraint . Setting $$\label{}
\Delta_\alpha\eql |u_\alpha |\,,\qquad \theta_\alpha\eql \mathop{\rm Arg}(u_\alpha)\,,$$ we find that indeed $$\label{eqISabjm}
\cals I_D^\text{ABJM}(\fn_\alpha(u),\fq_\alpha(u),u_\alpha)\eql S_{\rm BH}^\text{STU}(z_i)\,.$$ The twisted index on the left hand side is evaluated using , which corresponds to the analytic continuation defined by .
By comparing the magnetic fluxes and the electric charges with the ones obtained from using and in STU supergravity, we find that $$\label{eqnnabjm}
\fn_\alpha\eql 2g\,|\fg-1|\,m_{\alpha-1}\,,$$ and $$\label{eqqqabjm}
\fq_{\alpha}\eql {2\sqrt 2\over 3} \,g\,|\fg-1|\,N^{3/2}\,q_{\alpha-1}\,,$$ where $\alpha=1,\ldots,4$. The relations and agree with the ones proposed in [@Benini:2016rke]. This confirms the BHZ conjecture for the ABJM SCFT and the STU black holes.
We have verified - by evaluating both sides numerically for a large number of randomly chosen values of the near horizon scalars, $z_i$. The result is that the extremized index, the magnetic fluxes and electric charges for the particular branch for the square-root in defined by agree with the entropy, the fluxes and the charges on the supergravity side in the entire domain of the scalars, $|z_i|<1$. This includes values of $z_i$ for which the supergravity solution may not be regular, such as when the entropy is negative.
The mABJM SCFT and W supergravity {#sselecchar}
---------------------------------
We have seen in Sections \[sec:mabjmdind\] and \[sec:manjmditc\] that the extremized dyonic twisted index in the mABJM SCFT theory could be obtained in two ways. On the one hand, one can work entirely within the mABJM theory, which yields the extremized index , the magnetic fluxes and the electric charges corresponding to [*three*]{} gauge fields for the unbroken ${\rm U}(1)^3$ global symmetry. On the other hand, one can start with the ABJM SCFT and extremize the dyonic twisted index while imposing two constraints on the fugacities, where the second constraint comes from the mass deformation from ABJM to mABJM. This leads to the same result for the extremized index as in . However, now there are four magnetic fluxes and four electric charges corresponding to the ${\rm U}(1)^4$ global symmetry in ABJM. The mass deformation fixes one of the magnetic fluxes, which together with the topological twist condition, can be used to determine the real parts of the two Lagrange multipliers to find the full agreement between the remaining three magnetic fluxes in both calculations. However, there remains an ambiguity in the solution for the electric charges due to the shift symmetry . In this section we will compare the field theory results with the supergravity calculations in Section \[subsec:dyonicBPS\] and, in particular, clarify the ambiguity of the electric charges found in Section \[sec:manjmditc\].
### The dyonic twisted index and the entropy {#sec:mabjmentropy}
To obtain the mapping between the fugacities $u_2,u_3,u_4$ in mABJM and the near horizon scalars, $z_1,z_2,z_3$, in our supergravity model, we note that the cubic constraint is equivalent to $u_1=1$ in . Then $u_2,u_3,u_4$ automatically satisfy . One can also arrive at this mapping by the following change of variables. Observe that the Möbius transformation, $$\label{ztildedef}
z_i\qquad \longrightarrow\qquad \tilde z_i\eql {1-z_i\over 1+z_i}\,,\qquad \Re \,\tilde z_i>1\,,$$ turns the cubic constraint into $$\label{ncubconstr}
\tilde{\cals C}\equiv \tilde z_i\tilde z_2\tilde z_3 -\tilde z_1-\tilde z_2-\tilde z_3\eql 0\,,$$ which can be rewritten in the following suggestive form, $$\label{}
{1\over \tilde z_2\tilde z_3}+{1\over \tilde z_1\tilde z_3}+{1\over \tilde z_1\tilde z_2}\eql 1\,.$$ Hence, it is natural to define $$\label{construsub}
u_{i}\eql {\tilde z_{i-1}\over\tilde z_1\tilde z_2\tilde z_3}\,,\qquad i=2,3,4\,,$$ which reproduces modulo the constraint .
Using the substitution , we find that the extremized twisted index in mABJM and the entropy given by and are the same, $$\label{}
\cals I^\text{mABJM}_{D}(\fn_i(u),\fq_i(u);u_i)\eql S_\text{BH}(z_i)\qquad \text{for}\qquad \cals C(z_i)\eql 0\,,$$ in the entire domain of the scalars, $|z_i|<1$, where the cubic contraint is satisfied.
### Duality for the magnetic fluxes and electric charges {#sec422}
An initial puzzle when comparing the magnetic fluxes and the electric charges in mABJM with the ones in the dual (W-) supergravity is that the latter appears to have four vector fields, while there are only three fluxes and three charges in and , respectively. The resolution is that at the W-critical point, which is the gravity dual for mABJM, one of the vector fields, $$\label{}\label{massA}
A^{(m)}\eql A^0-A^1-A^2-A^3\,,$$ becomes massive and must be set to zero. This leaves us with three vector fields, $A^1$, $A^2$ and $A^3$ and the corresponding three magnetic fluxes, $n_i$, and three electric charges, $q_i$, in to compare.
Using the map between the constrained scalars, $z_i$, and the fugacities, $u_i$, we find the same relation between the magnetic fluxes in mABJM and their gravity duals, $$\label{nabjmcom}
\fn_i\eql 2g\,|\fg-1|\,n_{i-1}\,,\qquad i=2,3,4\,.$$ The comparison of the electric charges is more subtle. We must first impose the massive condition in the Maxwell Lagrangian to find the scalar matrices $\widetilde{\cals I}_{ij}$ and $\widetilde{\cals R}_{ij}$ for the vector fields, $A^i$, $$\label{}
\begin{split}
\widetilde{\cals I}_{ij} & \eql \cals I_{ij}+\cals I_{00}+\cals I_{i0}+\cals I_{0j}\,,\qquad
\widetilde{\cals R}_{ij} \eql \cals R_{ij}+\cals R_{00}+\cals R_{i0}+\cals R_{0j}\,,\\
\end{split}$$ and then calculate the electric charges, $\tilde q_i$, using . Comparing with the field theory charges in , which we denote by $\tilde \fq_i$, we find $$\label{qabjmcomp}
\tilde \fq_i\eql {2\sqrt 2\over 3}g\,|\fg-1|\,N^{3/2}\,\tilde q_{i-1}\,,\qquad i=2,3,4\,,$$ which is the same as . As before, the comparison is carried out by a numerical substitution and both and hold for all scalars, $z_i$, satisfying the cubic constraint.
### More on electric charges {#sec:eleccharg}
We have argued in Section \[sec:manjmditc\] that the dyonic twisted index and the magnetic fluxes in mABJM could be obtained unambiguously by performing simultaneously the mass deformation and the topological twist in ABJM. However, the resulting electric charges were determined only up to the 1-parameter shift symmetry . One way to fix that symmetry was to set one of the charges to zero, see , to obtain a complete agreement with the purely mABJM charges.
Another possibility, which we will discuss now, is to compare the four electric charges with the four electric charges in our supergravity model that are present before imposing the massive constraint on the vector fields. With all four vector fields present, the corresponding charges, $q_\alpha$, are given by . By a direct substitution, we find that $$\label{qabjmtwoL}
\begin{split}
\fq_1-\nu_1 & \eql {2\sqrt 2\over 3}g\,|\fg-1|\,N^{3/2}\,q_0\,,\\
\fq_i+\nu_1 & \eql {2\sqrt 2\over 3}g\,|\fg-1|\,N^{3/2}\,q_{i-1}\,,\qquad i=2,3,4\,.
\end{split}$$ Hence for $\nu_1=0$ we find the same relation between the charges as in . Given , which followed from the reality of the extremized twisted index , we see that a complete match between mass-deformed, topologically twisted ABJM theory and our supergravity model requires that both Lagrange multipliers, $\lambda_1$ and $\lambda_2$ , in be real. Once again, the agreement between the field theory and the dual supergravity charges holds for all allowed values of the near horizon scalars.
Magnetic black holes {#sec:magneticBH}
====================
So far we have constructed a large family of supersymmetric AdS$_2\times \Sigma_{\fg}$ solutions which can be interpreted as near horizon limits of supersymmetric dyonic black holes in our gauged supergravity model. In addition, we have shown that the entropy associated with these near-horizon backgrounds is the same as the large $N$ limit of the dyonic topologically twisted index of the mABJM SCFT. Now we try to be more explicit and focus on a class of supersymmetric solutions of our model that do not have electric charges. This allows for a much more explicit analysis of the BPS equations. In particular, in addition to the near-horizon solutions discussed above, we are able to find fully-fledged black hole backgrounds.
To this end we modify the Ansatz employed in Section \[sec:ansatz\] by setting the electric charge parameters, $e_{\alpha}$, in to zero. It is consistent then to take the four complex scalars $z_{1,2,3}$ and $z$ to have constant phases. As discussed around , the BPS equations impose then that the Riemann surface is hyperbolic, so we set $\kappa=-1$ in this section.
We find it convenient to use the following reparametrization of the complex scalar fields, cf. and , $$\label{thezss}
z\eql \tanh\chi\,e^{\,\rm{i}\,\psi}\,,\qquad z_i\eql \tanh\lambda_i\,e^{\,{\rm i}\,\varphi_i}\,,\quad i=1,2,3\,,$$ where $\chi,\lambda_i\geq 0$, $2\pi > \psi,\varphi_i\geq 0$ are real valued fields.
Since we are interested in solutions that can be asymptotic to the ${\rm SU(3)}\times {\rm U(1)}$ invariant vacuum, we choose the following phases in , cf. , $$\label{psivarphi}
\psi = \frac{\pi}{2}\;, \qquad \varphi_1=\varphi_2=\varphi_3=\pi\;.$$ The metric for the black hole solutions of interest takes the following form $$\label{AnsatzAdS2r}
ds^2 = e^{2f(r)}(-dt^2+dr^2) + e^{2h(r)} ds^2_{\Sigma_{\fg}}\;,$$ with the same metric on the Riemann surface as in . The four real scalars are in general functions of the radial variable, $\chi(r)$ and $\lambda_i(r)$.
With this Ansatz at hand, one can analyze the supersymmetric variations and of the $\cals N=8$ gauged supergravity and find BPS equations for the metric functions and the scalars. To write these equations in a compact form, we find it convenient to introduce the following positive variables:[^23] $$\label{xlambda}
x_i\equiv e^{2\lambda_i}\,,\qquad i=1,2,3\,,$$ as well as the “real superpotential” (see , , and ) $$\label{eq:W}
\mathcal{W} \equiv \mathfrak{W}|_{2\psi=\varphi_i=\pi}= \frac{2 x_{1} x_{2} x_{3} \sinh ^2(\chi )-\cosh ^2(\chi ) (x_{1} x_{2} x_{3}+x_{1}+x_{2}+x_{3})}{ 2\sqrt{x_{1}x_2x_3}} \, .$$ Then the potential in can be written as $$\mathcal{P} = \frac{1}{2}\left(\frac{\partial \mathcal{W}}{\partial\chi}\right)^2+\frac{1}{2}\sum_{i=1}^3\left(\frac{\partial \mathcal{W}}{\partial\lambda_i}\right)^2-\frac{3}{2}\,\mathcal{W}^2\;,$$ and the BPS equations are given by $$\label{magnBPS}
\begin{split}
\frac{df}{dr} &= \frac{g}{\sqrt{2}} e^f\, \mathcal{W} - e^{f-2h} \,\mathcal{H} \, , \qquad\qquad \frac{dh}{dr} = \frac{g}{\sqrt{2}} e^f \,\mathcal{W} + e^{f-2h} \,\mathcal{H} \, ,\\[6 pt]
\frac{d\chi}{dr} &= -\frac{g}{\sqrt{2}}\, e^f\, \frac{\partial \mathcal{W}}{\partial \chi} \, , \hspace{77 pt} \frac{d\lambda_i}{dr} = -\frac{g}{\sqrt{2}} \,e^f\, \frac{\partial \mathcal{W}}{\partial \lambda_i} - e^{f-2h}\, \frac{\partial \mathcal{H}}{\partial \lambda_i} \,,
\end{split}$$ where $$\label{eq:H}
\mathcal{H} = \frac{1}{2 \sqrt{2\,x_{1} x_{2} x_{3}}} \left(m_0+m_{1} x_{2} x_{3}+m_2 x_{1} x_{3}+m_3 x_{1} x_{2}\right)\;.$$ In addition to the equations in , one has to impose the constraints and on the magnetic fluxes.
Our goal is to find supersymmetric black hole solutions with regular horizons to the BPS equations , , and to analyze their entropy.
${\rm AdS}_2\times \Sigma_\fg$ solutions {#AdS2xSigmag}
----------------------------------------
We begin with a classification of the possible ${\rm AdS}_2\times \Sigma_\fg$ solutions, which should correspond to the near-horizon limits of the supersymmetric black holes of interest. To this end, we take the familiar Ansatz for the metric and real, constant scalar fields, $x_i$ and $\chi$. For that radial dependence of the metric functions and the scalars, the BPS equations reduce to a set of algebraic equations, which can be solved following the procedure outlined in Section \[subsec:dyonicBPS\].
Just as before, for solutions with nonzero[^24] $\chi$, the scalars must obey the cubic constraint , which now takes the form $$\label{sumxmxc}
x_1 +x_2 +x_3 \eql x_1 x_2 x_3\,.$$ This is the same as the constraint for real $z_i$, cf. .
Using , the magnetic fluxes can be expressed in terms of the scalars as follows: $$\label{solal}
\begin{split}
m_1 & \eql {m_0\over x_1x_2+x_1x_3+x_2x_3}\Big(1+{x_2\over x_3}+{x_3\over x_2}\Big)\,,\\[6 pt]
m_2 & \eql {m_0\over x_1x_2+x_1x_3+x_2x_3}\Big(1+{x_3\over x_1}+{x_1\over x_3}\Big)\,,\\[6 pt]
m_3 & \eql {m_0\over x_1x_2+x_1x_3+x_2x_3}\Big(1+{x_1\over x_2}+{x_2\over x_1}\Big)\,,
\end{split}$$ where $m_0$ satisfies . The other constraint on the magnetic fluxes is then automatically satisfied modulo the cubic constraint .
Ideally, one would like to invert to find the scalar fields in terms of the magnetic fluxes, that should be thought of as the physical parameters specifying a solution. However, the inversion is tedious and not very insightful, so we choose to write our solutions as above in terms of the scalar fields.
Using , we can solve for $f_0$, $h_0$ and the scalar, $\chi$, in terms of the scalars, $x_i$: $$\label{fhchi}
\begin{split}
e^{2h_0} &= \frac{1}{2g^2}\frac{1}{x_1 x_2+x_2 x_3+x_3 x_1} \left(x_1 x_2 x_3-\frac{1}{x_1}-\frac{1}{x_2}-\frac{1}{x_3}\right) \, , \\[6 pt]
{\rm csch}^2\chi &= 1+{x_1^2+x_2^2+x_3^2\over x_1x_2+x_1x_3+x_2x_3}\,,\\[6 pt]
e^{f_0} &= \frac{1}{g \sqrt{2 x_1 x_2 x_3}} \,,
\end{split}$$ where $x_i$ are constrained by . These equations can also be obtained directly from , , and by restricting the complex scalars, $z_i$, as in and .
### AdS$_2$ solution space {#sec:solspace}
![The range of magnetic fluxes giving rise to regular AdS$_2\times\Sigma_{\fg}$ solutions.[]{data-label="fig:positivity"}](solspace_no_mesh.pdf){width="3.5in"}
Although we have solved the algebraic BPS equations, we still need to analyze for what range of the magnetic fluxes, $m_{\alpha}$, we have a regular well-defined horizon. By that we mean a solution for which the scalars $\chi$ and $\lambda_i$, as well as $f_0$ and $h_0$, are real.
The magnetic fluxes, $m_\alpha$, and the scalar fields, $x_i$, are related by . The scalar fields must be positive, $x_i>0$, and are constrained by . For $\kappa=-1$, we have $m_0>0$ (see ) and thus the constraint , combined with the positivity of $x_i$, leads to the following region in the $m_{1,2,3}$ magnetic flux space: $$\label{CPWconstr}
m_1+m_2+m_3\eql m_0 \,,\qquad \frac{m_i}{m_0} > 0\,.$$ This region is the triangle shown in Figure \[fig:positivity\]. As noted above, the relation between the magnetic fluxes in combined with ensures that the constraint is satisfied. We have checked numerically that for every value of the magnetic fluxes inside the region specified by , there is a corresponding regular AdS$_2\times \Sigma_{\fg}$ solution, that is the scalars $x_i$ and $\chi$, and the metric functions $e^{f_0}$ and $e^{h_0}$ are real and positive. For the magnetic fluxes at the boundary of the triangle (both the edges and the vertices) in Figure \[fig:positivity\], one has to analyze the regularity of the solution with more care since the relations become singular. By solving the algebraic BPS equations directly, we find that there are no regular AdS$_2\times \Sigma_{\fg}$ solutions for these “boundary” values of $m_{\alpha}.$
It is instructive to compare the region of the allowed magnetic fluxes in Figure \[fig:positivity\] with the region in which the STU model regular magnetic black holes discussed in [@Benini:2015eyy] exist. Using the results in Section \[sssec:STUBH\] and the positivity of the scalars, $x_i$, one can show that for the purely magnetic STU model black holes with hyperbolic horizons, i.e. $\kappa=-1$, the magnetic fluxes must obey the inequality $$\label{STUmagineq}
(m_0+m_1-m_2-m_3)(m_0-m_1+m_2-m_3)(m_0-m_1-m_2+m_3) < 0\;.$$ When evaluated on the surface $m_0=m_1+m_2+m_3$ relevant for our discussion, reduces to the constraint $m_1m_2m_3<0$. Given and , we thus conclude that the magnetic STU model black holes have magnetic fluxes that lie outside the orange region in Figure \[fig:positivity\].
### Black hole entropy {#subsec:genBHmag}
For general magnetic fluxes, it is a non-trivial exercise to write the entropy as a function of the magnetic fluxes since one has to invert the algebraic equations . Therefore we adopt a different strategy. The key observation is that one can use the relation between the field theory and supergravity magnetic fluxes to show that the extremized values of the R-charges, $\bDelta _{i}$, given in , are related to the scalars, $x_i$, obtained by solving , by $$\label{deltax}
\bDelta _{i+1} = \frac{x_i}{x_1 x_2 x_3} \, .$$ This relation is the same as evaluated for real scalars and fugacities. Thus we can use and to express the entropy of the black hole in terms of the scalars, $x_i$. Then the relation allows us to compare the entropy to the topologically twisted index for the mABJM SCFT as written in . Implementing this procedure leads to the following expression for the entropy of a general magnetic black hole in our model: $$\label{SBHgenmag}
S_\text{BH} = \frac{3\sqrt{3}}{2}\frac{1}{x_1 x_2+x_2 x_3+x_3 x_1} \left(x_1 x_2 x_3-\frac{1}{x_1}-\frac{1}{x_2}-\frac{1}{x_3}\right)(\fg-1) F_{S^3} \,,$$ where we used , , and . Now, we can compare to the twisted index of the mABJM SCFT in . Indeed, using the relation , we find that the black hole entropy is equal to the topologically twisted index for all magnetic black hole solutions.
In Section \[subsec:magnexpl\], we computed explicitly the twisted index of the mABJM SCFT for a particular choice of magnetic fluxes. Let us now describe the supergravity dual to this setup. To this end we set $$\label{SU2a}
m_1 = m_2 = m_0\, \fn \, .$$ The remaining charges $m_0$ and $m_3$ are then fixed in terms of the constant $\fn$ by and . Thus, for a given choice of $\Sigma_{\fg}$, we are left with a one-parameter family of AdS$_2$ solutions. The positivity constraints imply that in order to have a regular horizon we should have $0<\fn<1/2$. Combining with implies that the scalar fields should obey $$\frac{x_2}{x_3}+\frac{x_3}{x_2} = \frac{x_1}{x_3} + \frac{x_3}{x_1} \,.$$ This equation has two solutions: $$\text{Branch 1: } x_1 x_2 = x_3^2 \, , \qquad \text{Branch 2: } x_1=x_2 \, .$$ Solving for $x_i$ in terms of $\fn$ breaks each of the branches in two more branches[^25] which we denote by the subscript $\pm$. For Branch 1$_\pm$ we find $$\begin{aligned}
\label{x1}
x_1 = \frac{1-\mathfrak{n}\pm\sqrt{(1+\mathfrak{n}) (1-3\mathfrak{n})}}{2 \mathfrak{n}^{3/2}} \, , \quad x_2 = \frac{1-\mathfrak{n}\mp\sqrt{(1+\mathfrak{n}) (1-3\mathfrak{n})}}{2 \mathfrak{n}^{3/2}} \, , \quad x_3 = \frac{1}{\sqrt{\mathfrak{n}}}\, ,\end{aligned}$$ and for Branch 2$_\pm$ we find $$\label{x2}
x_1 = x_2 = \frac{1}{\sqrt{\fn \mp \sqrt{(1+\fn)(\fn-1/3)}}} \, , \quad x_3 = \frac{2}{x_1-x_1^{-1}} \, .$$ The scalar, $\chi$, as well as the metric constants, $f_0$ and $h_0$, for each of the four branches can be determined by plugging the expressions for $x_i$ above in . Finally we have to impose that for each of the branches $\chi$ is real and $x_i$, $e^{2f_0}$ and $e^{2h_0}$ are positive. This restricts the range of the flux parameter, $\fn$, as follows $$\text{Branch 1$_\pm$:} \quad 0<\fn<\frac{1}{3} \, , \qquad \text{Branch 2$_\pm$:} \quad \frac{1}{3}<\fn<\frac{1}{2} \, .$$ Note that for each value of $\fn$, there are two corresponding near-horizon solutions. Thus fixing $\fn$ does not specify a unique black hole solution – one should additionally provide the scalar and metric functions profiles. In the IR this amounts to selecting a $\pm$ branch, while in the UV one should specify the falloff conditions on the scalar fields. In Section \[numerical\] we construct numerically the full black hole solution for Branch 1$_+$ and one can do the same for the other branches.
Having inverted , we are ready to compute the entropy for this class of near-horizon backgrounds. To do this we combine , , , , and to find $$\begin{aligned}
\label{entropySU2}
\text{Branch 1$_\pm$: } S_\text{BH} &= \frac{3\sqrt{3\fn}(1-\fn)}{2} (\fg-1) F_{S^3} \, , \\
\text{Branch 2$_\pm$: } S_\text{BH} &=\frac{3}{2\sqrt{2}} \frac{(1 - 2\fn)\left(1+3 \fn^2 \pm(1-3 \fn) \sqrt{(1+\fn) \left(\fn-\frac{1}{3}\right)}\right)}{\sqrt{(1 - 2\fn)\left(1-3 \fn^2 \mp(1-3 \fn) \sqrt{(1+\fn) \left(\fn-\frac{1}{3}\right)}\right)}} (\fg-1) F_{S^3} \, ,\end{aligned}$$ Comparing these supergravity results to the field theory computation we find that the black hole entropy and the twisted index agree perfectly.
One might wonder why did we venture into such an explicit analysis of this particular class of magnetic black hole horizons when we have already shown in Section \[sselecchar\] that the topologically twisted index matches the black hole entropy for a more general class of dyonic black holes. The key point we want to stress here is that the supersymmetric black holes are parametrized by their electric and magnetic charges and one has to carefully study the allowed values of these charges for which a regular black hole horizon exists. Unfortunately, the algebraic equations that determine the supersymmetric AdS$_2$ solutions are complicated and do not allow for an analytic solution of this problem. The example studied in this section reveals explicitly the somewhat involved branch structure of the space of regular black holes parametrized by the electric and magnetic charges. The successful comparison between the topologically twisted index and the black hole entropy in Section \[sselecchar\] was somewhat implicit and did not allow for such an insight.
The universal solution {#universal}
----------------------
There is a special type of solution to the BPS equations for which the scalars do not flow as a function of the radial coordinate. This is the supergravity dual to the *universal solution* described in Section \[ssec:twistedindex\], which arises from a topological twist purely along the superconformal R-symmetry. This black hole solution was discussed in [@Romans:1991nq; @Caldarelli:1998hg] in the context of minimal four-dimensional gauged supergravity and its universality was emphasized recently in [@ABCMZ; @BC].[^26] The near-horizon limit of this solution is part of the class of solutions described in Section \[AdS2xSigmag\] – specifically it is obtained by setting $x_i=\sqrt{3}$ and $\fn=1/3$ in . Since the scalar fields do not flow, we set them at their Warner AdS$_4$ vacuum values . Note that this is consistent with the BPS equations in . Using , , and with $\fn=1/3$ leads to the following magnetic fluxes $$\label{}
m_0=3m_1=3m_2=3m_3 = -\frac{\kappa }{2g} \,.$$ For these values of the scalar fields and charges, we find that $\cals W$ and $\cals H$ take on the following constant values: $$\cals W_{*} = -3^{3/4} \,, \qquad \cals H_{*} = -\frac{\kappa}{\sqrt{2}\,3^{3/4}g} \,.$$ To find the metric functions it is useful to trade the radial coordinate $r$ in for a new coordinate $r'$ implicitly defined by $$e^{2f}dr^2=e^{-2f}dr'^2 \,.$$ In the new radial variable, the BPS equations for the metric functions read $$\begin{aligned}
\label{BPSrho}
\frac{d f}{d r'} = \frac{g}{\sqrt{2}} \,e^{-f} \,\mathcal{W} - e^{-f-2h} \,\mathcal{H} \, , \qquad \frac{d h}{d r'} = \frac{g}{\sqrt{2}} \,e^{-f}\, \mathcal{W} + e^{-f-2h} \,\mathcal{H} \, .\end{aligned}$$ One can check that $$\label{constmot}
\mathcal{J}\equiv \frac{g}{\sqrt{2}} \, e^{-f+h}\,\mathcal{W}+ e^{-f-h} \,\mathcal{H} \, ,$$ is a constant of motion for the system , $$\label{Ih1}
{d \cals J\over d r'}\eql 0\,.$$ Moreover, using , $\mathcal{J}$ can be written as $$\label{Ih2}
\mathcal{J} = \frac{d e^{h}}{d r'} \, .$$ Combining and , we find that $e^{h}=c_1 r'+c_0$, where $c_1$ and $c_0$ are integration constants. Plugging this into the BPS equation for $h$, , and solving for $f$ gives $$e^f=\frac{1}{c_1}\Big(\frac{g}{\sqrt{2}}\,\mathcal{W}\, e^h+\mathcal{H}\, e^{-h}\Big) \, .$$ Finally, using those results in the metric, we find $$\label{universalBH}
ds^2 = -\left(\frac{\rho}{L_{\text{AdS}_4}}+\frac{\kappa L_{\text{AdS}_4}}{2\rho}\right)^2 dt'^2 + \left(\frac{\rho}{L_{\text{AdS}_4}}+\frac{\kappa L_{\text{AdS}_4}}{2\rho}\right)^{-2} d\rho^2 + \rho^2 ds_{\Sigma_\fg}^2 \,,$$ where $L_{\text{AdS}_4}$ is the scale of the Warner AdS$_4$ vacuum defined in , $\rho \equiv c_1 r' +c_0$ and $t'\equiv t/c_1$.
For $\kappa=0$, i.e. $\fg=1$, the gauge field vanishes and the solution is simply AdS$_4$ in Poincaré coordinates. For $\kappa=+1$, i.e. $\fg=0$, the metric has a naked singularity at $\rho=0$. For $\kappa=-1$, we find a hyperbolic black hole with a regular horizon at $\rho_0=L_{{\rm AdS}_4}/\sqrt{2}$. The metric is normalized such that in the UV, i.e. $\rho \rightarrow \infty$, it approaches AdS$_4$ with a hyperbolic boundary and radius $L_{\text{AdS}_4}$. In the IR, i.e. $\rho \rightarrow \rho_0$, the metric approaches $$ds^2 = \frac{L_{\text{AdS}_4}^2}{4} \left( ds^2_{\text{AdS}_2} + 2 ds^2_{\Sigma_\fg} \right) \, .$$ This near-horizon solution is part of the larger class of AdS$_2\times\Sigma_\fg$ solutions described in Section \[AdS2xSigmag\]. The entropy for the hyperbolic black hole can thus be obtained from by setting $\fn=1/3$. Then one finds $$\label{SugraUS}
S_\text{BH} = (\fg-1) F_{S^3}^{{\rm mABJM}} \, ,$$ where $F_{S^3}^{{\rm mABJM}}$, the free energy for the AdS$_4$ Warner vacuum with $S^3$ as asymptotic boundary, is given in . Thus we find an exact match between the universal results: from supergravity and from field theory.
Numerical black hole solutions {#numerical}
------------------------------
In this section we present a numerical analysis of the BPS equations . We will do so for the choice of magnetic fluxes $m_1=m_2=m_0 \fn$ discussed in Section \[subsec:genBHmag\] and furthermore restrict to a scalar field profile that corresponds to Branch 1$_+$ in . There is no obstruction for repeating the same analysis for general choices of magnetic fluxes and branches.
We find it useful to define $p\equiv f+h$ and use it as the new radial variable. Taking the sum of the BPS equations for $f$ and $h$ in allows us to write $$\frac{d}{d r} = \frac{d p}{d r} \frac{d }{d p} = \sqrt{2}\, e^f g\, \mathcal{W} \frac{d}{d p} \, .$$ The BPS equations in terms of $p$ reduce to the following system of five first order non-linear ODEs for the functions $h(p)$, $\chi(p)$ and $\lambda_i(p)$: $$\begin{split}
\frac{d h}{d p} &= \frac{1}{2}\left(1 +\sqrt{2} \frac{ \cals H}{g \cals W} e^{-2h}\right) \, ,\\[6 pt]
\frac{\partial \chi}{\partial p} &= -\frac{1}{2 \mathcal{W}} \frac{\partial \mathcal{W}}{\partial \chi} \, , \\[6 pt]
\frac{\partial \lambda_i}{\partial p} &= -\frac{1}{2g\mathcal{W}} \left(g\frac{\partial \mathcal{W}}{\partial \lambda_i} + \sqrt{2}e^{-2h} \frac{\partial \mathcal{H}}{\partial \lambda_i} \right) \, .
\end{split}$$ Once these equations are solved, the function $f(p)$ can be found using the identity $f=p-h(p)$.
In order to perform the numerical analysis, it is most convenient to specify the boundary conditions in the IR at the AdS$_2\times \Sigma_{\fg}$ horizon. This is where we choose to restrict ourselves to Branch 1$_+$ by taking the following IR boundary conditions: $$\begin{aligned}
\label{IRcond}
\lambda_1^\text{IR} &= \frac{1}{2}\log\left(\frac{1-\mathfrak{n}+\sqrt{(1+\mathfrak{n}) (1-3\mathfrak{n})}}{2 \mathfrak{n}^{3/2}}\right) \, , & \lambda_3^\text{IR} &= \frac{1}{2}\log\left(\frac{1}{\sqrt{\mathfrak{n}}}\right) \, , \\[6 pt] \label{IRcondd}
\lambda_2^\text{IR} &= \frac{1}{2}\log\left(\frac{1-\mathfrak{n}-\sqrt{(1+\mathfrak{n}) (1-3\mathfrak{n})}}{2 \mathfrak{n}^{3/2}}\right) \, , & \chi^\text{IR} &= \textrm{arctanh}{\sqrt{\fn}} \, , \\[6 pt]
h^\text{IR} &= \frac{1}{2}\log\left(\frac{\sqrt{\fn}(1-\fn)}{2g^2}\right) \,,\end{aligned}$$ where we used , (with $\pm \rightarrow+$) and . We note that in the radial coordinate $p$ the IR AdS$_2$ region is at $p\to -\infty$ and the UV AdS$_4$ is at $p\to \infty$. We use a numerical implementation in Mathematica by starting with these initial conditions in the IR and numerically integrating towards the UV. To move away from the near-horizon solution we perturb the scalar fields slightly from their IR values. However, arbitrary perturbations will generally result in singular solutions.
To find the allowed perturbations that produce regular asymptotically AdS$_4$ solutions let us define the fields $\phi_n$ as $$\label{initial}
\lambda_i = \lambda_i^{\text{IR}}+\phi_{\lambda_i}\, , \quad \chi = \chi^{\text{IR}} + \phi_\chi \, , \quad h = h^\text{IR} + \phi_h \, ,$$ and expand the BPS equation to first order in $\phi_n$. This produces the following set of linear equations $$\frac{\partial \phi_n}{\partial p} = M_{nm} \phi_m \, , \quad m,n \in \{\lambda_1,\lambda_2,\lambda_3,\chi,h\} \, ,$$ where $M_{nm}$ is a matrix that depends on the IR values of the scalar and metric fields. Negative eigenvalues of the matrix $M_{nm}$ correspond to directions in field space which lead to singular solutions. Therefore we have to choose the deformation in in the direction along the positive eigenvalues. The matrix $M_{mn}$ always has two negative and three positive eigenvalues in the allowed range of $\fn$, i.e. $0<\fn<1/3$, as shown in Figure \[fig:Meig\].
![Eigenvalues, $\Lambda_m$, of $M_{nm}$ as a function of $\mathfrak{n}$.[]{data-label="fig:Meig"}](meig){width=".5\textwidth"}
Note that two eigenvalues degenerate at $\mathfrak{n}=\sqrt{5}-2$. Since this is an irrational number it cannot correspond to a properly quantized magnetic flux and we do not discuss it further. This leads us to set up the initial conditions as with $$\phi_n = \epsilon_{(1)} \varphi_{n}^{(1)} + \epsilon_{(2)} \varphi_{n}^{(2)} +\epsilon_{(3)} \varphi_{n}^{(3)} \, ,$$ where $\epsilon_{(1)}$, $\epsilon_{(2)}$, $\epsilon_{(3)}$, are small parameters, i.e. $\epsilon_{(a)}\ll1$ for $a=1,2,3$, and $\varphi^{(1)}_n$, $\varphi^{(2)}_n$, $\varphi^{(3)}_n$, are orthonormal eigenvectors corresponding to the three positive eigenvalues.
Figures \[fig:scalar12\] and \[fig:chilambda12\] showcase the numerical analysis for $\mathfrak{n}=1/4$, $|\epsilon_{(a)}| \approx 10^{-4}$, the IR value of the radial coordinate $p_{\rm IR}=-10$ and the UV value at around $p_{\rm UV}=28$. The Warner point is numerically unstable, but by finely tuning $\epsilon_{(a)}$, one can get very close to the Warner fixed point as indicated in Figure \[fig:chilambda12\] and Figure \[fig:scalar12\]. Figure \[fig:scalar12\] shows that the scalar fields take on the Warner values and stay there longer as we tune more finely towards the Warner AdS$_4$ vacuum. Eventually however, the flow will always move back to the ${\rm SO(8)}$ AdS$_4$ vacuum which is numerically stable. As visible from Figure \[fig:chilambda12\], the solutions shown in Figure \[fig:scalar12\] lie very close to the Warner fixed point. Rather than keeping $\mathfrak{n}$ fixed and taking different initial conditions, we can also vary $\fn$ to produce Figure \[fig:chilambdan\].
![Examples of scalar profiles for $\fn=1/4$ fine tuned to approach the Warner AdS$_4$ fixed point. The dashed lines correspond to the fixed point values for the scalars given in .[]{data-label="fig:scalar12"}](scalar12){width=".7\textwidth"}
![Examples of scalar profiles for $\fn=1/4$ projected onto the $\chi,\lambda_1$ plane. The parameters correspond to those of Figure \[fig:scalar12\].[]{data-label="fig:chilambda12"}](chilambda12){width=".6\textwidth"}
For the sake of brevity of the presentation here we focused on a detailed analysis of the numerical solutions that asymptote to the $1_+$ branch. However, we have also found similar numerical solutions for the $1_{-}$ and $2_{\pm}$ branches discussed in Section \[subsec:genBHmag\]. In fact, the solutions for the $1_{-}$ branch are identical to the ones for the $1_{+}$ branch upon an interchange of the scalar fields $\lambda_1$ and $\lambda_2$.
![Examples of scalar profiles for various values of $\fn$ ranging from $\fn=\frac{1}{3}$ to $\fn=\frac{1}{30}$, projected onto the $\chi,\lambda_1$ plane. The dashed line corresponds to $\chi^\text{IR}(\lambda_1^\text{IR})$ using and .[]{data-label="fig:chilambdan"}](chilambdanlegend){width=".8\textwidth"}
It is clear from Figure \[fig:chilambda12\] that for a fixed AdS$_2\times \Sigma_\fg$ near horizon geometry there is a one real parameter family of black hole solutions that asymptote to the ${\rm SO(8)}$ AdS$_4$ vacuum of the four-dimensional supergravity. Only a single member of that family of solutions, illustrated by the red curve in Figure \[fig:chilambda12\], asymptotes to the Warner AdS$_4$ vacuum. This may be viewed as a violation of black hole uniqueness and is a feature absent in the known asymptotically black holes solutions of four-dimensional $\mathcal{N}=2$ gauged supergravity without hypermultiplets. We believe that the violation of black hole uniqueness is due to the presence of the hypermultiplet scalar field that has a non-trivial profile in the black hole solution and is charged under the ${\rm U(1)}_{m}$ gauge field.
A different perspective on this continuous family of solutions is offered by the dual holographic QFT. The ${\rm SO(8)}$ AdS$_4$ vacuum is dual to the ABJM SCFT. In the field theory setup discussed in Section \[sec:CFT\] we are deforming this SCFT in two ways. We turn on the superpotential mass term and in addition we perform the partial topological twist described in Section \[ssec:twistedindex\]. These two deformations are relevant and are thus associated with dimensionful parameters, namely the mass, $m$, and the length scale, $\ell_{\Sigma}$, of the Riemann surface, $\Sigma_\fg$.[^27] Therefore we should expect a one-parameter family of RG flows emerging from the ABJM SCFT in the UV labelled by the dimensionless parameter $m\ell_{\Sigma}$. The one-parameter family of supersymmetric black holes we have constructed is the holographic dual realization of this family of RG flows. A very similar picture was presented in [@Bobev:2014jva] for a deformation of the four-dimensional $\mathcal{N}=4$ SYM, which is a combination of a superpotential mass term and a partial topological twist.
Conclusions {#sec:Conclusions}
===========
In this paper we studied the topologically twisted index of the mABJM SCFT which can be thought of as an interacting IR fixed point arising from the ABJM theory deformed by an $\mathcal{N}=2$ preserving mass term. We exploited the gravitational dual of the mABJM SCFT to construct static supersymmetric black hole solutions of the maximal four-dimensional $\rm SO(8)$ $\mathcal{N}=8$ gauged supergravity. In addition, we showed explicitly that the planar limit of the topologically twisted index is equal to the entropy of these black holes to leading order in $N$. Our results can be viewed as an extension of the results and conjectures in [@Benini:2015eyy; @Benini:2016rke] which employed the topologically twisted index of the ABJM theory to account for the entropy of supersymmetric asymptotically AdS$_4$ black holes in the STU model of four-dimensional $\mathcal{N}=2$ gauged supergravity. A distinct feature of our new supergravity solutions is that they can be viewed as solutions to a particular four-dimensional $\mathcal{N}=2$ gauged supergravity coupled to three vector multiplets and one hyper multiplet. Thus our solutions constitute rare examples of supersymmetric asymptotically AdS$_4$ black holes with non-trivial profiles for hyper multiplet scalars. Our work opens several interesting avenues to explore.
We have provided an exhaustive classification of the supersymmetric dyonic AdS$_2$ solutions in our supergravity model. However we have discussed fully-fledged black holes only when the electric charges vanish. Given that we have numerous dyonic AdS$_2$ backgrounds with entropy that matches exactly the topologically twisted index of the dual SCFT, it is natural to expect that one can construct full black hole solutions that interpolate between the AdS$_4$ Warner vacuum and these near-horizon AdS$_2$ geometries. Due to the complexity of the BPS equations in our supergravity model, it is likely that such black hole solutions can only be constructed numerically.
In our analysis we focused on static solutions of the BPS equations. It is known that there are also rotating supersymmetric black holes asymptotic to AdS$_4$ (see, e.g., [@Cacciatori:2009iz]). It is natural to conjecture that our supergravity truncation also contains similar solutions and it would be very interesting to construct them explicitly and to understand their entropy microscopically.
Finally, it should be possible to construct non-supersymmetric black hole solutions in our supergravity theory. A large class of such solutions was found in [@Klemm:2012yg; @Chow:2013gba] in the STU model. One might hope that similar methods can be applied to our truncation, although the presence of hypermultiplet scalars may complicate the construction.
It would be interesting to uplift our black hole solutions to M-theory by combining the uplifts of the STU-model [@Cvetic:1999xp; @Azizi:2016noi] with the uplift of the Warner critical point [@Corrado:2001nv]. Such M-theory backgrounds may further elucidate the structure of supersymmetric wrapped M2-branes and potentially allow for constructing generalizations of our solutions. A particular generalization may proceed as follows. It was pointed out in [@Corrado:2001nv] that the CPW AdS$_4$ vacuum of eleven-dimensional supergravity can be generalized by changing the topology of the internal squashed $S^7$. The SU(3) invariance of the Warner AdS$_4$ vacuum is realized in eleven dimensions by an explicit $\mathbb{CP}^2$ submanifold of $S^7$. It was observed in [@Corrado:2001nv] that one can substitute this four-manifold with any other Kähler-Einstein four-manifold $\mathcal{M}_4$ while still preserving $\mathcal{N}=2$ supersymmetry. If this manifold has at least one ${\rm U(1)}$ isometry some of our black hole solutions can probably be generalized. We expect that $\mathcal{M}_4 = \mathbb{CP}^1 \times \mathbb{CP}^1$ should be the simplest generalization to study.[^28] The universal flow solution discussed in Section \[universal\] should be easy to find for any $\mathcal{M}_4$.
Finally, we would like to point out that our consistent truncation of the four-dimensional $\mathcal{N}=8$ gauged supergravity may find other applications in the context of holography. One potentially fruitful avenue for further study is to look for Euclidean solutions that are asymptotic to the Warner AdS$_4$ vacuum with an $S^3$ boundary and have non-trivial scalar profiles. These backgrounds should be generalizations of the solutions discussed in [@Freedman:2013ryh] and can be viewed as a holographic analog of the $F$-maximization procedure applied to the mABJM SCFT. More precisely, these putative supergravity solutions should describe massive deformations of the mABJM $\mathcal{N}=2$ SCFT on $S^3$ that break conformal invariance but preserve supersymmetry. It would be interesting to construct these backgrounds explicitly.
We are grateful to P. Marcos Crichigno, Fririk Gautason, Alessandra Gnecchi, Adolfo Guarino, Kiril Hristov, and Alberto Zaffaroni for interesting discussions. The work of NB is supported in part by an Odysseus grant G0F9516N from the FWO. The work of VSM is supported by a doctoral fellowship from the Fund for Scientific Research - Flanders (FWO) and in part by the European Research Council grant no. ERC-2013-CoG 616732 HoloQosmos. NB and VSM are also supported by the KU Lueven C1 grant ZKD1118 C16/16/005, by the Belgian Federal Science Policy Office through the Inter-University Attraction Pole P7/37, and by the COST Action MP1210 The String Theory Universe. KP was supported in part by DOE grant DE-SC0011687. KP would like to thank the Instituut voor Theoretische Fysica, KU Leuven for hospitality and support during the initial and final stages of this project.
Conventions {#appconv}
===========
Throughout this paper we consider smooth Riemann surfaces $\Sigma_{\fg}$ of genus $\fg$. We put a constant curvature metric on $\Sigma_{\fg}$ of the form $$\label{hdef}
ds^{2}_{\Sigma_{\fg}}=H(x,y)^2\,(dx^{2}+dy^{2})\,,\qquad {H(x,y)} = \begin{cases}\displaystyle \frac{2}{1 + x^2 +y^2} & \text{for } \fg = 0 \,,\\ \sqrt{2\pi} & \text{for } \fg = 1\,, \\ \displaystyle{1\over y} & \text{for } \fg >1\;.\end{cases}$$ The volume form $$\label{volSigma}
{\rm vol}_{\Sigma_{\fg}}\equiv H^2\,dx\wedge dy\,,$$ integrates to: $$\label{normtg}
\int_{\Sigma_\fg}{\rm vol}_{\Sigma_{\fg}}=2\pi\eta_{\Sigma} \,, \qquad \eta_{\Sigma} =\begin{cases}
2|\fg-1| &\qquad\text{for $\fg\neq 1$}\,,\\
1 &\qquad\text{for $\fg= 1$}\,.
\end{cases}$$ The normalized curvature of $\Sigma_{\fg}$ is denoted by $\kappa=1,0$ and $-1$ for $\fg=0$, $\fg=1$, and $\fg>1$, respectively. We also use the following locally defined potential, $\omega_{\Sigma_\fg}$, for the volume form: $$\label{potSigma}
\omega_{\Sigma_\fg}\eql \begin{cases}\displaystyle
{2(xdy-ydx)\over 1+x^2+y^2} & \text{for~$\fg=0$}\,,\\
\pi(xdy-ydx) & \text{for~$\fg=1$}\,,\\\displaystyle
{dx\over y} & \text{for~$\fg>1$}\,,
\end{cases}\qquad \quad d\omega_{\Sigma_\fg}={\rm vol}_{\Sigma_{\fg}} \,.$$ The AdS$_2$ metric is $$\label{}
ds^2_\text{AdS$_2$}\eql {1\over r^2}({-dt^2+dr^2})\,,$$ with the volume form and the potential given by $$\label{volAdS}
\text{vol}_\text{AdS$_2$}\eql{dt\wedge dr\over r^2}\,,\qquad \omega_\text{AdS$_2$}\eql {dt\over r}\,.$$ To conform with the prevailing custom, we use different index conventions in the field theory and the supergravity sections of the paper. In Section \[sec:CFT\], the indices labelling the R-charges and fugacities are $\alpha,\beta =1,\ldots,4$ and $i,j=2,3,4$, while in Sections \[sec:Sugra\] and \[sec:magneticBH\] the range of the same indices now labelling the dual scalar fields and the vector potentials is $\alpha,\beta=0,\ldots,3$ and $i,j=1,2,3$. In other words, $$\label{}
\alpha^\text{QFT}\eql \alpha^\text{SUGRA}+1\,,\qquad i^\text{QFT}\eql i^\text{SUGRA}+1\,,$$ and so on.
The U(1)$^2$-invariant truncation {#appendixA}
=================================
In this appendix we present a $\rm U(1)^2$-invariant truncation of the de Wit-Nicolai $\cals N=8$ gauged supergravity in four dimensions [@deWit:1982bul], where $\rm U(1)^2$ is the Cartan subgroup of the standard $\rm SU(3)\subset SO(6)\subset SO(8)$. On the one side, this truncation can be viewed as a generalization of the $\rm SU(3)$-invariant truncation originally studied in [@Warner:1983vz] and recently, in more detail, in [@Bobev:2010ib]. On the other side, it generalizes the $\rm U(1)^4$-invariant truncation, where $\rm U(1)^4$ is the Cartan subgroup of $\rm SO(8)$, to the STU-model [@Behrndt:1996hu; @Duff:1999gh; @Cvetic:1999xp]. The resulting theory is a matter coupled $\cals N=2$ gauged supergravity specified by the geometric data that naturally arise from the two simpler truncations. To determine those data, we use the same method as in Appendix B of [@Bobev:2010ib], that is we compare judiciously chosen terms in supersymmetry variations and in the action of the truncated $\cals N=8$ theory with those in $\cals N=2$ supergravity. We work here with the original formulation of gauged $\cals N=2$ supergravity as given in [@Andrianopoli:1996cm; @Andrianopoli:1996vr].
Let $T_{12}$, $T_{34}$, $T_{56}$ and $T_{78}$ denote the four standard Cartan generators of $\rm SO(8)$, where $T_{ij}$ is the generator of rotation in the $(ij)$-plane with charge one. Then the two symmetry generators are $$\label{Csu3gen}
{1\over 2}(T_{12}-T_{34})\,,\qquad {1\over \sqrt 3}(T_{12}+T_{34}-2\,T_{56})\,,$$ under which the 8 gravitini, $\psi_\mu{}^i$, and the corresponding supersymmetries, $\epsilon^i$, of the full theory transform with the charges $$\label{decom8}
\bfs 8_v\quad \longrightarrow\quad (1,\coeff 1 {\sqrt{3}})+ (1,-\coeff 1 {\sqrt{3}})+ (-1,\coeff 1 {\sqrt{3}})+ (-1,-\coeff 1 {\sqrt{3}})+(0,\coeff 2 {\sqrt 3})+(0,-\coeff 2 {\sqrt 3})+(0,0)+(0,0)\,.$$ The two invariant gravitini and supersymmetries are, respectively, the chiral $\psi^{7,8}$ and $\epsilon^{7,8}$ and their complex conjugates $\psi_{\,7,8}$ and $\epsilon_{7,8}$ of opposite chirality.
The unbroken gauge symmetry is given by the commutant of the generators in $\rm SO(8)$. Clearly, it is the Cartan subgroup, $\rm U(1)^4$, of $\rm SO(8)$. The corresponding gauge fields, $A^{ij}$, are the same as in the STU-model and consist of the graviphoton and three gauge fields in vector multiplets. We find it convenient to work in the same symplectic frame as in [@Duff:1999gh], which is specified by the following canonical gauge fields, $A^\alpha$, $\alpha=0,\ldots,3$, $$\label{Avfields}
\begin{split}
A^{12} & \eql {1\over 2}(A^0+A^1-A^2-A^3)\,,\qquad
A^{34} \eql {1\over 2}(A^0-A^1+A^2-A^3)\,,\\
A^{56} & \eql {1\over 2}(A^0-A^1-A^2+A^3)\,,\qquad
A^{78} \eql {1\over 2}(A^0+A^1+A^2+A^3)\,.
\end{split}$$ In the symmetric gauge, the scalar 56-bein of the $\cals N=8$ supergravity is given by $$\label{56bein}
\cals V\eeql \left(\begin{matrix}
u_{ij}{}^{IJ} & v_{ijKL} \\ v^{klIJ} & u^{kl}{}_{KL}
\end{matrix}\right)\eql \exp \left(\begin{matrix}
0 & \phi_{ijkl}\\
\phi^{ijkl} & 0
\end{matrix}\right)\in {\rm E_{7(7)}/SU(8)}\,,$$ where $$\label{selfphi}
\phi_{ijkl}\eql {1\over 24}\epsilon_{ijklmnpq}\phi^{mnpq}\,,\qquad \phi^{ijkl}\eql(\phi_{ijkl})^*\,,$$ are completely antisymmetric complex self-dual scalar fields. We find that the $\rm U(1)^2$-invariant nonvanishing $\phi_{ijkl}$’s are given by $$\label{}
\begin{split}
\phi_{1278}& \eql -{1\over 2}\, \lambda_1\, e^{\i\,\varphi_1}\,,\qquad \phi_{3478}\eql -{1\over 2} \,\lambda_2\, e^{\i\,\varphi_2}\,,\qquad \phi_{5678}\eql -{1\over 2}\, \lambda_3 \,e^{\i\,\varphi_3}\,,\\[3 pt]
\phi_{1234} & \eql -{1\over 2}\, \lambda_3 \,e^{-\i\,\varphi_3}\,,\qquad \phi_{1256}\eql -{1\over 2}\,\lambda_2\, e^{-\i\,\varphi_2}\,,\qquad \phi_{3456}\eql -{1\over 2}\, \lambda_1\, e^{-\i\,\varphi_1}\,,\\[3 pt]
\phi_{1357} & \eql -\phi_{1467}\eql -\phi_{2367}\eql -\phi_{2457}\eql {1\over 4}(\chi_1\cos\psi_1+\i\,\chi_2\sin\psi_2)\,,\\[3 pt]
\phi_{1367} & \eql \phi_{1457}\eql \phi_{2357}\eql -\phi_{2467}\eql -{1\over 4}(\chi_1\sin\psi_1-\i\,\chi_2\cos\psi_2)\,,\\[3 pt]
\phi_{1368}& \eql \phi_{1458}\eql \phi_{2358}\eql -\phi_{2468}\eql -{1\over 4}(\chi_1\cos\psi_1-\i\,\chi_2\sin\psi_2)\,,\\[3 pt]
\phi_{1358}& \eql -\phi_{1468}\eql -\phi_{2368}\eql -\phi_{2458}\eql -{1\over 4}(\chi_1\sin\psi_1+\i\,\chi_2\cos\psi_2)\,,
\end{split}$$ where the five complex fields, $$\label{thezs}
z_i\eql {\lambda_i\tanh|\lambda_i|\over |\lambda_i|}\,e^{\i\,\varphi_i}\,,\qquad i=1,2,3\,,$$ and $$\label{thezetas}
\zeta_r\eql {\chi_r\tanh\sqrt{\chi_1^2+\chi_2^2}\over \sqrt{\chi_1^2+\chi_2^2}}\,e^{\i\,\psi_r}\,,\qquad r=1,2\,,$$ parametrize the special Kähler manifold, $\cals M_{V}$, of three vector multiplets and the quaternionic Kähler manifold, $ \cals M_{H}$, of the universal hypermultiplet, respectively. The two Kähler manifolds are $$\label{}
\cals M_{V}\times \cals M_{H} \eql \left[{\rm SU(1,1)\over U(1)}\right]^3\times {\rm SU(2,1)\over SU(2)\times U(1)}\,,$$ with the standard metrics $$\label{}
g_{z_i\bar z_j}dz_id\bar z_j\eql {dz_1d\bar z_1\over (1-|z_1|^2)^2}+{dz_2d\bar z_2\over (1-|z_2|^2)^2}+{dz_3d\bar z_3\over (1-|z_3|^2)^2}\,,$$ and $$\label{}
g_{\zeta_i\bar\zeta_j}d\zeta_id\bar\zeta_j\eql {d\zeta_1d\bar\zeta_1+d\zeta_2d\bar\zeta_2\over 1-|\zeta_1|^2-|\zeta_2|^2}+{(\zeta_1 d\bar\zeta_1+\zeta_2 d\bar\zeta_2)(\bar\zeta_1d\zeta_1+\bar\zeta_2 d\zeta_2)\over (1-|\zeta_1|^2-|\zeta_2|^2)^2}\,,$$ and the corresponding Kähler potentials $$\begin{aligned}
\label{KvP}
K_{V} & \eql -\log\big[ (1-|z_1|^2)(1-|z_2|^2)(1-|z_3|^2)\big]\,,\\[3 pt]
K_{H} & \eql -\log(1-|\zeta_1|^2-|\zeta_2|^2)\,.\label{KhP}\end{aligned}$$ Note that $\cals M_V$ is invariant under, $\rm U(1)^4$, and is the scalar manifold of the STU-model. In turn, by construction, $\cals M_H$ is invariant under the symmetries and in fact under the full $\rm SU(3)$. It is thus the same as the hypermultiplet in the $\rm SU(3)$-invariant truncation [@Bobev:2010ib].
As expected, there are eight invariant chiral spin-1/2 fields: $$\label{spin12v}
\chi^{127}\,,\quad \chi^{128}\,,\quad \chi^{347}\,,\quad \chi^{348}\,,\quad \chi^{567}\,,\quad \chi^{568}\,,$$ $$\label{spin12h}
\chi^{135}\eql -\chi^{146}\eql -\chi^{236}\eql -\chi^{245}\,,\qquad
\chi^{246}\eql -\chi^{136}\eql -\chi^{145}\eql -\chi^{235}\,,$$ and their complex conjugates. By examining the supersymmetry variations of the scalar fields,[^29] one can check that belong to the three vector multiplets, while to the hypermultiplet.
Similarly, from the supersymmetry variations of the vector fields into spin-3/2 fields, we read off the symplectic sections $$\label{defofLa}
L^\alpha \eql e^{K_V/2} X^\alpha\,,\qquad \alpha=0,\ldots,3\,,$$ where the holomorphic sections, $X^\alpha$, are explicitly given by $$\label{theXs}
\begin{split}
X^0 & \eql {1\over 2\sqrt 2}\,(1-z_1)(1-z_2)(1-z_3)\,,\qquad
X^1 \eql {1\over 2\sqrt 2}\,(1-z_1)(1+z_2)(1+z_3)\,,\\
X^2 & \eql {1\over 2\sqrt 2}\,(1+z_1)(1-z_2)(1+z_3)\,,\qquad
X^3 \eql {1\over 2\sqrt 2}\,(1+z_1)(1+z_2)(1-z_3)\,,\\
\end{split}$$ and $K_V$ is the Kähler potential . The specific normalization in is fixed by imposing identities of the special Kähler geometry and by matching terms in the $\cals N=8$ and $\cals N=2$ actions.
As a consistency check we verify that the vectors $f_{z_j}{}^\alpha$, that follow from the supersymmetry variations of the vector fields into spin-1/2 fields, are indeed given by $$\label{thefs}
f_{z_j}{}^\alpha\equiv D_{z_j}L^\alpha\eql \big(\partial_{z_j}+{1\over 2}\partial_{z_j}K_V\big)\,L^\alpha,$$ where the derivative on the right hand side is the usual Kähler covariant derivative.
The prepotential in the STU-model as a function of the holomorphic sections is $$\label{prep}
F\eql -2\,\i \sqrt{X^0X^1X^2X^3}\,,$$ It is determined by solving $$\label{derprep}
F\eql {1\over 2}X^\alpha F_\alpha\,,\qquad F_\alpha\eql {\partial F\over\partial X^\alpha}\,,$$ and requiring that $F$ be homogenous of degree two as a function of $X^\alpha$’s. The overall normalization can be verified by the matching of Maxwell actions (see, below), in which the coupling of the scalars, $z_i$, to the gauge fields is given by the second derivatives of the prepotential, $$\label{twoderprep}
F_{\alpha\beta}\equiv {\partial^2F\over \partial X^\alpha\partial X^\beta}\,.$$ One should note that a priori the prepotential and its derivatives and have a sign ambiguity, in particular when evaluated as functions of the scalars, $z_i$. That ambiguity is removed by setting $$\label{}
\sqrt{X^0X^1X^2X^3}\eql {1\over 8}(1-z_1^2)(1-z_2^2)(1-z_3^2)\,,$$ which follows from the corresponding $\cals N=8$ expressions.
The gauging in the $\cals N=2$ supergravity is determined by the action of the gauge symmetries on the scalar manifolds. As we have already noted above, there are just two $\rm U(1)$’s that act nontrivially on $\cals M_H$, and hence the gauging is the same as in the $\rm SU(3)$-invariant truncation. By comparing with (2.38) and (2.40) in [@Bobev:2010ib], we find the Killing vectors, $K_\alpha$, corresponding to the gauge fields $A^\alpha$ in , $$\label{}
\begin{split}
K_0 & \eql 2\,\i\, \zeta_1\partial_{\zeta_1}+\i\,\zeta_2\partial_{\zeta_2}+\text{c.c.}\,,\\[6 pt]
K_1 & \eql K_2\eql K_3\eql -\i\,\zeta_2\partial_{\zeta_2}+\text{c.c.}\,.
\end{split}$$ The corresponding moment maps, $P_\alpha=(P_\alpha ^1,P_\alpha ^2,P_\alpha ^3)$, can be read-off from (B.39) and (B.40) in [@Bobev:2010ib] and become quite simple if one sets one of the hyperscalars to zero. In particular, for $\zeta_1=0$ and $\zeta_2=z$, we have $$\label{}
P^\alpha\eql P^2_\alpha\eql 0\,,\qquad P^3_0\eql -{1-2 \,|z|^2\over 1-|z|^2}\,,\qquad P^3_1\eql P^3_2\eql P^3_3\eql -{1\over 1-|z|^2}\,.$$ This completes the list of geometric data for the $\cals N=2$ supergravity that arises from this truncation.
As a consistency check we verify explicitly that the $\cals N=8$ bosonic action reduces to the canonical $\cals N=2$ action for the invariant fields. The latter reads $$\label{N2act}
\begin{split}
e^{-1}\cals L_{\cals N=2}& \eql {1\over 2}\,R-g_{z_i\bar z_j}\partial_\mu z_i\partial^\mu\bar z_j-g_{\zeta_i\bar\zeta_j}\nabla_\mu \zeta_i\nabla^\mu\bar\zeta_j\\[6 pt]
& \quad +{1\over 4}\Big(\ \cals I_{\alpha\beta}F_{\mu\nu}^\alpha F^{\beta\,\mu\nu}-\cals R_{\alpha\beta}F_{\mu\nu}^\alpha \widetilde F^{\beta\,\mu\nu}\Big)-g^2\,\cals P\,,
\end{split}$$ where $$\label{}
\nabla_\mu \zeta_i\eql \partial_\mu \zeta_i+g A_\mu^\alpha K_\alpha{}^{\zeta_i}\,,$$ is the covariant derivative of the scalar fields, $\cals R_{\alpha\beta}$ and $\cals I_{\alpha\beta}$ are, respectively, the real and imaginary part of the matrix $$\label{}
\cals N_{\alpha\beta}\eql \overline F_{\alpha\beta}+2 \,\i\,{(\Im F_{\alpha\gamma})(\Im F_{\beta\delta})X^\gamma X^\delta\over (\Im F_{\gamma\delta})X^\gamma X^\delta}\,,\qquad \cals N_{\alpha\beta}\eql \cals R_{\alpha\beta}+\i\,\cals I_{\alpha\beta}\,,$$ and[^30] $$\label{}
\cals P \eql 4\, g_{ab} K_\alpha{}^{a}K_\beta{}^{b}\overline L^\alpha L^\beta+g^{z_i\bar z_j}f_{z_i}{}^\alpha f_{\bar z_j}{}^\beta P_\alpha\cdot P_\beta-3\, \overline L^\alpha L^\beta P_\alpha\cdot P_\beta\,.$$ is the scalar potential.
For the vector fields with constant curvatures as in , the Maxwell equations reduce to the following system of algebraic equations: $$\label{Maxabs}
\begin{split}
-{(1+|\zeta_1|^2)|\zeta_2|^2\over (1-|\zeta_1|^2-|\zeta_2|^2)^2}\,m_0+{(1-|\zeta_1|^2)|\zeta_2|^2\over(1-|\zeta_1|^2-|\zeta_2|^2)^2}\,(m_1+m_2+m_3) & \eql 0\,,\\[6 pt]
{|\zeta_1|^2 (4-|\zeta_2|^2)+|\zeta_2|^2\over (1-|\zeta_1|^2-|\zeta_2|^2)^2}\,m_0-{(1+|\zeta_1|^2)|\zeta_2|^2\over (1-|\zeta_1|^2-|\zeta_2|^2)^2}\,(m_1+m_2+m_3) & \eql 0\,,
\end{split}$$ with the same equations for the electric parameters, $e_\alpha$.
Derivation of the near horizon BPS equations {#appendixB}
============================================
C.1. The general set-up {#c.1.the-general-set-up .unnumbered}
-----------------------
In this appendix we outline the main steps of the truncation of the fermion supersymmetry variations in $\mathcal{N}=8$ gauged supergravity [@deWit:1982bul] to the ${\rm U}(1)^2$-invariant sector, and the derivation of the BPS equations for the dyonic black holes used in Section \[subsec:dyonicBPS\]. The main difference with the similar truncations discussed previously, such as the STU-model in [@Duff:1999gh] or the $\SU(3)\times {\rm U}(1)^\text{W}$-invariant truncation in [@Ahn:2000aq] or [@Corrado:2001nv], is the presence of a nontrivial electric field, which precludes futher truncation to real scalar fields. At the same time, we simplify our calculation by restricting to the ${\rm U(1)}^3$-invariant bosonic fields of the ${\rm AdS}_2\times\Sigma_\fg$ near horizon black hole Ansatz in Section \[sec:ansatz\]. Hence we use the metric with constants $f_0$ and $h_0$, constant scalars, $z_i$ and $z$, $i=1,2,3$, and constant electric and magnetic fluxes .
The spin-1/2 and spin-3/2 supersymmetry variations of the $\cals N=8$ theory are given by: $$\begin{aligned}
\label{spin12}
\delta\chi^{ijk}& \eql -\cals A_\mu{}^{ijkl}\gamma^\mu\epsilon_l+{3\over 2}\gamma^{\mu\nu}\overline F{}_{\mu\nu}^{-[ij}\epsilon^{k]}
-2g A_{2l}{}^{ijk}\epsilon^l\,,\\[6 pt]
\delta\psi_\mu{}^i & \eql 2D_\mu\epsilon^i+{1\over 4}\sqrt 2 \,\overline{ F}{}^-_{\nu\rho}{}^{ij}\gamma^{\nu\rho}\gamma_\mu\epsilon_j+\sqrt 2 g A_1{}^{ij}\gamma_\mu\epsilon_j\,,\label{spin32}\end{aligned}$$ and their complex conjugates. We refer the reader to [@deWit:1982bul] for the definitions and explicit formulae for the covariantized scalar kinetic tensor, $\cals A_\mu{}^{ijkl}$, and the scalar $A$-tensors, $A_1{}^{ij}$ and $A_{2l}{}^{ijk}$, which are constructed from the scalar 56-bein .
The gauge fields enter the variations and both through the “bare” potential, $A^{IJ}$, in the scalar kinetic tensor and the covariant derivative, as well as through[^31] $\overline F{}^-{}^{ij}$, which are the anti-self-dual field strengths “dressed” with the scalar fields. They can be expressed in terms of the field strengths, $F^{IJ}$, by solving the following system of equations: $$\label{defFbarm}
F^-{}^{IJ}\eql (u_{ij}{}^{IJ}+v_{ijIJ})\,\overline F{}^-{}^{ij}\,,\qquad F^{-}{}^{IJ}\eql {1\over 2}(F^{IJ}-\i\,*F^{IJ})\,.$$ The symmetry of the truncation guarantees that $\overline F{}^-{}^{ij}$, as an $\rm SO(8)$ tensor, has the same structure as the vector potential, $A^{IJ}$, in , with $A^\alpha$ replaced with $\overline F{}^{-\alpha}$. However, in the absence of a closed form general solution to , one has to perform the calculation explicitly. The result simplifies if we use the following linear combinations: $$\label{}
\begin{split}
\overline F{}^{-0}+\overline F{}^{-1}+\overline F{}^{-2}+\overline F{}^{-3}
& \eql -{\i\over 2}\, \cfb_0 \,(e^{2f_0}\,\text{vol}_{\rm AdS_2}+\i\,e^{2h_0}\,\text{vol}_{\Sigma_\fg}) \,,\\
\overline F{}^{-0}+\overline F{}^{-1}-\overline F{}^{-2}-\overline F{}^{-3} & \eql -{\i\over 2}\, \cfb_1 \,(e^{2f_0}\,\text{vol}_{\rm AdS_2}+\i\,e^{2h_0}\,\text{vol}_{\Sigma_\fg})\,,\\
\overline F{}^{-0}-\overline F{}^{-1}+\overline F{}^{-2}-\overline F{}^{-3}& \eql -{\i\over 2}\, \cfb_2 \,(e^{2f_0}\,\text{vol}_{\rm AdS_2}+\i\,e^{2h_0}\,\text{vol}_{\Sigma_\fg})\,,\\
\overline F{}^{-0}-\overline F{}^{-1}-\overline F{}^{-2}+\overline F{}^{-3}& \eql -{\i\over 2}\, \cfb_3 \,(e^{2f_0}\,\text{vol}_{\rm AdS_2}+\i\,e^{2h_0}\,\text{vol}_{\Sigma_\fg})\,,
\end{split}$$ that also arise in the supersymmetry variations below. Then the complex constants, $\cfb_\alpha$, are related to the electric and magnetic parameters, $e_\alpha$ and $m_\alpha$, in by $$\label{Smatact}
S_{\alpha\beta}\cfb_\beta \eql e^{-2h_0}m_\alpha+\i\,e^{-2f_0}e_\alpha\,,$$ where $$\label{Smatrix}
\begin{split}
S_{\alpha0} & \eql {1\over 2}\sqrt 2\, L^\alpha\,,\\
S_{\alpha i} & \eql -{1\over 2}\sqrt 2 (1-|z_i|^2) \overline{D_{z_i}L^\alpha}\,,\qquad \alpha=0,\ldots,3\,,\quad i=1,\ldots,3\,.
\end{split}$$ The symplectic sections, $L^\alpha$, have been defined in and their Kähler covariant derivative in .
Finally, it has been observed in [@Halmagyi:2013sla] that the BPS equations for near horizon black holes in $\cals N=2$ supergravity coupled to hypermultiplets must be supplemented by Maxwell equations, which impose massive constraints on the electric and magnetic parameters. The same holds in our model. Indeed, setting $\zeta_1=0$ and $\zeta_2=z$ in we find that, cf. , $$\label{mconstr}
\begin{split}
e^{(m)}& \equiv e_0-e_1-e_2-e_3\eql 0\,,\\[6 pt]
m^{(m)} & \equiv m_0-m_1-m_2-m_4\eql 0\,.
\end{split}$$ Implementing those constraints from the start simplifies the derivation of the BPS equations considerably.
The spinor fields, $\chi^{ijk}$ and $\psi_\mu{}^i$, and the supersymmetry parameters, $\epsilon^i$, in and are $\gamma^5$-chiral. Since we are using a real representation of the $\gamma$-matrices, which makes $\gamma^5$ to be pure imaginary, the complex conjugation, which lowers/raises the $\SU(8)$ indices $i,j,k,\ldots$, changes the $\gamma^5$-chirality, for example $$\label{proj5}
\gamma^5 \epsilon^j\eql \epsilon^j\,,\qquad \gamma^5\epsilon_j\eql -\epsilon_j\,,\qquad \gamma^5\equiv \i\,\gamma^0\gamma^1\gamma^2\gamma^3\,.$$ In particular, implies $$\label{proj23}
\gamma^2\gamma^3\epsilon^j\eql -\i\,\gamma^0\gamma^1\epsilon^j\,,\qquad
\gamma^2\gamma^3\epsilon_j\eql \i\,\gamma^0\gamma^1\epsilon_j\,.$$ It follows from that there are two U(1)$^2$-invariant supersymmetry parameters, $\epsilon^7$ and $\epsilon^8$. In the following we set $$\label{noninveps}
\epsilon^1\eql\ldots\eql \epsilon^6\eql 0\,,$$ and relabel $\epsilon^7$ and $\epsilon^8$ as $\epsilon^1$ and $\epsilon^2$, respectively.
The condition for a supersymmetric solution is that and vanish. Here, we are interested in solutions for which the Killing spinors, $\epsilon^i$, of unbroken supersymmetries are constant along the Riemann surface and the usual Killing spinors along AdS$_2$. In the coordinate system we are using, this means that $\epsilon^i$ do not depend on $t$, $x$ and $y$ and satisfy $$\label{raddep}
\partial_r\epsilon^i\eql -{1\over 2r}\epsilon^i\,,\qquad i=1,2\,,$$ along the radial coordinate, $r$. In addition to the spinors obeying there are also the “conformal Killing spinors” in AdS$_2$ dual to the S-type supercharges in the 1d superconformal quantum mechanics.
C.2The spin-1/2 variations {#c.2the-spin-12-variations .unnumbered}
--------------------------
After imposing in , the only nonvanishing variations are for the U(1)$^2$-invariant fields in and . Setting the variations of to zero and using , yields three pairs of equations of the form $$\label{spin12va}
\begin{split}
\cfb_i \epsilon^1+2\,\i\,g\,\fF_i\,\gamma^0\gamma^1\epsilon^2 & \eql 0\,, \\[6 pt]
\cfb_i \epsilon^2-2\,\i\,g\,\fF_i\,\gamma^0\gamma^1\epsilon^1 & \eql 0\,,\qquad i=1,\ldots,3\,,
\end{split}$$ where $$\label{}
\begin{split}
\fF_1 & \eql e^{K_S/2} \Big[{|z|^2\over 1-|z|^2}(1-\bar z_1)(1-z_2)(1-z_3)+{2\over 1-|z|^2}(\bar z_1-z_2z_3) \Big]\,, \end{split}$$ with $\fF_2$ and $\fF_3$ obtained by the other two cyclic permutations of $z_1$, $z_2$ and $z_3$. Equations reduce to the projector $$\label{proj12}
\gamma^0\gamma^1\epsilon^1\eql \i\,\xi \,\epsilon^2\,,\qquad \gamma^0\gamma^1\epsilon^2\eql -\i\,\xi \, \epsilon^2\,,\qquad \xi=\pm 1\,,$$ and three BPS equations $$\label{BPSfi}
\cfb_i\eql -2\xi g\,\fF_i\,,\qquad i=1,\ldots,3.$$ Using the massive constraints , the variations of simplify to $$\label{Gvars12}
\begin{split}
\fGG\,\Big[{z-\bar z\over 1-|z|^2}\,\epsilon^1-\i\,{z+\bar z\over 1-|z|^2}\epsilon^2\Big]\eql 0\,,\\[6 pt]
\fGG\,\Big[{z+\bar z\over 1-|z|^2}\,\epsilon^1-\i\,{z-\bar z\over 1-|z|^2}\epsilon^2\Big]\eql 0\,,\\
\end{split}$$ where $$\label{}
\fGG\eql e^{K_S/2}\big[(1-z_1)(1-z_2)(1-z_3)+2(z_1z_2z_3-1)\big]\,.$$ Clearly, vanish identically when we turn-off the hypermultiplet and thus are absent in the STU-model. For a nontrivial hypermultiplet, $z\not=0$, they imply the BPS equation $$\label{Gequation}
\fGG\eql 0\,,$$ which is a cubic constraint on the scalars, $z_i$.
C.3The spin-3/2 variations {#c.3the-spin-32-variations .unnumbered}
--------------------------
We now turn to the spin-3/2 variations . Using , and in the variations $\gamma^2\delta\psi_x{}^{7,8}+\gamma^3\delta\psi_y{}^{7,8}$, we find $$\label{a1eqs}
m_0\eql \kappa\,{\xi\over 2g}\,,$$ where $\kappa=\pm 1$ or 0 is the normalized curvature of $\Sigma_\fg$, see Appendix \[appconv\]. The difference of the two variations yields $$\label{Phi0eqs}
\cfb_0\eql 2g\xi\,\overline \fW\,,$$ where $$\label{}
\fW\eql e^{K_V/2}\Big[{|z|^2\over 1-|z|^2}(1-z_1)(1-z_2)(1-z_3)+{2\over 1-|z|^2}(z_1z_2z_3-1)\Big]\,.$$ The variations, $\delta\psi^{7,8}_r$, along the radial directions, assuming , give $$\label{}
\begin{split}
\gamma^0\epsilon^1 & \eql \sqrt 2\,\i\,g\xi\,e^{f_0}\fW\,\epsilon_2\,,\\[6 pt]
\gamma^0\epsilon^2 & \eql - \sqrt 2\,\i\,g\xi\,e^{f_0}\fW\,\epsilon_1\,.\\
\end{split}$$ Taken together with their complex conjugates, they yield the projector $$\label{projLa}
\gamma^0\epsilon^1\eql e^{\i\,\Lambda}\epsilon_2\,,\qquad \gamma^0\epsilon^2\eql -e^{\i\,\Lambda}\epsilon_1\,,$$ and the BPS equation $$\label{apsolf0}
e^{-f_0}\eql - \sqrt 2\,\i\,g\xi\,e^{\i\,\Lambda}\,\fW\,,$$ where $\Lambda$ is a constant.
Finally, using all the projectors as well as and , the variations $\delta\psi_t{}^{7,8}$ set $$\label{e0eqss}
e_0\eql 0\,.$$ This concludes our truncation of the supersymmetry variations and .
C.4Summary and comments {#c.4summary-and-comments .unnumbered}
-----------------------
We have shown that the truncation of the $\cals N=8$ supersymmetry variations and the Maxwell equations resulted in:
- Four real equations , and for the electric and magnetic parameters $e_\alpha$ and $m_\alpha$.
- Four complex equations and for the scalar dressed components, $\Phi_\alpha$, of the fluxes.
- One complex equation for the metric constant, $f_0$, and the phase $\Lambda$.
- A complex cubic constraint for the scalars, $z_i$.
Using the geometric data of the corresponding $\cals N=2$ supergravity derived in Appendix \[appendixA\], we have verified that our BPS equations above agree, modulo differences in conventions, with those derived for the near horizon black holes in general $\cals N=2$ $d=4$ gauged supergravities coupled to hypermultiplets in [@Halmagyi:2013sla]. In fact, a comparison with the $\cals N=2$ formulae suggests some simplifications. In particular, we have $$\label{}
\fF_i\eql -(1-|z_i|^2)\,D_{z_i}\fW\,,\qquad i=1,2,3\,.$$ Those identities turn out useful for solving the BPS equations in Section \[subsec:dyonicBPS\] using some standard identities of the special Kähler geometry [@Ceresole:1995ca; @Andrianopoli:1996vr] and to rewrite them as attractor equations in Appendix \[appattractor\].
Finally, note that the equations above are invariant under [@Halmagyi:2013sla] $$\label{}
(e_\alpha,m_\alpha,\Lambda,\xi)\qquad \longrightarrow\qquad (-e_\alpha,-m_\alpha,\Lambda+\pi,-\xi)\,,$$ so that we may set $\xi=-1$ for convenience.
The attractor equations {#appattractor}
=======================
In Section \[sec:DyonicComp\], we obtained a match between the black hole entropy and topologically twisted index by explicitly solving the BPS equations and extremization equations for the supergravity scalar fields and field theory fugacities, respectively. In this section we show an alternative method to achieve a match between the twisted index and the black hole entropy. In particular, we will solve a subset of the BPS equations for $e^{2h_0}$ as a function of the scalar fields and the electric and magnetic charges and show that the remaining BPS equations imply that $e^{2h_0}$ is extremized with respect to the scalar fields. This allows for a comparison with the topologically twisted index and its extremization with respect to the fugacities and Lagrange multipliers. This procedure is the same as the AdS$_4$ black hole attractor mechanism discussed in [@Benini:2016rke; @DallAgata:2010ejj; @Hristov:2010ri; @Gnecchi:2013mta; @Halmagyi:2013sla; @Klemm:2016wng].
It is crucial for our analysis to work with the electric charges, $q_\alpha$, defined in . As explained in Sections \[sec422\] and \[sec:eleccharg\], equation leads to different electric charges depending on whether the massive condition is imposed in the Maxwell Lagrangian before or after varying with respect to $F^\alpha$. In this appendix we choose the latter and as a consequence we will compare the resulting entropy with the field theory in Section \[sec:manjmditc\].
Using and , we can write as $$\label{fs}
\Phi_\alpha = 4 \,\i\, e^{-2h_0} \overline{S}_{\beta \alpha} \left(q_\beta - \overline{\mathcal{N}}_{\beta\sigma} m_\sigma\right) \,.$$ Combining with the identity $$L^\beta \mathcal{N}_{\alpha\beta} = M_\alpha \equiv e^{K_V/2} \frac{\partial F}{\partial X^\alpha} \,,$$ we can rewrite the BPS equation as $$\label{set1}
\widehat{\mathcal{Z}} \equiv \sqrt{2} e^{-2h_0} \left(L^\alpha q_\alpha - M_\alpha m_\alpha\right) + {\rm i} \, g \, \fW = 0 \,.$$ With the use of the identity [@Andrianopoli:1996vr] $$D_{z_i}\left(L^\alpha \mathcal{N}_{\alpha\beta}\right) =\left( D_{z_i} L^\alpha \right) \overline{\mathcal{N}}_{\alpha\beta} \,,$$ we can furthermore write as $$\label{set2}
D_{z_i} \widehat{\mathcal{Z}} = 0 \,.$$ Since only $\fW$ depends explicitly on $z$, we can similarly write as $$\label{set3}
\frac{\partial \widehat{\mathcal{Z}}}{\partial z} = 0 \,.$$ Now, we note that the equations , and imply the following suggestive set of equations $$\label{e2h0rewrite}
\begin{split}
e^{2h_0} = \sqrt{2}\, {\rm i}\, \frac{L^\alpha q_\alpha - M_\alpha m_\alpha}{ g \,\fW} \,, \qquad\qquad \frac{\partial e^{2h_0}}{\partial z_i} = \frac{\partial e^{2h_0}}{\partial z} = 0 \,,
\end{split}$$ where $e^{2h_0}$ is a function of the electric charges, $q_\alpha$, magnetic fluxes, $m_\alpha$, and the scalars, $z_i$ and $z$. The BPS equations thus imply an extremization procedure for the metric coefficient $e^{2h_0}$ as a function of the scalar fields. Using and implementing the relation between $L^\alpha$ and $u_\alpha$ to write $$\label{e2h0expl}
e^{2h_0} = \frac{1}{g} \frac{\sum_{\alpha=1}^4 \sqrt{u_1u_2u_3u_4}\,{ \frac{m_{\alpha-1}}{u_{\alpha}}}-{\rm i} \,u_\alpha q_{\alpha-1}}{2u_1-\frac{1}{1-z \bar z}\left(u_1-u_2-u_3-u_4\right)} \,.$$ We are now in a position to see how our equation manipulations above pay off. Note that it follows trivially from that the $u_\alpha$’s satisfy $\sum_{\alpha=1}^{4} u_\alpha = 2$. In addition the last equation in combined with implements the massive constraint $u_1=u_2+u_3+u_4$. Implementing the massive constraint in we can write $$\label{e2h0nice}
e^{2h_0} = \frac{1}{2g} \sum_{\alpha=1}^4 \left(\sqrt{u_1u_2u_3u_4}\,\frac{m_{\alpha-1}}{u_{\alpha}}-{\rm i}\, u_\alpha q_{\alpha-1} \right) \,.$$ After identifying the field theory and supergravity charges as in and we observe that the entropy with $e^{2h_0}$ replaced by takes the same functional form as the twisted index . Equation then implies that the entropy is extremized with respect to the $u_\alpha$, ensuring that the same extremization principle applies to both the topologically twisted index and the black hole entropy.
We have not yet discussed the BPS constraints - on the charges. In supergravity we start with four magnetic fluxes and four electric parameters which satisfy the above four constraints. While the constraints act linearly on the electric parameters $e_\alpha$, they act non-linearly on the electric charges $q_\alpha$. In field theory we start off with four magnetic charges and four electric charges and implement the two constraints and on the magnetic charges. The constraints and are equivalent to the constraints and . One more constraint is imposed on the electric charges by imposing the index to be real. Indeed, the BPS constraints - are crucial to ensure that $e^{2h_0}$ is real. However, there is no further constraint on the electric charges in the field theory and there is in fact a shift symmetry which allows us to shift the electric charges by a free parameter as in . The supergravity computation has thus fixed the shift symmetry in a particular way. Explicit comparison shows that the shift symmetry is fixed such that the Lagrange multipliers $\lambda_1,\lambda_2$ in Section \[sec:manjmditc\] are real, i.e. $\nu_1=\nu_2=0$.
In conclusion, both in supergravity and field theory we are evaluating the same expression subject to the same extremization equations. Imposing the Lagrange multipliers to be real then ensures that also the constraints on the charges coincide. We can thus conclude that the black hole entropy and the extremized topologically twisted index are equal.
[^1]: Here we are focusing on black holes and do not discuss higher-dimensional black branes.
[^2]: See also [@ABCMZ; @Guarino:2017eag; @Guarino:2017pkw; @Hosseini:2017fjo; @Benini:2017oxt] for an extension of these results to asymptotically AdS$_4$ black holes in massive IIA string theory.
[^3]: See [@Liu:2017vll; @Jeon:2017aif; @Liu:2017vbl] for recent attempts to account for subleading corrections in $N$.
[^4]: The ABJM theory has a global symmetry group of rank 4 and thus the magnetic fluxes, R-charges and fugacities are labelled by $\alpha=1,2,3,4$. The mass deformation in mABJM reduces the rank of the symmetry group to 3 and the parameters are labelled by $i=2,3,4$.
[^5]: We expect that most of our results should hold for more general values of $k$.
[^6]: As usual, the R-charge of an $\cals N=2$ chiral supermultiplet is defined as the R-charge of its lowest component which is a complex scalar.
[^7]: See [@Freedman:2013ryh] where a discussion on $F_{S^3}$ for the ABJM theory as a function of $\Delta_a$ can be found. Note also that we define the free energy as $F_{S^3} = -\log Z_{S^3}$, where $Z_{S^3}$ is the supersymmetric partition function of the theory on $S^3$.
[^8]: One can always choose a gauge in which the monopole operators have vanishing R-charge, see for example [@Jafferis:2011zi] for a discussion on this.
[^9]: See, Appendix \[appconv\] for our conventions.
[^10]: We use the shorthand notation $\Delta_{A_1,A_2} \to \Delta_{1,2}$ and $\Delta_{B_1,B_2} \to \Delta_{3,4}$.
[^11]: Note that due to the quantization condition $\fn_\alpha\in\ZZ$, the universal twist is well defined only on Riemann surfaces for which $\fg$ is odd.
[^12]: Due to the quantization condition $\fn_{i} \in \mathbb{Z}$, we find that the universal twist for the mABJM theory is well defined only on Riemann surfaces for which $\fg$ is a multiple of 4.
[^13]: The restriction used in [@Benini:2016rke] to avoid the square-root sign ambiguity in is $0<\Re\, u_\alpha<2$.
[^14]: In this section, the indices $i,j,\ldots$ run over the set $2,3,4$. In particular, $\epsilon_{234}=1$, etc.
[^15]: $C(\tau)=C(\vartheta)|_{\theta_1=0}$ and $S(\tau)=S(\vartheta)|_{\theta_1=0}$.
[^16]: A four-dimensional $\mathcal{N}=2$ supergravity with the same mater content was recently used in [@Hosseini:2017fjo; @Benini:2017oxt; @Kim:2018cpz] to construct asymptotically AdS$_4$ black holes which admit uplifts to massive type IIA supergravity. The supergravity theory we study here differs from the one in [@Hosseini:2017fjo; @Benini:2017oxt; @Kim:2018cpz] by the type of gauging performed on the vector multiplets.
[^17]: The four dimensional metric, $g_{\mu\nu}$, has signature $(-+++)$ and $e=\sqrt{-\det g_{\mu\nu}}$.
[^18]: The dual field strength is $ \widetilde F^{\alpha}_{\mu\nu}={1\over 2}\eta_{\mu\nu}{}^{\lambda\sigma}F^\alpha_{\lambda\sigma}$, where $\eta_{0123}=e$.
[^19]: In the sense that $\nabla_{z_i}\cals V=\partial_z\cals V=0$.
[^20]: See, the following equations in Appendix \[appendixB\]: (i) , and ; (ii) and ; (iii) ; (iv) . We set $\xi=-1$.
[^21]: Note that $D_{z_i}= e^{-K_V/2}\partial_{z_i}e^{K_V/2}= e^{K_V/2}\nabla_{z_i}e^{-K_V/2}$, see .
[^22]: One can verify explicitly that $${\overline \Pi}\,\Xi-\Pi\,\overline\Xi \eql 48\,\mathcal{C}+ \sum_{i=1}^3(1-|z_i|^2)\,\mathcal{C}\partial_{\bar{z}_i}\overline{\Pi}+4\sum_{i=1}^3\mathcal{C}\,\partial_{\bar{z}_i}\bar{\mathcal{C}}-6\sum_{i,j=1}^3\mathcal{C}\,\partial_{\bar{z}_i}\partial_{\bar{z}_j}\bar{\mathcal{C}}-{\rm c.c.}\;.$$ Hence ${\overline \Pi}\,\Xi$ is real for $\cals C=\bar{\cals C}=0$ as expected from the discussion above.
[^23]: Notice that for real $z_i$ we have $x_i=\tilde{z}_i$ where $\tilde{z}_i$ are defined in .
[^24]: As discussed above, setting $\chi=0$ reduces our supergravity model to the STU model and thus the supersymmetric black hole solutions with $\chi=0$ reduce to the ones discussed in [@Benini:2015eyy].
[^25]: The $x_i$ must be positive so we discard solutions where the $x_i$ take negative values.
[^26]: See also [@Guarino:2017jly] for a recent discussion of this universal solution.
[^27]: We have fixed $\ell_{\Sigma}$ throughout this paper by normalizing the curvature of $\Sigma_\fg$ to $\kappa$.
[^28]: See Section 4.3 of [@Jafferis:2011zi] for a field theory discussion of this type of generalization.
[^29]: For supersymmetry variations in the $\cals N=8$ theory, see (3.1)-(3.5) and Section 5 in [@deWit:1982bul]. For the supersymmetry variations in $\cals N=2$ supergravity, see, e.g., (4.18)-(4.25) in [@Andrianopoli:1996vr].
[^30]: The “dot” denotes the summation over the vector index of the moment maps.
[^31]: Note that the “bar” here does not mean complex conjugation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Gorenstein liaison seems to be the natural notion to generalize to higher codimension the well-known results about liaison of varieties of codimension 2 in projective space. In this paper we study points in ${\mathbb P}^3$ and curves in ${\mathbb P}^4$ in an attempt to see how far typical codimension 2 results will extend. While the results are satisfactory for small degree, we find in each case examples where we cannot decide the outcome. These examples are candidates for counterexamples to the hoped-for extensions of codimension 2 theorems.'
author:
- Robin Hartshorne
title: |
Some Examples of Gorenstein Liaison\
in Codimension Three
---
[*Subject Classification:*]{} 14H50, 14M05, 14M06, 14M07
For curves in projective three-space ${\mathbb P}_k^3$, the theory of liaison, or linkage, is classical, and is now a well-understood framework for the classification of algebraic space curves [@MDP], [@RHClassIII]. This theory has been successfully extended to schemes of codimension $2$ in any projective space ${\mathbb P}^n$ [@BBM], [@R], [@Nollet], [@M Ch. 6]. Recently a number of efforts have been made to find a suitable extension of these results in codimension $\ge 3$ [@M]. Traditional liaison uses complete intersections to link one scheme to another. In codimension $2$, the property of being a complete intersection is equivalent to being arithmetically Gorenstein [@Serre]. Thus there are two natural ways to generalize. It appears that complete intersection liaison is too fine a relation to give analogous results in higher codimension. Thus attention has been focussed on Gorenstein liaison, and a number of recent results have created an optimistic attitude that much of the codimension $2$ case will carry over naturally to higher codimension [@M], [@KMMNP], [@CM], [@N], $\ldots$. The purpose of this paper is to give some examples of Gorenstein liaison for points in ${\mathbb P}^3$ and for curves in ${\mathbb P}^4$, which suggest that the situation in codimension $\ge 3$ may be more complicated than was initially suspected.
For points in ${\mathbb P}^3$, we show first that any set of $n$ points in general position in a plane or on a nonsingular quadric surface can be obtained from a single point by a sequence of ascending Gorenstein biliaisons (see §1 below for definitions of these terms). Thus any set of $n$ points in a plane or on a quadric surface is glicci (in the Gorenstein liaison class of a complete intersection). On a nonsingular cubic surface, we can still show that any set of $n$ points in general position is glicci, but we have to use ascending and descending biliaisons and simple liaisons to prove this. For a set of $n$ points in general position in ${\mathbb P}^3$, we show that for $n
\le 19$ it is glicci, but we are unable to prove this for $n=20$. Thus a set of 20 points in general position in ${\mathbb P}^3$ becomes a candidate for a possible counterexample to the hope that all ACM schemes are glicci.
In ${\mathbb P}^4$, various classes of ACM curves have been shown to be glicci, in particular, determinantal curves and ACM curves lying on general smooth ACM rational surfaces in ${\mathbb P}^4$ [@KMMNP]. We show that any general ACM curve of degree $\le 9$ or degree $10$ and genus $6$ in ${\mathbb P}^4$ is glicci. Then we study ACM curves of degree $20$ and genus $26$. There are determinantal curves of this degree and genus, but we show that a general curve in the irreducible component of the Hilbert scheme containing the determinantal curves cannot be obtained by ascending Gorenstein biliaison from a line. We do not know if this curve is glicci, so we propose it as a candidate for an example of an ACM curve that is not glicci.
For curves in ${\mathbb P}^4$, we consider the set of curves with Rao module $k$, i.e., of dimension $1$ in $1$ degree only. We call such a curve [*minimal*]{} if its Rao module occurs in degree $0$. We show that there are minimal curves of every degree $\ge 2$. Then we give examples that suggest that there are curves in the liaison class of two skew lines that cannot be reached by ascending Gorenstein biliaison from a minimal curve; and that there are other curves with Rao module $k$ that are not in the liaison class of two skew lines. We will describe the examples and the evidence for these statements below, but in most cases we cannot prove anything.
I hope that further study of these examples and others will establish whether these guesses are correct or not, and help clarify some of the major questions concerning Gorenstein liaison in codimension $\ge 3$.
I would like to thank Juan Migliore for his book [@M], which clearly sets out the case for Gorenstein liaison, and which stimulated this research. I would also like to thank him and Rosa Miró–Roig and Uwe Nagel for sharing their unpublished papers with me. Lastly, I would like to thank the referee for many helpful suggestions, and in particular for an idea that led to a great improvement of Example $4.6$.
Basic Results and Questions
===========================
Let $V_1$ and $V_2$ be two equidimensional closed subschemes without embedded components, of the same dimension $r$ in ${\mathbb P}_k^n$, the $n$-dimensional projective space over an algebraically closed field $k$. We say $V_1$ and $V_2$ are [*linked by a complete intersection scheme $X$*]{}, if $X$ is a complete intersection scheme of dimension $r$ containing $V_1$ and $V_2$, and if $$\begin{aligned}
{\mathcal I}_{V_2,X} &\cong &{\mathop{\mathrm{Hom}}\nolimits}({\mathcal O}_{V_1},{\mathcal O}_X), \mbox{
and} \\
{\mathcal I}_{V_1,X} &\cong &{\mathop{\mathrm{Hom}}\nolimits}({\mathcal O}_{V_2},{\mathcal O}_X).\end{aligned}$$ Using the language of generalized divisors on Gorenstein schemes [@GD $4.1$] we can say equivalently $V_1$ and $V_2$ are linked by a complete intersection $X$ if and only if there is a complete intersection scheme $S$ of dimension $r+1$ containing $V_1$ and $V_2$, such that $X$ is an effective divisor in the linear system $|mH|$ on $S$ for some $m > 0$, where $H$ is the hyperplane section of $S$, and $V_2 = X - V_1$ as generalized divisors on $S$.
The equivalence relation generated by complete intersection linkage is called [*$CI$-liaison*]{}. If the equivalence can be accomplished by an even number of links, we speak of [*even $CI$-liaison*]{}.
A scheme $X$ in ${\mathbb P}^n$ is called [*arithmetically Gorenstein*]{} (AG) if its homogeneous coordinate ring $R/I_X$ is a Gorenstein ring, where $R =
k[x_0,\dots,x_n]$ is the homogeneous coordinate ring of ${\mathbb P}^n$, and $I_X$ is the (saturated) homogeneous ideal of $X$. If, in the first definition above, we require that $X$ be an arithmetically Gorenstein scheme, instead of a complete intersection, then we say that $V_1$ and $V_2$ are [*linked by an AG scheme*]{}. The equivalence relation generated by this kind of linkage is called [*Gorenstein liaison*]{} (or [*$G$-liaison*]{} for short); if the equivalence can be accomplished by an even number of $G$-links, then we speak of [*even Gorenstein liaison*]{}.
One way of obtaining AG schemes is as follows. Let $S$ be an [*arithmetically Cohen–Macaulay*]{} (ACM) scheme in ${\mathbb P}^n$ (this means that the homogeneous coordinate ring $R/I_S$ is a Cohen–Macaulay ring). Assume also that $S$ satisfies the property $G_1$ (Gorenstein in codimension $1$ [@GD p. 291]), so that we can use the language of generalized divisors. Then any effective divisor $X$ in the linear system $|mH-K|$ on $S$, where $m
\in {\mathbb Z}$, $H$ is the hyperplane section, and $K$ is the canonical divisor, is an arithmetically Gorenstein scheme [@M $4.2.8$].
Suppose now that $V_1$ and $V_2$ are divisors on an ACM scheme $S$ of dimension $r+1$ satisfying $G_1$, let $X$ be an effective divisor in the linear system $|mH-K|$ for some $m$, and suppose that $V_2 = X - V_1$ as generalized divisors on $S$. Then it is easy to see that $V_1$ and $V_2$ are linked by the AG scheme $X$ (cf. proof of [@GD $4.1$] and note that since $S$ satisfies $G_1$, $X$ is an almost Cartier divisor on $S$ [@GD p. 301]). In this case we will say that $V_1$ and $V_2$ are [*strictly $G$-linked*]{}. We do not know whether the equivalence relation generated by strict $G$-linkages is equivalent to the $G$-liaison defined above, so we will call it [*strict $G$-liaison*]{}, and if it is accomplished in an even number of steps, [*strict even $G$-liaison*]{}.
Combining two strict $G$-linkages gives the following result.
[**Proposition 1.1**]{} [@KMMNP $5.14$], [@M $5.2.27$]. [*Let $V_1$ and $V_2$ be effective divisors on an [*ACM*]{} scheme $S$ satisfying $G_1$. Suppose that $V_2 \in |V_1 +hH|$ for some $h \in {\mathbb Z}$. Then $V_1$ and $V_2$ can be strictly $G$-linked in two steps.*]{}
Note that even though the statement in [@M $5.2.27$] requires $S$ smooth, the proof given there works for $S$ ACM satisfying $G_1$ if one takes into account that any divisor $X \in |mH-K|$ is almost Cartier [@GD $2.5$].
The proposition above motivates the following definition. In the situation of $(1.1)$, we say that $V_2$ is obtained by an [*elementary Gorenstein biliaison of height $h$*]{} from $V_1$ [@M $5.4.7$]. Because of the proposition, an elementary $G$-biliaison is a strict even $G$-liaison. If $h
\ge 0$, we call the biliaison [*ascending*]{}.
Now we can state some of the main questions raised by trying to generalize codimension $2$ results to higher codimension.
[**Question 1.2.**]{} a) [*Does strict Gorenstein liaison generate the same equivalence relation ($G$-liaison) as Gorenstein liaison?*]{}
b\) [*Do the elementary Gorenstein biliaisons generate the same equivalence relation as even $G$-liaison?*]{}
In codimension 2, both questions reduce to $CI$-liaison, for which the answers to parts a) and b) are both yes [@GD $4.1$, $4.4$].
[**Question 1.3.**]{} a) [*Is every [*ACM*]{} subscheme of ${\mathbb P}^n$ in the $G$-liaison class of a complete intersection (in which case we say it is [*glicci*]{})?*]{}
b\) [*Can every glicci scheme be obtained by a finite sequence of ascending elementary $G$-biliaisons from a scheme (of the same dimension) of degree $1$?*]{}
In codimension $2$, part a) is the classical theorem of Gaeta [@G], [@PS], $\ldots$. Part b) seems likely to be true, though I do not know a reference. Note in b) it would be equivalent to ask for ascending elementary $G$-biliaisons starting with any complete intersection scheme. In higher codimension, many special cases of a) have been shown to be true [@M], [@KMMNP], [@N], [@CM]. Closely related is the theorem of Migliore and Nagel [@MN] that every ACM subscheme $X$ of ${\mathbb P}^n$ has a flat deformation to a glicci scheme, and there is also a glicci scheme with the same Hilbert function as $X$.
For the following questions we limit the discussion to curves (locally Cohen–Macaulay schemes of dimension $1$) for simplicity. For a curve $C
\subseteq {\mathbb P}^n$ we define its [*Rao module*]{} to be the finite length graded $R$-module $M = \oplus_{l \in {\mathbb Z}} H^1({\mathcal I}_C(l))$, where ${\mathcal I}_C$ is the ideal sheaf of $C$. It is easy to see that even $G$-liaison preserves the Rao module, up to shift of degrees [@M $5.3.3$].
[**Question 1.4.**]{} [*Does the Rao module characterize the even $G$-liaison class of a curve? In other words, if $C$ and $C'$ are two curves with $M_{C'}
\cong M_C(h)$ for some $h \in {\mathbb Z}$, are $C$ and $C'$ in the same even $G$-liaison class? (In codimension $2$, this is the well-known theorem of Rao [@R].)*]{}
Now we come to the problem of the structure of an even $G$-liaison class. Let $C \subseteq {\mathbb P}^n$ be a curve, and let ${\mathcal L}$ be the class of all curves $C'$ in the even $G$-liaison class of $C$. The Rao modules of curves in ${\mathcal L}$ are all isomorphic up to shift. As long as the Rao module is not zero (which is equivalent to saying the curves are not ACM), one knows that there is a minimal leftward shift of $M$ that can occur [@M $1.2.8$]. We denote by ${\mathcal L}_0$ the subset of ${\mathcal L}$ consisting of those curves with the leftmost possible shift of the Rao module, and we call these [*minimal curves*]{}. Let ${\mathcal L}_h$ denote the set of curves with Rao module shifted $h$ places to the right from ${\mathcal
L}_0$, for each $h\ge 0$. Then ${\mathcal L} = \cup {\mathcal L}_h$ for $h\ge
0$, and each one of these ${\mathcal L}_h $ for $h \ge 0$ is nonempty [@M $1.2.8$].
In codimension $2$ a biliaison class ${\mathcal L}$ satisfies the [*Lazarsfeld–Rao property*]{} [@M $5.4.2$]. It says that a) ${\mathcal L}_0$ is a single irreducible family of curves, and b) any curve $C
\in {\mathcal L}_h$ can be obtained by a finite sequence of ascending biliaisons from a minimal curve, plus if necessary a deformation with constant cohomology within the class ${\mathcal L}_h$. (But even for curves in ${\mathbb P}^3$ it is not known if these deformations are necessary [@MDP IV, $5.4$, p. 93].) Easy examples show that in codimension $3$, ${\mathcal L}_0$ need not consist of a single irreducible family of curves [@M $5.4.8$]. So we rephrase the question somewhat.
[**Question 1.5.**]{} a) [*Describe the set ${\mathcal L}_0$ of minimal curves in an even $G$-liaison class.*]{}
b\) [*Can every curve in ${\mathcal L}_h$ for $h> 0$ be obtained from a minimal curve by a finite sequence of ascending elementary $G$-biliaisons, followed possibly by a flat deformation within the family ${\mathcal L}_h$?*]{}
An optimist might hope for positive answers to all these questions. However, the examples we give below suggest that many of the answers may be no.
Points in ${\mathbb P}^3$
=========================
Any scheme of dimension zero is ACM, so in this section we will address Question $1.3$. The study of arbitrary zero-schemes, even in ${\mathbb P}^3$, becomes quite complicated, so we will direct our attention to sets of reduced points in general position. In this section the phrase “a general $X$ has property $Y$” will mean that there is a nonempty Zariski open subset of the family of all $X$’s having the property $Y$.
We begin with points in ${\mathbb P}^2$. In this case it is known from the theorem of Gaeta [@M $6.1.4$] that any zero scheme in ${\mathbb P}^2$ is [*licci*]{} (in the liaison class of a complete intersection), but we give a slightly more precise statement for a general set of points and at the same time we illustrate in a simple case the technique we will use in the later propositions.
[**Proposition 2.1.**]{} [*A set of $n$ general points in ${\mathbb P}^2$ can be obtained from a single point by a sequence of ascending elementary biliaisons.*]{}
[**Proof.**]{} By induction on $n$. For $n = 1$ there is nothing to prove. For $n = 2$, any $2$ points lie on a line $L$. A single biliaison of height $1$ on $L$ reduces $2$ points to $1$ point. For $n = 3,4,5$, a set of $n$ reduced points, no three on a line, lies on a nonsingular conic. These points can be obtained by an elementary biliaison of height $1$ or $2$ from a set of $1$ or $2$ points, and we are done by induction.
In general, let $n\ge 3$. Then there is an integer $d \ge 2$ such that $\frac
{1}{2} (d-1)(d+2) < n \le \frac {1}{2} d(d+3)$. Since curves of degree $d$ in ${\mathbb P}^2$ form a linear system of dimension $\frac {1}{2} d(d+3)$, any set of $n$ points will lie on a curve of degree $d$. Since the nonsingular curves form a Zariski open subset of the family of all curves, a set of $n$ [*general*]{} points in ${\mathbb P}^2$ (in the sense mentioned above) will lie on a nonsingular curve $C$ of degree $d$, and will form a set of $n$ general points on $C$. The genus of $C$ is $g = \frac {1}{2} (d-1)(d-2)$. We use the fact that on a nonsingular curve of genus $g$, any divisor of degree $\ge g$ is effective. Let $D$ be the divisor of $n$ general points on $C$, and let $H$ be the hyperplane class on $C$. We define a divisor $D'$ as follows. $$D' = \left\{ \begin{array}{ll}
D-H &\mbox{if $n = \frac {1}{2} (d-1)(d+2) + 1$} \vspace{1\jot} \\
D-2H &\mbox{if $\frac {1}{2} (d-1)(d+2) + 2 \le n \le \frac {1}{2} d(d+3)$.}
\end{array} \right.$$ Then we verify that in either case, the degree of $D'$ is $\ge g$, so that the divisor $D'$ is effective, and secondly that $n' = {\mathop{\mathrm{deg}}\nolimits}D' \le \frac {1}{2}
(d-1)(d+2)$.
Now, by induction on $d$, a general set of $n'$ points can be obtained from a single point by ascending biliaisons. Since any $D$ as above bilinks down to a $D'$, it follows that by bilinking up $n'$ general points on $C$, we obtain $n$ general points on $C$, as required.
[**Proposition 2.2.**]{} [*A set of $n$ general points on a (fixed) nonsingular quadric surface $Q \subseteq {\mathbb P}^3$ can be obtained from a single point by a finite number of ascending elementary $G$-biliaisons on $Q$.*]{}
[**Proof.**]{} The method is analogous to the proof of $(2.1)$, except that now we use both types of ACM curves on $Q$. For $n=1$, there is nothing to prove. For $n = 2$, we put $2$ points on a twisted cubic curve, and then move them by linear equivalence (a biliaison of height $0$) until they lie on a line on $Q$. On that line, we obtain $2$ points by a biliaison of height $1$ from $1$ point. For $n = 3$, the points lie on a conic, and come from $1$ point by biliaison. For $n = 4,5$, the points lie on a twisted cubic curve, and reduce by biliaison to $1$ or $2$ points.
Now suppose $n\ge 6$. Then there is an integer $a \ge 2$ such that either
- $a^2 + a \le n \le a^2 + 2a$, or
- $a^2 + 2a + 1 \le n \le a^2 + 3a + 1$.
In case i) we consider the complete intersection curve $C$ of bidegree $(a,a)$ on $Q$. It has degree $2a$ and genus $g = (a-1)^2$ and moves in a linear system of dimension $a^2 + 2a$. Hence $n$ general points lie on a smooth such curve $C$, forming a divisor $D$. The divisor $D' = D - H$ on $C$ has degree $n' = n-2a$. Since $n' \ge a^2 - a > g$, the divisor $D'$ is effective. On the other hand $n' \le a^2$, so $n'$ falls in the range i) for $a-1$ unless $n'
= a^2$, in which case it falls in range ii) for $a-1$. By induction on $a$, a set of $n'$ general points can be obtained by ascending elementary $G$-biliaisons from a point, so also can $n$ points.
In case ii), we consider the ACM curve $C$ of bidegree $(a,a+1)$ on $Q$. It has degree $d = 2a+1$, genus $g = a(a-1)$, and moves in a linear system of dimension $a^2 + 3a + 1$ on $Q$. So $n$ general points form a divisor $D$ on a nonsingular such curve $C$. The divisor $D' = D-H$ has degree $n' = n - 2a - 1
\ge a^2 > g$, so $D'$ is effective. On the other hand $a^2 \le n' \le a^2 + a
+ 1$, which is range ii) for $a-1$. So by induction on $a$ again, we can obtain $D'$ by biliaisons from a point, and $D$ by a single elementary $G$-biliaison for $D'$ on $C$.
[**Corollary 2.3.**]{} [*A set of $n$ general points on a nonsingular quadric surface $Q$ in ${\mathbb P}^3$ is in the strict even $G$-liaison class of a point. In particular, it is glicci.*]{}
[**Proposition 2.4.**]{} [*A set of $n$ general points on a (fixed) nonsingular cubic surface $S$ in ${\mathbb P}^3$ is in the same strict Gorenstein liaison class as a point on $S$.*]{}
[**Proof.**]{} A curve $C$ of degree $d$ and genus $g$ on $S$ moves in a linear system of dimension $d+g-1$. If a set of $n$ general points is to form a divisor $D$ on $C$, we need $n\le d+g-1$. In that case the linear system $D-H$ on $C$ has degree $\le g-1$, and hence may not be effective. Thus we cannot use Gorenstein biliaisons on the cubic surface. Instead, we will use strict Gorenstein liaison by AG divisors in the linear systems $|mH-K|$ on ACM curves $C$ on $S$.
There are four types of smooth ACM curves on $S$, obtained by biliaison (of curves) on $S$ from the line, the conic, the twisted cubic, and the hyperplane class $H$, which is a plane cubic curve. For $a \ge 1$ the four types are
- $d = 3a - 2$, $g = \frac {1}{2} (3a^2 - 7a + 4)$
- $d = 3a - 1$, $g = \frac {1}{2} (3a^2 - 5a + 2)$
- $d = 3a$, $g = \frac {1}{2} (3a^2-3a)$
- $d = 3a$, $g = \frac {1}{2} (3a^2 - 3a + 2)$.
In each case, one of these curves $C$ with degree $d$ and genus $g$ (we say [*type*]{} $(d,g)$) moves in a linear system of dimension $d+g-1$. On an ACM curve $C$ of type $(d,g)$, we will consider only divisors of degree $n$, where $g \le n \le d+g-1$. If $n$ and $n'$ are both in this range, and if $n+n' =
{\mathop{\mathrm{deg}}\nolimits}(mH-K)$ for some $m$, then a strict Gorenstein liaison by AG divisors in the linear system $|mH-K|$ will transform general divisors of degree $n$ to general divisors of degree $n'$ and vice versa. To explain this in more detail, let $Z$ be a set of $n$ general points on $S$. If $n \le d + g - 1$, then $Z$ is contained in a curve $C$ as above. If $n' = {\mathop{\mathrm{deg}}\nolimits}(mH-K) - n \ge
g$, then there is an effective divisor $Z'$ of degree $n'$ such that $Z + Z'
\in |mH-K|$. Thus $Z$ and $Z'$ are linked. The same arguments work in reverse, starting with $Z'$, assuming $n' \le d + g - 1$ and $n \ge g$. Hence there are Zariski open subsets $U$ (resp. $U'$) of the set of all subsets of $n$ (resp. $n'$) points of $S$ such that each $Z \in U$ is linked to a $Z'
\in U'$ and vice versa. If one of these is already known to be in the strict $G$-liaison class of a point, we conclude so is the other. We will write $n
\leftrightarrow n'$ by $mH - K$ on $(d,g)$.
For $n \le 8$, we use the following liaisons.
- $1 \leftrightarrow 3$ by $H-K$ on $(4,1)$
- $2 \leftrightarrow 6$ and $3 \leftrightarrow 5$ by $2H-K$ on $(5,2)$
- $6 \leftrightarrow 8$ by $3H - K$ on $(6,3)$
- $4 \leftrightarrow 8$ and $5 \leftrightarrow 7$ by $3H-K$ on $(6,4)$
- $6 \leftrightarrow 7$ by $3H-K$ on $(7,5)$.
These liaisons show that any set of $n\le 8$ general points on $S$ is in the strict $G$-liaison class of a point. Note that we must use ascending [*and*]{} descending liaisons to accomplish this. For example, if $n = 2$, the links go $2 \rightarrow 6 \rightarrow 7 \rightarrow 5 \rightarrow 3 \rightarrow 1$.
For $9 \le n \le 17$, we use the following links.
- $9 \leftrightarrow 11$ by $4H-K$ on $(7,5)$
- $7 \leftrightarrow 13$ and $8 \leftrightarrow 12$ by $4H-K$ on $(8,7)$
- $12 \leftrightarrow 17$ and $13 \leftrightarrow 16$ by $5H-K$ on $(9,9)$
- $10 \leftrightarrow 17$; $11 \leftrightarrow 16$; $12 \leftrightarrow
15$; and $13 \leftrightarrow 14$ by $5H-K$ on $(9,10)$.
Using these links a general set of $9 \le n \le 17$ points is linked down to a set of $7$ or $8$ points treated above.
For $n \ge 18$, we find an integer $a$ such that $n_0 = \frac {3}{2} a(a-1) \le
n < n_1 = \frac {3}{2} (a+1)a$. We divide these $n$’s into six ranges.
- $n_0 \le n \le n_0+2$
- $n_0+3 \le n \le n_0 + a-1$
- $n_0 + a$, $n_0 + a + 1$
- $n_0 + a + 2 \le n \le n_0 + 2a - 2$
- $n_0 + 2a - 1$, $n_0 + 2a$
- $n_0 + 2a + 1 \le n \le n_0 + 3a - 1$.
In range $A$, we do
- $n_0 \leftrightarrow n_0 + 2$ by $(2a-2)H-K$ on type (i) curve.
- $n_0 + 1 \leftrightarrow n_0 + 3a-1$ and $n_0 + 2 \leftrightarrow n_0
+ 3a - 2$ by $(2a-1)H-K$ on type (iv).
In range $B$, we do
- $n_0 + t \leftrightarrow n' < n_0$ by $(2a-2)H-K$ on (i).
In range $C$, we do
- $n_0 + t \leftrightarrow n' < n_0$ by $(2a-2)H-K$ on (ii).
In range $D$, we do links by $(2a-1)H-K$ on types (ii) and (iv). If $a$ is odd, $a = 2k+1$, start with $m = n_0 +3k +2$, link on type (iv). Then alternate linkages on type (ii) and (iv). This covers all values of $n$ in range $D$, starting in the middle and spiraling outward, until finally we land in range $E$. If $a$ is even, $a = 2k$, start with $m = n_0+3k$, and do a link on type (ii) first, then alternate (iv) and (ii).
In range $E$, link by $(2a-1)H-K$ on type (iv), to land in range $C$.
In range $F$, link by $(2a-1)H-K$ on type (iii) to land in range $B$ or $C$.
In summary, ranges $B$ and $C$ link down to $n' < n_0$ so are ok by induction. Ranges $E$ and $F$ link down to ranges $B$ and $C$. Range $D$ spirals up and down until it lands in range $E$; and finally range $A$ links up to range $F$. So for example, if $n = 18$ (range $A$), the links go $18 \rightarrow 20
\rightarrow 28 \rightarrow 22 \rightarrow 16 \rightarrow 13 \rightarrow 7$, which we did earlier. If $n = 54$ (range $D$), the links go $54 \rightarrow 55
\rightarrow 53 \rightarrow 56 \rightarrow 52 \rightarrow 40 \rightarrow 35
\rightarrow 27 \rightarrow 23 \rightarrow 15 \rightarrow 12 \rightarrow 8
\rightarrow 6$, treated above.
[**Corollary 2.5.**]{} [*A set of $n$ general points on a smooth cubic surface in ${\mathbb P}^3$ is glicci.*]{}
[**Corollary 2.6.**]{} [*A set of $n \le 19$ general points in ${\mathbb P}^3$ is glicci.*]{}
[**Proof.**]{} Indeed, since the cubic surfaces in ${\mathbb P}^3$ form a linear system of dimension $19$, a set of $n \le 19$ general points lie on a smooth cubic surface, and we can apply $(2.4)$. Using $(2.2)$ we can also see that a set of $n\le 9$ general points can be obtained from a single point by ascending elementary $G$-biliaisons. However, we can get a stronger result, and another proof of $(2.6)$ by another method.
[**Proposition 2.7.**]{} [*A set of $n \le 19$ general points in ${\mathbb
P}^3$ is in the strict Gorenstein liaison class of a point. Furthermore, if $n \ne 17,19$, it can be obtained by a sequence of ascending elementary $G$-biliaisons from a point.*]{}
[**Proof.**]{} Again we use ACM curves $(d,g)$ in ${\mathbb P}^3$, but now we need to know how many general points of ${\mathbb P}^3$ lie on such a curve. Call this number $m(d,g)$. This is no longer an elementary question, because the families of these curves form a Hilbert scheme, not a linear system. The question was studied in Perrin’s thesis [@P], and depends on semi-stability of the normal bundle. Here are his results, for the ACM curves we need:
$(d,g)$ $m$ reference in [@P]
--------- ----- -----------------------
(1,0) 2
(2,0) 3
(3,0) 6
(4,1) 8
(5,2) 9
(6,3) 12 0.15 in p. 66
(7,5) 14 0.15 in p. 87
(8,7) 16 0.15 in p. 10; p. 116
(9,9) 18 0.15 in p. 87
(10,11) 20 0.15 in p. 66
In this table $m$ has the naive value $2d$, except for $(2,0)$, a conic, which lies in a plane, so can pass through at most $3$ general points, and $(5,2)$, which lies on a quadric surface, so can pass through at most $9$ general points.
To prove our result, we use the following biliaisons and liaisons.
- $1 \leftrightarrow 2$ biliaison on $(1,0)$
- $1 \leftrightarrow 3$ biliaison on $(2,0)$
- $1,2,3 \leftrightarrow 4,5,6$ biliaison on $(3,0)$
- $3,4 \leftrightarrow 7,8$ biliaison on $(4,1)$
- $4 \leftrightarrow 9$ biliaison on $(5,2)$
- $4,5,6 \leftrightarrow 10,11,12$ biliaison on $(6,3)$
- $6,7 \leftrightarrow 13,14$ biliaison on $(7,5)$
- $8,9 \leftrightarrow 15,16$ biliaison on $(8,7)$
- $9 \leftrightarrow 18$ biliaison on $(9,9)$
- $12 \leftrightarrow 17$ liaison by $5H-K$ on $(9,9)$
- $11 \leftrightarrow 19$ liaison by $5H-K$ on $(10,11)$.
[**Remark 2.8.**]{} If we consider a set of 20 general points in ${\mathbb P}^3$, none of the above methods works. They do not lie on a cubic surface, so we cannot apply $(2.4)$. They do form a divisor $D$ on an ACM curve $(10,11)$, but $D-H$ has degree $10$, less than the genus, so it may not be effective. Liaison by $5H-K$ would give a divisor of degree $10$, which may not be effective. Liaison by $6H-K$ gives another general divisor of degree $20$, so we get nowhere.
It is conceivable that some upward liaison may eventually lead to a zero-scheme that can then be linked back down to a point. Or perhaps there are other AG schemes in ${\mathbb P}^3$ besides the ones of the form $mH-K$ on ACM curves that we have been using.
On the other hand, it may simply be that 20 general points in ${\mathbb P}^3$ are not in the $G$-liaison class of a point, so we propose this as a potential counterexample to Question $1.3$. If this is so, the cone over these 20 points would be an ACM curve in ${\mathbb P}^4$ that is not glicci.
[**Remark 2.9.**]{} In [@KMMNP $3.1$], the authors prove that any standard determinantal scheme is glicci. Taking the $t \times t$ minors of a $t \times
(t+2)$ matrix of linear forms in ${\mathbb P}^3$ gives a zero-dimensional determinantal scheme of degree $\frac {1}{6} (t+2)(t+1)t$, which is glicci by the above result. For $t = 1,2$, any set of $1$ or $4$ points in ${\mathbb
P}^3$ is determinantal. However, for $t = 3,4$, the dimension calculation in [@KMMNP $10.3$] show that determinantal sets of $10$ points have codimension $3$ in zero-schemes of degree $10$, and determinantal sets of $20$ points have codimension $15$ among zero-schemes of degree $20$.
ACM curves in ${\mathbb P}^4$
=============================
In the literature, a number of special cases of ACM curves have been shown to be glicci [@CM], [@KMMNP]. In this section we begin a systematic study of ACM curves of small degree in ${\mathbb P}^4$. We show that any general ACM curve of degree $\le 9$, or a general ACM curve of degree $10$ and genus $6$ can be obtained by ascending Gorenstein biliaisons from a line. On the other hand, we show that there is an irreducible component of the Hilbert scheme of curves of degree $20$ and genus $26$ whose general member is a smooth ACM curve that cannot be obtained by ascending Gorenstein biliaisons from a line. We propose this curve as a candidate for a possible counterexample to the question whether every ACM curve is glicci.
We start by finding a lower bound on the genus of an ACM curve in ${\mathbb
P}^4$.
[**Proposition 3.1.**]{} [*Let $C$ be a nondegenerate (i.e., not contained in a hyperplane) [*ACM*]{} curve in ${\mathbb P}^4$, of degree $d$ and arithmetic genus $g$. Then $d \ge 4$ and $g \ge G_{\mbox{\em min}}(d)$, where*]{} $$G_{\mbox{min}}(d) = (s-1)d - \binom{s+2}{3} - \binom{s+2}{4} + 1,$$ [*and $s \ge 2$ is the unique integer for which $\binom{s+2}{3} \le d <
\binom{s+3}{3}$. Furthermore, if $g = G_{\mbox{\em min}}(d)$, then $s =
s_0(c)$, the least degree of a hypersurface containing $C$.*]{}
[**Proof.**]{} The simplest way to see this is to consider the $h$-vector of the curve [@M §$1.4$]. This is a sequence of positive integers $c_0 = 1$, $c_1,c_2,\dots,c_r$, which determine the degree and genus of the curve according to the formulae $$d = \sum_{i=0}^r c_i,\hskip 0.5 in g = \sum_{i=2}^r (i-1)c_i.$$ The $c_i$ measure the Hilbert function of a graded ring $R = k[x_0,x_1,x_2]/J$ of finite length, since $C$ has codimension $3$. Since $R$ is a quotient of a polynomial ring in three variables, we have $c_i \le \binom{i+2}{2}$ for $i \ge
1$. The hypothesis $C$ nondegenerate implies $c_1 = 3$. Thus $d \ge 4$. For a given value of $d$,, the genus will be minimized by making each $c_i$ as large as possible for $i = 2,3,\dots$. Thus for $4 \le d < 10$ the minimum $g$ is attained by the $h$-vector $1,3,d-4$, with genus $g = d-4$. For $10 \le d <
20$, the minimum genus comes from the $h$-vector $1,3,6,d-10$, with $g = 2d -
14$. For the general case, a short calculation with binomial coefficients gives the formula above.
The least degree $s_0(c)$ of a hypersurface containing $C$ can be read from the $h$-vector as the least $i$ for which $c_i < \binom{i+2}{2}$. For the $h$-vectors giving the minimum genus, this is just the number $s$ defined above.
[**Remark 3.2.**]{} It seems reasonable to expect that for each $d \ge 4$ and $g
= G_{\mbox{min}}(d)$, the set of ACM curves of degree $d$ and genus $g$ in ${\mathbb P}^4$ should form an open subset of an irreducible component of the Hilbert scheme of curves in ${\mathbb P}^4$, and that a general such curve should be nonsingular, but we do not know how to prove this.
[**Notation 3.3.**]{} We will be dealing with curves on certain rational ACM surfaces in ${\mathbb P}^4$, so here we fix some terminology and notation.
The [*smooth cubic scroll*]{} $S$ is obtained by blowing up one point $P \in
{\mathbb P}^2$, and embedding in ${\mathbb P}^4$ by the complete linear system $H = 2l-e$, where $l$ is the total transform of a line in ${\mathbb P}^2$, and $e$ is the class of the exceptional divisor $E$. One knows that $\mbox{Pic } S
= {\mathbb Z} \oplus {\mathbb Z}$, generated by $l,e$. We denote the divisor class $al-be$ by $(a;b)$.
The [*Del Pezzo surface*]{} $S$ is obtained by blowing up five points $P_1,\dots,P_5$, no three collinear in ${\mathbb P}^2$, and embedding in ${\mathbb P}^4$ by $H = 3l - \Sigma e_i$. In this case $\mbox{Pic } S =
{\mathbb Z}^6$, and we denote the divisor class $al - \Sigma b_ie_i$ by $(a;b_1,\dots,b_5)$. If some $b$’s are repeated, we denote that with an exponent. Thus in the discussion of $(8,4)$ curves below, the divisor class $(5;2^2,1^3)$ means $(5;2,2,1,1,1)$.
A [*Castelnuovo surface*]{} $S$ in ${\mathbb P}^4$ is a smooth surface of degree $5$ and sectional genus $2$. It can be obtained by blowing up $8$ points $P_0,P_1,\dots,P_7$ in ${\mathbb P}^2$ and embedding by the linear system $H = (4;2,1^7)$ (see [@Okonek]). If the points $P_i$ are no three collinear and no $6$ on a conic, we call it a [*general*]{} Castelnuovo surface. Here $\mbox{Pic } S = {\mathbb Z}^9$, and we denote the divisor class $al - \Sigma b_ie_i$ by $(a;b_i)$.
A [*Bordiga surface*]{} is a smooth surface of degree $6$ and sectional genus $3$. It can be obtained by blowing up ten points $P_1,\dots,P_{10}$ in ${\mathbb P}^2$ and embedding by the linear system $H = (4;1^{10})$ [@Okonek]. If the points $P_i$ are such that no three are collinear, no $6$ on a conic, and no $10$ on a cubic curve, we call it a [*general*]{} Bordiga surface.
[**Proposition 3.4.**]{} a) [*If $C$ is an integral nondegenerate [*ACM*]{} curve of degree $d \le 9$ in ${\mathbb P}^4$, the degree-genus pair $(d,g)$ must be one of the following: $(4,0)$, $(5,1)$, $(6,2)$, $(7,3)$, $(8,4)$, $(8,5)$, $(9,5)$, $(9,6)$, $(9,7)$.*]{}
b\) [*For each $(d,g)$ pair in*]{} a), [*the set of nonsingular nondegenerate curves in ${\mathbb P}^4$ forms an open subset of an irreducible component of the Hilbert scheme, and*]{}
c\) [*For each $(d,g)$ pair as above, the general such curve is [*ACM*]{} and can be obtained by ascending Gorenstein biliaisons from a line.*]{}
[**Proof.**]{} a) A lower bound on $g$ is given by $(3.1)$; an upper bound is given by the Castelnuovo bound for the genus of an integral curve (see, e.g., Rathmann [@Rath]). This list gives all possible values of $g$ between the lower and upper bounds.
b\) For $g = d-4$, the irreducibility is given by a theorem of Ein [@Ein]. For $(d,g) = (8,5)$, $C$ is the canonical embedding of a non-hyperelliptic curve of genus $5$, so the family is irreducible. For $(d,g) = (9,6)$ th curve is a non-trigonal curve of genus $6$, embedded by a linear system $D = K-P$, so the family is irreducible (I am indebted to E. Drozd for this observation). For $(d,g) = (9,7)$, the family is irreducible by a theorem of Harris [@EH].
c\) We do a case-by-case analysis.
For $(d,g) = (4,0)$, up to automorphisms of ${\mathbb P}^4$, there is just one rational normal curve $C$ of degree $4$. It lies on a smooth cubic scroll $S$, having divisor class $(2;0)$. If $H$ denotes the hyperplane class $(2;1)$, then $C-H = (0;-1)$, which is a line. Thus $C$ is obtained by an ascending Gorenstein biliaison from a line on the surface $S$.
For $(d,g) = (5,1)$, suppose given a smooth nondegenerate $(5,1)$ curve $C$ in ${\mathbb P}^4$. Let $C_0$ be the abstract elliptic curve, and let $D_0$ be the divisor corresponding to ${\mathcal O}_C(1)$. Then $C$ is obtained by embedding $C_0$ with the complete linear system $|D_0|$. Choose $F$ a divisor of degree $3$ on $C_0$, and use $|F|$ to embed $C_0$ as a nonsingular cubic curve $C_1$ in ${\mathbb P}^2$. Choose a point $P \in C_1$. Blow up $P$ in ${\mathbb P}^2$ and embed by the linear system $H = 2l - e$ to get a nonsingular cubic scroll $S$ in ${\mathbb P}^4$. The image of $C_1$ will be a $(5,1)$ curve $C_2 \subseteq {\mathbb P}^4$, obtained by embedding $C_0$ with the linear system $|2F-P|$. By adjusting the choice of $P$, we may arrange that $D_0 \sim 2F-P$. Then $C$ and $C_2$ will differ by an automorphism of ${\mathbb P}^4$. We conclude that $C$ lies on a smooth cubic scroll $S'$, and has divisor class $(3;1)$ on $S'$. Then $C-H = (1;0)$ is a conic. The conic in turn can be obtained ascending Gorenstein biliaison from a line on a plane. Thus $C$ is obtained by ascending Gorenstein biliaisons from a line.
A similar argument shows that every smooth nondegenerate $(6,2)$ curve $C$ in ${\mathbb P}^4$ appears as a divisor of type $(4;2)$ on a smooth cubic scroll. Then $C-H = (2;1)$ is a twisted cubic curve, which can in turn be obtained by an ascending Gorenstein biliaison on a quadric surface in ${\mathbb P}^3$.
The case $(d,g) = (7,3)$ is a little more complicated. We will consider only a smooth nondegenerate non-hyperelliptic $(7,3)$ curve $C$ in ${\mathbb P}^4$. Let $C_0$ be the abstract curve of genus $3$, and $D_0$ the divisor giving the embedding $C$. Let $C_1$ be the canonical embedding of the non-hyperelliptic curve $C_0$ as a smooth plane quartic curve in ${\mathbb P}^2$. Choose five points $P_1,\dots,P_5$ on $C_1$ with no $3$ collinear. Blow up $P_1,\dots,P_5$ and embed by $3l - \Sigma e_i$ to get a Del Pezzo surface $S$ in ${\mathbb
P}^4$. The image of $C_1$ is then a smooth $(7,3)$ curve $C_2 \subseteq S$ with divisor class $(4;1^5)$, which is an embedding of $C_0$ by the divisor $3K
- \Sigma P_i$. We would like to choose the $P_i$ so that $3K - \Sigma P_i \sim
D_0$, i.e., $\Sigma P_i \sim 3K - D_0$.
[**Case 1.**]{} The divisor $3K - D_0$ can be represented by $5$ points, no three collinear. In this case we find that $C$ is contained in a Del Pezzo surface $S'$, with divisor class $(4;1^5)$. Then $C-H = (1;0^5)$, which is a twisted cubic curve, so $C$ can be obtained from a line by ascending Gorenstein biliaisons.
[**Case 2.**]{} If Case 1 does not occur, one sees easily that $3K-D_0 \sim K +
P$ for some point $P$. In this case we take $S$ to be the smooth cubic scroll obtained by blowing up $P$. Then $C_2$ is a divisor of type $(4;1)$ on $S$, and is an embedding of $C_0$ by $2K - P \sim D_0$. So $C$ lies on a cubic scroll, and $C - 2H = (0;-1)$ is a line.
Note the two types of non-hyperelliptic $(7,3)$ curves can be distinguished by the property that the Case 1 curves have only finitely many trisecants, while the Case 2 curves have infinitely many trisecants.
For a smooth $(8,4)$ curve $C$ we use a different technique. For this curve, $h^0({\mathcal I}_C(2)) = 2$, so $C$ is contained in a unique complete intersection surface $S = F_2 \cap F'_2$. Since the family of $(8,4)$ curves is irreducible, and since a general complete intersection surface $S = F_2 \cap
F'_2$ is a smooth Del Pezzo surface containing a smooth $(8,4)$ curve $C$ in the divisor class $(5;2^2,1^3)$, we conclude that a general such $C$ lies on a smooth Del Pezzo surface $S$, with divisor class $(5;2^2,1^3)$. Then $C-H =
(2;1^2,0^3)$, which is a nondegenerate smooth $(4,0)$ curve in ${\mathbb
P}^4$. Thus using the case of $(4,0)$ above, we conclude that $C$ can be obtained by ascending Gorenstein biliaisons from a line.
The case $(9,6)$ is similar to $(8,4)$, because again $h^0({\mathcal I}_C(2)) =
2$, we conclude that a general smooth $(9,6)$ curve $C$ lies on a Del Pezzo surface $S$ with divisor class $(6;2^4,1)$. In this case $C-2H = (0;0^4,-1)$ which is a line.
A smooth nondegenerate $(8,5)$ curve $C$ in ${\mathbb P}^4$ is the canonical embedding of a non-hyperelliptic genus $5$ curve. According to the theorem of Petri [@S-D] if the curve is not trigonal, then $C$ is the complete intersection of three quadric hypersurfaces $C = F_2 \cap F'_2 \cap F''_2$. Let $S = F_2 \cap F'_2$. Then $C$ is the divisor $2H$ on $S$, and $C-H$ is an elliptic quartic curve in ${\mathbb P}^3$, which can be obtained from a conic by biliaison on a quadric surface. Thus $C$ is obtained by ascending Gorenstein biliaisons from a line.
If the curve is trigonal, then $C$ lies on a smooth cubic scroll $S = F_2 \cap
F'_2 \cap F''_2$. It has divisor class $(5;2)$, so $C - 2H$ is $(1;0)$, a conic, and we conclude again. Note in this case $C = H - K$ is arithmetically Gorenstein, even though it is not a complete intersection.
For $(d,g) = (9,5)$, arguments like the ones above show that we can embed the general genus $5$ curve as a plane quintic with a double point, and thus obtain a general $(9,5)$ curve $C$ on a Castelnuovo surface with divisor class $(5;2,1^7)$. Then $C - H = (1,0^8)$ is a $(4,0)$ curve and we use our earlier result.
Finally, every smooth $(9,7)$ curve $C$, as a curve of maximal genus, lies on a cubic surface $S$ as the divisor $3H$, by a theorem of Harris [@EH]. Then $C-2H$ is a twisted cubic curve, and we are done.
[**Corollary 3.5.**]{} [*A general smooth [*ACM*]{} curve of degree $d \le 9$ in ${\mathbb P}^4$ is glicci.*]{}
[**Remark 3.6.**]{} The glicciness of ACM curves lying on general ACM surfaces in ${\mathbb P}^4$ and of integral ACM curves of degree $\le 7$ was already proven in [@KMMNP §8]. Our contribution is to show that a general smooth ACM curve of degree $\le 9$ actually does lie on a smooth rational ACM surface, and to check the possibly stronger property that they can be obtained from a line by ascending Gorenstein biliaisons (cf. Question $1.3$b).
Note that our proof actually shows [*every*]{} smooth curve with $(d,g) =
(4,0),(5,1),(6,2)$ is ACM and can be obtained by ascending Gorenstein biliaisons from a line. The same can be said for $(7,3)$ curves, by extending the analysis above: one can show that a hyperelliptic $(7,3)$ cuve lies on a cubic scroll or the cone over a twisted cubic curve, and in both cases is obtained from a smooth $(4,0)$ curve by biliaison.
For the next case of $(8,4)$ curves, the situation is more complicated. There are smooth hyperelliptic $(8,4)$ curves on a cubic scroll, but they are not ACM. Since an ACM $(8,4)$ curve lies on a unique complete intersection surface $S = F_2 \cap F'_2$, to study [*all*]{} smooth $(8,4)$ curves, one would presumably have to study the possible singular surfaces $S$. One approach is to use Riemann–Roch on the surface to show that the divisor $C-H$ is effective, but then one has to deal with not necessarily irreducible $(4,0)$ curves.
The analysis becomes increasingly complex for the remaining cases, so we do not know if c) holds for all smooth ACM curves of the given degree and genus.
[**Remark 3.7.**]{} For $d \ge 10$, the family of smooth non-degenerate ACM curves of given $(d,g)$ in ${\mathbb P}^4$ may not be irreducible. The first example is $(d,g) = (10,9)$, for which there are two different families of such curves lying on smooth cubic scrolls.
[**Example 3.8.**]{} We consider smooth $(10,6)$ curves in ${\mathbb P}^4$. Note that all the curves in Proposition $3.4$, being ACM of degree $\le 9$, are contained in quadric hypersurfaces, since their hyperplane section is $\le 9$ points and is contained in a quadric surface of ${\mathbb P}^3$. The case $(d,g) = (10,6)$ is the first case where there are smooth ACM curves not contained in a quadric hypersurface.
By the theorem of Ein [@Ein], the family of smooth $(10,6)$ curves in ${\mathbb P}^4$ is irreducible. To show that a general $(10,6)$ curve is ACM, it suffices, by semicontinuity, to exhibit one. A divisor of type $(5;1^{10})$ on a general Bordiga surface $S$ is the transform of a plane quintic curve, which can be taken to be smooth, so we get a smooth $(10,6)$ curve $C$ on $S$. For this curve $C-H = (1;0^{10})$ is a smooth $(4,0)$ curve, which is ACM, so $C$ is also ACM. Note however that the curve just described is not general in the variety of moduli of curves of genus $6$, because it has a $g_5^2$: a representation as a plane quintic curve.
Next, let $C_0$ be an abstract curve of genus $6$, with general moduli. Then $C_0$ admits a birational representation as a plane curve $C_1 \subseteq
{\mathbb P}^2$ with four nodes $P_1,P_2,P_3,P_4$, no three collinear [@ACGH]. Choose six additional points $P_5,\dots,P_{10}$ on $C_1$ in general position. Blow up $P_1,\dots,P_{10}$ to obtain a Bordiga surface $S$, containing the proper transform $C_2 \subseteq S$ of $C_1$. Then $C_2$ is a smooth $(10,6)$ curve in ${\mathbb P}^4$ with general moduli. Since the curve has genus $6$, by varying the choice of the six points $P_5,\dots,P_{10}$, we can obtain any general divisor class on $C_2$ as its hyperplane section. We conclude that the general $(10,6)$ curve $C$ in ${\mathbb P}^4$ is contained in a general Bordiga surface $S$ with divisor class $(6;2^4,1^6)$. Then $C-H
= (2;1^4,0^6)$, which is a smooth $(4,0)$ curve, so $C$ can be obtained by ascending Gorenstein biliaisons from a line.
To study the $(10,6)$ curves in more detail, we note that as a general Bordiga surface $S$, there are eight divisor classes (up to permutation of the $P_i$) containing $(10,6)$ curves. They are $$\begin{aligned}
D_1 &= &(5;1^{10}) \\
D_2 &= &(6;2^4,1^6) \\
D_3 &= &(7;2^9,0) \\
D_4 &= &(7;3,2^6,1^3) \\
D_5 &= &(8;3^3,2^6,1) \\
D_6 &= &(8;4,2^9) \\
D_7 &= &(9;3^6,2^4) \\
D_8 &= &(10;3^{10}).\end{aligned}$$ Of these $D_3$ and $D_6$ have Rao module $k$. They will be discussed in the next section. The remaining $6$ cases are ACM. The first three of these, $D_1$, $D_2$, and $D_4$, can be obtained by Gorenstein biliaison from $(4,0)$ curves on $S$. However, $D_5 - H$, $D_7 - H$, and $D_8 - H$ are not effective divisors so these curves cannot be obtained by Gorenstein biliaison on this surface $S$.
Using the arithmetically Gorenstein divisor $3H - K$ on $S$, of degree $20$, the divisor class $D_i$ is Gorenstein-linked to $D_{8-i}$. It follows that $D_5$, $D_7$, $D_8$ are glicci (as observed in [@KMMNP §8]). However, we do not know whether or not these curves may be obtained by ascending Gorenstein biliaison on some other surface.
[**Example 3.9.**]{} For our last example, we will study ACM $(20,26)$ curves in ${\mathbb P}^4$. Note that all the curves in the earlier part of this section, plus all the ACM curves lying on rational ACM surfaces in ${\mathbb
P}^4$, which were proved to be glicci in [@KMMNP §8], lie on cubic hypersurfaces in ${\mathbb P}^4$. So by analogy with our findings for points in ${\mathbb P}^3$ in §2 above, we might expect that all ACM curves contained in cubic hypersurfaces in ${\mathbb P}^4$ would be glicci. This also suggests that in looking for counterexamples to ACM $\Rightarrow$ glicci (Question $1.3$), we should look at curves not contained in a cubic hypersurface.
The first example of an ACM curve in ${\mathbb P}^4$ not contained in a cubic hypersurface will have $h$-vector $1,3,6,10$ (cf. proof of $3.1$). It has degree $20$ and genus $26$. For existence of such curves, we let $C$ be the determinantal curve defined by the $4 \times 4$ minors of a $4 \times 6$ matrix of general linear forms. A general such curve will be smooth, ACM, of degree $20$ and genus $26$. The family of such determinantal curves has dimension $\le 69$, by [@KMMNP 10.3].
The method of [@KMNP 3.7] shows that $C$ is linearly equivalent to $H+K$ on an ACM surface $S$ in ${\mathbb P}^4$, of degree $10$ and sectional genus $11$, defined by the $4 \times 4$ minors of a $4 \times 5$ matrix of general linear forms, where $H$ denotes the hyperplane section of $S$, and $K$ denotes the canonical class of $S$. Furthermore, a similar argument using [@KMNP 3.1] shows that the curve $C_0$ defined by the $3 \times 3$ minors of a $3 \times 5$ matrix of linear forms will be linearly equivalent to $K$ on $S$. This latter curve $C_0$ also appears in the divisor class $2H_0+K_0$ on the surface $S_0$ defined by $3 \times 3$ minors of a $3 \times 4$ matrix of linear forms. (I am grateful to J. Migliore for pointing out the paper [@KMNP] and explaining to me how to obtain these linear equivalences.)
Now $S_0$ is just the Bordiga surface, and $C_0$ is an ACM $(10,6)$ curve, discussed earlier. By the linear equivalence $C_0 \sim 2H_0 + K_0$ on $S_0$ we recognize that $C_0$ is in the class $(5;1^{10})$, which we called $D_1$ in $(3.8)$ above. These are isomorphic to plane curves of degree $5$ and thus are not general in the moduli of curves of genus $6$.
Since $C_0$ can be obtained by ascending Gorenstein biliaison from a line, and since $C \sim C_0 + H = K + H$ as the ACM surface $S$, we conclude that the determinantal $(20,26)$ curve $C$ can also be obtained by ascending Gorenstein biliaison from a line.
Next, I claim that the only way to obtain an ACM $(20,26)$ curve $D$ in ${\mathbb P}^4$ by ascending Gorenstein biliaison is from an ACM $(10,6)$ curve $C_1$ on an ACM surface $S_1$ of degree $10$ and sectional genus $11$, as $D
\sim C_1 + H$ on $S_1$. Indeed, suppose that $D \sim C_1 + H$ on some ACM surface $S_1$. Then $C_1$ is an ACM curve of type $(d_1,g_1)$ in ${\mathbb
P}^4$, while $H$ is an ACM curve of type $(d_2,g_2)$ in ${\mathbb P}^3$. From this we get $(20,26) = (d_1 + d_2, g_1 + g_2 + d_1 -1)$. For each $d_1$ (resp. $d_2$) we know the minimum possible genus of an ACM curve in ${\mathbb
P}^4$ (resp. ${\mathbb P}^3$)—cf. $3.1$. Looking at these, we find that $g_1 + g_2 + d_1 - 1 > 26$ in all cases except $(d_1,g_1) = (10,6)$ and $(d_2,g_2) = (10,11)$. Thus any $(20,26)$ curve $D$ that can be obtained by ascending Gorenstein biliaison must lie on a surface $S_1$ of degree $10$ and sectional genus $11$.
Now we look at the dimensions of some families of $(20,26)$ curves. By [@KMMNP $10.3$], the family of determinantal curves $C$ as above has dimension $\le 69$. On the other hand, each component of the Hilbert scheme of $(20,26)$ curves in ${\mathbb P}^4$ has dimension $\ge 5d + 1-g = 75$. So we see immediately that a general element of an irreducible component of $H_{20,26}$ cannot be determinantal. However, there may be other $(20,26)$ curves $C'$ on $S$, not determinantal themselves, but linearly equivalent to $C$, obtainable by ascending Gorenstein biliaison on $S$.
So let us find the dimension of the complete linear system $|C|$ on $S$. From the exact sequence $${\mathcal O} \rightarrow {\mathcal O}_S \rightarrow {\mathcal O}_S(C)
\rightarrow {\mathcal O}_C(C)
\rightarrow 0$$ we see that $\dim_S|C| = h^0({\mathcal O}_C(C)) = C^2 + 1-g + h^1({\mathcal
O}_C(C))$. We also have a resolution of ${\mathcal O}_S$ $$0 \rightarrow {\mathcal O}_{{\mathbb P}^4}(-5)^4 \rightarrow {\mathcal
O}_{{\mathbb P}^4} (-4)^5 \rightarrow {\mathcal O}_S \rightarrow 0$$ coming from its matrix representation. From this we find $h^2({\mathcal O}_S)
= 4$ and $p_a(S) = 4$. On the surface $S$ we have $H^2 = {\mathop{\mathrm{deg}}\nolimits}S = 10$. From the adjunction formula for $H$, which is a $(10,11)$ curve, we find $H\cdot K =
10$. And from the formula [@AG p. 434] for surfaces in ${\mathbb
P}^4$, we find $K^2 = 5$. Now $C = H + K$, so we get $C^2 = 35$. Also, since $C = H + K$, from the Kodaira Vanishing Theorem we have $H^1({\mathcal O}_S(C))
= H^2({\mathcal O}_S(C)) = 0$. Thus $h^1({\mathcal O}_C(C)) \cong
h^2({\mathcal O}_S) = 4$. So we find $$\dim_S|C| = 35 + 1 - 26 + 4 = 14.$$ The family of ACM surface $S$ has dimension $60$ (for example by the formula of Ellingsrud [@E]), so we find that the family of $(20,26)$ curves in ${\mathbb P}^4$ that can be linearly equivalent to $C$ on such a surface $S$ has dimension $\le 74$. In particular, a general curve in an irreducible component of $H_{20,26}$ does not arise in this way.
There may also be other linear equivalence classes of $(20,26)$ curves on $S$ of larger dimension. Indeed, this is what does happen with the $(10,6)$ curves studied in $(3.8)$ above: the curves linearly equivalent to the determinantal curves were all of type $D_1$ on the Bordiga surface, and these curves are not general in the moduli of genus $6$ curves, while a general genus $6$ curve appears as an ACM curve in a different linear system $D_2$ on the Bordiga surface. So we must see if something analogous happens with the $(20,26)$ curves.
First, we look on a general ACM surface $S$ of degree $10$ and sectional genus $11$. According to a theorem of Lopez [@Lop III.4.2], $\mbox{Pic } S =
{\mathbb Z} \oplus {\mathbb Z}$ generated by $H$ and $K$. We look for divisors $mH + nK$ with degree $20$ and genus $26$. There are only two possibilities: $C = H + K$ or $C' = 4H - 2K$. In the latter case we compute $C'{}^2 = 20$. Therefore, by Clifford’s theorem, $h^0({\mathcal O}_{C'}(C')) -
1 \le 10$, so $\dim_S|C'| \le 11$. Thus the family of such curves $C'$ in ${\mathbb P}^4$ has dimension $\le 71$. So we see that a general ACM $(20,26)$ curve in ${\mathbb P}^4$ cannot lie on a [*general*]{} ACM surface $S$ of degree $10$ and sectional genus $11$.
Now let us estimate the dimension of a family of smooth $(20,26)$ curves $D$, general in an irreducible component of $H_{20,26}$ containing the determinantal curves $C$ above, and lying on a non-general ACM surface $X$ of degree $10$ and sectional genus $11$. We will make use of the Clifford index of a curve.
Recall that the [*gonality*]{} of a curve $C$ is the least $d$ for which there exists a linear system $g^1_d$ on the curve. The [*Clifford index*]{} of the curve is the minimum of $d-2r$, taken over all linear systems $g^r_d$ with $r
\ge 1$ and $0 < d \le g-1$. For most curves, the Clifford index is equal to $\mbox{gon}(C) - 2$, computed by a $g^1_d$. Curves for which this is not so are [*Clifford exceptional*]{} curves, and have been studied by Martens [@Mar] and Eisenbud et al. [@ELMS].
If $C$ is the determinantal $(20,26)$ curve studied above, then $C \sim H + K$ on the surface $S$. The hyperplane section $H$ is a $(10,11)$ curve in ${\mathbb P}^3$, obtained as $C_0 + H_0$ on a nonsingular quartic surface in ${\mathbb P}^3$. Here $C_0$ is a nonhyperelliptic $(6,3)$ curve having gonality $3$; $H_0$ is a plane quartic curve, also having gonality $3$. Hence, by [@CIA], $H$ has gonality $\ge 6$. (In fact the gonality is equal to $6$ because $H$ must have a $4$-secant.) The curve $K$ on $S$ is the determinantal $(10,6)$ curve discussed above, isomorphic to a plane quintic curve, with gonality $4$. So applying [@CIA] again, we find the gonality of $C$ is $\ge 10$. It follows from the study of Clifford exceptional curves in [@Mar] and [@ELMS] that $C$ is not exceptional, so we conclude $\mbox{Cliff } C \ge 8$. (On the other hand, the linear system $|K|$ on $S$ cuts out a $g_{15}^3$ on $C$, so $\mbox{Cliff } C \le 9$. I suspect $\mbox{Cliff } C = 9$, but don’t know how to prove that.) Since we are considering a curve $D$ that is general in an irreducible component of $H_{20,26}$ containing $C$, we may assume also that $\mbox{Cliff } D \ge 8$.
Now we consider a smooth $(20,26)$ curve $D$ with $\mbox{Cliff } D \ge 8$, contained in a smooth ACM surface of degree $10$ and sectional genus $11$ in ${\mathbb P}^4$, and we want to estimate the dimension of the linear system $|D|$ on $S$. As above, we find $$\dim_S|D| = D^2 + 1 - g + h^1({\mathcal O}_D(D)).$$ Since $D$ is a $(20,26)$ curve, the adjunction formula gives $$D^2 + D \cdot K = 50.$$ Let us denote $D \cdot K$ by $b$. Then $D^2 = 50 - b$. On the other hand, let us consider the linear system $|D \cdot K|$ on $D$. It has dimension $a-1$, where $a = h^0({\mathcal O}_D(K))$. Since $K_D = (D+K)\cdot D$, we also have $h^1({\mathcal O}_D(D)) = a$. The linear system $|D\cdot K|$ thus has dimension $a-1$ and degree $b$. Our hypothesis $\mbox{Cliff } D \ge 8$ thus implies $b - 2a + 2 \ge 8$, or $b \ge 2a + 6$.
Now we can compute $$\begin{aligned}
\dim_S|D| &= &D^2 + 1 - g + h^1({\mathcal O}_D(D)) \\
&= &50 - b + 1 - 26 + a \\
&= &25 + a - b.\end{aligned}$$ Since $b \ge 2a + 6$, we find $$\dim_S|D| \le 19 - a.$$ Now from the exact sequence $$0 \rightarrow {\mathcal O}_X(K-D) \rightarrow {\mathcal O}_X(K) \rightarrow
{\mathcal O}_D(K) \rightarrow 0,$$ we find $a = h^0({\mathcal O}_D(K)) \ge h^0({\mathcal O}_X(K)) = h^2({\mathcal
O}_X) = 4$. Thus $$\dim_S|D| \le 15.$$ On the other hand, our surface $S$ is not general, so it moves in a family of dimension at most $59$, so the dimension of the family of curves that arise in this way is at most $74$.
In conclusion, we see that there exists an irreducible component of the Hilbert scheme $H_{20,26}$ of smooth $(20,26)$ curves in ${\mathbb P}^4$ (namely one containing the determinantal curves) whose general member is an ACM curve that does not lie on an ACM surface $S$ of degree $10$ and sectional genus $11$, and so cannot be obtained by ascending Gorenstein biliaison from a line. We propose this curve as a possible candidate for a counterexample to ACM $\Rightarrow$ glicci (Question $1.3$).
Curves in ${\mathbb P}^4$ with Rao module $k$
=============================================
Let ${\mathcal M}$ be the set of all locally CM curves in ${\mathbb P}^4$ with Rao module $k$ (i.e., of dimension one in one degree only). One knows that the Rao module must occur in a nonnegative degree [@M $1.3.11$(b)], and that there are curves with Rao module $k$ in degree $0$ (e.g., two skew lines). So we denote by ${\mathcal M}_h$ the set of curves with Rao module $k$ in degree $h$, and note that ${\mathcal M} = \cup_{h \ge 0} {\mathcal M}_h$.
Let ${\mathcal L} \subseteq {\mathcal M}$ be the subset of those curves in the $G$-liaison class of two skew lines, and let ${\mathcal L}_h = {\mathcal L}
\cap {\mathcal M}_h$. Then ${\mathcal L} = \cup_{h \ge 0} {\mathcal L}_h$, and the curves in ${\mathcal L}_0$ are the [*minimal*]{} curves defined in §1 above.
In this section we will study the curves in ${\mathcal M}$, with a view to elucidating Questions $1.4$ and $1.5$ above.
[**Proposition 4.1.**]{} a) [*${\mathcal M}_0$ contains curves of every degree $d \ge 2$.*]{}
b\) [*For each $d \ge 2$, the set of curves in ${\mathcal M}_0$ of degree $d$ forms an irreducible family, whose general member is the disjoint union $C
= C' \cup L$ of a plane curve $C'$ of degree $d-1$ and a line $L$, not meeting the plane of $C'$.*]{}
c\) [*Every curve in ${\mathcal M}_0$ is in the $G$-liaison class of two skew lines, i.e. ${\mathcal L}_0 = {\mathcal M}_0$.*]{}
[**Proof.**]{} a) The case of two skew lines in ${\mathbb P}^3$ is well-known [@MDP Example $6.2$, p. 34]. For $d \ge 3$, let $C = C' \cup L$ as described in b). Clearly $h^0({\mathcal O}_C) = 2$, so $h^1({\mathcal I}_C) =
1$. On the other hand, since $C'$ and $L$ are contained in disjoint sublinear spaces of ${\mathbb P}^4$, it is clear that $H^0({\mathcal O}_{{\mathbb P}^4}(n)) \rightarrow
H^0({\mathcal O}_C(n))$ is surjective for $n \ge 1$, so $C \in {\mathcal M}_0$.
b\) I claim any degree $2$ curve $C$ in ${\mathbb P}^4$ with $M = k$ lies in ${\mathbb P}^3$. If the curve is reduced, it is two lines, hence in a ${\mathbb P}^3$. If it is not reduced, then it is a double structure on a line $L$, and we have an exact sequence $$0 \rightarrow {\mathcal L} \rightarrow {\mathcal O}_C \rightarrow {\mathcal
O}_L \rightarrow 0$$ where ${\mathcal L}$ is an invertible sheaf on $L$. Then there is a surjective map $u: {\mathcal I}_L/{\mathcal I}_L^2 \rightarrow {\mathcal L} \rightarrow
0$, and ${\mathcal L} \cong {\mathcal O}_L(a)$ for some $a$. Since ${\mathcal
I}_L/{\mathcal I}_L^2 \cong {\mathcal O}_L(-1)^3$, the map $u$ is given by three sections of ${\mathcal O}_L(a+1)$. If $a = -1$, we get a double line in a plane. If $a=0$, there is a linear form $x$ killed by $u$, so $C$ lies in the ${\mathbb P}^3$ defined by $x=0$. If $a > 0$, then the exact sequence $$H^0({\mathcal O}_L(-1))^3 \rightarrow H^0({\mathcal O}_L(a)) \rightarrow
H^1({\mathcal I}_C) \rightarrow 0$$ shows the Rao module is bigger than $k$.
Thus a curve of degree $2$ with $M=k$ lies in a ${\mathbb P}^3$, so these form an irreducible family whose general member is two skew lines.
So now let $d \ge 3$. Then $C$ cannot be contained in ${\mathbb P}^3$, because of the Lazarsfeld–Rao property for curves in ${\mathbb P}^3$, so $h^0({\mathcal O}_C) = 2$, because of the Rao module, and $h^0({\mathcal
O}_C(1)) = 5$. Let $A = H_*^0({\mathcal O}_C)$. This is a graded $S$-algebra, where $S = k[x_0,x_1,x_2,x_3,x_4]$, and in particular $A_0$ is a $2$-dimensional $k$-algebra. We consider two cases.
[**Case 1.**]{} $A_0$ is reduced, hence isomorphic to $k \times k$ as a $k$-algebra. Then $A_0$ contains two orthogonal idempotents $e',e''$, such that $e' + e'' = 1$, $e'{}^2 = e'$, $e''{}^2 = e''$, and $e'e'' = 0$. Hence $C$ is the disjoint union of two curves $C',C''$, defined by the vanishing of $e',e''$, respectively. Let $H',H''$ be the linear spans of the curves $C',C''$. Then $h^0({\mathcal O}_C(1)) = h^0({\mathcal O}_{H'}(1)) +
h^0({\mathcal O}_{H''}(1)) = 5$. So one of these, say $H'$, is a plane, and the other, $H''$ is a line $L$. Thus $C'$ is a plane curve in $H'$, and $C =
C' \cup L$ as required. Note that $H',H''$ do not meet since $h^0({\mathcal
O}_{{\mathbb P}^4}(1)) = h^0({\mathcal O}_{H'}(1)) + h^0({\mathcal
O}_{H''}(1))$.
[**Case 2.**]{} $A_0$ is non-reduced, in which case it is isomorphic to the ring $k[\epsilon]/(\epsilon^2)$. Let $f \in A_0$ be a nonzero element with $f^2 =
0$. Now $A_1 \cong S_1$ is the $k$-vector space generated by $x_0,x_1,x_2,x_3,x_4$. Multiplication by $f$ on $A_1$ is a nilpotent linear map with $f^2 = 0$. Furthermore, since $C$ is locally CM, the kernel of $f$ acting on $A_1$ must have dimension $\le 3$. Otherwise $f$ would be supported at a point. So $f$ has rank $\ge 2$. Now from the structure of nilpotent transformations it follows (after a linear change of coordinates) that $fx_0 =
x_2$, $fx_1 = x_3$, $fx_2 = fx_3 = fx_4 = 0$. Hence we can identify the $S$-algebra $A$ as $$A \cong S[f]/((f^2,fx_0-x_2,fx_1-x_3,fx_2,fx_3,fx_4) + I_C)$$ where $I_C \subseteq S$ is the homogeneous ideal of $C$.
Now let $H'$ be the plane $x_2 = x_3 = 0$, and let $C'$ be the curve obtained from $C \cap H$ by removing its embedded points, if any. Then there is an exact sequence $$0 \rightarrow {\mathcal L} \rightarrow {\mathcal O}_C \rightarrow {\mathcal
O}_{C'} \rightarrow 0.$$ Since $C'$ is a plane curve, $h^0({\mathcal O}_{C'}) = 1$, and so $h^0({\mathcal L}) = 1$. Furthermore note that the image of $f$ in ${\mathcal
O}_{C'}$ is annihilated by $x_0,x_1,x_2,x_3,x_4$, hence is $0$. So $f$ generates $h^0({\mathcal L})$. Now $f$ is annihilated by $x_2,x_3,x_4$, so it has support on the line $L: x_2 = x_3 = x_4 = 0$. Thus ${\mathcal L}$ is an ${\mathcal O}_L$-module, it is torsion-free since $C$ is locally CM, and contains the submodule ${\mathcal O}_L$ generated by $f$. Hence ${\mathcal L}
\cong {\mathcal O}_L$, generated by $f$.
Now it is clear that $C$ consists of the plane curve $C'$ of degree $d-1$, containing the line $L$, plus a multiplicity two structure on $L$ with $p_a =
-1$. This is the limit of a flat deformation of the disjoint unions $C' \cup
L$ described above, as the skew line $L$ approaches a line in the curve $C'$.
So the curves in ${\mathcal M}_0$ of any degree $d \ge 2$ form an irreducible family.
c\) Let $C \in {\mathcal M}_0$ have degree $d$. The case $d = 2$ in ${\mathbb
P}^3$ is well-known, so we may assume $d \ge 3$. First consider the disjoint union $C = C' \cup L$ as in b). Take a hyperplane ${\mathbb P}^3$ containing $L$ and meeting the plane $H'$ of $C'$ in a line $L'$, skew to $L$, and not a component of $C'$. Let $Q$ be a nonsingular quadric surface in that ${\mathbb P}^3$ containing $L$ and $L'$. Then $S = H' \cup Q$ is an ACM surface of degree $3$ in ${\mathbb P}^4$. Note that its negative canonical divisor $-K$ consists of a conic in $H'$ plus a divisor of bidegree $(1,2)$ on $Q$, meeting $L'$ in the same two points as the conic, where $(1,0)$ is the class of $L$. (Here we leave some details to the reader.) Now given $C$, there is an AG divisor $X$ in the linear system $(d-3)H-K$ on $S$ containing $C$ [@M $4.2.8$]. The linked curve $D$ is a divisor of bidegree $(d-3,d-1)$ on $Q$, which is in the biliaison class of two skew lines on $Q$. Thus $C$ is in ${\mathcal L}_0$.
In the special case where $C$ is a plane curve $C'$ containing a line $L$, plus a double structure on $L$ as above, we use exactly the same construction, except that now the hyperplane ${\mathbb P}^3$ meets $H'$ in $L$, and the quadric surface $Q$ contains the double structure on $L$. The same liaison works, using the theory of generalized divisors [@GD].
[**Remark 4.2.**]{} The fact that ${\mathcal L}_0$ is not a single irreducible family was observed by Migliore [@M $5.4.8$], who gave the example of a curve of degree $3$ in ${\mathcal L}_0$. His student Lesperance [@Les] has independently proved $3.1$a), c) in the case of reduced curves. Lesperance has also shown [@Les $4.5$] that for other Rao modules, the set of minimal curves of given degree need not be irreducible. Thus $(4.1b)$ is special to the case of Rao module $M=k$.
[**Example 4.3.**]{} Let $C$ be a smooth curve of degree $5$ and genus $0$ in ${\mathbb P}^4$, not contained in any ${\mathbb P}^3$. It is the projection of the rational normal curve $\Gamma$ in ${\mathbb P}^5$ from a point not lying on any secant line of $\Gamma$. A little elementary geometry shows that $C$ has a unique trisecant $E$. If $C$ meets $E$ in three distinct points, then the three points of intersection of $C$ and $E$ determine a unique isomorphism (of abstract ${\mathbb P}^1$’s) from $C$ to $E$ fixing those three points. Let $S$ be the surface formed as the closure of the set of lines joining corresponding points of $C$ and $E$. Then $S$ is a rational cubic scroll in ${\mathbb P}^4$.
On $S$, our rational quintic $C$ has divisor class $(4;3)$. The linear system $C-H = (2;2)$ contains a disjoint union of two rulings of the surface $S$. Hence $C$ is obtained by one elementary $G$-biliaison from two skew lines. In particular, $C \in {\mathcal L}_1$.
If $E$ is a degenerate trisecant, i.e., a tangent line meeting the curve again, or an inflectional tangent, we can still show $C \in {\mathcal L}_1$ as follows. The smooth $(5,0)$ curves in ${\mathbb P}^4$ form an irreducible family, so $C$ is a specialization of the general type described above. Hence $C$ must lie on a cubic surface in ${\mathbb P}^4$. It cannot lie on a reducible surface, since $C$ is not in ${\mathbb P}^3$. The only other irreducible cubic surface is the cone over a twisted cubic curve, and that surface contains no smooth $(5,0)$ curves. Hence $C$ is on a smooth rational cubic scroll, and the previous argument applies.
[**Example 4.4.**]{} We consider smooth curves of type $(6,1)$ (degree $6$ and genus $1$) in ${\mathbb P}^4$, not contained in any ${\mathbb P}^3$. Then $h^1({\mathcal I}_C(1)) = 1$ and $h^0({\mathcal I}_C(2)) \ge 3$.
[**Case 1.**]{} If three quadric hypersurfaces containing $C$ intersect in a surface, then that surface must be a cubic rational scroll $S$ (reason: the degree of $S$ must be $\le 3$; $C$ is not contained in a plane or a quadric surface, and there is no $(6,1)$ curve on the cone over a twisted cubic curve). In this case $C = (3;0)$ on $S$ and $C-H = (1;-1)$, which contains the disjoint union of a conic and a line. Thus $C$ is in ${\mathcal L}_1$ and is obtained by a single elementary $G$-biliaison from a curve of degree $3$ in ${\mathcal L}_0$. This curve $C$ has infinitely many trisecants, formed by the rulings of $S$.
[**Case 2.**]{} Three quadric hypersurfaces containing $C$ will intersect in a complete intersection curve $X$ of degree $8$ and genus $5$. The residual intersection $D$ will be a curve of degree $2$. $D$ cannot be a plane curve, because then it would meet $C$ in $5$ points, and projection from the plane of $D$ would be a birational map of $C$ to a line, which is impossible. Hence $D$ is two skew lines or a nonplanar double structure on a line. By reason of the genus of $X$, $D$ will be either two trisecants of $C$ or a single trisecant. Note also that these are all the trisecants of $C$, because any trisecant of $C$ must be contained in each quadric hypersurface containing $C$, hence in $X$.
[**Case 2a.**]{} An example of a $(6,1)$ curve with two trisecants can be obtained on a Del Pezzo surface, as a divisor of type $(3;1^3,0^2)$. In this case $C - H = (0;0^3,-1^2)$, which is a disjoint union of two lines, so $C$ is in ${\mathcal L}_1$ and is obtained by one elementary $G$-biliaison from the minimal curve of degree $2$ in ${\mathcal L}_0$. This curve $C$ has two trisecants, the lines $F_{45} = (1;0^3,1^2)$ and $G = (2;1^5)$.
[**Case 2b.**]{} An example of a $(6,1)$ curve with one trisecant can be obtained as follows. We project the Veronese surface $V$ in ${\mathbb P}^5$ from a point in a plane containing a conic of $V$, so as to obtain a quartic surface $S$ in ${\mathbb P}^4$ with a double line $L$. A general cubic curve in ${\mathbb P}^2$ gives a $(6,1)$ curve in $V$ meeting the conic in three points that project to distinct points of the line $L$ in $S$. Thus the image $C \subseteq S$ of this curve will be a smooth $(6,1)$ curve having $L$ as a trisecant. Now the surface $S$ is smooth except for a double line and two pinch points, hence locally $CM$. Its general hyperplane section is an integral curve in ${\mathbb P}^3$ of degree $4$, arithmetic genus $1$, with one node. This is a complete intersection in ${\mathbb P}^3$, hence $S$ is a complete intersection of two quadric hypersurfaces in ${\mathbb P}^4$ [@M $1.3.3$], so it must contain every trisecant of $C$. But $S$ contains no lines except $L$, so $C$ has a unique trisecant. Since $C$ is linked to a double structure on $L$, $C$ is in the $CI$-liaison class of two skew lines, so $C$ is in ${\mathcal L}_1$. Note that $C-H$ is not effective on $S$, so $C$ cannot be obtained by an elementary Gorenstein biliaison on $S$. However, it seems likely that $C$ also lies on a normal singular Del Pezzo surface on which it can be obtained by an elementary Gorenstein biliaison from two skew lines.
Thus we see that any smooth nondegenerate $(6,1)$ curve in ${\mathbb P}^4$ is in ${\mathcal L}_1$. The family of all such curves in ${\mathbb P}^4$ is irreducible [@Ein]. The general type with two trisecants (Case 2a) is obtained by an elementary Gorenstein biliaison from a curve of degree $2$ in ${\mathcal L}_0$, while the special type (Case 1) with infinitely many trisecants is obtained by Gorenstein biliaison from a curve of degree $3$ in ${\mathcal L}_0$.
[**Example 4.5.**]{} We consider nonsingular degree $7$ genus $2$ curves in ${\mathbb P}^4$, not contained in any ${\mathbb P}^3$. The family $H_{7,2}$ of all of these curves is irreducible, by Ein [@Ein]. We see $h^1({\mathcal
I}_C(1)) = 1$, and there exist such curves with Rao module $k$ on a Del Pezzo surface (see below), so by semicontinuity, the general such curve has Rao module $k$, i.e., it is in ${\mathcal M}_1$.
Next, note that $h^0({\mathcal I}_C(2)) \ge 2$. If $h^0({\mathcal I}_C(2)) >
2$, then the intersection of three quadric surfaces would either be a curve of degree $8$, and then $C$ would be linked to a line, hence ACM, which is impossible; or it would be a surface of degree $3$, but there are no $(7,2)$ curves on surfaces of degree $3$ in ${\mathbb P}^4$. Hence $h^0({\mathcal
I}_C(2)) = 2$, and $h^1({\mathcal I}_C(2)) = 0$, so by Castelnuovo–Mumford regularity, $C \in {\mathcal M}_1$. Thus all curves of $H_{7,2}$ are in ${\mathcal M}_1$.
Now look on a Del Pezzo surface $S$, and let $C = (4;2,1^3,0)$. Then $C$ is a smooth $(7,2)$ curve, and $C - H = (1;1,0^3,-1)$ is a disjoint union of a conic and a line. Thus $C \in {\mathcal L}_1$, and $C$ is obtained by an elementary Gorenstein biliaison from a curve of degree $3$ in ${\mathcal L}_0$. Note that $C$ has exactly four mutually skew trisecants, namely the lines $F_{25}$, $F_{35}$, $F_{45}$, and $G$ on $S$.
If $C$ is any smooth $(7,2)$ curve, we have seen that $h^0({\mathcal I}_C(2)) =
2$. Let $S$ be the complete intersection surface $F_2 \cdot F'_2$ of two quadric surfaces containing $C$. Then $S$ is uniquely determined by $C$. It is a surface of degree $4$, with sectional genus $1$, but it may be singular. However, it must be irreducible, and hence has at most a line of singular points. Therefore $C$ is an almost Cartier divisor on $S$, and we can apply the theory of generalized divisors. There is an exact sequence $$0 \rightarrow {\mathcal O}_S \rightarrow {\mathcal L}(C) \rightarrow
\omega_C(1) \rightarrow 0$$ [@GD $2.1$], making use of the fact that $\omega_S = {\mathcal O}_S(-1)$. Twisting by $-1$, and taking cohomology, we obtain $$0 \rightarrow H^0({\mathcal O}_S(-1)) \rightarrow H^0({\mathcal L}(C-H))
\rightarrow H^0(\omega_C) \rightarrow H^1({\mathcal O}_S(-1)),$$ where $H$ denotes the hyperplane class on $S$. The two outside groups are $0$, and $H^0(\omega_C)$ has dimension $2$, so $H^0({\mathcal L}(C-H)) \ne 0$. This shows that $C-H$ is effective on $S$. It is a divisor of degree $3$, and must be in ${\mathcal L}_0$, so we see that any $C$ in ${\mathcal M}_1$ is in ${\mathcal L}_1$, and is obtained by an elementary Gorenstein biliaison from a curve of degree $3$ in ${\mathcal L}_0$.
For an example of a special $(7,2)$ curve, let $S_0 = {\mathbb P}^1 \times
{\mathbb P}^1$. Let $\Gamma$ be a line of bidegree $(1,0)$, and fix an involution $\sigma$ on $\Gamma$. Take $\vartheta$ to be the linear system of those curves of bidegree $(1,2)$ on $S_0$ meeting $\Gamma$ in a pair of the involution $\sigma$. Then $\vartheta$ maps $S_0$ to a surface $S$ of degree $4$ in ${\mathbb P}^4$ with a double line $L_0$ (the image of $\Gamma$). If $C_0$ is a general curve of bidegree $(2,3)$ on $S_0$, then the image $C$ of $C_0$ in $S$ is a smooth $(7,2)$ curve meeting $L_0$ in three points. It has four trisecants, namely the double line $L_0$ and the three rulings (images of $(0,1)$ curves in $S_0$) that meet $L_0$ at the points where $C$ meets $L_0$. This curve is different from the general ones described above, because three of the trisecants meet the fourth one. Because of the general result above, $C$ must arise by an elementary Gorenstein biliaison on $S$, but in this case the curve of degree $3$ in ${\mathcal L}_0$ will be a nonreduced curve containing a double structure on the line $L_0$.
Next we look at a general Castelnuovo surface $S'$. On this surface, there are three different kinds of smooth $(7,2)$ curves, distinguished by their self-intersections, namely $$\begin{array}{rll}
C_1 &= &(4;2,1^5,0^2) \\
C_2 &= &(5;2^4,1^3,0) \\
C_3 &= &(5;1^4,2^4)
\end{array} \hskip 0.5 in
\begin{array}{rll}
C_1^2 &= &7 \\
C_2^2 &= &6 \\
C_3^2 &= &5.
\end{array}$$ Of these $C_1$ is obtained by an elementary Gorenstein biliaison on $S'$ from two skew lines, while $C_2 - H$ and $C_3 - H$ are not effective. Since we have seen above that every smooth $(7,2)$ curve arises by elementary Gorenstein biliaisons from a degree $3$ curve in ${\mathcal L}_0$, this gives examples of curves that may be obtained by two different routes by elementary Gorenstein biliaisons from curves of two different degrees in ${\mathcal L}_0$.
In fact, I claim that every general $(7,2)$ curve arises also as a curve of type $C_1$ on a smooth Castelnuovo surface. To prove this in detail is rather long, so I will just give a sketch. Start with a smooth $(7,2)$ curve $C$ on a smooth Del Pezzo surface $S$, say $C = (4;2,1^3,0)$ as before. Choose a twisted cubic curve $D$ and a conic $\Gamma$ so that $C + D + \Gamma = 3H$. (For example $D = (3;1^4,2)$ and $\Gamma = (2;0,1^4)$.) Let $\Pi$ be the plane containing $\Gamma$. Then the two quadric hypersurfaces containing $C$ meet $\Pi$ in $\Gamma$, so a linear combination of them contains $\Pi$. So we may assume $S = F_2 \cdot F'_2$ where $F_2$ contains $\Pi$. By construction, there are cubic hypersurfaces $F_3$ containing $C+D+\Gamma$. Such an $F_3$ will meet $\Pi$ in $\Gamma$ plus a line. Adjusting $F_3$ by a linear form times $F'_2$, we may assume that $F_3$ contains $\Pi$. Now $F_2 \cdot F_3 =
\Pi \cup S'$, where $S'$ is an ACM surface of degree $5$, hence a Castelnuovo surface. Now one can verify that $C$ on $S'$ is a curve with self-intersection $7$, like $C_1$ above, and that $C_1-H$ is effective and represented by a curve of degree $2$ in ${\mathcal L}_0$.
In conclusion, we see that every smooth $(7,2)$ curve is in ${\mathcal L}_1$, and can be obtained by elementary Gorenstein biliaison from ${\mathcal L}_0$, in general by two different routes. This is in contrast to the $(6,1)$ case above, where the curves are divided into two types, distinguished by which component of ${\mathcal L}_0$ they arise from.
[**Example 4.6.**]{} We consider smooth $(10,6)$ curves in ${\mathbb P}^4$ (cf.Example $3.8$ above).
By Riemann–Roch applied to ${\mathcal O}_C(1)$ we see that an ACM $(10,6)$ curve is nonspecial. Also we see that ${\mathcal O}_C(1)$ is special if and only if ${\mathcal O}_C(1)$ is a canonical divisor, and this is equivalent to $h^1({\mathcal I}_C(1)) = 1$. The $(10,6)$ curves in ${\mathbb P}^4$ with ${\mathcal O}_C(1)$ special are all projections of the canonical curves of genus $6$ in ${\mathbb P}^5$. Since these form an irreducible family, we see that their projections, the canonical $(10,6)$ curves in ${\mathbb P}^4$, form an irreducible family, and they all have $h^1({\mathcal I}_C(1)) = 1$. A general such curve has Rao module $k$ in degree $1$, i.e., it is in ${\mathcal
M}_1$. To see this, by semicontinuity, it is sufficient to exhibit one such. Let $C = (7;2^9,0)$ on a general Bordiga surface. This curve is the image of a plane septic curve with $9$ double points, embedded in ${\mathbb P}^4$ by the linear system of quartics passing through the double points (and one further point). These are adjoint curves to $C$ and so cut out the canonical linear series. Hence $C$ is a canonical curve. Now $C-H = (3;1^9,-1)$ is the disjoint union of a plane cubic curve and a line, which is a degree $4$ curve in ${\mathcal L}_0$. Hence $C \in {\mathcal L}_1$ is obtained by an elementary Gorenstein biliaison from ${\mathcal L}_0$.
Now let us consider $(10,6)$ curves in ${\mathcal M}_2$, i.e., with Rao module $k$ in degree $2$. Examples of such can be found on a general Castelnuovo surface $S$, for example $C_1 = (6;3,2,1^6)$ and $C_2 = (6;2^4,1^4)$. Note that the curves $C_1$ are trigonal, while the curves $C_2$ can have general moduli. Both types can be obtained by Gorenstein biliaison on $S$, since $C_1
- H = (2;1^2,0^6)$ and $C_2 - H = (2;0,1^3,0^4)$ are both $(5,0)$ curves, hence in ${\mathcal L}_1$.
An examination of curves of minimal genus in ${\mathcal L}_1$ and ACM curves of minimal genus in ${\mathbb P}^3$ shows that the only way to obtain a $(10,6)$ curve in ${\mathcal M}_2$ by Gorenstein biliaison is from a $(5,0)$ curve in ${\mathcal L}_1$ on a surface of degree $5$ and sectional genus $2$ in ${\mathbb P}^4$, like the Castelnuovo surfaces.
Another example of a $(10,6)$ curve in ${\mathcal M}_2$ is obtained by the curve $C$ formed by the intersection of a smooth quintic elliptic scroll $V$ with a hypersurface $F$ of degree $2$. If we take $F$ to be a smooth quadric hypersurface, then by Klein’s theorem [@AG II.Ex. $6.5$d] it contains no surfaces of odd degree, so $C$ cannot be obtained by ascending Gorenstein biliaison from a curve in ${\mathcal L}_1$.
However one can show that $C$ is in the $G$-liaison class of two skew lines by the following method, suggested by the referee. First note that two general cubic hypersurfaces $F_3,F'_3$ containing $V$ will link $V$ to be a Veronese surface $W$ in ${\mathbb P}^4$. Thus $C$ is linked by the complete intersection $F_2 \cap F_3 \cap F'_3$ to a curve $C' \subseteq W$, which is $W
\cap F_2$. The curve $C'$ is an $(8,3)$ curve, obtained from a plane curve of degree $4$ by the $2$-uple embedding of ${\mathbb P}^2$ and projection to ${\mathbb P}^4$.
Now $W$ is not an ACM surface, but if we take a hyperplane section $\Gamma = W
\cap {\mathbb P}^3$, then $\Gamma$ is a $(4,0)$ curve in ${\mathbb P}^3$. It is contained in a unique nonsingular quadric surface $Q \subseteq {\mathbb
P}^3$, and the union $W \cup Q$, meeting along $\Gamma$, will be an ACM surface of degree $6$ in ${\mathbb P}^4$. We regard $C' \subseteq W$ as a curve on the surface $W \cup Q$. Now one can show (I leave some details to the reader) that $2H-K-C'$ on the surface $W \cup Q$ (where $H,K$ denote the hyperplane section and canonical divisor) is a curve $D \cup \Gamma'$, where $\Gamma'$ is a $(4,0)$ curve in ${\mathbb P}^3$, and $D$ is a conic, not in ${\mathbb P}^3$, meeting $\Gamma'$ in two points. Since $2H-K$ is an arithmetically Gorenstein curve on $W \cup Q$, we have thus linked $C$ to $C'$ and then to $D \cup
\Gamma'$.
For the last step, we take a quadric surface $Q'$ containing $D$ and meeting the quadric $Q$ (which contains $\Gamma'$) in a conic. Then $Q \cup Q'$ is a complete intersection quartic surface in ${\mathbb P}^4$, and on $Q\cup Q'$, $D
\cup \Gamma' - H$ is $2$ skew lines.
Thus $C$ is an example of a curve with Rao module $k$, that cannot be obtained by ascending Gorenstein biliaison from a minimal curve, and yet is in the $G$-liaison class of $2$ skew lines.
[**Example 4.7.**]{} For our last example, we consider smooth $(11,7)$ curves in ${\mathcal M}_2$. To construct such curves on a general Bordiga surface $S$, take $C = (6;2^3,1^7)$. This is an $(11,7)$ curve, and $C-H = (2;1^3,0^7)$ is a smooth $(5,0)$ curve on $S$. Since $(5,0)
\in {\mathcal L}_1$ by $(4.3)$ above, we see that $C \in {\mathcal L}_2$, and is obtained from a minimal curve by two elementary Gorenstein biliaisons.
Next, I claim the only way to obtain an $(11,7)$ curve in ${\mathcal L}_2$ by two elementary $G$-biliaisons is the one just described. Indeed, the curves of minimal genus in ${\mathcal L}_1$ of degrees $4$ to $7$ are $(4,0)$, $(5,0)$, $(6,1)$, $(7,2)$. The minimal genus of ACM curves in ${\mathbb P}^3$ of complementary degree are $(7,5)$, $(6,3)$, $(5,2)$, $(4,1)$, which will give rise respectively to curves $(11,8)$, $(11,7)$, $(11,8)$, $(11,9)$ in ${\mathcal L}_2$. So an $(11,7)$ curve obtained by elementary $G$-biliaisons must be on the Bordiga surface or its specialization.
Since $h^0({\mathcal O}_C(1)) = 5$, we see that ${\mathcal O}_C(1)$ is nonspecial, so we can compute the dimension of the Hilbert scheme of $(11,7)$ curves in ${\mathbb
P}^4$ (which is irreducible by Ein [@Ein]), by the usual formula $5d+1-g$. Thus the Hilbert scheme has dimension $49$.
Now let us count the curves obtained by the construction above. The Bordiga surface moves in a family of dimension $36$ (use, for example, the formula of Ellingsrud [@E]). To find the dimension of the linear system $|C|$ on $S$, we use the exact sequence $$0 \rightarrow {\mathcal O}_S \rightarrow {\mathcal O}_S(C) \rightarrow {\mathcal O}_C(C)
\rightarrow 0.$$ Thus $\dim_S |C| = h^0({\mathcal O}_C(C))$. For the curve $C$ of type $(6;2^3,1^7)$ mentioned above, we find $C^2 = 17$, so the divisor $C^2$ is nonspecial on $C$, and by Riemann–Roch, $h^0({\mathcal O}_C(C)) = C^2 + 1 - g
= 11$. Thus the dimension of the family of all curves of this type on Bordiga surfaces is $\le 11 + 36 = 47$. In particular, these curves are not general among all $(11,7)$ curves.
But we wish to show more, namely that a general $(11,7)$ curve does not lie on a Bordiga surface. So suppose now that $C$ is any $(11,7)$ curve on a Bordiga surface. I claim $C^2 \le 17$. Indeed, we have $2g - 2 = C^2 + C.K$. On the Bordiga surface let $C = (a;b_1,\dots,b_{10})$. The canonical divisor $K$ can be written $K = (1;0^{10}) - H$. So $2g - 2 = C^2 + a - d$, and $C^2 = 2g - 2
+ d - a = 23 - a$. But in order to get a curve of genus $7$, we must have $a \ge 6$. Thus $C^2 \le 17$. Then the same argument as above shows that $h^0({\mathcal O}_C(C)) \le 11$, and we get the same dimension count, unless ${\mathcal O}_C(C)$ is a special divisor. But in that case $C^2 \le 12$, and by Clifford’s theorem $h^0({\mathcal O}_C(C)) \le 7$. Thus a general $(11,7)$ curve does not lie on a Bordiga surface.
Next, observe that for any $(11,7)$ curve in ${\mathbb P}^4$, $h^0({\mathcal O}_C(2))
= 16$, so necessarily $h^1({\mathcal I}_C(2)) \ge 1$. Since we have constructed curves $C$ with $h^1({\mathcal I}_C(2)) = 1$ and the other $h^1({\mathcal I}_C(n)) = 0$ for $n \ne 2$, we conclude by semicontinuity that the general $(11,7)$ curve in ${\mathbb P}^4$ has Rao module $k$ in degree $2$, i.e., it lies in ${\mathcal M}_2$. Since the general such curve does not lie on a Bordiga surface, by the above remarks, it cannot be obtained by ascending Gorenstein biliaisons from ${\mathcal L}_0$.
It is conceivable that the general $(11,7)$ curve is linked by some ascending and descending $G$-liaisons to two skew lines, but this seems unlikely, so we propose the general $(11,7)$ curve in ${\mathbb P}^4$ as a possible curve with Rao module $k$, not in the $G$-liaison class of two skew lines.
Conclusion
==========
The examples presented in this paper would lead me to expect that for ACM schemes of codimension $\ge 3$, some may be obtained by elementary Gorenstein biliaisons from a scheme of degree one; a broader class may be obtained by ascending and descending $G$-liaisons from a scheme of degree one; but that a general ACM scheme of high degree may not be in the $G$-liaison class of a complete intersection. Example $3.9$ shows that at least one of the Questions $1.3$a, $1.3$b has no for an answer. Namely, for the general ACM $(20,26)$ curve in ${\mathbb P}^4$, we must have either
- it is ACM and not glicci, or
- it is glicci, but cannot be obtained by ascending Gorenstein biliaisons from a curve.
For curves in ${\mathbb P}^n$, $n\ge 4$, with a given Rao module $M$, I would expect that the minimal curves form an infinite union of irreducible families; some curves in the family may be obtained by a sequence of ascending elementary Gorenstein biliaisons from a minimal curve; a larger class may be obtained by ascending and descending Gorenstein liaisons from a minimal curve; but that a general curve of high degree and genus with Rao module $M$ is not in the $G$-liaison class of a minimal curve. Example $4.6$ gives an example of a smooth $(10,6)$ curve with Rao module $k$, that is in the Gorenstein liaison class of a minimal curve, but cannot be obtained by ascending Gorenstein biliaison from a minimal curve. Example $4.7$ shows that either Question $1.4$ has no for an answer, or the deformation is necessary in Question $1.5$b. Indeed, for the general $(11,7)$ curve in ${\mathbb P}^4$ we must have either
- it has Rao module $k$, but is not in the $G$-liaison class of two skew lines, or
- it is in the liaison class of two skew lines, but cannot be obtained from a minimal curve by ascending Gorenstein biliaisons.
Based on this evidence, I would expect a no answer to Questions $1.3$, $1.4$, $1.5b$. I have no idea about Question $1.2$ since in this paper I used only strict Gorenstein liaisons and biliaisons. Also the question of even or odd liaison has not been addressed here, since it is irrelevant for ACM schemes and curves with Rao module $k$. This is a question that merits further study.
[10]{}
Arbarello, E.; Cornalba, M.; Griffiths, P. A.; and Harris, J., “Geometry of Algebraic Curves, I”. Springer, New York (1985).
Ballico, E.; Bolondi, G.; and Migliore, J. C., [*The Lazarsfeld–Rao problem for liaison classes of two-codimensional subschemes of ${\mathbb P}^n$*]{}, Amer. J. Math. [**113**]{} (1991), 117–128.
Bolondi, G., and Migliore, J. C., [*The Lazarsfeld–Rao property on an arithmetically Gorenstein variety*]{}, Manus. Math. [**78**]{} (1993), 347–368.
Casanellas, M., and Miró–Roig, R. M., [*Gorenstein liaison of curves in ${\mathbb P}^4$*]{}, J. Algebra [**230**]{} (2000), 656–664.
Casanellas, M., and Miró–Roig, R. M., [*$G$-liaison of arithmetically Gorenstein schemes*]{}, (preprint).
Casanellas, M., and Miró–Roig, R. M., [*Gorenstein liaison of divisors on standard determinantal schemes and on rational normal scrolls*]{}, (preprint).
Ein, L., [*The irreducibility of the Hilbert scheme of smooth space curves*]{}, in “Algebraic Geometry,” Bowdoin 1985, Proc. Symp. Pure Math. [**46**]{} AMS (1987), 83–87.
Eisenbud, D., and Harris, J., “Curves in Projective Space”, Les Presses de l’Université de Montréal, Montreal (1982).
Eisenbud, D.; Lange, H.; Martens, G.; and Schreyer, F.-O., [*The Clifford dimension of a projective curve*]{}, Compositio Math. [**72**]{} (1989), 173–204.
Ellingsrud, G., [*Sur le schéma de Hilbert des variétés de codimension $2$ dans ${\mathbb P}^e$ à cône de Cohen–Macaulay*]{}, Ann. Scient. ENS [**8**]{} (1975), 423–432.
Gaeta, F., [*Sulle curve sghembe algebriche di residuale finito*]{}, Ann. di Mat. IV [**27**]{} (1948), 177–241.
Hartshorne, R., [*Classification of algebraic space curves, III*]{}, in “Algebraic Geometry and its Applications,” ed. C. L.Bajaj, Springer (1994), 113–120.
Hartshorne, R., [*Generalized divisors on Gorenstein schemes*]{}, $K$-Theory [**8**]{} (1994), 287–339.
Hartshorne, R., “Algebraic Geometry”, Springer, New York (1977).
Hartshorne, R., [*Clifford index of ACM curves in ${\mathbb P}^3$*]{}, Rend. Sem. Mat. Fis. Milano (to appear).
Kleppe, J.; Migliore, J.; Miró–Roig, R.; Nagel, U.; and Peterson, C., [*Gorenstein liaison, complete intersection liaison invariants, and unobstructedness*]{}, Memoirs, A.M.S. (to appear).
Kreuzer, M.; Migliore, J.; Nagel, U.; and Peterson, C., [*Determinantal schemes and Buchsbaum–Rim sheaves*]{}, J. Pure Appl. Algebra [**150**]{} (2000), 155–174.
Lesperance, J., [*Gorenstein liaison of some curves in ${\mathbb P}^4$*]{} (preprint).
Lopez, A., [*Noether–Lefschetz theory and the Picard group of projective surfaces*]{}, Memoirs AMS [**89**]{} (1991), n${}^{\mbox{o}}$ 438.
Martens, G., [*Ueber den Clifford–Index algebraischer Kurven*]{}, Crelle J. [**336**]{} (1982), 83–90.
Martin–Deschamps, M., and Perrin, D., [*Sur la classification des courbes gauches*]{}, Astérisque [**184–185**]{} (1990).
Migliore, J. C., [*Introduction to liaison theory and deficiency modules*]{}, Birkhäuser, Progress in Mathematics [**165**]{} (1998).
Migliore, J., [*Gorenstein liaison via divisors*]{}, (preprint).
Migliore, J., and Nagel, U., [*Monomial ideals and the Gorenstein liaison class of a complete intersection*]{}, (preprint).
Nagel, U., [*Even liaison classes generated by Gorenstein linkage*]{}, J. Algebra [**209**]{} (1998), 543–584.
Nollet, S., [*Even linkage classes*]{}, Trans. Amer. Math. Soc. [**348**]{} (1996), 1137–1162.
Okonek, C., [*Moduli reflexiver Garben und Flächen von kleinem Grad in ${\mathbb P}^4$*]{}, Math. Zeit. [**184**]{} (1983), 549–572.
Perrin, D., [*Courbes passant par $m$ points généraux de ${\mathbb P}^3$*]{}, Soc. Math. France, Mémoire [**28/29**]{} (1987).
Peskine, C., and Szpiro, L., [*Liaison des variétés algébriques, I*]{}, Invent. Math. [**26**]{} (1974), 271–302.
Rao, A. P., [*Liaison equivalence classes*]{}, Math. Ann. [**258**]{} (1981), 169–173.
Rathmann, J., [*The genus of curves in ${\mathbb P}^4$ and ${\mathbb P}^5$*]{}, Math. Zeit. [**202**]{} (1989), 525–543.
Saint-Donat, B., [*On Petri’s analysis of the linear system of quadrics through a canonical curve*]{}, Math. Ann. [**206**]{} (1973), 157–175.
Serre, J.-P., [*Sur les modules projectifs*]{}, Sém. Dubreil–Pisot (1960/61).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We derive a novel computational scheme for functional Renormalization Group (fRG) calculations for interacting fermions on 2D lattices. The scheme is based on the exchange parametrization fRG for the two-fermion interaction, with additional insertions of truncated partitions of unity. These insertions decouple the fermionic propagators from the exchange propagators and lead to a separation of the underlying equations. We demonstrate that this separation is numerically advantageous and may pave the way for refined, large-scale computational investigations even in the case of complex multiband systems. Furthermore, on the basis of speedup data gained from our implementation, it is shown that this new variant facilitates efficient calculations on a large number of multi-core CPUs. We apply the scheme to the $t$,$t''$ Hubbard model on a square lattice to analyze the convergence of the results with the bond length of the truncation of the partition of unity. In most parameter areas, a fast convergence can be observed. Finally, we compare to previous results in order to relate our approach to other fRG studies.'
address:
- |
Institute for Theoretical Solid State Physics, RWTH Aachen University,\
D-52074 Aachen, Germany
- |
Jülich Supercomputing Centre, Forschungszentrum Jülich GmbH,\
D-52425 Jülich, Germany
- 'Aachen Institute for Advanced Study in Computational Engineering Science (AICES), RWTH Aachen University, D-52056 Aachen, Germany'
- |
JARA-FIT, Jülich Aachen Research Alliance - Fundamentals of\
Future Information Technology
- 'Institut für Theoretische Physik, Universität zu Köln, D-50937 Cologne, Germany'
author:
- 'J. Lichtenstein'
- 'D. Sánchez de la Peña'
- 'D. Rohe'
- 'E. Di Napoli'
- 'C. Honerkamp'
- 'S. A. Maier'
bibliography:
- 'bibliography.bib'
date: 'July 29, 2016'
title: 'High-performance functional renormalization group calculations for interacting fermions'
---
functional Renormalization Group ,truncated partition of unity ,interacting fermions ,hybrid parallelization ,Hubbard model
Introduction
============
Having passed through a process of refinement and development that has taken more than two decades, the functional Renormalization Group (fRG) can be rightfully seen as an unbiased method for investigating interacting Fermi systems. In the medium term, it is conceivable to use the fRG methods to investigate existing materials in terms of their ground state properties with [*quantitative precision*]{} regarding energy scales and parameter ranges, as well as to discover new materials with superior features. While the development process is far from being completed, with this paper we contribute in pushing the fRG method forward into this direction.
Our main focus is to show how one can capture the wavevector dependence of effective two-fermion vertices in a physically meaningful and numerically advantageous way. The fRG—as we use it in this paper—is based on the Wetterich equation [@Wetterich1993] that describes the evolution of an effective action. Since this equation results in a full hierarchy of differential equations encompassing all orders of the vertex functions, a truncation is necessary in order to ensure solvability. As defined in the review paper by Metzner et al. [@Metzner2012], we build on the level-2 truncation of the fRG equation hierarchy which has become an often-used standard. In addition to that, self-energies are neglected in the current state of the method development for the sake of simplicity. Moreover, in most cases further approximations are indispensable, for example, a discretization of the Brillouin zone (BZ) using a grid of momentum sampling points. Within the first fRG studies on Fermi systems [@Zanchi1998; @Halboth2000; @Honerkamp2001a]—those addressed the 2D Hubbard model—the BZ was discretized by using the so-called *Fermi surface patching* scheme, where the vertex dependences on the radial parts of momenta are neglected. This scheme was further extended to multiband models and applied to systems like iron superconductors [@Wang2009; @Thomale2009; @Lichtenstein2014] or single- and multilayer graphene [@Scherer2012; @Kiesel2012; @Scherer2012a; @Pena2014].
In terms of a method development, the *exchange parametrization fRG*—as introduced in Ref. and used, for example, in Refs. —can be seen as the next important step. Within that scheme, the two-particle coupling function, which generally depends on three external momenta due to the conservation of total momentum, is decomposed into three single-channel functions, where every channel only depends strongly on one momentum. As a consequence, this parametrization—which can be used for the dependence on frequencies in a similar way [@Karrasch2008]—softens the scaling of the number of coupled differential equations with respect to the momentum grid point number. While in the Fermi surface patching one has to deal with a third power scaling behavior, in the exchange parametrization the scaling becomes almost linear. Although there is some freedom in defining the three channels, usually they can be interpreted as being responsible for charge, spin and pairing fluctuations respectively. Hence, the exchange parametrization allows to describe the potentially complex momentum structure of the effective interaction in a fashion that is physically easier to interpret.
Besides the exchange parametrization fRG, the *singular-mode fRG (SMFRG)* [@Wang2012] was introduced as another scheme to capture the momentum dependence. Similar to the exchange parametrization fRG, the SMFRG scheme distinguishes between three different channels, but it uses other quantities to represent those and it treats the feedback between these channels differently.[^1] In this paper we build on the two last-named schemes and take the next step in the development of the fRG method. In Sec. \[sec:tufrg\] we present a step-by-step derivation of a new fRG variant that combines features of both the exchange parametrization fRG and the SMFRG, and we argue that the new variant is numerically beneficial compared to the exchange parametrization fRG. In order to point out the major distinction between the latter and the new scheme, we name our newly developed variant *truncated unity fRG (TUfRG)*. We comment on the relation of the TUfRG to the SMFRG in \[sec:rel\_to\_smfrg\].
In order to enlarge the application area of the fRG to more complex systems, method development has to focus on two parallel research directions: equation parametrization and parallel implementation. While it is crucial to develop new meaningful parametrizations and approximations for the flow equations, it is equally relevant to explore parallelization and performance strategies enabling the efficient use of massively parallel computing architectures. For instance, as shown in Ref. , fRG can profit highly from a sophisticated hybrid parallelization that can achieve a speedup of several orders of magnitude. In Sec. \[sec:perform\], we discuss our algorithmic choices, their parallel implementation and the speedup gained by TUfRG when running on parallel computing platforms.
The TUfRG contains an additional approximation compared to the exchange parametrization fRG—namely the insertion of a truncated partition of unity. This insertion leads to a simplification of the integrals involved, which are computationally the most challenging tasks in the fRG calculation. In Sec. \[sec:phase\_hubb\], we check the quality of this approximation by applying the scheme to the well studied [@LeBlanc2015; @Honerkamp2001; @Giering2012; @Eberlein2015; @Yamase2016] $t$-$t'$ Hubbard model. Furthermore, we provide a more analytic view on this aspect in \[sec:app-smooth-ff\].
The TUfRG equations {#sec:tufrg}
===================
Since there are comprehensive descriptions of the fermionic fRG already given in other works (for a recent review, see e.g. Refs. ), we will not explain the basic concept of this method and the derivation of the hierarchy of flow equations in detail, but briefly mention the important equations to bring the reader up to speed with our notation. Afterwards, we will derive the TUfRG from the fermionic fRG equations. For further classification we explain that the TUfRG equations are strongly related to the ones from the SMFRG [@Wang2012] and point out the conceptual differences in \[sec:rel\_to\_smfrg\].
fRG Flow equations {#sec:su2-inv}
------------------
In the following, we consider an effective action of the form $$\Gamma [\bar{\psi},\psi] = \int \! d\xi \, \bar{\psi}(\xi) \, Q (k) \, \psi (\xi) +\Gamma^{(4)} [\bar{\psi},\psi] \, ,$$ where $\xi=(k,\sigma)$, with $k =(k_0,\mathbf{k})$, denotes a collection of frequency ($k_0$), momentum ($\mathbf{k}$) and spin projection ($\sigma$) quantum numbers. As described in Refs. , the two-particle interaction $$\begin{aligned}
\Gamma^{(4)} [\bar{\psi},\psi] & = \frac{1}{4} \int \! d \xi_1 \dots d \xi_4 \, f(\xi_1,\xi_2,\xi_3,\xi_4) \\
& \qquad \times \bar{\psi} (\xi_4) \, \bar{\psi} (\xi_3) \, \psi (\xi_2) \, \psi (\xi_1) \end{aligned}$$ of a charge conserving and SU(2) invariant theory can be parametrized with one spin-independent coupling function $ V(k_1,k_2,k_3) $ according to $$\begin{aligned}
f (\xi_1,\xi_2,\xi_3,\xi_4) & =
\left[ V (k_1,k_2,k_3) \, \delta_{\sigma_1, \sigma_4}
\delta_{\sigma_2, \sigma_3} - V (k_2,k_1,k_3) \, \delta_{\sigma_1, \sigma_3} \delta_{\sigma_2, \sigma_4} \right] \\
& \quad \times \delta (k_1 +k_2 -k_3 -k_4) \, ,\end{aligned}$$ where the $\delta$-function ensures momentum and energy conservation.
The quadratic part of the effective action is diagonal in spin-space, which implies $$G (\xi_1,\xi_2) = \delta_{\sigma_1,\sigma_2} \delta(k_1-k_2) \, G (k_1) \quad \text{and} \quad
\Sigma (\xi_1,\xi_2) = \delta_{\sigma_1,\sigma_2} \delta(k_1-k_2) \, \Sigma (k_1)$$ for the one-particle propagator $G$ and the 1PI self-energy $\Sigma$. Once a regulator is added to the propagator, we can derive the fRG flow equations. More explicitly, we replace $G$ by a function $G^\Lambda$ in a way that we get $G^\Lambda\to0$ for $\Lambda\to\infty$ and $G^\Lambda\to G$ for $\Lambda\to 0$. This results in differential equations of 1PI vertex functions with respect to the regularization scale, which we call $\Lambda$ in this paragraph. The flow equation corresponding to the 1PI self-energy reads $$\label{eqn:sigma-su2fl}
\dot{\Sigma} (k) = \int\! dp \,\, S(p) \, \left[ V (p,k,p) - 2 V (k,p,p) \right] \, ,$$ with the single-scale propagator $S$ [@Metzner2012]. Note that in the following we will not mark dependences on the regularization scale with a superscript $\Lambda$ in order to simplify the notation. A derivative with respect to this scale is denoted as a dot above the affected quantity. The scale derivative of the coupling function $ V $ consists of three parts $$\dot{V} (k_1,k_2,k_3) = \mathcal{T}_\mathrm{pp} (k_1,k_2,k_3)
+ \mathcal{T}^\mathrm{cr}_\mathrm{ph} (k_1,k_2,k_3) + \mathcal{T}^\mathrm{d}_\mathrm{ph} (k_1,k_2,k_3) \, .$$ The particle-particle contribution $$\label{eqn:pp-su2fl}
\mathcal{T}_\mathrm{pp} = - \int \! d p \, \left[ \partial_\lambda G(p) \, G(k_1+k_2-p) \right]
V (k_1,k_2,p) \, V(k_1+k_2-p,p,k_3)$$ and the crossed particle-hole part $$\label{eqn:phcr-su2fl}
\mathcal{T}^\mathrm{cr}_\mathrm{ph} = - \int \! d p \, \left[ \partial_\lambda G(p) \, G(p+k_3-k_1) \right]
V (k_1,p+k_3-k_1,k_3) \, V(p,k_2,p+k_3-k_1)$$ can each be represented by one diagram (see Fig. \[fig:diags-fleq\]).
Particle-particle diagram $ \mathcal{T}_\mathrm{pp}$\
![The five diagrams driving the flow of the two-particle interaction $ V(k_1,k_2,k_3)$ of an U(1) and $ \mathrm{SU} (2) $ invariant theory. For the closed loops in these diagrams, a scale derivative is implicit. (These figures have been taken from Ref. .)[]{data-label="fig:diags-fleq"}](Tpp.eps "fig:"){width="0.23\linewidth"}\
Crossed particle-hole diagram $ \mathcal{T}_\mathrm{ph}^\mathrm{cr}$\
![The five diagrams driving the flow of the two-particle interaction $ V(k_1,k_2,k_3)$ of an U(1) and $ \mathrm{SU} (2) $ invariant theory. For the closed loops in these diagrams, a scale derivative is implicit. (These figures have been taken from Ref. .)[]{data-label="fig:diags-fleq"}](Tph-cr.eps "fig:"){width="0.23\linewidth"}\
Direct particle-hole diagrams $ \mathcal{T}_\mathrm{ph}^\mathrm{d}$\
![The five diagrams driving the flow of the two-particle interaction $ V(k_1,k_2,k_3)$ of an U(1) and $ \mathrm{SU} (2) $ invariant theory. For the closed loops in these diagrams, a scale derivative is implicit. (These figures have been taken from Ref. .)[]{data-label="fig:diags-fleq"}](Tph-d.eps "fig:"){width="0.48\linewidth"}
Vertex corrections and particle-hole screening, however, are accounted for by the direct particle-hole diagrams depicted in Fig. \[fig:diags-fleq\]: $$\begin{aligned}
\notag
\mathcal{T}^\mathrm{d}_\mathrm{ph} = \int \! d p & \,\, \left[ \partial_\lambda G(p) \, G(p+k_2-k_3) \right]
\left[ 2 V (k_1,p+k_2-k_3,p) \, V(p,k_2,k_3) \right. \\ \notag
& - V (k_1,p+k_2-k_3,k_1+k_2-k_3) \, V(p,k_2,k_3) \\ \label{eqn:phd-su2fl} & \left.
- V (k_1,p+k_2-k_3,p) \, V(p,k_2,p+k_2-k_3) \right] \, .\end{aligned}$$
Channel decomposition
---------------------
Let us now recapitulate the channel decomposition of Karrasch *et al.* [@Karrasch2008] for the frequency- and by Husemann and Salmhofer [@Husemann2009] for the momentum-dependence of the vertices. In these approaches, the coupling function $ V(k_1,k_2,k_3) $ is decomposed into contributions resulting from three different channels. More precisely, we have $$\begin{aligned}
V \left(k_1,k_2,k_3 \right) & = V^{(0)}_{k_1,k_2,k_3} - \Phi^\mathrm{SC}_{k_1+k_2,\frac{k_1-k_2}{2},\frac{k_4-k_3}{2}} + \Phi^\mathrm{M}_{k_1-k_3,\frac{k_1+k_3}{2},\frac{k_2+k_4}{2} } \\ &
\quad + \frac{1}{2} \Phi^\mathrm{M}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2} } - \frac{1}{2} \Phi^\mathrm{K}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2} } \, ,\end{aligned}$$ with $ V^{(0)} $ being the bare interaction, and $ \Phi^{\mathrm{SC}} $, $ \Phi^\mathrm{M} $, and $ \Phi^\mathrm{K} $ representing the coupling functions of the pairing, the magnetic, and the forward scattering channel, respectively. The first argument of each channel accounts for the dependence on the total ($\mathbf{k_1}+\mathbf{k_2}$) or on one of the transfer momenta ($\mathbf{k_1}-\mathbf{k_3}$ and $\mathbf{k_3}-\mathbf{k_2}$). These momentum combinations appear inside the fermionic loops from Fig. \[fig:diags-fleq\] and label the most important momentum dependences at weak coupling. Regarding the other two (weak) momentum dependences of each channel, we have chosen a convention that is more symmetric than in Refs. . These single-channel coupling functions are generated during the flow according to the following equations $$\begin{aligned}
\label{eqn:flow-PhiP}
\dot{\Phi}^\mathrm{SC}_{k_1+k_2,\frac{k_1-k_2}{2},\frac{k_4-k_3}{2}} & = - \mathcal{T}_\mathrm{pp} \left(k_1,k_2,k_3 \right) \\ \notag
\dot{\Phi}^\mathrm{M}_{k_1-k_3,\frac{k_1+k_3}{2},\frac{k_2+k_4}{2}} & = \mathcal{T}_\mathrm{ph}^\mathrm{cr} \left(k_1,k_2,k_3 \right) \\ \notag
\dot{\Phi}^\mathrm{K}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2}} & = - 2 \mathcal{T}_\mathrm{ph}^\mathrm{d} \left(k_1,k_2,k_3 \right) + \mathcal{T}_\mathrm{ph}^\mathrm{cr} \left(k_1,k_2,k_1+k_2-k_3 \right) \, .\end{aligned}$$ At the formal level, the channel decomposition may be performed in a different way. Instead of $\Phi^\mathrm{M}$ and $\Phi^\mathrm{K}$, the particle-hole channels are accounted for by the coupling functions $\Phi^\mathrm{C}$ and $\Phi^\mathrm{D}$, which flow according to $$\begin{aligned}
\label{eqn:flow-PhiC}
\dot{\Phi}^\mathrm{C}_{k_1-k_3,\frac{k_1+k_3}{2},\frac{k_2+k_4}{2}} & = \mathcal{T}_\mathrm{ph}^\mathrm{cr} \left(k_1,k_2,k_3 \right) \\ \label{eqn:flow-PhiD}
\dot{\Phi}^\mathrm{D}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2}} & = \mathcal{T}_\mathrm{ph}^\mathrm{d} \left(k_1,k_2,k_3 \right) \end{aligned}$$ and enter in the effective interaction as $$\begin{aligned}
\notag
V \left(k_1,k_2,k_3 \right) & = V^{(0)}_{k_1,k_2,k_3} - \Phi^\mathrm{SC}_{k_1+k_2,\frac{k_1-k_2}{2},\frac{k_4-k_3}{2}} + \Phi^\mathrm{C}_{k_1-k_3,\frac{k_1+k_3}{2},\frac{k_2+k_4}{2}} \\ \label{eqn:simple-deco} &
\quad + \Phi^\mathrm{D}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2}} \, .\end{aligned}$$ This latter form of the channel decomposition corresponds to the parametrization of the interaction used in Refs. . In the following, we will work with the latter variant, while magnetic and forward scattering channels can be obtained as $$\begin{aligned}
\Phi^\mathrm{M}_{k_1-k_3,\frac{k_1+k_3}{2},\frac{k_2+k_4}{2}} &= \Phi^\mathrm{C}_{k_1-k_3,\frac{k_1+k_3}{2},\frac{k_2+k_4}{2}} \\
\Phi^\mathrm{K}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2}} & = - 2 \Phi^\mathrm{D}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2}}\\
& \quad + \Phi^\mathrm{C}_{k_3-k_2,\frac{k_1+k_4}{2},\frac{k_2+k_3}{2}} \, .\end{aligned}$$
Projection onto exchange propagators
------------------------------------
Let us now describe the dependence of the coupling functions on the weak momenta through a complete set of form factors $\{\,f_m (\mathbf{k})\,\}$ that are square integrable on the BZ. For the particle-particle channel, we can project the single-channel coupling function $ \Phi^\mathrm{SC}$ onto a matrix $\mathbf{P} (l) =\hat{P} \left[ \Phi^\mathrm{SC} \right] (l) $ of exchange propagators. The matrix elements then read $$\label{eqn:P_phi_sc}
P_{m,n} (l) = \hat{P} \left[ \Phi^\mathrm{SC} \right]_{m,n} (l) =\left. \int \! d\mathbf{k} \, d\mathbf{k'} \, f^*_m(\mathbf{k}) \, f_n(\mathbf{k'}) \, \Phi^\mathrm{SC}_{l,k,k'} \right|_{k_0=k_0'=0} \, ,$$ and the single-channel coupling function is recovered as $$\label{eqn:app-PhiP}
\Phi^\mathrm{SC}_{l,k,k'} \approx \sum_{m,n} f_m(\mathbf{k}) \, f^*_n(\mathbf{k'}) \, P_{m,n} (l) \, .$$ On a formal level, the momentum dependences are then reproduced exactly, while the frequency dependences contained in $k$ and $k'$ are neglected which is expressed by the approximately-equal sign in Eqn. \[eqn:app-PhiP\]. Additionally, in a numerical calculation, one will be forced to truncate the infinite sum over the form factors. Note that, if the underlying lattice structure corresponds to a Bravais lattice, the form factors can always be chosen to be real valued in momentum representation. Hence, we will leave out the asterisks from Eqs. \[eqn:P\_phi\_sc\] and \[eqn:app-PhiP\] in the following.
Similarly to the particle-particle channel, the exchange propagators of the particle-hole channels are defined as $$\begin{aligned}
\label{eqn:C_phi_C}
\mathbf{C} (l) &=\hat{C} \left[ \Phi^\mathrm{C} \right] (l)\, ,\\
\label{eqn:D_phi_D}
\mathbf{D} (l) &=\hat{D} \left[ \Phi^\mathrm{D} \right] (l)\end{aligned}$$ and the corresponding single-channel coupling functions read in exchange parametrization $$\begin{aligned}
\label{eqn:app-PhiC}
\Phi^\mathrm{C}_{l,k,k'} \approx \sum_{m,n} f_m(\mathbf{k}) \, f_n(\mathbf{k'}) \, C_{m,n} (l) \, ,\\ \label{eqn:app-PhiD}
\Phi^\mathrm{D}_{l,k,k'} \approx \sum_{m,n} f_m(\mathbf{k}) \, f_n(\mathbf{k'}) \, D_{m,n} (l) \, .\end{aligned}$$
The flow equations for the exchange propagators are obtained by applying the projection operations from Eqs. (\[eqn:P\_phi\_sc\]), (\[eqn:C\_phi\_C\]), and (\[eqn:D\_phi\_D\]) to the respective diagrams in the right-hand sides of Eqs. (\[eqn:flow-PhiP\])-(\[eqn:flow-PhiD\]). This yields $$\begin{aligned}
\label{eqn:expar-flowP}
\dot{\mathbf{P}} (l) &= - \hat{P} \left[ \mathcal{T}_\mathrm{pp} \right] (l)\, ,\\ \label{eqn:expar-flowC}
\dot{\mathbf{C}} (l) &=\hat{C} \left[ \mathcal{T}_\mathrm{ph}^\mathrm{cr} \right] (l)\, ,\\ \label{eqn:expar-flowD}
\dot{\mathbf{D}} (l) &=\hat{D} \left[ \mathcal{T}_\mathrm{ph}^\mathrm{d} \right] (l)\, ,\end{aligned}$$ where the projection operators applied to a test function $F$ read: $$\begin{aligned}
\label{eqn:proj-P}
\hat{P} \left[ F \right]_{m,n} (l) & = \left. \int \! d\mathbf{k} \, d\mathbf{k'} \,
f_m (\mathbf{k}) \,f_n (\mathbf{k'}) \, F \left(\frac{l}{2}+k,\frac{l}{2}-k,\frac{l}{2}-k'\right) \right|_{k_0=k_0'=0} \, ,\\ \label{eqn:proj-C}
\hat{C} \left[ F \right]_{m,n} (l) & = \left. \int \! d\mathbf{k} \, d\mathbf{k'} \,
f_m (\mathbf{k}) \,f_n (\mathbf{k'}) \, F \left(k+\frac{l}{2},k'-\frac{l}{2},k-\frac{l}{2}\right) \right|_{k_0=k_0'=0} \, ,\\ \label{eqn:proj-D}
\hat{D} \left[ F \right]_{m,n} (l) & = \left. \int \! d\mathbf{k} \, d\mathbf{k'} \,
f_m (\mathbf{k}) \,f_n (\mathbf{k'}) \, F \left(k+\frac{l}{2},k'-\frac{l}{2},k'+\frac{l}{2}\right) \right|_{k_0=k_0'=0} \, .\end{aligned}$$
![Diagrams contributing to the flow of $B_1$. Solid lines correspond to fermionic and wiggly ones to exchange propagators. Fermion-boson vertices (our form factors $f_m$) are represented by empty circles. $B_2$ denotes a linear combination of exchange propagators that differ from $B_1$.[]{data-label="fig:expar-diags"}](prop-ren.ti.eps "fig:"){width=".62\linewidth"}\
Propagator renormalization\
![Diagrams contributing to the flow of $B_1$. Solid lines correspond to fermionic and wiggly ones to exchange propagators. Fermion-boson vertices (our form factors $f_m$) are represented by empty circles. $B_2$ denotes a linear combination of exchange propagators that differ from $B_1$.[]{data-label="fig:expar-diags"}](vert-corr.ti.eps "fig:"){width=".55\linewidth"}\
Vertex correction\
![Diagrams contributing to the flow of $B_1$. Solid lines correspond to fermionic and wiggly ones to exchange propagators. Fermion-boson vertices (our form factors $f_m$) are represented by empty circles. $B_2$ denotes a linear combination of exchange propagators that differ from $B_1$.[]{data-label="fig:expar-diags"}](boxes.ti.eps "fig:"){width=".28\linewidth"}\
Box diagrams
Eqs. (\[eqn:expar-flowP\])-(\[eqn:expar-flowD\]) can be seen as flow equations of the exchange-parametrization fRG (see Refs. for details on this scheme). By substituting the decomposed interaction of Eq. (\[eqn:simple-deco\]) in the flow equations (\[eqn:expar-flowP\])-(\[eqn:expar-flowD\]), one obtains three different kinds of diagrams governing the flow of the exchange propagator $B_1 \in \{P,C,D\}$ (see Fig. \[fig:expar-diags\]). In the propagator renormalization diagrams, bosonic lines (exchange propagators) corresponding to $B_1$ appear outside the loops. Apart from a $B_1$ line outside the loops, vertex correction diagrams have one bosonic line inside, which corresponds to a linear combination $B_2$ of bosonic propagators. In the box diagrams, both bosonic lines correspond to $B_2$ and appear inside the loops. In a numerical implementation of the exchange-parametrization fRG, bosonic lines inside the loops pose a challenge. Since these exchange propagators may be sharply peaked close to an instability, they can significantly enhance the computational cost of the loop integrals. In the following we show that one can separate the bosonic from the fermionic lines to make loop integrations a computationally easier task.
Insertion of truncated partitions of unity {#sec:trunc_uni}
------------------------------------------
We continue with the derivation of the TUfRG equations by inserting partitions of unity of the form factor basis $$1=\int \! d\mathbf{p'} \, \delta(\mathbf{p}-\mathbf{p'})=\int \! d\mathbf{p'} \, \sum_m \, f_m(\mathbf{p'}) f_m(\mathbf{p})
\label{eqn:unity}$$ on both sides of the two fermion loops in Eqs. (\[eqn:expar-flowP\])-(\[eqn:expar-flowD\]). These equations can be rewritten as $$\begin{aligned}
\label{eqn:matmul-P}
\dot{\mathbf{P}} (l) &= \mathbf{V}^P (l)\, \dot{\boldsymbol\chi}^\mathrm{pp} (l) \, \mathbf{V}^P (l)\, ,\\ \label{eqn:matmul-C}
\dot{\mathbf{C}} (l) &= - \mathbf{V}^C (l)\, \dot{\boldsymbol\chi}^\mathrm{ph} (l) \, \mathbf{V}^C (l)\, ,\\ \label{eqn:matmul-D}
\dot{\mathbf{D}} (l) &= 2 \mathbf{V}^D (l)\, \dot{\boldsymbol\chi}^\mathrm{ph} (l) \, \mathbf{V}^D (l) -
\mathbf{V}^C (l)\, \dot{\boldsymbol\chi}^\mathrm{ph} (l) \, \mathbf{V}^D (l) - \mathbf{V}^D (l)\, \dot{\boldsymbol\chi}^\mathrm{ph} (l) \, \mathbf{V}^C (l) \, ,\end{aligned}$$ where $$\begin{split} \label{eqn:chis}
\chi^\mathrm{pp}_{m,n} (l) &= \int \! dp \, G \left( \frac{l}{2} + p \right) \, G \left( \frac{l}{2} - p \right) \, f_m (\mathbf{p}) \, f_n (\mathbf{p}) \, ,\\
\chi^\mathrm{ph}_{m,n} (l) &= \int \! dp \, G \left(p+ \frac{l}{2} \right) \, G \left(p- \frac{l}{2} \right) \, f_m (\mathbf{p}) \, f_n (\mathbf{p}) \,
\end{split}$$ and $$\begin{aligned}
\label{eqn:proj_VP}\mathbf{V}^P \left(l \right) & = \hat{P} \left[ V^{(0)} \right] (l) - \mathbf{P} (l) + \hat{P} \left[ \Phi^\mathrm{C} \right] (l) + \hat{P} \left[ \Phi^\mathrm{D} \right] (l) \, ,\\
\label{eqn:proj_VC}\mathbf{V}^C \left(l \right) & = \hat{C} \left[ V^{(0)} \right] (l) - \hat{C} \left[ \Phi^\mathrm{SC} \right] (l) + \mathbf{C} (l) + \hat{C} \left[ \Phi^\mathrm{D} \right] (l) \, ,\\
\label{eqn:proj_VD}\mathbf{V}^D \left(l \right) & = \hat{D} \left[ V^{(0)} \right] (l) - \hat{D} \left[ \Phi^\mathrm{SC} \right] (l) + \hat{D} \left[ \Phi^\mathrm{C} \right] (l) + \mathbf{D} (l) \end{aligned}$$ are the three different projections from Eqs. (\[eqn:proj-P\])-(\[eqn:proj-D\]) applied to the two-particle interaction. Via Eqs. (\[eqn:app-PhiP\]), (\[eqn:app-PhiC\]), and (\[eqn:app-PhiD\]), the exchange propagators are inserted back into the flow equations (\[eqn:matmul-P\])-(\[eqn:matmul-D\]), which results in a closed system of differential equations. The bosonic propagators are now separated from the loop integrations at the cost of introducing the projections (\[eqn:proj\_VP\])-(\[eqn:proj\_VD\]). For instance, the third contribution of $\mathbf{V}^P (l)$ can be expressed as $$\begin{aligned}
\hat{P} \left[ \Phi^\mathrm{C} \right]_{m,n} (l) & \approx \int \! d\mathbf{k} \, d\mathbf{k'} \, f_m(\mathbf{k}) \, f_n(\mathbf{k'}) \notag \\
& \qquad \times \, \left. \sum_{m',n'} \, f_{m'}\left(\frac{\mathbf{l}+\mathbf{k}-\mathbf{k'}}{2}\right) \, f_{n'}\left(\frac{\mathbf{l}-\mathbf{k}+\mathbf{k'}}{2}\right) \,
C_{m',n'}(k'+k) \right|_{k_0=k'_0=0}
\label{eqn:proj_mom} \\
& = \sum_{\mathbf{R_1},\mathbf{R_2},\mathbf{R_3}} \, \sum_{m',n'} \, f_m\left(-\frac{\mathbf{R_1}}{2}+\frac{\mathbf{R_2}}{2}-\mathbf{R_3}\right) \,
f_n\left(\frac{\mathbf{R_1}}{2}-\frac{\mathbf{R_2}}{2}-\mathbf{R_3}\right) \notag \\
& \qquad\qquad \times \, f_{m'}\left(\mathbf{R_1}\right) \, f_{n'}\left(\mathbf{R_2}\right) \,
C_{m',n'}(\mathbf{R_3},k_0=0) \, e^{-i\frac{1}{2}\mathbf{l}\cdot(\mathbf{R_1}+\mathbf{R_2})}
\label{eqn:proj_pos}\end{aligned}$$ in momentum and position space.
Let us summarize which steps need to be performed in order to calculate the increment of the interaction in the TUfRG scheme:
1. Project ${\mathbf{P}} (l)$, ${\mathbf{C}} (l)$, ${\mathbf{D}} (l)$ and the bare interaction to other channels in order to obtain ${\mathbf{V}}^P (l)$, ${\mathbf{V}}^C (l)$, and ${\mathbf{V}}^D (l)$ according to Eqs. (\[eqn:proj\_VP\])-(\[eqn:proj\_VD\]). \[enu:project-new\]
2. Calculate the form factor projected fermionic loops $\dot{\boldsymbol\chi}^\mathrm{pp} (l)$ and $\dot{\boldsymbol\chi}^\mathrm{ph} (l)$ in Eq. (\[eqn:chis\]). \[enu:chis-new\]
3. Evaluate $\dot{\mathbf{P}} (l)$, $\dot{\mathbf{C}} (l)$, and $\dot{\mathbf{D}} (l)$ by performing matrix multiplications in the form factor basis \[see Eqs. (\[eqn:matmul-P\])-(\[eqn:matmul-D\])\]. \[enu:matmul-new\]
Fig. \[fig:steps-new\] graphically represents the calculation steps that are listed above.
![Graphical representation of the steps in the calculation of the increment for the interaction in the TUfRG approach. The upper diagram corresponds to the projection in step \[enu:project-new\].) and the lower one to steps \[enu:chis-new\].) and \[enu:matmul-new\].). The brick-stones in the lower part correspond to the *full* interaction projected to the respective channel with $B_1= P$, $C$, or $D$. Dashed lines correspond to simple contractions and *not* to fermionic or bosonic propagators.[]{data-label="fig:steps-new"}](new-project.ti.eps "fig:"){width=".35\linewidth"}\
[$\downarrow$]{}\
![Graphical representation of the steps in the calculation of the increment for the interaction in the TUfRG approach. The upper diagram corresponds to the projection in step \[enu:project-new\].) and the lower one to steps \[enu:chis-new\].) and \[enu:matmul-new\].). The brick-stones in the lower part correspond to the *full* interaction projected to the respective channel with $B_1= P$, $C$, or $D$. Dashed lines correspond to simple contractions and *not* to fermionic or bosonic propagators.[]{data-label="fig:steps-new"}](new-matmul.ti.eps "fig:"){width=".76\linewidth"}
For implementing the TUfRG flow we require
a) a suitable grid for the bosonic frequencies and momenta and
b) a finite set of form factors $f_m (\mathbf{k})$.
While an inappropriate choice of the former may cause severe parametrization errors, the form factor expansion should be truncated in a way that avoids large projection errors. Generically, form factors corresponding to fixed bond lengths on the direct lattice seem appropriate, as it is likely and in fact known from previous studies (e.g. ) that only small bond lengths (or slowly varying form factors) contribute significantly.
Benefits from the truncated partitions of unity {#sec:benefits}
-----------------------------------------------
From a formal point of view, our approach is nothing else than the standard exchange parametrization method with an additional approximation. Namely, we have inserted truncated partitions of unity in the form factor basis in order to pull bosonic lines out of the loops.[^2] This additional approximation is depicted in Fig. \[fig:insert-one\] for a vertex-correction diagram. A suitable truncation of the form factor expansion is likely to contain more terms than in the exchange parametrization studies in Refs. . However, we still expect fast convergence with increasing maximal bond length. (For a more detailed discussion, see Section \[sec:phase\_hubb\] and \[sec:app-smooth-ff\].)
![Insertion of a unity operator in the form factor basis into a vertex-correction diagram.[]{data-label="fig:insert-one"}](mod-vert-corr.ti.eps){width=".84\linewidth"}
The insertion of truncated unity partitions generates a computational advantage in calculating the right hand side of the flow equations due to the decomposition of the loop integrals. As a consequence of separating the bosonic from the fermionic lines, the integrands only consist of a product of two fermionic propagators and two slowly varying form factors instead of two interaction channels that can be strongly peaked close to a phase transition. Since integrations over structures with sharp features are numerically expensive, the replacement by smooth functions makes the loop integration an easier task. This simplification is done at the cost of adding the projection operations \[enu:project-new\].). As can be seen from Eq. (\[eqn:proj\_mom\]), these additional tasks consist of two nested momentum integrals involving a product of four form factors and one exchange propagator. If the form factors correspond to fixed bond lengths, the calculation can be done most efficiently in position space [@Wang2012]. In this case, the form factors translate to superpositions of Kronecker deltas that limit the appearing sums to a fixed upper bond length (see Eq. (\[eqn:proj\_pos\])). This upper length is of course directly related to the truncation length of the form factor basis.
With these preliminary considerations, the projection step can be implemented in terms of evaluating overlaps of Kronecker-deltas and performing Fourier transforms of the exchange propagators for a finite set of vectors in position space. Moreover, it should be mentioned that the exchange propagator is the only object that depends on the fRG scale: all the other components (e.g. the non-vanishing Kronecker delta combinations and Fourier exponentials) stay the same for the whole fRG flow. Such a simplified dependence enables us to calculate these scale-independent parts only once and to use the result at all scales instead of repeating the same calculation at every step. We implemented the reuse of precomputed projection data in the code version for studies on the honeycomb lattice [@Pena2016]. In a calculation using $3217$ sampling points for each dependence on momentum $\mathbf{l}$ in Eqs. (\[eqn:proj\_VP\])-(\[eqn:proj\_VD\]), the recycling of data caused speedups of $2.6$ and $1.7$ in the case of truncations after $7$ and $13$ form factors respectively. Although the total size of those data is of the order of some gigabytes and this part of the code is not yet optimized in terms of cache lines and load balancing, the computation time for the projections can be reduced by storing precomputed data.[^3] In case of a well behaved loop integrand, the projection process needs the major part of the computation time. Then, the storing of data can accelerate the whole fRG flow significantly. In terms of performance, the latter case might be seen as the optimal area of application for the TUfRG.
Towards high-performance fRG {#sec:perform}
============================
Despite the physically motivated truncations that enter the TUfRG, the development of a computationally efficient method, that significantly reduces the time-to-solution while providing meaningful predictions of ground state properties, relies on the usage of high-performance computers. In the last decade the evolution of the building blocks of large computing architectures moved from single-core CPUs to compute nodes with multiple cores, where large numbers of them are interconnected in complex and heterogeneous networks. As a consequence of this evolution, it is only natural that a modern fRG implementation should be able to make use of a large number of compute cores in order to maximally exploit the parallelism of modern computing platforms. To this purpose, in our implementation, we make extensive use of the directive-based OpenMP as well as the Message Passing Interface (MPI) API, which are the most used standards for achieving shared memory and distributed memory parallelization, respectively.
A clear advantage of the TUfRG method lies in the fact that bosonic lines have been completely pulled out of the loops. Consequently, the integrals in step \[enu:chis-new\].) are generically more well behaved than in the exchange-parametrization approach (see Section \[sec:benefits\]). Furthermore, all the loop integrations for different form factor and bosonic momentum combinations are completely independent from each other[^4], and so embarrassingly parallel. In this step, communication is only necessary to share the results of the integration tasks between the different MPI processes. When compared to the time spent in computations, the communication overhead is negligible. In our current implementation we use MPI for distributing the bosonic momenta across the available compute nodes, i.e. the largest computation unit whose constituents share one block of memory, and OpenMP for parallelizing the form factor combinations of each bosonic momentum. To perform a single integration, for fixed form factor indices and momentum, we use the adaptive quadrature routine `DCUHRE` [@Berntsen1991].
For our implementation of step \[enu:project-new\].) we used the position space representation from Eq. in the variant that avoids large memory consumptions. Here we decided to accept a longer runtime for this part of the calculation, since in the $t$-$t'$ Hubbard model at van Hove filling step \[enu:chis-new\].) consumes the major part of the total computation time. As in step \[enu:chis-new\].) we use a hybrid parallelization, where the sum over $m'$ and $n'$ from the right-hand side of Eq. is distributed over the nodes with MPI, and the different components of the exchange propagators regarding the form factor indices $m$ and $n$ on the left-hand side are calculated in parallel using OpenMP.
The matrix multiplications in \[enu:matmul-new\].) are of minor relevance in terms of compute time and their implementation is therefore not optimized yet. It is based on nested *for*-loops, where those over external indices are parallelized using OpenMP. For calculations using up to $128$ nodes, the compute time needed by this step is negligible compared to the ones of steps \[enu:project-new\].) and \[enu:chis-new\].), due to the shared memory parallelism of the *for*-loop iterations.
As explained above, by performing the steps \[enu:project-new\].) - \[enu:matmul-new\].) we calculate the derivative of the exchange propagators with respect to the regularization scale. Since we aim to obtain these propagators at lower scales, we are left with solving ordinary differential equations (ODEs) of order one, which is a standard task that we have implemented with the use of an explicit ODE solver from the ‘Odeint’ library [@Ahnert2011].
![Speedup gained from shared memory parallelization relative to serial execution. The data have been produced using one node of a general purpose cluster with $24$ physical cores. For thread numbers higher than $24$ more than one thread is executed on one physical core.[]{data-label="fig:speedup_threads"}](speedup_threads.eps){width="55.00000%"}
![Speedup gained from distributed memory parallelization relative to an execution on one node. The data have been produced using a general purpose cluster with $24$ cores per node.[]{data-label="fig:speedup_nodes"}](speedup_nodes.eps){width="55.00000%"}
We analyzed our implementation of the TUfRG in terms of parallelization speedup using the JURECA compute cluster, which is located at the Jülich Supercomputing Centre. Every node consists of two Intel Xeon E5-2680 v3 Haswell CPUs with $12$ cores each, working at $2.5\, \textnormal{GHz}$. Due to simultaneous multithreading (SMT), every JURECA node supports $48$ threads in total. While there are nodes with larger main memory available within the cluster, we only used nodes with $128\,\textnormal{GB}$ as this appeared to be more than enough for our needs. For the following tests, we apply the implementation to the $t$-$t'$ Hubbard model at van Hove filling with model parameters $t'=-0.3\,t$ and $U=3.0\,t$. In this first application of the scheme, we neglect the self-energy feedback and use the $\Omega$-cutoff [@Husemann2009] as regulator. The form factor basis is chosen in a way that every element corresponds to a certain bond length[^5] and the truncation of that basis is introduced by an upper limit in the bond length. More precisely, a truncation at the $n$th nearest neighbor means that we only take form factors into account that correspond to the $n$th nearest neighbor bonds or to shorter bonds. Fig. \[fig:speedup\_threads\] shows how the runtime for one calculation of the ODEs’ right-hand side decreases with increasing number of threads running on one node, or in other words it shows the performance of our shared memory parallelization. In this context the quantity ‘speedup’ can be understood as the quotient of the runtime using the reference configuration, i.e. one thread, and the runtime using the respective number of threads. For these data we used a truncation of the form factor basis at the fourth nearest neighbor and compared to the data from Section \[sec:phase\_hubb\] we have reduced the density of sampling points for bosonic momenta in order to get a runtime below $24$ hours for the serial execution. As it can be seen from Fig. \[fig:speedup\_threads\], the speedup scales well with the number of threads in the region where the number of threads is less or equal to $24$ and each thread runs exclusively on a physical core. With $24$ threads and one thread per core we achieve a speedup of $19.5$, while by harvesting the additional underlying hardware parallelism when putting two threads on a physical core we arrive at a speedup of $24.5$.
At the next level we can further enhance this result by using the distributed memory parallelization. In Fig. \[fig:speedup\_nodes\] the speedup is plotted against the number of nodes, where the point of reference is the runtime when using $48$ threads on one node. Since we want to analyze the performance of our implementation under production conditions, we now choose the same resolution of the bosonic momenta as in Section \[sec:phase\_hubb\] (see Fig. \[fig:q-mesh\]) and a truncation at the fifth nearest neighbor. Due to huge consumptions of time, these conditions have not been feasible for the program executions for Fig. \[fig:speedup\_threads\]. However, this does not diminish the validity of our analysis, since an increase of the resolution will lead to an enlarged number of parallelizable work packages, which in turn rather supports parallelizability. Our implementation scales very well up to $64$ nodes and for $128$ nodes we still find a very reasonable speedup of $98.4$, as shown in Fig. \[fig:speedup\_nodes\].
The $t$-$t'$ Hubbard model as a test case {#sec:phase_hubb}
=========================================
![The discretization of the momentum dependence is shown for the particle-particle exchange propagator $\mathbf{P}(l)$ (left part, 992 grid points) and for the particle-hole exchange propagators $\mathbf{C}(l)$ and $\mathbf{D}(l)$ (right part, 6632 grid points).[]{data-label="fig:q-mesh"}](q-mesh_pp.eps "fig:"){width="42.00000%"} ![The discretization of the momentum dependence is shown for the particle-particle exchange propagator $\mathbf{P}(l)$ (left part, 992 grid points) and for the particle-hole exchange propagators $\mathbf{C}(l)$ and $\mathbf{D}(l)$ (right part, 6632 grid points).[]{data-label="fig:q-mesh"}](q-mesh_ph.eps "fig:"){width="42.00000%"}
We have applied our implementation of the TUfRG to the $t$-$t'$ Hubbard model on the square lattice, which is well studied [@LeBlanc2015; @Honerkamp2001; @Giering2012; @Eberlein2015; @Yamase2016] but still contains some regions in the parameter space with unclear ground state properties. The single particle dispersion is given as $$\epsilon(\mathbf{k}) = -2\, t\, (\cos(k_x)+\cos(k_y)) - 4\, t'\, \cos(k_x)\, \cos(k_y) - \mu
\label{eqn:hubb-disp}$$ which contains three free parameters in general. In addition a fourth parameter in the Hubbard model is given by the onsite density-density interaction strength $U$. A simultaneous rescaling of these four parameters will leave the physics of the system invariant, but will rescale all physical energies. To take this into account we measure all energies relative to the parameter $t$ and leave the value of $t$ undefined. Furthermore, we restrict ourselves to van Hove filling $\mu = 4\,t'$ and use $U=3\,t$ which leaves us with only one free parameter $t'$.
Fig. \[fig:q-mesh\] shows the discretization of the momentum space that we have used for the calculation of the exchange propagators. Inside the areas of high grid point density we have expected strong peaks of the exchange propagator values that need to be resolved more accurately. Those areas have been chosen according to the results of previous studies on this model and can also be motivated by simple single channel deliberations.
![The critical scales for different truncation (bond) lengths of the form factor basis are plotted against the second nearest neighbor hopping $t'$ for the $t$-$t'$ Hubbard model at van Hove filling with $U=3.0\,t$. Nearest neighbor bonds that correspond to the different truncations are shown in the inset. The gray bars separate the parameter regimes of the three observed instabilities, those are spin density wave (SDW) with $\mathbf{l}\sim(\pi,\pi)$, $d$-wave superconductivity ($d$SC) with zero total momentum and ferromagnetism (FM).[]{data-label="fig:conv_test"}](convergence_test.eps){width="85.00000%"}
In order to investigate the applicability of the insertion of truncated unity partitions, we have checked how the results change with increasing truncation length. To this end, we have performed the fRG flow in the parameter range from $t'=-0.10\,t$ to $t'=-0.45\,t$ with different truncations of the form factor basis. Starting from a truncation at the first nearest neighbor, we have successively increased the number of form factors until a sixth nearest neighbor truncation. The calculations have been started at an initial scale two orders of magnitude larger than the bandwidth and have been stopped when the largest absolute value of the exchange propagators has become one order of magnitude higher than the bandwidth. It turns out that the results do not depend on the precise values of the initial scale and the largest exchange propagator component. For each data point, the ground state type has been determined by the largest exchange propagator value by means of the corresponding channel (pairing, magnetic or charge), ordering vector, and form factor symmetry. Fig. \[fig:conv\_test\] shows the critical scales as functions of $t'$ and exhibits three different kinds of ground states. Spin density wave (SDW) and ferromagnetic (FM) instabilities manifest in the magnetic channel with an $s$-symmetry and ordering vector $\mathbf{l}\sim(\pi,\pi)$ and $\mathbf{l}=(0,0)$, respectively. The observed $d$-wave superconductivity ($d$SC) appears in the pairing channel with zero total momentum. By reason of clarity the transitions between the different phases are only shown sketchily in this plot. However, the transition values of $t'$ turn out to change very mildly with the truncation length for both transitions.
Focusing on the SDW regime, it becomes obvious that the critical scales are nearly unchanged by increasing the number of form factors. This shows that the important feedback from other channels—which lowers the critical scales compared to those from single channel calculations—is already contained in the TUfRG using a nearest neighbor truncation. Within the $d$SC regime we find a change of the critical scales by including the fourth nearest neighbor form factors. But this correction is small until we do not get too close to the phase transition to FM. Moreover, it can be seen from Fig. \[fig:conv\_test\] that numerically the last-named parameter region around $t'\approx-0.34\,t$ is the most difficult one. These difficulties are directly connected to the fact that the nature of this transition is highly unclear in general. There are fRG studies that find a quantum critical point between the two phases [@Honerkamp2001; @Giering2012] while others—like the present one—do not indicate such a phenomenon [@Husemann2009]. By further decreasing $t'$ we enter the FM region, where a larger jump occurs between the scales of a first and a second nearest neighbor truncation. Beyond that, the inclusion of longer bond form factors has only a small impact on the results. Taken together we find a fast convergence of the critical scales—more precisely the influence of form factors beyond second nearest neighbor is rather low—within the investigated parameter regime except for values of $t'$ close to the phase transition between $d$SC and FM.
Besides comparing different truncations with each other, a comparison to findings from other studies is necessary to validate the insertion of a truncated partition of unity into the flow equations. In Fig. 6 of Ref. and in curve (i) in the left part of Fig. 3 of Ref. the same model has been studied using exchange parametrization fRG in the same truncation of the flow equation hierarchy. The findings from those studies are very similar to our results in both the SDW and the $d$SC parameter regime, but in the region with a FM ground state those scales are closer to our results from first nearest neighbor truncation than to the higher order findings. Generally, it is not surprising that critical scales are reduced by increasing interchannel feedback—which is the consequence of taking more form factors into account—when the influence of the subleading ordering tendency ($d$SC) on the dominant one (FM) has destructive character. Hence, our higher order truncation results can be seen as quantitative corrections to the critical scales from the two previous investigations within the FM region.
Conclusion
==========
We have derived the TUfRG equations on the basis of an exchange parametrization in Sec. \[sec:tufrg\]. As argued in Sec. \[sec:benefits\], the loop integration in the TUfRG is much easier from a computational viewpoint than in a conventional exchange parametrization approach. This advantage has been obtained at the cost of adding a projection task which turns out to be of minor importance for the total computation time in many cases and can be accelerated by reusing precomputed data in the other cases. The convenience in numerics originates from a separation of fermionic and exchange propagators that at the same time simplifies the parallelization of the program code. Benefits from a hybrid parallelization in terms of speedup have been illustrated in Sec. \[sec:perform\]. As a consequence of accelerating the calculation by using many compute cores, it has been possible to access a large set of form factors in the TUfRG approach. Most importantly, a high-performance implementation combined with an efficient parametrization of the coupling function, as it is done in the TUfRG—originating from its relation to the exchange parametrization fRG—, should make applications to complex multiband systems possible.
A convergence check for the case of the $t$-$t'$ Hubbard model at van Hove filling has shown that the results converge fast with the number of form factors except for parameters close to the phase transition between $d$-wave superconductivity and ferromagnetism. In addition, we have seen a good agreement with results from previous exchange parametrization studies when using a comparable set of form factors. It has further been possible to take more form factors into account for producing results of higher accuracy.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Q.H. Wang and M. Salmhofer for discussions. Numerical experiments have been conducted within the JUBE workflow environment [@Luhrs2016] which has greatly facilitated data generation, management and analysis. The authors gratefully acknowledge the computing time granted by JARA-HPC and provided on the supercomputer JURECA at Jülich Supercomputing Centre (JSC). Furthermore, the German Research Foundation (DFG) is acknowledged for support via RTG 1995 and SPP 1459.
Relation to the SMFRG {#sec:rel_to_smfrg}
=====================
In the following we describe the relation of the TUfRG equations to the ones used in the SMFRG, which was introduced in Ref. . Instead of exchange propagators, the core objects now are three complementary approximate representations of the two-particle interaction $V (k_1,k_2,k_3)$: $$\begin{aligned}
V \left(k_1,k_2,k_3 \right) &\approx \sum_{m,n} f_m \left(\frac{\mathbf{k}_1-\mathbf{k}_2}{2} \right) \,f_n \left(\frac{\mathbf{k}_4-\mathbf{k}_3}{2} \right) \, V^P_{m,n} \left(k_1+k_2 \right) \, ,\\
V \left(k_1,k_2,k_3 \right) &\approx \sum_{m,n} f_m \left(\frac{\mathbf{k}_1+\mathbf{k}_3}{2} \right) \,f_n \left(\frac{\mathbf{k}_2+\mathbf{k}_4}{2} \right) \, V^C_{m,n} \left(k_1-k_3 \right) \, ,\\
V \left(k_1,k_2,k_3 \right) &\approx \sum_{m,n} f_m \left(\frac{\mathbf{k}_1+\mathbf{k}_4}{2} \right) \,f_n \left(\frac{\mathbf{k}_2+\mathbf{k}_3}{2} \right) \, V^D_{m,n} \left(k_3-k_2 \right) \, .\end{aligned}$$ The form of these three representations is reminiscent of the exchange parametrization of the single-channel coupling functions in Eqs. (\[eqn:app-PhiP\])-(\[eqn:app-PhiD\]). However, we are dealing with three different approximations of the same coupling function in the present case, and it depends on the context which one is used.
Formally, the bond length representations $$\begin{aligned}
\mathbf{V}^P \left(k_1+k_2 \right) & = \hat{P} \left[ V \right] (k_1 + k_2) \, ,\\
\mathbf{V}^C \left(k_1-k_3 \right) & = \hat{C} \left[ V \right] (k_1 - k_3) \, ,\\
\mathbf{V}^D \left(k_3-k_2 \right) & = \hat{D} \left[ V \right] (k_3 - k_2)\end{aligned}$$ can be projected out from $V$ by applying the projection rules (\[eqn:proj-P\])-(\[eqn:proj-D\]). From these relations one can directly see that those are the same objects as the ones calculated in Eqs. (\[eqn:proj\_VP\])-(\[eqn:proj\_VD\]) as part of the TUfRG procedure. In contrast to Refs. , the conventions used here render the matrices $\mathbf{V}^P$, $\mathbf{V}^C$, and $\mathbf{V}^D$ hermitian due to the Pauli principle and positivity. Their flow arises from a projection of all five one-loop diagrams of Fig. \[fig:diags-fleq\] to the respective channel: $$\begin{aligned}
\dot{\mathbf{V}}^P \left(l \right) & = \hat{P} \left[ \mathcal{T}_\mathrm{pp} + \mathcal{T}_\mathrm{ph}^\mathrm{cr} + \mathcal{T}_\mathrm{ph}^\mathrm{d} \right] (l) \, ,\\
\dot{\mathbf{V}}^C \left(l \right) & = \hat{C} \left[ \mathcal{T}_\mathrm{pp} + \mathcal{T}_\mathrm{ph}^\mathrm{cr} + \mathcal{T}_\mathrm{ph}^\mathrm{d} \right] (l) \, ,\\
\dot{\mathbf{V}}^D \left(l \right) & = \hat{D} \left[ \mathcal{T}_\mathrm{pp} + \mathcal{T}_\mathrm{ph}^\mathrm{cr} + \mathcal{T}_\mathrm{ph}^\mathrm{d} \right] (l) \, .\end{aligned}$$ This is nothing but the derivatives of Eqs. (\[eqn:proj\_VP\])-(\[eqn:proj\_VD\]) with respect to the regularization scale: $$\begin{aligned}
\label{equ:VP_dot}
\dot{\mathbf{V}}^P \left(l \right) & = - \dot{\mathbf{P}} (l) + \hat{P} \left[ \dot{\Phi}^\mathrm{C} \right] (l) + \hat{P} \left[ \dot{\Phi}^\mathrm{D} \right] (l) \, ,\\
\label{equ:VC_dot}
\dot{\mathbf{V}}^C \left(l \right) & = - \hat{C} \left[ \dot{\Phi}^\mathrm{SC} \right] (l) + \dot{\mathbf{C}} (l) + \hat{C} \left[ \dot{\Phi}^\mathrm{D} \right] (l) \, ,\\
\label{equ:VD_dot}
\dot{\mathbf{V}}^D \left(l \right) & = - \hat{D} \left[ \dot{\Phi}^\mathrm{SC} \right] (l) + \hat{D} \left[ \dot{\Phi}^\mathrm{C} \right] (l) + \dot{\mathbf{D}} (l) \, .\end{aligned}$$ Scale derivatives of single channel coupling functions $\Phi^\mathrm{SC}$, $\Phi^\mathrm{C}$ and $\Phi^\mathrm{D}$ can be expressed by derivatives of exchange propagators as implied by Eqs. (\[eqn:app-PhiP\])-(\[eqn:app-PhiD\]). Using Eqs. (\[eqn:matmul-P\])-(\[eqn:matmul-D\]) we obtain a closed system of differential equations for $\mathbf{V}^P$, $\mathbf{V}^C$, and $\mathbf{V}^D$. While in the TUfRG the exchange propagators are the central objects that are stored during the whole fRG calculation, in the SMFRG only the derivatives can be known. Those values have to be calculated in every SMFRG step, since they are needed temporarily for calculating the right-hand sides of Eqs. (\[equ:VP\_dot\])-(\[equ:VD\_dot\]).
![Graphical representation of the steps in the calculation of the increment for the interaction in the SMFRG approach. The upper diagram corresponds to steps \[enu:chis\].) and \[enu:matmul\].) and the lower one to the projection in step \[enu:project\].). The brick-stones in the upper part correspond to the *full* interaction projected to the respective channel with $B= P$, $C$, or $D$. Dashed lines correspond to simple contractions and *not* to fermionic or bosonic propagators.[]{data-label="fig:steps-smfrg"}](matmul.ti.eps "fig:"){width=".76\linewidth"}\
[$\downarrow$]{}\
![Graphical representation of the steps in the calculation of the increment for the interaction in the SMFRG approach. The upper diagram corresponds to steps \[enu:chis\].) and \[enu:matmul\].) and the lower one to the projection in step \[enu:project\].). The brick-stones in the upper part correspond to the *full* interaction projected to the respective channel with $B= P$, $C$, or $D$. Dashed lines correspond to simple contractions and *not* to fermionic or bosonic propagators.[]{data-label="fig:steps-smfrg"}](project.ti.eps "fig:"){width=".35\linewidth"}
In summary, calculating the increment in the SMFRG flow of the interaction involves the following steps (for a graphical representation, see Fig. \[fig:steps-smfrg\]):
1. Calculate the form factor projected fermionic loops $\dot{\boldsymbol\chi}^\mathrm{pp} (l)$ and $\dot{\boldsymbol\chi}^\mathrm{ph} (l)$ in Eq. (\[eqn:chis\]). \[enu:chis\]
2. Evaluate $\dot{\mathbf{P}} (l)$, $\dot{\mathbf{C}} (l)$, and $\dot{\mathbf{D}} (l)$ by performing matrix multiplications in the form factor basis \[see Eqs. (\[eqn:matmul-P\])-(\[eqn:matmul-D\])\]. \[enu:matmul\]
3. Project $\dot{\mathbf{P}} (l)$, $\dot{\mathbf{C}} (l)$, and $\dot{\mathbf{D}} (l)$ to other channels and obtain $\dot{\mathbf{V}}^P (l)$, $\dot{\mathbf{V}}^C (l)$, and $\dot{\mathbf{V}}^D (l)$ according to Eqs. (\[equ:VP\_dot\])-(\[equ:VD\_dot\]). \[enu:project\]
From a computational viewpoint these steps are the same as the ones from Sec. \[sec:trunc\_uni\], just in a different order of execution. Therefore, the computational effort is the same for both schemes.
Taken together, the main difference between TUfRG and SMFRG is the choice of the core objects. Within the TUfRG framework we permanently store the exchange propagators during the flow, as it is done in exchange parametrization studies like . In contrast, the three different projections of $V$ play this role within the SMFRG. From Eqs. (\[eqn:proj\_VP\])-(\[eqn:proj\_VD\]) we can directly obtain the projected $V$s from the exchange propagators in the TUfRG scheme and thus should be able to recover the SMFRG results. Vice versa, it would be necessary to invert those equations in order to extract the TUfRG results from the SMFRG. This is nothing but a matrix inversion, which is a very expensive task, since the dimension of that matrix is equal to the total number of components of the projected couplings or of the exchange propagators, respectively. E.g., in the case of a truncation at the fifth nearest neighbor and the same momentum resolution as in Section \[sec:phase\_hubb\] one would need to invert a matrix with a dimension of about $8.9 \times 10^6$.
Let us now compare the applicability of the projected couplings and the exchange propagators in the context of recovering the full coupling function, which is needed for self-energy calculations, e.g.. On the side of the TUfRG there is a unique way of calculating the single-channel coupling functions from the exchange propagators and from those the full coupling. The three strong momentum dependences, one from each channel, directly enter the resulting object, since it is just a sum of the three channels. On the side of the SMFRG there are three different possibilities to undo the projection, i.e. we can use $\mathbf{V}^P$, $\mathbf{V}^C$, or $\mathbf{V}^D$ to get $V$. The result will strongly depend on the choice we have used for the calculation, since the projected couplings only contain one of the three important momentum dependencies while the other two are smoothened out by the projection process and can not be recovered due to a truncated form factor basis.
With the last two paragraphs in mind and in view of the fact that the numerical effort is the same, we see a clear advantage in using the exchange propagators as central objects instead of the projected couplings.
Applicability of the truncation of form factor unity operators {#sec:app-smooth-ff}
==============================================================
In the following, an argument for the applicability of the insertion of truncated unity partitions in vertex-correction and box diagrams is given.
For this purpose, let us rewrite the rule (\[eqn:proj-D\]) for the projection to the direct particle-hole channel in the following way: $$\label{eqn:internal-proj}
\hat{D} \left[ F \right]_{m,n} (l) = \left. \int \! d\mathbf{k'} \, f_n (\mathbf{k'}) \, \, \hat{E}_D \left[F \right]_m (l,k') \right|_{k'_0=0} \, ,$$ where the external projections $$\label{eqn:external-proj}
\hat{E}_D \left[ F \right]_m (l,k') = \left. \int \! d\mathbf{k} \,
f_m (\mathbf{k}) \, F \left(k+\frac{l}{2},k'-\frac{l}{2},k'+\frac{l}{2}\right) \right|_{k_0=0}$$ have been defined in a similar way as in Ref. . While the contractions of the coupling function $F$ with the two form factors $f_m (\mathbf{k})$ and $f_n (\mathbf{k'})$ have been treated on equal footing in Eq. (\[eqn:proj-D\]), the contraction with the form factor $f_n (\mathbf{k'})$ resulting from the insertion of a truncated partition of unity is performed *after* the external projection $\hat{E}_D [F]_m (l,k') $ in Eq. (\[eqn:internal-proj\]). In the following, we will argue that, for external form factors corresponding to short bond lengths, the external projections $ \hat{E}_D [F]_m (l,k') $ vary slowly in $\mathbf{k}'$. The sum over the inserted form factors $f_n (\mathbf{k'})$ can therefore be truncated after a certain bond length, since contributions from longer bonds vanish.
Here, we focus on the feedback of the particle-particle and the crossed particle-hole channels on the direct particle-hole channel. We therefore consider Eq. (\[eqn:external-proj\]) with $$\begin{aligned}
F (k_1,k_2,k_3) &= \sum_{c,c'} f_c \left( \frac{\mathbf{k}_1 - \mathbf{k}_2}{2} \right) \, f_{c'} \left( \frac{\mathbf{k}_4 - \mathbf{k}_3}{2} \right) \, P_{c,c'} \left( k_1 + k_2 \right) \\
& \left. \quad + \sum_{c,c'} f_c \left( \frac{\mathbf{k}_1 + \mathbf{k}_3}{2} \right) \, f_{c'} \left( \frac{\mathbf{k}_2 + \mathbf{k}_4}{2} \right) \, C_{c,c'} \left( k_1 - k_3 \right) \right|_{k_4=k_1+k_2-k_3} \, .\end{aligned}$$ However, this example is generic for the feedback between different interaction channels. For the external projection, we then have $$\begin{aligned}
\hat{E}_D \left[ F \right]_m (l,k') &= \int \! d\mathbf{k} \, f_m (\mathbf{k}) \left[
\sum_{c,c'} f_c \left( \frac{\mathbf{k} - \mathbf{k'} +\mathbf{l}}{2} \right) \, f_{c'} \left( \frac{\mathbf{k} - \mathbf{k'} - \mathbf{l}}{2} \right) \, P_{c,c'} \left( k + k' \right) \right. \\
& \left. \qquad + \sum_{c,c'} f_c \left( \frac{\mathbf{k} + \mathbf{k'} + \mathbf{l}}{2} \right) \, f_{c'} \left( \frac{\mathbf{k} + \mathbf{k'} - \mathbf{l}}{2} \right) \, C_{c,c'} \left( k - k' \right) \right]_{k_0=0} \, .\end{aligned}$$ Apparently, if the exchange propagators $P$ and $C$ are slowly varying functions of the total and transfer momenta, the dependence of $\hat{E}_D \left[ F \right]_m (l,k')$ on $\mathbf{k'}$ is smooth. But also if these exchange propagators are strongly peaked, the strong dependence on $\mathbf{k'}$ is washed out by the above convolution-like integral. This can be more clearly seen after a substitution of the integration variable: $$\begin{aligned}
\hat{E}_D \left[ F \right]_m (l,k') &= \int \! d\mathbf{u} \, \left. f_m (\mathbf{u} - \mathbf{k'})
\sum_{c,c'} f_c \left( \frac{\mathbf{u} - 2\mathbf{k'} +\mathbf{l}}{2} \right) \, f_{c'} \left( \frac{\mathbf{u} - 2\mathbf{k'} - \mathbf{l}}{2} \right) \, P_{c,c'} \left( u \right) \right|_{u_0=k'_0} \\
&+ \int \! d\mathbf{u} \, \left. f_m (\mathbf{u} + \mathbf{k'})
\sum_{c,c'} f_c \left( \frac{\mathbf{u} + 2\mathbf{k'} + \mathbf{l}}{2} \right) \, f_{c'} \left( \frac{\mathbf{u} + 2\mathbf{k'} - \mathbf{l}}{2} \right) \, C_{c,c'} \left( u \right) \right|_{u_0=-k'_0} \, .\end{aligned}$$ Let us now consider form factors $$f_r (\mathbf{q}) = \sum_\mathbf{R} f_r (\mathbf{R}) \, e^{-i \mathbf{q} \cdot \mathbf{R}}$$ corresponding to fixed bond lengths. This means that $f_r (\mathbf{R}) $ only contributes for a fixed value of $| \mathbf{R} |$, which is given by $r$. We then can easily observe that the external projection $\hat{E}_D \left[ F \right]_m (l,k')$ varies smoothly with $\mathbf{k'}$. More formally, in the internal projection $\hat{D} [F]_{m,n} (l)$ in Eq. (\[eqn:internal-proj\]), contributions with values of $n$ that correspond to a large bond length will be absent. In particular, the projection integral vanishes once the bond length of $f_n (\mathbf{k'})$ exceeds the sum of the maximal bond lengths of the form factors labeled with $m$, $c$, and $c'$.
This argument can be straightforwardly carried over to all other inter-channel feedback contributions in the vertex-correction and box diagrams in the fRG flow equations. Therefore, a numerically tractable truncation of the unity operators inserted between internal fermionic and bosonic lines should be applicable.
Useful symmetries {#sec:symmetries}
=================
The spin-independent coupling function $V(k_1,k_2,k_3)$ obeys the relation $$V( k_1,k_2,k_3) = V(k_2,k_1,k_1+k_2-k_3)$$ stemming from the Pauli principle and the positivity of the action corresponding to the Hermiticity of the Hamiltonian requires $$V( k_1,k_2,k_3) = V(\hat{k}_1+\hat{k}_2-\hat{k}_3,\hat{k}_3,\hat{k}_2)^\ast \, ,$$ where $ \hat{k} = (-k_0,\mathbf{k}) $ (see Appendix A of Ref. for a discussion). In the channel decomposed form this results in $$\begin{aligned}
\Phi^\mathrm{SC}_{l,q,q'} &= \Phi^\mathrm{SC}_{l,-q,-q'} = \Phi^\mathrm{SC}_{l,q',q}\, ,\\
\Phi^\mathrm{C}_{l,q,q'} &= \Phi^\mathrm{C}_{-l,q',q} = \Phi^\mathrm{C}_{l,q',q}\, ,\\
\Phi^\mathrm{D}_{l,q,q'} &= \Phi^\mathrm{D}_{-l,q',q} = \Phi^\mathrm{D}_{-l,q,q'} \, ,\end{aligned}$$ where the first equality sign in each line corresponds to the Pauli principle and the second one to positivity. Note that, for the latter, we have already exploited that all coupling functions are mapped to their complex conjugates under frequency inversion. Using these relations with regard to exchange propagators it can be shown that the matrices $\mathbf{P}$, $\mathbf{C}$, and $\mathbf{D}$ are symmetric (cf. Ref. ): $$B_{m,n} (l) = B_{n,m} (l) \, , \quad \text{where} \quad B \in \{ P,C,D\} \, .$$ In addition, the constraints $$\begin{aligned}
P_{m,n} (l) &= \pi_m \pi_n \, P_{m,n} (l) \, , \\
C_{m,n} (-l) &= C_{m,n} (l) \, ,\\
D_{m,n} (-l) &= D_{m,n} (l)\end{aligned}$$ hold, where $\pi_m$ denotes the parity eigenvalue of the $m$-th form factor, i.e., $ f_m (- \mathbf{k}) = \pi_m \, f_m (\mathbf{k})$. If the form factors are chosen to transform with irreducible representations of the point group, some matrix elements of $\mathbf{P}$, $\mathbf{C}$, and $\mathbf{D}$ vanish at points of high symmetry due to Schur’s second lemma. [@Maier2013] Under frequency inversion, all three exchange propagators behave as $$B_{m,n} (\hat{l}) =B_{m,n} (l)^\ast \, , \quad \text{where} \quad B \in \{ P,C,D\} \, .$$
For the aforementioned possibility of calculating the scale independent parts of the projection operations at the start of the flow and storing them, exploiting symmetries can reduce the substantial increase in memory usage. Such projection operations can be expressed as $$\hat{B}\left[ \Phi^\mathrm{B'} \right]_{m,n}(l)=\sum_{\substack{m',n' \\ \mathbf{q}}} A_{m,n,m',n'}^{B,B'}(\mathbf{q}, \mathbf{l}) B'_{m',n'}(\mathbf{q},q_0=0)\, ,$$ where—compared to Eqn. \[eqn:proj\_pos\]—we have replaced the exchange propagator in position space by its Fourier series. $\mathbf{A}^{B,B'}$ is constant during the flow, and corresponds to $$\begin{aligned}
A_{m,n,m',n'}^{P,C}(\mathbf{q}, \mathbf{l}) =
\sum_{\mathbf{R_1},\mathbf{R_2},\mathbf{R_3}} \, &f_m\left(-\frac{\mathbf{R_1}}{2}+\frac{\mathbf{R_2}}{2}-\mathbf{R_3}\right) \,
f_n\left(\frac{\mathbf{R_1}}{2}-\frac{\mathbf{R_2}}{2}-\mathbf{R_3}\right) \,\\
\times \ &f_{m'}\left(\mathbf{R_1}\right) \, f_{n'}\left(\mathbf{R_2}\right) \,
e^{i \mathbf{q} \mathbf{R_3}} \, e^{-i\frac{1}{2}\mathbf{l}\cdot(\mathbf{R_1}+\mathbf{R_2})}\end{aligned}$$ in the example case of feedback from the $C$ channel to the $P$ channel. The objects $\mathbf{A}^{B,B'}$ are all real valued, and the 6 possible combinations of $B,B'$ are related in pairs $$\begin{aligned}
\mathbf{A}^{P,C}&=\pi_n \mathbf{A}^{P,D}\, ,\\
\mathbf{A}^{C,P}&=\pi_{n'} \mathbf{A}^{D,P}\, ,\\
\mathbf{A}^{D,C}&= \mathbf{A}^{C,D}\end{aligned}$$ so that only three of them have to be computed and stored. Furthermore, each of these three objects has given symmetries with respect to exchange of form factor indices or momentum inversion. For the previous example those symmetries can be written as $$\begin{aligned}
A_{n,m,m',n'}^{P,C}(\mathbf{q}, \mathbf{l})&= \pi_m \pi_n A_{m,n,m',n'}^{P,C}(\mathbf{-q}, \mathbf{l})\, ,\\
A_{m,n,n',m'}^{P,C}(\mathbf{q}, \mathbf{l})&= \pi_m \pi_n A_{m,n,m',n'}^{P,C}(\mathbf{-q}, \mathbf{l})\, ,\\
A_{m,n,m',n'}^{P,C}(\mathbf{-q}, \mathbf{-l})&= \pi_m \pi_n \pi_{m'} \pi_{n'} A_{m,n,m',n'}^{P,C}(\mathbf{q}, \mathbf{l})\, .\end{aligned}$$ The last-named equation can further be generalized by using the point group’s symmetry operations. In a first step for each operation $\hat S$ a matrix has to be set up that contains prefactors of the linear combination of form factors necessary to compose $f_m\left(\hat S \, \mathbf k \right)$, i.e., $$f_m\left(\hat S\, \mathbf k \right) \, = \, \sum_{\tilde m} \, C^{\hat S}_{m,\tilde m} \, f_{\tilde m}\left(\mathbf k \right) \, .$$ As a second step we can relate the elements of $\mathbf{A}^{B,B'}$ to the symmetry transformed ones, e.g., $$A_{m,n,m',n'}^{P,C}(\hat S \, \mathbf{q}, \hat S \, \mathbf{l})= \sum_{\tilde m,\tilde n,\tilde m',\tilde n'} C^{\hat S}_{m,\tilde m} \, C^{\hat S}_{n,\tilde n} \, C^{\hat S}_{m',\tilde m'} \, C^{\hat S}_{n',\tilde n'} \, A_{\tilde m,\tilde n,\tilde m',\tilde n'}^{P,C}(\mathbf{q}, \mathbf{l})\, .$$
The definitions of the form factor projected loops from Eqn. \[eqn:chis\] directly indicate that the matrices $\boldsymbol{\chi}^\mathrm{pp}$ and $\boldsymbol{\chi}^\mathrm{ph}$ are symmetric $$\begin{aligned}
\chi^\mathrm{pp}_{n,m}(l) \, &= \, \chi^\mathrm{pp}_{m,n}(l) \, , \\
\chi^\mathrm{ph}_{n,m}(l) \, &= \, \chi^\mathrm{ph}_{m,n}(l) \, ,\end{aligned}$$ and an inversion of the integration variable shows that in the particle-particle case only elements corresponding to form factors with the same parity may be nonzero $$\chi^\mathrm{pp}_{m,n}(l) \, = \, \pi_m \, \pi_n \, \chi^\mathrm{pp}_{m,n}(l) \, .$$ The usage of point group symmetries yields $$\chi^\mathrm{pp/ph}_{m,n} \left( \hat S \, \mathbf{l},l_0 \right) \, = \, \sum_{\tilde m,\tilde n} \, C^{\hat S}_{m,\tilde m} \, C^{\hat S}_{n,\tilde n} \, \chi^\mathrm{pp/ph}_{\tilde m,\tilde n} \left( \mathbf{l},l_0 \right) \, .$$
[^1]: In a first step the contributions are calculated in a single-channel fashion: only intra-channel contributions are evaluated. The inter-channel feedback is done in a second step by a projection of the single-channel results onto the other two.
[^2]: Formally, our scheme reproduces the original one-loop flow equations of Sec. \[sec:su2-inv\] for a complete (infinite) set of form factors.
[^3]: See \[sec:symmetries\] for details on how symmetries can be used for minimizing the memory consumption.
[^4]: Note that the same is true for the bosonic frequencies. However, in the current implementation we neglect the frequency dependence of the exchange propagators and focus on the zero frequency terms.
[^5]: This still leaves some freedom for the specification of the form factors. In the current implementation these basis functions are chosen to transform according to the irreducible representations of the $C_{4v}$ point group, i.e., they fulfill $s$-, $p$-, $d$-, and $g$-symmetry, respectively.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In an attempt to generalize general relativity, we propose a new Hermitian theory of gravity. Space-time is generalized to space-time-momentum-energy and both the principles of general covariance and equivalence are extended. The theory is endowed with a Hermitian metric on a complex manifold. The Hermitian metric contains, apart from the symmetric metric, an anti-symmetric part, which describes dynamical torsion. The causality structure is changed in a way such that there is a minimal time for events to be in causal contact and a maximal radius for a non-local instantaneous causally related volume. The speed of light can exceed the conventional speed of light in non-inertial frames and accelerations are bounded. We have indications that the theory of Hermitian gravity yields general relativity at large scales and a theory equivalent to general relativity at very small scales, where the momenta and energies are very large. As an example, we study cosmology in Hermitian gravity, where matter is described by two scalar fields. While at late times Hermitian gravity reproduces the standard cosmological FLRW models, at early times it differs significantly: quite generically the Universe of Hermitian cosmology exhibits a bounce where a maximal expansion rate (Ricci curvature) is attained. Moreover, we prove that no cosmological constant is permitted at the classical level within our model of Hermitian cosmology.'
author:
- Christiaan Mantz
- Tomislav Prokopec
date: '**'
title: Hermitian Gravity and Cosmology
---
Motivation
==========
According to Albert Einstein’s principle of relativity the laws of physics are independent of system of reference. The principle of equivalence states that an observer cannot tell whether he is accelerating or placed in a gravitational field. If we follow Einstein’s principle of relativity closely one could argue that there must be a similar principle of equivalence between rotating observers and observers placed in a torsion field. A torsion field is a gravitational field which causes observers to rotate [@Mao:2006bb]. Theories of generalized gravity, in which dynamical torsion is present, have been proposed, using the standard principle of covariance. The theory proposed by Moffat [@Moffat:1978tr] has unsatisfactory properties [@Janssen:2006jx][@Damour:1991ru], which caused many to give up dynamical torsion.
In this article we propose a theory of dynamical torsion, by not only extending the principle of equivalence, but also by extending the principle of covariance. We generalize space-time to space-time-momentum-energy, imposing the reciprocity symmetry as a symmetry between space-time and momentum-energy, as was suggested by Max Born in 1938 [@Born1938Reciprocity]. Max Born’s original motivation behind the idea that the laws of physics should be invariant under the reciprocity transformation was that position and momentum operators of quantum mechanics obey the reciprocity symmetry transformation. Hence a theory unifying quantum mechanics and gravity should also be invariant under the reciprocity transformation. Our aim is to formulate this new theory, incorporating the reciprocity symmetry and simultaneously satisfying an extended principle of general covariance. We will leave the quantum aspect of the theory for future work.
The Reciprocity Principle
-------------------------
According to Max Born the laws of physics are invariant under the reciprocity transformation, which is given by $$\begin{aligned}
\label{eq:reciprocity transformation}
x^\mu \rightarrow p^\mu
\phantom{halloda}
p^\mu \rightarrow - x^\mu,\end{aligned}$$ where $x^\mu$ and $p^\mu$ are the four vectors ($ct,\vec{x}$) and ($\frac{E}{c},\vec{p}$), respectively. The components of the angular momentum $$\begin{aligned}
\nonumber%\label{eq:}
x_\mu p_\nu - p_\mu x_\nu
=
M_{\mu\nu}\end{aligned}$$ are indeed invariant under the reciprocity transformation. Since torsion couples to angular momentum [@Mao:2006bb], it seems natural to demand that a new theory describing dynamical torsion should be invariant under the reciprocity transformation.
Note that, when quantizing this new theory, which we leave for future work, the commutation relations from quantum mechanics $$\begin{aligned}
{\hat{x}}^\mu {\hat{p}}_\nu - {\hat{p}}_\nu {\hat{x}}^\mu
=
i\hbar \delta^\mu_\nu
\label{eq:commutation relations}\end{aligned}$$ are also invariant under the reciprocity relation.
General Relativity and the Reciprocity Principle
------------------------------------------------
The theory of general relativity describes our universe at large scales (at the moment we are not considering cosmological issues such as dark matter and dark energy) and it generalizes classical mechanical ideas as orbits, instead of wave functions, in order to describe particles. The four dimensional line element $$\begin{aligned}
\label{eq:line element spacetime}
d s^2
=
g_{\mu\nu}d x^\mu d x^\nu\end{aligned}$$ is a fundamental notion in the theory of general relativity. It is clear that general relativity and the way distances are determined (\[eq:line element spacetime\]) breaks the reciprocity symmetry (\[eq:reciprocity transformation\]). Demanding that the theory, unifying quantum mechanics and general relativity, should respect the principle of reciprocity, we can state a four dimensional momentum-energy line element $$\begin{aligned}
\label{eq:line element MomentumEnergy}
d \sigma^2
=
\gamma^{\mu\nu}d p_\mu d p_\nu,\end{aligned}$$ which should dominate over the space-time line element (\[eq:line element spacetime\]), whenever the momenta are very large compared to this position length scale.
According to the classical laws the momentum $p^\mu$ is given by $m \dot{x^\mu}$, which corresponds to the tangent vector of the path taken. The idea of having a tangent space at each point of the manifold, corresponding to the physical idea of the momentum as tangent vector, is clearly only applicable in the classical realm of physics, when momenta are small compared to distances. For the sake of brevity, Max Born called this scale, at which the theory of general relativity is valid the molar world [@Born1938Reciprocity], while he called the “small world”, which is described by the momentum energy line element (\[eq:line element MomentumEnergy\]) the nuclear world. The world, which lies in between these worlds on the energy-momentum and space-time scales, is familiarly called the quantum world, which is the realm of quantum gravity.
Since general relativity is governed by Einstein’s equations $$\begin{aligned}
\nonumber%\label{eq:}
R_{\mu\nu} - \frac{1}{2}g_{\mu\nu}R - \Lambda g_{\mu\nu}
=
\kappa T_{\mu\nu},\end{aligned}$$ we can state via the principle of reciprocity the reciprocal Einstein equations $$\begin{aligned}
\nonumber%\label{eq:}
P^{\mu\nu} - \frac{1}{2}\gamma^{\mu\nu}P - \Lambda'\gamma^{\mu\nu}
=
\kappa'{T'}^{\mu\nu},\end{aligned}$$ which are supposed to govern the momentum-energy curvature of the nuclear world.
We now have a vague idea of how the theory should behave in certain limits and the principles it should obey, namely the principle of reciprocity and the generalized principle of general covariance. Our goal is now to construct a theory which obeys all these limits and principles. The principle of general covariance suggests that there should exist a space-time-momentum-energy line element that specifies a corresponding space-time-momentum-energy interval, which is absolute in the sense that all observers would agree on it; interpreting the momentum-energy coordinates as coordinates, specifying non inertial frames, all relatively non inertial moving observers should agree on the measured space-time-momentum-energy interval. While our proposal extends the standard covariance principle, proposed by Einstein in 1905 [@Einstein1905SR], it necessarily breaks this principle of covariance by introducing energy-momentum into the distance measurements. Space-time then becomes a relative space with respect to observers moving non inertially with respect to each other and becomes absolute only in the limit of relatively inertial moving observers.
Almost Complex Structure
========================
Clearly we need a metric on a manifold to mathematically describe space-time-momentum-energy curvature. In order to build a theory which is reciprocal in momentum and space and at the same time reduces to the theory of general relativity in the molar limit, we need complex manifolds with a Hermitian metric. This is the case because the Hermitian metric is defined such that it is invariant under the reciprocity transformation (\[eq:reciprocity transformation\]). Any $2d$ dimensional manifold, with a $d$ $x$ and $d$ $y$ coordinates, locally admits a tensor field $J$ [@Nakahara:2003nw], which maps the tangent space of the manifold into itself, $J_p : T_p M \rightarrow T_p M $, in the following manner $$\begin{aligned}
\label{eq:reciprocity transformation by J operation}
J\left(x^\mu\right) \rightarrow y^\mu
\phantom{hallo}
J\left(y^\mu\right) \rightarrow - x^\mu,\end{aligned}$$ where the index $\mu$ runs from $0$ to $d-1$. It is clear that this map is equivalent to the reciprocity transformation (\[eq:reciprocity transformation\]), if the $y$ coordinate is interpreted as the energy-momentum coordinate. The map $J$, also known as the “almost complex structure” operator, may be defined globally on a complex manifold and then it specifies completely the complex structure of the manifold. A metric C, which is invariant under the action of this $J$ operator in the following way $$\begin{aligned}
\nonumber%\label{eq:}
C_p \left( J_p Z , J_p W \right)
=
C_p \left( Z , W \right)
\,,\end{aligned}$$ is a Hermitian metric, where $Z,W \in T_p M$ and $T_p M$ is the complexified tangent space [@Nakahara:2003nw]. The action of the almost complex structure operator on the basis vectors of the complexified tangent space follows from the definitions of the almost complex structure map and these basis vectors $$\begin{aligned}
\label{eq:almost complex structure z zbar operation}
J_p\left(\frac{\partial}{\partial z^\mu}\right)
=
i \frac{\partial}{\partial z^\mu}
\phantom{halloda}
J_p\left(\frac{\partial}{\partial \bar{z}^\mu}\right)
=
- i \frac{\partial}{\partial \bar{z}^\mu}.\end{aligned}$$ Consider a full complex metric $$\begin{aligned}
\label{complex metric}
C
=
C_{\mu\nu} d z^{\mu} \otimes d z^{\nu}
&+&
C_{\mu\bar\nu} d z^{\mu} \otimes d z^{\bar\nu}
\\\nonumber
+
C_{\bar\mu\nu} d z^{\bar\mu} \otimes d z^{\nu}
&+&
C_{\bar\mu\bar\nu} d z^{\bar\mu} \otimes d z^{\bar\nu}
\,.\end{aligned}$$ A Hermitian metric is a complex metric which has – as a consequence of the reciprocity symmetry – vanishing $C_{\mu\nu}$ and $C_{\bar\mu\bar\nu}$ components: $$\begin{aligned}
\nonumber%\label{eq:hermitean metric}
C
=
C_{\mu\bar\nu} d z^{\mu} \otimes d z^{\bar\nu}
+
C_{\bar{\mu}\nu} d z^{\bar{\mu}} \otimes d z^{\nu},\end{aligned}$$ where barred indices $z^{\bar{\mu}} \equiv {\bar{z}}^{\mu}$ denote complex conjugation. The basic definitions of complex manifolds do not differ too much from the usual definitions of a manifold, except for the fact that the complex manifold is locally homeomorphic to the complex space $\mathbb{C}^m$ and the coordinate transformations are holomorphic and hence satisfy the Cauchy-Riemann equations.
The Hermitian Metric
====================
We can write the Hermitian line element in eight dimensional form $$\begin{aligned}
\nonumber
ds^2 &=& d\boldsymbol{z}^T\cdot \boldsymbol{C} \cdot d\boldsymbol{z}
=
d\boldsymbol{z}^m\boldsymbol{C}_{mn}d\boldsymbol{z}^n
\\\label{eq:line element hermitian metric subsection}
&=&
(dz^{\mu},
dz^{\bar{\mu}})
\left(\begin{array}{cc}
0 & C_{\mu\bar{\nu}}\\
C_{\bar{\mu}\nu} & 0
\end{array}\right)
\left(\begin{array}{c}
dz^{\nu}\\
dz^{\bar{\nu}}
\end{array}\right),\end{aligned}$$ where the Latin indices can take the values $0,1,...,$$d-1$,$\bar{0},\bar{1},...,\overline{d-1}$, where the Greek indices can take the values $0,1,...,d-1$ and where the number $d$ is the complex dimension of the complex manifold. The entries of the metric $C_{mn}$ are functions of holomorphic and anti antiholomorphic vielbeins defined as follows [^1] $$\begin{aligned}
\label{eq:vielbein complex metric in terms of vielbeins}
&C_{\mu\bar{\nu}}
=
e(z)_{\mu}^{\phantom{\mu}a}\eta_{ab}e(\bar{z})_{\bar{\nu}}^{\phantom{\bar{\nu}}b}\\\nonumber
&C_{\bar{\mu}\nu}
=
e(\bar{z})_{\bar{\mu}}^{\phantom{\bar{\nu}}a} \eta_{a b}e(z)_{\nu}^{\phantom{\nu}
b},\end{aligned}$$ where $\eta_{ab} \equiv $ diag$(-1,1,1,1)$. For completeness, we also quote the other two elements of the full complex metric (\[complex metric\]), $$\begin{aligned}
\label{eq:complex metric in terms of vielbeins:2}
C_{\mu{\nu}}
&=&
e(z)_{\mu}^{\phantom{\mu}a}\eta_{ab}e({z})_{{\nu}}^{\phantom{\bar{\nu}}b}
\\\nonumber
C_{\bar{\mu}\bar\nu}
&=&
e(\bar{z})_{\bar{\mu}}^{\phantom{\bar{\nu}}a}
\eta_{a b}e(\bar z)_{\bar\nu}^{\phantom{\nu}
b}
\,.\end{aligned}$$ Note that holomorphy of the vielbeins $e_\mu^{\,a}=e_\mu^{\,a}(z)$ implies the reality condition for the line element, $$\begin{aligned}
\label{eq:reality condition of the line element}
{ds^2}^\dag = ds^2,\end{aligned}$$ since it implies the following definition for complex conjugation of the metric, $$\begin{aligned}
\nonumber
{C_{\bar{\mu}\nu}}^* \equiv \overline{ C_{\bar{\mu}\nu}} =
C_{\mu\bar{\nu}}.\end{aligned}$$ We can also write the Hermitian line element in its familiar four dimensional form $$\begin{aligned}
\label{eq:Hermitian familiar form line element}
ds^2
=
2d z^T \cdot C \cdot d \bar{z}
=
2dz^\mu C_{\mu\bar{\nu}}d z^{\bar{\nu}},\end{aligned}$$ which is a logical extension of the complex inner product $\langle
w,v\rangle = \bar{w_i}v_i $. Note that, in this familiar form (\[eq:Hermitian familiar form line element\]), the Hermitian metric is actually Hermitian in the usual way, $C= C^\dagger $, since the line element is real. We shall see that the eight dimensional notation (\[eq:line element hermitian metric subsection\]), through which the Hermitian metric takes its symmetric form, $ \bm C^T = \bm C $, is very handy for obtaining the equations of motion for this theory. We can define the $z^\mu$ and ${\bar{z}}^{\bar{\mu}}$ coordinates in terms of $x^\mu$ and $y^{\check{\mu}}$, as follows[^2] $$\begin{aligned}
\nonumber
z^\mu
%&=&
=
\frac{1}{\sqrt{2}}(x^\mu+ i y^{\check{\mu}})
\phantom{hallo}
%{\bar{z}}^{\bar{\mu}}
% =
%\frac{1}{\sqrt{2}}(x^\mu - i y^{\check{\mu}})
%\\\nonumber
\frac{\partial}{\partial z^\mu}
&=&
\frac{1}{\sqrt{2}}\Big(\frac{\partial}{\partial x^\mu}
- i \frac{\partial}{\partial y^{\check{\mu}}}\Big)
%\phantom{hal}
%\frac{\partial}{\partial {\bar{z}}^{\bar{\mu}}}
%=
%\frac{1}{\sqrt{2}}\Big(\frac{\partial}{\partial x^\mu}
% + i \frac{\partial}{\partial y^{\check{\mu}}}\Big)\end{aligned}$$ and their complex conjugates. This implies the following decomposition of complex vielbeins in their real, ${e_R}^a_\mu$, and imaginary, ${e_I}^a_{\check{\mu}}$, parts in the following manner $$\begin{aligned}
\nonumber
e_a^\mu
&=&
{e_R}_a^\mu + i {e_I}_a^{\check{\mu}}\,,
\phantom{hallo}
\overline{e_a^{\mu}}
=
e_a^{\bar{\mu}} = {e_R}_a^\mu - i{e_I}_a^{\check{\mu}}
\\\label{eq:viebein decompositions}
e^a_\mu
&=&
{e_R}^a_\mu - i {e_I}^a_{\check{\mu}}\,,
\phantom{hallo}
\overline{e^a_{\mu}}
=
e^a_{\bar{\mu}} = {e_R}^a_\mu + i{e_I}^a_{\check{\mu}}.\end{aligned}$$ Vielbeins are holomorphic functions, and thus transform as holomorphic vectors (we consider only the transformation of the Greek indices for this purpose), $$\begin{aligned}
\label{tetrad:coord.transformations}
&&e_\mu^a(z^\nu) \rightarrow
\tilde e_\mu^a(w^\nu) = \frac{\partial z^\alpha(w^\nu)}
{\partial w^\mu}e_\alpha^a(z^\rho)
\\\nonumber
&&e^\mu_a(z^\nu) \rightarrow
\tilde e^\mu_b(w^\nu) = \frac{\partial w^\mu(z^\nu)}
{\partial z^\alpha}e^\alpha_b(z^\rho)
\,.
%\nonumber\end{aligned}$$ The holomorphy of vielbeins, $$\begin{aligned}
\nonumber
\partial/\partial z^{\bar \mu}e_\nu^a
=
1/\sqrt{2}[\partial/\partial{x^{\mu}}+i\partial/\partial y^{\check\mu}]
[{e_R}_\nu^a + i {e_I}_\nu^a] =0
,\end{aligned}$$ then implies the Cauchy-Riemann equations, $$\frac{\partial{e_R}_\nu^a }{\partial x^{\mu}}
= \frac{\partial{e_I}_\nu^a }{\partial y^{\check\mu}}
\,,\qquad
\frac{\partial{e_I}_\nu^a }{\partial x^{\mu}}
= -\frac{\partial{e_R}_\nu^a }{\partial y^{\check\mu}}
\,.
\label{eq:CauchyRiemann}$$ The Cauchy-Riemann equations (\[eq:CauchyRiemann\]) then imply that, as a consequence of holomorphy, the tetrads are effectively functions of [*four*]{} independent coordinates, even though they are defined on an [*eight*]{} dimensional manifold. Thanks to the holomorphy symmetry, the number of physical degrees of freedom of our eight dimensional theory is reduced to that of a four dimensional theory, as required by observations. Conversely, the knowledge of a complex tetrad (both the real and imaginary parts of the tetrad must be known) projected onto the $y^{\check\mu}=0$ hypersurface $e_\mu^a(x^\nu,0)$ plus the holomorphy symmetry allows for the unique reconstruction of the full eight dimensional dynamics. This feat is achieved by the simple replacement: $$e_\mu^a(x^\nu,0)\rightarrow e_\mu^a(\sqrt{2}z^\nu,0)
\equiv e_\mu^a(z^\nu)
\,. \nonumber$$ In this sense our Hermitian gravity is a [*holographic*]{} theory. [^3] Note that this is true only when tetrads are both complex and holomorphic.
Through the inverse relations for the vielbein $e^{\mu}_a
e_{\nu}^a = \delta^{\mu}_{\nu}$ we obtain $
C^{\mu\bar{\epsilon}} C_{\bar{\epsilon}\nu}
=
\delta^{\mu}_{\nu}
$, which is in eight dimensional notation equivalent to $$\begin{aligned}
\label{eq:hermitian metric inverse relations}
\bm C^{me} \bm C_{en}
=
\bm \delta^{m}_{n}.\end{aligned}$$ We can now rotate the line element from $z^\mu,\bar{z}^{\bar{\mu}}$ space to $x^\mu,y^{\check{\mu}}$ space, obtaining $$\begin{aligned}
\nonumber%\label{eq:line element rotated hermitian metric
%in terms of B field}
ds^2
&=&
d\boldsymbol{x}^T\cdot \boldsymbol{g} \cdot d\boldsymbol{x}
=
d\boldsymbol{x}^m\boldsymbol{g}_{mn}d\boldsymbol{x}^n
\\
&=&
(d x^{\mu},
d y^{\check{\mu}})
\left(\begin{array}{cc}
g_{\mu\nu} & g_{\mu\check{\nu}}\\
g_{\check{\mu}\nu} & g_{\check{\mu}\check{\nu}}
\end{array}\right)
\left(\begin{array}{c}
d x^{\nu}\\
d y^{\check{\nu}}
\end{array}\right),
\label{eq:rotated line element:1}\end{aligned}$$ where the Latin indices take the values $0,1,...,d-1,\check{0}$,$\check{1},...,\check{d}-\check{1}$, where the Greek indices take only the values $0,1,...,d-1$ and where the number $d$ is again the complex dimension of our manifold. Note that the equation for the eight dimensional inverse (\[eq:hermitian metric inverse relations\]) holds also for the rotated metric, since its a tensorial equation. We can express the rotated Hermitian metric components, $\bm g_{mn}$, in terms of complex metric components, in terms of real and imaginary parts of the vielbein and in terms of the real symmetric metric $g_{\mu\nu}$ and real anti-symmetric torsion field $B_{\mu\nu}$, in the following manner $$\begin{aligned}
\nonumber%\label{eq:metric g hermitian in C's}
%\bm g_{mn}
%=
\frac{1}{2}
\left(\begin{array}{cc}
(C_{\bar{\mu}\nu}+C_{\mu\bar{\nu}}) & i(C_{\bar{\mu}\nu}-C_{\mu\bar{\nu}})\\[.1cm]
i(- C_{\bar{\mu}\nu}+C_{\mu\bar{\nu}}) & (C_{\bar{\mu}\nu}+C_{\mu\bar{\nu}})
\end{array}\right)
=
\left(\begin{array}{cc}
\phantom{-}g_{\mu\nu} & B_{\mu\nu}\\[.1cm]
-B_{\mu\nu} & g_{\mu\nu}
\end{array}\right)
\\\nonumber
=
\left(\begin{array}{cc}
({e_R}^a_{\mu} {e_R}^b_{\nu} + {e_I}^a_{\check{\mu}} {e_I}^b_{\check{\nu}})\eta_{ab} & (-{e_I}^a_{\check{\mu}} {e_R}^b_{\nu} + {e_R}^a_{\mu} {e_I}^b_{\check{\nu}})\eta_{ab}\\[.1cm]
({e_I}^a_{\check{\mu}} {e_R}^b_{\nu} - {e_R}^a_{\mu} {e_I}^b_{\check{\nu}})\eta_{ab} & ({e_R}^a_{\mu} {e_R}^b_{\nu} + {e_I}^a_{\check{\mu}}
{e_I}^b_{\check{\nu}})\eta_{ab}
\end{array}\right).
%\label{eq:rotated line element:2}\end{aligned}$$ Note that we have defined the eight dimensional rotated Hermitian metric to be symmetric, $\bm g = \bm g^T$. The rotated Hermitian line element is then given by $$\begin{aligned}
%\label{eq:eq:line element rotated hermitian metric
%in terms of B field}
d s^2
=
g_{\mu\nu}\left(d x^{\mu}d x^{\nu}
+
d y^{\mu}d y^{\nu}\right)
+
2B_{\mu\nu}d x^\mu d y^{\nu}.
\label{eq:rotated line element:3}\end{aligned}$$ Clearly this line element is equal to the Hermitian line element in its familiar form (\[eq:Hermitian familiar form line element\]), when using Einstein’s decomposition of the Hermitian line element, $$\begin{aligned}
\label{eq:Einstein's decomposition of the Hermitian
line element}
C_{\bar{\mu}\nu}
=
g_{\mu\nu} + i B_{\mu\nu},\end{aligned}$$ which basically rotates the Hermitian line element to $x^\mu,y^{\check{\mu}}$ space [@Einstein:1945eu]. Note that this decomposition exhibits Hermiticity explicitly, since $g$ is real and symmetric and $B$ is real and anti-symmetric. The inverse rotated metric can be expressed in terms of inverse vielbeins in the following manner $$\begin{aligned}
\label{eq:metric inverse g hermitian in vielbeins}
\boldsymbol{g}^{mn}
=
\left(\begin{array}{cc}
{e_R}^{\mu} {e_R}^{\nu} + {e_I}^{\check{\mu}} {e_I}^{\check{\nu}} & -{e_I}^{\check{\mu}} {e_R}^{\nu} + {e_R}^{\mu} {e_I}^{\check{\nu}}\\[.1cm]
{e_I}^{\check{\mu}} {e_R}^{\nu} - {e_R}^{\mu} {e_I}^{\check{\nu}} & {e_R}^{\mu} {e_R}^{\nu} + {e_I}^{\check{\mu}} {e_I}^{\check{\nu}}
\end{array}\right),\end{aligned}$$ where we have suppressed local indices, $a$ and $b$. Finally, there is a very handy way of looking at the real and imaginary parts of the vielbein, which allows us to derive the just stated objects and their relations in a very trivial manner. Consider the following holomorphic coordinates $w^\mu(z^{\gamma})
= \frac{1}{\sqrt{2}}(u^{\mu} + i v^{\check{\mu}})$ and $z^{\nu}(w^{\delta}) = \frac{1}{\sqrt{2}}(x^{\nu} + i
y^{\check{\nu}})$, living in over lapping coordinate patches on a complex manifold. We can perform a coordinate transformation $$\begin{aligned}
\nonumber%\label{eq:metric inverse g hermitian in vielbeins}
\left(\begin{array}{c}
d u^{\mu} \\
d v^{\check{\mu}}
\end{array}\right)
=
\left(\begin{array}{cc}
\frac{\partial u^{\mu}}{\partial x^{\nu} } & \frac{\partial u^{\mu}}{\partial y^{\check{\nu}}}\\[.1cm]
\frac{\partial v^{\check{\mu}}}{\partial x^{\nu} } & \frac{\partial v^{\check{\mu}}}{\partial y^{\check{\nu}}}
\end{array}\right)
\left(\begin{array}{c}
d x^{\nu} \\
d y^{\check{\nu}}
\end{array}\right),\end{aligned}$$ where there are $32$ independent components in the transformation matrix, because of the Cauchy-Riemann equations $$\begin{aligned}
\nonumber%\label{eq:metric inverse g hermitian in vielbeins}
\frac{du^{\mu}}{dx^{\nu} }
=
\frac{dv^{\check{\mu}}}{dy^{\check{\nu}}}
\phantom{halloda}
\frac{\partial u^{\mu}}{\partial y^{\check{\nu}}}
=
-\frac{\partial v^{\check{\mu}}}{\partial x^{\nu}}.\end{aligned}$$ We can identify the components of the transformation matrix with components of vielbeins, when the set $(u,v)$ forms an orthonormal basis, whenever the $(x,y)$ set forms a coordinate basis. The Cauchy-Riemann equations in terms of vielbeins then become[^4] $$\begin{aligned}
\label{eq:vielbein Cauchy Riemann relations}
{e_R}^{\mu}_a
\equiv
e^{\mu}_{\alpha}
=
e^{\check{\mu}}_{\check{\alpha}}
\phantom{hallo}
{e_I}^{\check{\mu}}_a
\equiv
e^{\check{\mu}}_{\alpha}
=
-e^{\mu}_{\check{\alpha}}.\end{aligned}$$ These definitions of the real and imaginary parts of the vielbeins are consistent with the definition of the decomposition of the complex vielbein in its real and imaginary parts (\[eq:viebein decompositions\]). With these definitions the derivation of the rotated metric tensor is immediate. Since in general relativity the metric is defined in terms of vielbeins as follows $$\begin{aligned}
\nonumber%\label{eq:metric inverse g hermitian in vielbeins}
\bm g^{mn} = \bm e^m_k \bm \eta^{kl}\bm e^n_l,\end{aligned}$$ the $\mu\check{\nu}$ component becomes $$\begin{aligned}
\nonumber%\label{eq:metric inverse g hermitian in vielbeins}
g^{\check{\mu}\nu}
=
e^{\check{\mu}}_k \bm\eta^{kl} e^{\nu}_l
=
e^{\check{\mu}}_\kappa \eta^{\kappa\lambda} e^{\nu}_\lambda
+
e^{\check{\mu}}_{\check{\kappa}} \eta^{\check{\kappa}\check{\lambda}} e^{\nu}_{\check{\lambda}},\end{aligned}$$ when defining the rotated flat metric to be $\bm\eta=$ diag$(-1,1,1,1,-1,1,1,1)$. Using the definitions of the real and imaginary parts of the vielbeins (\[eq:vielbein Cauchy Riemann relations\]), the metric component $g^{\check{\mu}\nu}$ is then given by $$\begin{aligned}
\nonumber%\label{eq:metric inverse g hermitian in vielbeins}
e^{\check{\mu}}_\kappa \eta^{\kappa\lambda} e^{\nu}_\lambda
-
e^{\mu}_{\kappa} \eta^{\kappa\lambda} e^{\check{\nu}}_{\lambda}
=
({e_I}^{\check{\mu}}_a{e_R}^{\nu}_b
-{e_R}^{\mu}_a{e_I}^{\check{\nu}}_b)\eta^{ab},\end{aligned}$$ which is consistent with the metric components (\[eq:metric inverse g hermitian in vielbeins\]) given earlier. We can now state the inverse relations for the real and imaginary components of the vielbeins. In ordinary eight dimensional relativity $$\begin{aligned}
\nonumber%\label{eq:metric inverse g hermitian in vielbeins}
\bm e^m_k\bm e^k_l = \bm \delta^m_l.\end{aligned}$$ The $\mu\lambda$ component then becomes $$\begin{aligned}
\label{eq:inverse real and I parts of vielbeins}
e^\mu_k e^k_\lambda
=
\delta^\mu_\lambda
=
e^\mu_{\kappa} e^{\kappa}_\lambda
+
e^\mu_{\check{\kappa}} e^{\check{\kappa}}_\lambda
=
{e_R}^{\mu}_a {e_R}_{\lambda}^a
+
{e_I}^{\check{\mu}}_a {e_I}_{\check{\lambda}}^a.\end{aligned}$$ We can see that $
C^{\mu\bar{\epsilon}}C_{\bar{\epsilon}\nu}
=
g^{\mu\epsilon}g_{\epsilon\nu}+B^{\mu\epsilon}B_{\epsilon\nu},
$ which is equal to the $\mu\nu$ component of $\bm g^{me}\bm
g_{en}$, is indeed equal to $\delta^\mu_\nu$, using the inverse relations of the real and imaginary parts of the vielbeins (\[eq:inverse real and I parts of vielbeins\]).
Flat Space
==========
The Hermitian line element in flat space becomes $$\begin{aligned}
ds^2
=
-(c d t)^2 + (d\vec{x})^2
-(dy^0)^2 + (d\vec{y})^2 .
\label{eq:line element:dHermitian}\end{aligned}$$ From now on we declare the $y$ coordinate to be the energy-momentum coordinate by defining[^5] $y^\mu \equiv p^\mu\frac{G_N}{c^3}$. The space-time-momentum-energy interval squared from the origin to a space-time-momentum-energy point $\bm x =
(ct,\vec{x},\frac{G_N}{c^4}E,\frac{G_N}{c^3}\vec{p})$ is given by $$\begin{aligned}
\label{eq:linelement spacetimemomentum energy flatspace her on-shell }
d^2\left(\bm 0;\bm x\right)
=
-(ct)^2 + (\vec{x})^2 +
\frac{G_N^2}{c^6}
\left[ (\vec{p})^2-\left(\frac{E}{c}\right)^2 \right],\end{aligned}$$ where ${G_N^2}/{c^6} $ suppresses the momentum-energy part by a factor on the order of $10^{-72} {s^2}/{kg^2}$, as it should do, since we do not observe any momentum-energy contributions at low energies.
![A light cone, modified by non inertial coordinate transformations, is being portrayed on a space-time-momentum-energy diagram, separating the regions that are in causal contact with each other, from the regions that are not. There is a nonlocal causally related volume element at the origin.[]{data-label="fig:LightConeher"}](LightConeher){width="\columnwidth"}
The group of transformations, which leaves the Hermitian metric (\[eq:Hermitian familiar form line element\]) invariant, is the $U(1,3)$ group; the elements $U$ of the $U(1,3)$ group satisfy by definition the relation $U^\dagger \eta U = \eta $. When considering only one space and one momentum dimension, instead of three of each, we can represent the elements of the $SU(1,1)$ group [@Low:2005tb], operating on the space-time-momentum-energy vector $(t,x,p,E)$, by $$\begin{aligned}
\nonumber
\Gamma(v,f,f_0)
=
\gamma(v,f,f_0)
\left(
\begin{array}{cccc}
1 & \frac{v}{c^2} & \frac{G_N^2 f}{c^8} & -\frac{G_N^2 f_0}{c^{9}} \\[0.1cm]
v & 1 & \frac{G_N^2 f_0}{c^7} & -\frac{G_N^2 f}{c^8} \\
f & -\frac{f_0}{c} & 1 & \frac{v}{c^2} \\
c f_0 & -f & v & 1 \\
\end{array}
\right),\end{aligned}$$ where $ \gamma(v,f,f_0) = \left(1 - \frac{v^2}{c^2} - \frac{G_N^2
f^2}{c^8} + \frac{G_N^2 f_0^2}{c^{8}}\right)^{-\frac{1}{2}}$ and where $f = \dot{p}$ and $f_0 = \frac{\dot{E}}{c}$. We have neglected two space and two momentum components for convenience, since the transformations between the spacial components are simply rotations in space. An overall phase factor can be added later if one wants to consider the $U(1,1)$ group instead of the $SU(1,1)$ group. Note that the $SU(1,3)$ group reduces to its subgroup the Lorentz group, the $SO(1,1)$ group, for $f_i$ and $f_0$ being zero. Thus, for inertial frames the theory clearly reduces to the laws of special relativity.
If we set $ds^2 = 0$ and there is no momentum-energy contribution, we know that $\dot{\vec{x}}= c$. Demanding again that $ds^2 = 0$, but now for non vanishing momentum energy contributions however, the space-time-momentum-energy interval (\[eq:linelement spacetimemomentum energy flatspace her on-shell \]) becomes $$\begin{aligned}
\label{eq:distance flatspace her on-shell}
-(ct)^2 + (\vec{x})^2 + \frac{G_N^2 p^2}{c^6} = 0,\end{aligned}$$ where $ p^2 = p^\mu p_\mu $ measures the energy-momentum in a space-time-momentum-energy hyper surface. Setting the space-time-momentum-energy interval to zero determines the causality boundary: the hypersurface specified by this condition determines the boundary of the causally related regions (for causally related events the ‘$=$’ sign in (\[eq:distance flatspace her on-shell\]) should be replaced by a ‘$\leq$’). Consider now the case when the momentum-energy contribution $p^2$ is negative. We can furthermore restrict the hypersurface by setting $t$ to zero. Solving for $\vec{x}$ we obtain $$\begin{aligned}
\nonumber
r_{\rm max}
=
\frac{G_N\sqrt{- p^2}}{c^3} ,\end{aligned}$$ where $r_{\rm max} = \sqrt{\vec x^2(t=0)}$ is the maximal radius for a spatial volume element which is simultaneously causally related in a nonlocal manner. This can be seen from figure (\[fig:LightConeher\]), which depicts this causality boundary. The existence of a momentum dependent maximal radius (often referred to as a minimal length scale) suggests that nature is inherently nonlocal for non-inertial observers. When $p^2$ is zero the standard light-cone is recovered. To get a feeling on how big this ‘violation of causality’ can be, let us consider a particle on the momentum-energy shell (corresponding to the traditional [*on-shell*]{} notion), in which case $p^2 = -m^2c^2$, such that $r_{\rm max} = G_Nm/c^2 = r_{\rm Sch}/2$, where $r_{\rm Sch}$ denotes the conventional Schwarzschild radius. Recall that our flat-space analysis is based on the geodesic equation and its integral the line element, and hence completely neglects the self-gravity of (elementary) particles. When the self-gravity effects are included however, we expect that the above-discussed violation of causality gets hidden by the Schwarzschild radius created by the particles in consideration. It would be of interest to consider in detail how the causality analysis gets modified when the self gravity of particles is taken into account.
Alternatively, we can write $r_{\rm max} = G_Nm/c^2 = (m/m_{\rm
Pl})\ell_{\rm Pl}$, where we introduced the conventional Planck mass $m_{\rm Pl} = \sqrt{c\hbar/G_N}\simeq 2.18\times 10^{-8}{\rm
kg}$ and the Planck length $\ell_{\rm Pl}=\sqrt{\hbar
G_N/c^3}\simeq 1.616\times 10^{-35}{\rm m}$, and where $\hbar$ denotes the Planck constant. Even though $r_{\rm max}$ is a purely classical quantity, when represented in terms of the Planck units, the Planck constant appears (which gets, of course, cancelled in the ratio $\ell_{\rm Pl}/m_{\rm Pl}$).
We can also specify our hypersurface differently by setting $\vec{x}$ to zero instead of $t$ (this is in the case of positive momentum-energy, $p^2$). Solving for $t$ we obtain $$\begin{aligned}
\nonumber
t_{\rm min} = \frac{G_N\sqrt{p^2}}{c^4} ,\end{aligned}$$ where $t_{\rm min}$ is the minimal time for which an event is causally related to past space-time-momentum-energy events, as can be seen from the space-time-momentum-energy diagram, figure \[fig:LightConeHer1\]. This means that $t_{\rm min}$ is the minimal time for which an event can influence future events. Equivalently, $t_{\rm min}$ is the minimal time for which an event can be influenced by past events. In the light of the above discussion of $r_{\rm max}$, we see that the minimal time $t_{\rm
min}$ shown in figure \[fig:LightConeHer1\] exists only for particles moving off-shell, and hence these particles are strictly speaking not classical.
The definitions of space-time-momentum-energy past and future are natural generalizations of space-time past and future; the regions that are in causal contact are given again by events removed from the origin by space-time-momentum-energy intervals smaller or equal then zero, $d s^2 \leq 0$. We have thus shown that in the limit of weak fields, Hermitian gravity exhibits simultaneously connected regions (maximal distance) for time like energy-momentum intervals and a breach of causality (minimal time) for space-like frame energy-momenta. For light-like frame energy-momenta, Hermitian gravity reproduces the standard light causality structure characterizing the weak field limit of general relativity.
![A light cone, modified by non inertial coordinate transformations, is being portrayed on a space-time-momentum-energy diagram separating the regions that are in causal contact each other, from the regions that are not. There is a minimal time interval for events to be in causal contact.[]{data-label="fig:LightConeHer1"}](LightConeHol){width="\columnwidth"}
We now calculate the phase velocity, using the space-time-momentum-energy line element (\[eq:distance flatspace her on-shell\]) $$\begin{aligned}
\label{eq:phase velocity hermitian case}
v_{\rm phase}
=
\frac{\|\vec r\,\|}{t}
= \sqrt{c^2 - \frac{G_N^2}{c^6}\frac{p^2}{t^2}}
\,.\end{aligned}$$ One can see that the phase velocity approaches the conventional speed of light $c$ for large $t$. For times smaller than the minimal time the phase velocity becomes imaginary and damping will occur. The group velocity for massless particles (${d s}/{dt} = 0$) becomes $$\begin{aligned}
\label{eq:groupvelocity hermtian case}
v_{g}
=
\sqrt{c^2 - \frac{G_N^2}{c^6}f^2},\end{aligned}$$ where the four force squared is given by $$\begin{aligned}
f^2
=
-\frac{1}{c^2}\left(\frac{d E}{d t}\right)^2
+
\left(\frac{d\vec{p}}{d t}\right)^2.
\label{eq:four force2}\end{aligned}$$ The maximal group velocity approaches the speed of light for small four forces squared $f^2$. The maximal group velocity must be real in order to facilitate propagation. The reality requirement implies $f^2\leq {c^4}/{G_N}$. This can be seen from the space-time-momentum-energy line element (\[eq:line element:dHermitian\]) divided by $(dt)^2$. Upon solving (\[eq:line element:dHermitian\]) for the four force squared (\[eq:four force2\]), we obtain $({G_N^2}/{c^6}) f^2
\equiv ({G_N^2}/{c^6})\left({d p}/{d t}\right)^2 =
c^2 - v_g^2 + \left({d s}/{d t}\right)^2 \leq c^2 - v_g^2\leq c^2$, where the inequalities follow from the observations that $(ds/dt)^2\leq 0$ and $v_g^2\geq 0$ (the reality condition on $v_g$ in Eq. (\[eq:groupvelocity hermtian case\])). This then implies $$\begin{aligned}
\nonumber
f^2
\leq f_{\rm max}^2
= \frac{c^8}{G_N^2}.\end{aligned}$$ There is no lower bound on $f^2$, since $\left({d s}/{d
t}\right)^2\leq 0$ can at least in principle be arbitrarily large and negative. This means that there is also no upper bound on the maximal group velocity (\[eq:groupvelocity hermtian case\]), as $f^2$ can in principle be very large and negative. While this may be true in principle, more realistically – in the flat space limit and in the absence of external forces – the flat space geodesic equation implies that $dz^\mu/d\tau = U^\mu = ({\rm
constant})^\mu$, from which we conclude that $f^2$ must be [constant]{}, such that the (maximal allowed) group velocity (\[eq:groupvelocity hermtian case\]) acquires a [*constant*]{} correction in flat spaces and in the absence of external forces. This type of corrections can play an important role in strongly curved space-times however, where strong gravitational forces exist, which are expected to induce large changes in $f^2$, and thus possibly [*superluminal*]{} propagation. The change of causality structure discussed in this section deserves a deeper analysis, since it might have important consequences for the physics of structure formation in the early universe [@Moffat:1992ud; @Barrow:1999jq].
The Equations of Motion
=======================
In order to write down the equations of motion for this Hermitian theory of gravity, one needs to know what the connection coefficients are. There are already connection coefficients which are called Hermitian connection coefficients [@Nakahara:2003nw]. We will derive different connection coefficients later, but in order to appreciate these newly obtained connection coefficients, we will consider the old ones first.
The Known Connection Coefficients
---------------------------------
One can derive the known Hermitian connection coefficients easily if one requires metric compatibility of the Hermitian metric and the fact that the holomorphic covariant derivative of an anti-holomorphic basis vector vanishes. The vanishing of he holomorphic covariant derivative of an anti-holomorphic basis vector is given by $$\begin{aligned}
\label{eq:covariant derivative known}
\nabla_\mu \frac{\partial}{\partial z^{\bar{\nu}}}
=
0
\phantom{halloda}
\nabla_{\bar{\mu}} \frac{\partial}{\partial z^{\nu}}
=
0.\end{aligned}$$ This implies that $\Gamma(\text{mixed indices}) = 0$, since the complex connection coefficients are usually defined as $$\begin{aligned}
\nonumber%\label{eq:}
\nabla_\mu \frac{\partial}{\partial z^{\bar{\nu}}}
=
\Gamma^{\epsilon}_{\mu\bar{\nu}} \frac{\partial}{\partial z^\epsilon}
=
0.\end{aligned}$$ If one then imposes metric compatibility on the Hermitian metric
\[eq:vielbein compatibilty\] $$\begin{aligned}
\label{eq:vielbein compatibilty a}
\nabla_\rho C_{\bar{\mu}\nu}
=
\partial_\rho C_{\bar{\mu}\nu}
-
C_{\bar{\mu}\lambda}\Gamma^\lambda_{\rho\nu}
=
0\end{aligned}$$ $$\begin{aligned}
\label{eq:vielbein compatibilty b}
\nabla_{\bar{\rho}} C_{\bar{\mu}\nu}
=
\partial_{\bar{\rho}} C_{\bar{\mu}\nu}
-
C_{\bar{\lambda}\nu}\Gamma^{\bar{\lambda}}_{\bar{\rho}\bar{\mu}}
=
0\end{aligned}$$
one can easily read off the Hermitian connection coefficients $$\begin{aligned}
\label{eq:connection known}
\Gamma^\lambda_{\rho\nu}
=
C^{\bar{\epsilon}\lambda}\partial_\rho C_{\nu\bar{\epsilon}}
\phantom{halloda}
\Gamma^{\bar{\lambda}}_{\bar{\rho}\bar{\nu}}
=
C^{\bar{\lambda}\epsilon}\partial_{\bar{\rho}}
C_{\epsilon\bar{\nu}}.\end{aligned}$$ If one looks carefully at the Hermitian metric compatibility equations (\[eq:vielbein compatibilty\]) one can easily see that these equations imply vielbein compatibility; one can obtain the Hermitian connection coefficients by imposing vielbein compatibility in the following manner $$\begin{aligned}
\label{eq:vielbein compat}
\nabla_\mu e_\nu
=
0
\phantom{ha}
\nabla_{\bar{\mu}} e_\nu
=
0
\phantom{ha}
\nabla_\mu e_{\bar{\nu}}
=
0
\phantom{ha}
\nabla_{\bar{\mu}} e_{\bar{\nu}}
=
0.\end{aligned}$$ Theories of Hermitian gravity, satisfying vielbein compatibility, have been proposed [@Chamseddine:2005at]. We shall see below that in such theories the geodesic equation is not obtained via an action principle. In an attempt to fix this problem we will weaken this vielbein compatibility condition and obtain different connection coefficients.
Note that the two independent components of the Riemann tensor, $$\begin{aligned}
\label{eq:Riemann known}
&&R^{\kappa}_{\lambda\bar{\mu}\nu}
=
\partial_{\bar{\mu}}\Gamma^{\kappa}_{\nu\lambda}
=
\partial_{\bar{\mu}}\left(C^{\bar{\epsilon}\kappa}\partial_{\nu}
C_{\lambda\bar{\epsilon}}\right)
\\\nonumber
&&R^{\bar{\kappa}}_{\bar{\lambda}\mu\bar{\nu}}
=
\partial_{\mu}\Gamma^{\bar{\kappa}}_{\bar{\nu}\bar{\lambda}}
=
\partial_{\mu}
\left(C^{\bar{\kappa}\epsilon}\partial_{\bar{\nu}}
C_{\epsilon\bar{\lambda}}\right)
\,,\end{aligned}$$ contain only first order derivatives, when assuming the Hermitian metric to be a product of a holomorphic and anti-holomorphic vielbein $C_{\bar{\mu}\nu} = e_{\bar{\mu}}^a (\bar{z}^\gamma)
\eta_{ab}e_{\nu}^b (z^\gamma)$. This implies that if one would attempt to write a complex equation of motion, analogues to Einstein’s equations, one would obtain a first order differential equation. This means that the space-time-momentum-energy curvature for this theory is non-dynamical.
The Hermitian Geodesic Equations
--------------------------------
The easiest way to derive the connection coefficients for Hermitian gravity is through varying the Hermitian line element. One then obtains the Hermitian geodesic equations from which one can read off the connection coefficients. Varying the Hermitian line element is a very easy exercise, when considering its eight dimensional form (\[eq:line element hermitian metric subsection\]). One then obtains the eight dimensional complex geodesic equations $$\begin{aligned}
\label{eq:geodesic equation arbitrary complex}
\boldsymbol{\ddot{z}}^r
+ \boldsymbol{\Gamma}^r_{mn} \boldsymbol{\dot{z}}^m \boldsymbol{\dot{z}}^n
= \boldsymbol{0},\end{aligned}$$ where the complex connection coefficients are given by $$\begin{aligned}
\label{eq:connection coefficients arbitrary complex}
\boldsymbol{ \Gamma}^r_{mn}
=
\frac{1}{2}\boldsymbol{C}^{re}(\bm\partial_m \boldsymbol{C}_{en}
+ \bm\partial_n \boldsymbol{C}_{me}-\bm\partial_e
\boldsymbol{C}_{mn}).\end{aligned}$$ The Hermitian metric is defined such that the $\mu\bar{\nu}$ component of ${\bm C}_{mn}$ is $C_{\mu\bar{\nu}}$, with vanishing unmixed components. The Hermitian geodesic equations are then given by $$\begin{aligned}
\nonumber
\ddot{z}^\rho
+
\Gamma^\rho_{\mu\nu}\dot{z}^{\mu}\dot{z}^{\nu}
+
\Gamma^\rho_{\bar{\mu}\nu}\dot{z}^{\bar{\mu}}\dot{z}^{\nu}
+
\Gamma^\rho_{\mu\bar{\nu}}\dot{z}^{\mu}\dot{z}^{\bar{\nu}}
+
\Gamma^\rho_{\bar{\mu}\bar{\nu}}\dot{z}^{\bar{\mu}}\dot{z}^{\bar{\nu}}
&=&
0
\\\nonumber
\ddot{z}^{\bar{\rho}}
+
\Gamma^{\bar{\rho}}_{\mu\nu}\dot{z}^{\mu}\dot{z}^{\nu}
+
\Gamma^{\bar{\rho}}_{\bar{\mu}\nu}\dot{z}^{\bar{\mu}}\dot{z}^{\nu}
+
\Gamma^{\bar{\rho}}_{\mu\bar{\nu}}\dot{z}^{\mu}\dot{z}^{\bar{\nu}}
+
\Gamma^{\bar{\rho}}_{\bar{\mu}\bar{\nu}}\dot{z}^{\bar{\mu}}\dot{z}^{\bar{\nu}}
&=&
0,
%\label{eq:geodesic eqs:unrot}\end{aligned}$$ where the connection coefficients are $$\begin{aligned}
\label{eq:Hermitian connection coefficients z basis 4 d}
\Gamma^\rho_{\mu\nu}
&=
\frac{1}{2}C^{\bar{\lambda}\rho}(\partial_{\mu} C_{\nu\bar{\lambda}} + \partial_{\nu} C_{\mu\bar{\lambda}})
\\\nonumber
\Gamma^\rho_{\bar{\mu}\nu}
&=
\frac{1}{2}C^{\bar{\lambda}\rho}(\partial_{\bar{\mu}} C_{\nu\bar{\lambda}} - \partial_{\bar{\lambda}} C_{\nu\bar{\mu}})
\\\nonumber
\Gamma^\rho_{\mu\bar{\nu}}
&=
\frac{1}{2}C^{\bar{\lambda}\rho}(\partial_{\bar{\nu}} C_{\mu\bar{\lambda}} - \partial_{\bar{\lambda}} C_{\mu\bar{\nu}})
\\\nonumber
\Gamma^{\bar{\rho}}_{\bar{\mu}\bar{\nu}}
&=
\frac{1}{2}C^{\bar{\rho}\lambda}(\partial_{\bar{\mu}} C_{\lambda\bar{\nu}} + \partial_{\bar{\nu}} C_{\lambda\bar{\mu}})
\\\nonumber
\Gamma^{\bar{\rho}}_{\mu\bar{\nu}}
&=
\frac{1}{2}C^{\bar{\rho}\lambda}(\partial_{\mu} C_{\lambda\bar{\nu}} - \partial_{\lambda} C_{\mu\bar{\nu}})
\\\nonumber
\Gamma^{\bar{\rho}}_{\bar{\mu}\nu}
&=
\frac{1}{2}C^{\bar{\rho}\lambda}(\partial_{\nu} C_{\lambda\bar{\mu}} - \partial_{\lambda} C_{\nu\bar{\mu}})
\\\nonumber
\Gamma^\rho_{\bar{\mu}\bar{\nu}}
&=
0\,,
\phantom{halloda}
\Gamma^{\bar{\rho}}_{\mu\nu}
=
0.\end{aligned}$$ These connection coefficients are Hermitian in the following sense, $\overline{\Gamma^{\rho}_{\mu\bar{\nu}}} =
{\Gamma^{\bar{\rho}}_{\bar{\mu}\nu}}^T$ and $\overline{\Gamma^{\rho}_{\mu\nu}} =
{\Gamma^{\bar{\rho}}_{\bar{\mu}\bar{\nu}}}^T$. In eight dimensional form the connection coefficients are symmetric, $
({\bm \Gamma^{r}_{mn}})= ({\bm\Gamma^{r}_{nm}})$, just as the Levi-Civita symbols in general relativity. The wisdom of the eight dimensional notation becomes apparent now; for any known equation of general relativity one can sum over the barred and unbarred indices and plug in the just derived connection coefficients. If one plugs in the connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]) into the Hermitian geodesic equations one obtains $$\begin{aligned}
\nonumber%\label{eq:}
\ddot{z}^\rho
+
\frac{1}{2}&C^{\rho\bar{\lambda}}
(\partial_{\mu} C_{\nu\bar{\lambda}} + \partial_{\nu} C_{\bar{\lambda}\mu})
\dot{z}^{\mu}\dot{z}^{\nu}
+
\\\nonumber
&C^{\rho\bar{\lambda}}
(\partial_{\bar{\nu}} C_{\mu\bar{\lambda}} - \partial_{\bar{\lambda}} C_{\bar{\nu}\mu})
\dot{z}^{\mu}\dot{z}^{\bar{\nu}}
=
0\end{aligned}$$ and its Hermitian conjugate. One obtains precisely this result, when varying the particle action (with a mass $m$), $$S=-m\int ds \nonumber$$ with respect to the real one dimensional parameter proper time $\tau$, where $ds$ represents the Hermitian line element (\[eq:Hermitian familiar form line element\]). Note that the known Hermitian connection coefficients (\[eq:connection known\]) cannot be derived from any variation principle and therefore, in that sense, cannot have any physical meaning.
From general relativity we know that the metric transforms as $$\begin{aligned}
\nonumber%\label{eq:}
\bm {C}_{mn}
\rightarrow
\bm C_{mn}
+
\bm \nabla_{m}\bm\Lambda_{n}
+
\bm \nabla_{n}\bm\Lambda_{m}.\end{aligned}$$ One can check that for example the $\mu\bar{\nu}$ component of this equation with the connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]) plugged in corresponds indeed to the transformation of the Hermitian metric $C_{\mu\bar{\nu}}$, where $z^{\mu}\rightarrow z^{\mu} +
\Lambda^{\mu}$ and $C_{\mu\bar{\nu}} \rightarrow C_{\mu\bar{\nu}}
+ \Lambda^{\epsilon}\partial_{\epsilon}C_{\mu\bar{\nu}}+
\Lambda^{\bar{\epsilon}}\partial_{\bar{\epsilon}}C_{\mu\bar{\nu}}$. This should strengthen our belief in the connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]). Note that via this procedure we can only obtain the coefficients with mixed indices. The unmixed coefficients can be obtained from the variation of the particle action with respect to the real variable proper time, because this variable breaks homomorphy in the sense that it depends on both $z^{\mu}$ and $z^{\bar{\mu}}$. Finally, since we can simply rotate any tensorial equation from $z^{\mu},z^{\bar{\mu}}$ space to $x^{\mu},y^{\check{\mu}} $ space, it is useful to have expressions for the rotated connection coefficients, such that we can just plug these rotated coefficients into the rotated equations. Although the connection transforms by definition as a connection, it clearly transforms as a $(1,2)$ tensor under these rotations. This is the case, because these rotations are just constant transformations. Hence we can just replace the complex metric, $\bm C_{mn}$, by the rotated metric, $\bm g_{mn}$, in the expression of the eight dimensional connection (\[eq:connection coefficients arbitrary complex\]) [@MantzThesis:2007]. We state two components of the rotated connection coefficients $$\begin{aligned}
\nonumber%\label{eq:}
\Gamma^{\rho}_{\mu\nu}
=
\frac{1}{2}g^{\rho\epsilon}
(\partial_{\mu} g_{\epsilon\nu} &+ \partial_{\nu} g_{\mu\epsilon} -\partial_{\epsilon} g_{\mu\nu})
\\\nonumber
+
\frac{1}{2}B^{\rho\check{\epsilon}}
(\partial_{\mu} B_{\check{\epsilon}\nu} &+ \partial_{\nu} B_{\mu\check{\epsilon}} -\partial_{\check{\epsilon}} g_{\mu\nu})
\\\nonumber
\Gamma^{\rho}_{\check{\mu}\nu}
=
\frac{1}{2}g^{\rho\epsilon}
(\partial_{\check{\mu}} g_{\epsilon\nu} &+ \partial_{\nu} B_{\check{\mu}\epsilon} -\partial_{\epsilon} B_{\check{\mu}\nu})
\\\nonumber
+
\frac{1}{2}B^{\rho\check{\epsilon}}
(\partial_{\check{\mu}} B_{\check{\epsilon}\nu} &+ \partial_{\nu} g_{\check{\mu}\check{\epsilon}} -\partial_{\check{\epsilon}}
B_{\check{\mu}\nu}).\end{aligned}$$
Torsion and Curvature
---------------------
The Hermitian torsion[^6] tensor $T$ and the Riemann tensor $R$ are defined as $$\begin{aligned}
\nonumber%\label{eq:Hermitian Einstein-Hilbert action}
T(Z,W)
=
\nabla_{Z}W - \nabla_{W}Z - [Z,W]
\\\nonumber
R(Z,W)V
=
\nabla_{Z} \nabla_{W}V - \nabla_{W} \nabla_{Z}V -
\nabla_{[Z,W]}.\end{aligned}$$ The covariant derivative acting on a basis vector yields $$\begin{aligned}
\nonumber%\label{eq:Hermitian Einstein-Hilbert action}
\bm \nabla_m \frac{\bm\partial}{\bm\partial \bm z^n}
=
\bm\Gamma^{e}_{m n}\frac{\bm\partial}{\bm\partial \bm z^e}.\end{aligned}$$ Note that the $\mu\bar{\nu}$ component of the covariant derivative acting on the basis vector is this time non-vanishing, unlike the $\mu\bar{\nu}$ component of the covariant derivative acting on the basis vector (\[eq:covariant derivative known\]), using the known coefficients (\[eq:connection known\]). We can now write the expressions for the components of the Hermitian torsion tensor $$\begin{aligned}
\nonumber%\label{eq:Hermitian Einstein-Hilbert action}
\bm T^{l}_{\phantom{l}mn}
&=&
\langle {\hat{\bm e}}^l,
{\bm T}({\hat{\bm e}}_{m},{\hat{\bm e}}_{n})\rangle
=
\langle{\hat{\bm e}}^l, {\bm\nabla_m} {\hat{\bm e}}_{n}
- {\bm\nabla_n} {\hat{\bm e}}_m
\rangle
\\\nonumber
&=&
\langle
{\hat{\bm e}}^l,
{\bm\Gamma^{b}_{mn}}{\hat{\bm e}}_b
% {\hat{\bm e}}_{b}
- {\bm\Gamma^{b}_{nm}}{\hat{\bm e}}_b
\rangle
=
{\bm \Gamma^{l}_{mn}} - {\bm\Gamma^{l}_{nm}}
=
0\end{aligned}$$ and the Hermitian Riemann tensor $$\begin{aligned}
\nonumber
{\bm R^{s}_{mln}} =
{\bm\partial_{l}}{\bm\Gamma^{s}_{nm}} -
{\bm\partial_{n}}{\bm\Gamma^{s}_{lm}}
+ {\bm\Gamma^{s}_{la}}{\bm\Gamma^{a}_{nm}} -
{\bm\Gamma^{s}_{na}}{\bm\Gamma^{a}_{lm}},\end{aligned}$$ with the connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]) plugged in. Hence the Hermitian Riemann tensor is Hermitian in the following sense $\overline{R_{\bar{\mu}\nu}} = {R_{\mu\bar{\nu}}}^T $. Therefore the Hermitian Ricci scalar is real, $ \overline{R} = R$.
Action principle for Hermitian gravity
--------------------------------------
The action for Hermitian gravity can be formulated as, $$\begin{aligned}
S[{\bm C},\psi_i] = S_{hg}[{\bm C}]+S_{c}[{\bm C}]
+ S_{M}[{\bm C},\psi_i]
\label{eq:complex action:full}\end{aligned}$$ where the pure gravity action is the following generalization of the Hilbert-Einstein action, $$\begin{aligned}
\nonumber%\label{eq:Hermitian Einstein-Hilbert action}
S_{hg}[{\bm C}]
= \frac{1}{16\pi G_N}\int d {\bm z}^8 \sqrt{\bm {\bm C}} ({\bm R}-2\Lambda)
\, ,\end{aligned}$$ where $
{\bm R}
=
{\bm C^{mn}}{\bm R_{mn}}
$ denotes the Ricci scalar, $\Lambda$ cosmological constant and ${\bm C^{mn}}$ denotes the full complex metric tensor (\[complex metric\]). For the reasons explained below, we impose the reciprocity symmetry only at the level of equations of motion (on-shell), which at the level of the action can be realized by a constraint. This of course means that physical quantities still respect the reciprocity symmetry. There is no unique way of imposing the reciprocity symmetry on the metric tensor. One reasonable choice is the following ‘particle’ action, $$\begin{aligned}
\label{eq:constraint action}
S_{c}[{\bm C}]
=
- M \int \Big[\lambda_1(C_{\mu\nu}dz^\mu dz^\nu
&+&
C_{\bar\mu\bar\nu}dz^{\bar\mu} dz^{\bar\nu})
\\\nonumber
+ \,\lambda_2(C_{\mu\bar\nu}dz^\mu dz^{\bar\nu}
&+&
C_{\bar\mu\nu}dz^{\bar\mu} dz^\nu)
\Big]^{1/2},\end{aligned}$$ where $M$ is a (mass) parameter and $\lambda_1$, $\lambda_2$ are Lagrange multipliers which break the symmetry between the holomorphic and Hermitian components of the complex metric (\[complex metric\]). Taking, for example, $\lambda_1=\lambda$ and $\lambda_2=1$, imposes $C_{\mu\nu}=0=C_{\bar\mu\bar\nu}$ and thus on-shell Hermiticity of the metric. Conversely, when $\lambda_1= 1$ and $\lambda_2=
\lambda$ imposes $C_{\mu\bar\nu}=0=C_{\bar\mu\nu}$, implying on-shell holomorphy of the metric tensor. The constraint action (\[eq:constraint action\]) does not break holomorphy of the full theory (\[eq:complex action:full\]) realized at the level of tetrads.
Just like the gravitational action, which obeys holomorphy at the level of vielbeins, we shall require that the matter action in (\[eq:complex action:full\]) consists of holomorphic matter fields. Namely, holomorphy reduces the large number of degrees of freedom of the full [*eight*]{} dimensional theory to an acceptable number of degrees of freedom of an effectively [*four*]{} dimensional theory, as observationally required. For simplicity here we consider a matter action for scalar fields, which we use extensively below when we study cosmology. We consider two holomorphic scalar fields $\phi$ and $\psi$, one with Hermitian and one with holomorphic kinetic term, with the action: $$\begin{aligned}
\label{eq:scalar action}
S_M[\phi,\psi] = \int d^8{\bm z}\sqrt{{\bm C}}\, {\cal L},\end{aligned}$$ where the lagrangian density is given by $$\begin{aligned}
\nonumber
{\cal L} = - \frac{\alpha}{2}{\bm C}^{mn}(\partial_m\Phi)^\dagger
\cdot\partial_n\Phi
- \frac{\beta}{2}{\bm C}^{mn}(\partial_m\Psi)^T\cdot\partial_n\Psi
- V\end{aligned}$$ where $$\Phi = \left({\phi\atop{\bar\phi}}\right)
\,,\qquad
\Psi = \left({\psi\atop{\bar\psi}}\right)
\,,
\label{eq:scalar action:fields}$$ where $\alpha$ and $\beta$ are constants and where $V=V(\Phi,\Psi)$ is a potential. Note that $\Phi^\dagger\cdot\Phi
= 2\phi\bar\phi$ and $\Psi^T\cdot\Psi = \psi^2+\bar\psi^2$. The constants $\alpha$ and $\beta$ can be absorbed in the fields $\phi$ and $\psi$ by the appropriate field redefinitions, except for the sign of $\alpha$, which is an invariant and thus can have physical relevance. For simplicity, we have assumed in Eq. (\[eq:scalar action\]) that the scalar fields do not couple to the Ricci scalar.
Varying the action (\[eq:complex action:full\]) results in the Hermitian Einstein-Hilbert equations of motion $$\begin{aligned}
{\bm G}_{mn} + \Lambda {\bm C}_{mn} &=& 8\pi G_N {\bm T}_{mn}
\nonumber\\
C_{\mu\nu} &=& 0 = C_{\bar\mu\bar\nu}
\,,
\label{eq:hermitian eom}\end{aligned}$$ where the second line equation is obtained by choosing $\lambda_1=\lambda$, $\lambda_2=1$ and varying the action (\[eq:constraint action\]) with respect to $\lambda$. As usual the following definitions hold for the Einstein tensor ${\bm G}_{mn}$ and the stress energy tensor ${\bm T}_{mn}$: $$\begin{aligned}
{\bm G}_{mn} &=& {\bm R}_{mn} - \frac 12 {\bm C}_{mn}{\bm R}
\,,\qquad {\bm R} = {\bm C}^{mn}{\bm R}_{mn}
\nonumber\\
{\bm T}_{mn}
&=& -\frac{2}{\sqrt{\bm C}}\frac{\delta S_M}{\delta {\bm C}^{mn}}
\,.
\label{Gmn:Tmn}\end{aligned}$$ This formulation of the theory guarantees the (contracted) Bianchi identity, which in the eight dimensional form reads, $$\begin{aligned}
\nonumber%\label{}
\bm\nabla^{m} \bm G_{mn}
=
0.
\label{Bianchi:8d}\end{aligned}$$ The proof is analogous to that in general relativity. As a consequence, the stress energy must be covariantly conserved, ${\bm \nabla}^m{\bm T}_{mn}=0$, just as desired. Note that imposing the reciprocity symmetry on the action (\[eq:complex action:full\]) (off-shell) would result in an over-constrained on-shell dynamics which would fail to satisfy the Bianchi identity (\[Bianchi:8d\]). We consider that as unacceptable, since that would imply nonconservation of the stress energy tensor, implying that energy would leak from our four dimensional space-time hypersurface into the energy-momentum directions.
The stress energy tensor corresponding to the scalar field action (\[eq:scalar action:fields\]) is just, $${\bm T}_{mn} = \alpha (\partial_m\Phi)^\dagger
\cdot\partial_n\Phi
+ \beta(\partial_m\Psi)^T\cdot\partial_n\Psi
+{\bm C}_{mn} {\cal L}
\,,
\label{Tmn:scalars}$$ where we used $\delta \sqrt{\bm C} = -\frac12 \sqrt{\bm C}{\bm C}_{mn}\delta {\bm C}^{mn}$.
When written in the four dimensional notation, Eqs. (\[eq:hermitian eom\]) reduce to,
$$\begin{aligned}
G_{\mu\nu} = R_{\mu\nu} &=& 8\pi G_N T_{\mu\nu}
\label{eq:Hermitian Einstein's equations:1}
\\
G_{\mu\bar\nu} + C_{\mu\bar\nu}\Lambda &=& 8\pi G_N T_{\mu\bar\nu}
\label{eq:Hermitian Einstein's equations:2}\end{aligned}$$
plus the corresponding Hermitian conjugate equations, where $C_{\mu\nu} = 0$ and $G_{\mu\bar\nu} = R_{\mu\bar\nu}-
\frac12C_{\mu\bar\nu}R$. We also have, $R_{\mu\nu}=R^\alpha_{\;\mu\alpha\nu}+R^{\bar\alpha}_{\;\mu\bar\alpha\nu}$ and $R_{\mu\bar\nu} = R^\alpha_{\;\mu\alpha\bar\nu}
+ R^{\bar\alpha}_{\;\mu\bar\alpha\bar\nu}$. Note that the holomorphic equation (\[eq:Hermitian Einstein’s equations:1\]) does not admit a cosmological term $\Lambda$. Indeed, $\Lambda$ is removed from (\[eq:Hermitian Einstein’s equations:1\]) by the on-shell reciprocity symmetry.
Metric Compatibility
--------------------
When working in the first order formalism [@MantzThesis:2007], in addition to Eqs. (\[eq:hermitian eom\]) one also obtains the metric compatibility equations, $$\begin{aligned}
\label{eq:metric compatibility equations}
{\bm \nabla_m} {\bm C_{nr}} = {\bm 0}.\end{aligned}$$ We list two components in four dimensional notation $$\begin{aligned}
\nabla_\rho C_{\mu\bar{\nu}} = 0
\,,\qquad
\nabla_\rho C_{\mu\nu} = 0.
\label{eq:Hermitian metric compatibility equations}\end{aligned}$$ One can check that the connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]) are consistent with the metric compatibility condition (\[eq:Hermitian metric compatibility equations\]): since $C_{\mu\nu}= 0$ and $\Gamma^{\bar{\rho}}_{\mu\nu} = 0$, we obtain that $\nabla_\rho
C_{\mu\nu} = 0.$ The mixed components[^7] of the metric compatibility condition $$\begin{aligned}
\nonumber%\label{eq:Hermitian metric compatibility equations}
\nabla_\rho C_{\mu\bar{\nu}}
=
\partial_\rho C_{\mu\bar{\nu}}
-
\Gamma^{\epsilon}_{\rho\mu}C_{\epsilon\bar{\nu}}
-
\Gamma^{\bar{\epsilon}}_{\rho\bar{\nu}}C_{\bar{\epsilon}\mu}\end{aligned}$$ vanish as well $$\begin{aligned}
\nonumber%\label{eq:Hermitian metric compatibility equations}
\partial_\rho C_{\mu\bar{\nu}}
-
\frac{1}{2}(\partial_{\rho} C_{\mu\bar{\nu}}
+ \partial_{\mu} C_{\rho\bar{\nu}})
-
\frac{1}{2}(\partial_{\rho} C_{\mu\bar{\nu}}
- \partial_{\mu} C_{\rho\bar{\nu}})
=
0.\end{aligned}$$ On the other hand, writing the Hermitian metric component in terms of vielbeins and using the Leibnitz rule for the covariant derivative, one can show that the connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]) do not imply the vielbein compatibility (\[eq:vielbein compat\]) discussed above. Indeed, writing Eq. (\[eq:Hermitian metric compatibility equations\]) as $$\begin{aligned}
\nonumber%\label{eq:}
\nabla_\rho (e_{\mu}) e_{\bar{\nu}}
+
e_{\mu}\nabla_\rho e_{\bar{\nu}}
=
0,\end{aligned}$$ implies that $
\nabla_\rho e_{\bar{\nu}}
=
-e_{\bar{\epsilon}}\Gamma^{\bar{\epsilon}}_{\rho\bar{\nu}}
- e_{\epsilon}\Gamma^{\epsilon}_{\rho\bar{\nu}}
\neq 0
$ is not zero for instance.
Counting Degrees of freedom
===========================
In order to get a glimpse of the general structure of Hermitian gravity, it might be educational to determine some properties of relevant objects, which are part of the theory. If we would like to know for instance, if we can always go to a freely falling frame, we can begin with counting the degrees of freedom of an arbitrary coordinate transformation of the metric tensor, in order to see if there are enough coordinate degrees of freedom in order to do so. If we then Taylor expand both sides of the coordinate transformation of the metric tensor $$\begin{aligned}
\label{eq:coordinate trans count}
{\tilde{C}}_{\bar{\mu}\nu}
=
\frac{\partial z^{\bar{\alpha}}}{\partial {\tilde{z}}^{\bar{\mu}}}
\frac{\partial z^{\beta}}{\partial {\tilde{z}}^{\nu}}
C_{\bar{\alpha}\beta},\end{aligned}$$ we can collect terms of a specific order of the expansion of both sides of the equation and equate these terms. We are Taylor expanding both sides of the coordinate transformation of the Hermitian metric (\[eq:coordinate trans count\]) around a point $p$ on the (smooth) manifold.
When considering the zeroth order terms of the expansion, $$\begin{aligned}
\label{eq:coordinate trans count}
{\tilde{C}}_{\bar{\mu}\nu} |_p
=
\frac{\partial z^{\bar{\alpha}}}{\partial {\tilde{z}}^{\bar{\mu}}}
\frac{\partial z^{\beta}}{\partial {\tilde{z}}^{\nu}}
C_{\bar{\alpha}\beta}\Big |_p,\end{aligned}$$ we have $16$ degrees of freedom at the left hand side of the equation, since the metric is Hermitian. The formula for the real degrees of freedom of a Hermitian matrix is $d^2$, where $d$ is the complex dimension of the manifold, which is 4 in this case. On the right hand side of the equation, there are 32 real degrees of freedom to transform to the flat space metric (there are $32$ degrees of freedom instead of 64, because the coordinate transformations are holomorphic, implying they satisfy the Cauchy-Riemann equations). Subtracting the two, we obtain $32-16 =
16$ degrees of freedom, which leave the flat space metric invariant. These 16 degrees of freedom are precisely the 16 degrees of freedom of the $U(1,3)$ group, which by definition leave the Hermitian flat space metric invariant.
Considering the following terms at first order of the expansion of the coordinate transformation (\[eq:coordinate trans count\]) $$\begin{aligned}
\nonumber%\label{eq:}
\bm {\tilde{\partial}}_e \bm {\tilde{C}}_{m n}|_p
+ \cdots
=
\frac{\bm \partial^2 \bm z^{a}}{\bm \partial \bm {\tilde{z}}^{e} \bm \partial \bm {\tilde{z}}^{m}}
\frac{\bm \partial \bm z^{b}}{\bm \partial \bm {\tilde{z}}^{n}}
\bm C_{a b}\Big |_p
+\cdots,\end{aligned}$$ we count $64$ real degrees of freedom on the left hand side and $80$ on the right. We obtain the $64$ real dimension in the following manner. The complex number of degrees of freedom for a Hermitian matrix is $\frac{1}{2}d^2$, where $d$ is again the complex dimension of the manifold, which is, as said before, $4$ in this case. The complex dimension of the partial derivative is $d$. When multiplying these numbers we obtain $32$ complex degrees of freedom, which is equivalent to $64$ real degrees of freedom. The $80$ degrees of freedom from the first factor of the term on the right hand side are obtained as follows. The numerator has $d$ complex degrees of freedom. The denominator has $\frac{1}{2}d(d+1)$ complex degrees of freedom, which is just the formula of a symmetric matrix, since partial derivatives commute. By multiplying these numbers together we obtain $4\cdot 10 = 40$, complex degrees of freedom, which is equivalent to $80$ real degrees of freedom. When subtracting these numbers we obtain $80 -
64 = 16$ real degrees of freedom. This means that we have $16$ degrees of freedom too many in order to transform to the free falling frame.
Finally, considering the following two terms at second order of the expansion of the coordinate transformation of the Hermitian metric (\[eq:coordinate trans count\]) $$\begin{aligned}
\nonumber%\label{eq:}
\bm {\tilde{\partial}}_e \bm {\tilde{\partial}}_f\bm {\tilde{C}}_{m n} |_p
+ \cdots
=
\frac{\bm \partial^3 \bm z^{a}}{\bm \partial \bm {\tilde{z}}^{e}
\bm \partial \bm {\tilde{z}}^{f} \bm \partial \bm {\tilde{z}}^{m}}
\frac{\bm \partial \bm z^{b}}{\bm \partial \bm {\tilde{z}}^{n}}
\bm C_{a b}\Big|_p
+\cdots,\end{aligned}$$ we have $160$ real degrees of freedom at the left hand side and $160$ on the right. We obtain the $160$ real degrees of freedom on the left hand side as follows. The complex degrees of freedom of the partial derivatives is $ \frac{1}{2}d(d+1) $ and the complex degrees of freedom of the metric is again $\frac{1}{2}d^2$. Multiplying these numbers together we obtain $10 * 8 = 80$ complex degrees of freedom, which is equivalent to $160$ real degrees of freedom. The $160$ real degrees of freedom on the right hand side are obtained as follows. The numerator has again dimension $d$. The denominator has dimension $\frac{1}{3!}d(d+1)(d+2)$. Multiplying these numbers together we obtain $4*20 = 80$ complex degrees of freedom, which is again equivalent to $160$ real degrees of freedom. When subtracting these two numbers we obtain $160 - 160 = 0$ degrees of freedom. This means that we have precisely enough degrees of freedom in order to obtain flat space at second order of the expansion. Hence there is no space-time-momentum-energy curvature in the theory of Hermitian gravity. This might appear as problematic since general relativity does contain space-time curvature, which we cannot get rid off by coordinate transformations. We shall see below that in the limit of projecting space-time-momentum-energy onto space-time we will obtain space-time curvature as an artifact of the limiting procedure.
Hermitian gravity is dynamical in the sense that there are second order derivatives, acting upon the dynamical variable, the vielbein or the metric. Consider the following independent components of the Hermitian Riemann tensor, $$\begin{aligned}
\label{eq:Riemann un mixed indices}
R^{\kappa}_{\lambda\mu\nu}
=
%\frac{1}{2}
C^{\bar{\epsilon}\kappa}\partial_{\lambda }\partial_{[\mu}C_{\nu]\bar{\epsilon}}
+
\text{first order derivatives},\end{aligned}$$ and its Hermitian conjugate. Unlike the components of the Riemann tensor (\[eq:Riemann known\]), constructed from the known complex connection coefficients (\[eq:connection known\]), we do have non-vanishing components of the Riemann tensor, which do contain second order derivatives. These components *do* enter the Hermitian Einstein equations (\[eq:Hermitian Einstein’s equations:2\]), although $C_{\mu\nu} = 0 $. These components of the Hermitian Einstein tensor act as constraints, such that we remain on the hypersurface, which specified by the reciprocity transformation.
The Limit to General Relativity
===============================
The limit of Hermitian gravity to the theory of general relativity is based on the assumption that the $y$ coordinate and its corresponding vielbein are small. When expanding these theories in powers of $y$ and its corresponding vielbein, we would hope to obtain the theory of general relativity at zeroth order of the expansion and meaningful corrections to the theory at linear order. We will see that this is not the case, since we will obtain corrections to general relativity at zeroth order. The easiest way to obtain the limit to general relativity is to expand the real and the imaginary parts of the vielbein in terms of the $y$ coordinate in order to collect the terms in orders of the $y$ coordinate and its corresponding vielbein, ${e_I}_{\check{\mu}}$, yielding
\[eq:vielbein expansion\] $$\begin{aligned}
\label{eq:expansion x vielbein}
{e_{R}}_\mu (x,y) &=& {e_{R}}_\mu (x) - y^\lambda
\partial_\lambda {e_{I}}_{\check{\mu}} + O(y^2)
\\\nonumber
&=& {e_{R}}_\mu (x) + O'(y^2)\end{aligned}$$ and $$\begin{aligned}
\label{eq:expansion y vielbein}
{e_{I}}_{\check{\mu}} (x,y) = {e_{I}}_{\check{\mu}} (x) + y^\lambda
\partial_\lambda {e_{R}}_\mu (x)
+ O(y^2).\end{aligned}$$
These expansions contain sufficient information in order to obtain the limit to general relativity. We will, however, also expand the rotated metric components and a component of the rotated connection coefficients. Using the expansions of the real and imaginary parts of the vielbeins (\[eq:vielbein expansion\]), the rotated metric components up to second order of the $y$ coordinate and its corresponding vielbein are $$\begin{aligned}
\nonumber%\label{}
g_{\mu\nu}(x,y)
=
g_{\mu\nu}(x) + O(y^2),\end{aligned}$$ $$\begin{aligned}
\nonumber%\label{}
g_{\mu\check{\nu}}(x,y)
=
g_{\mu\check{\nu}}(x) + y^{\lambda}\left(\partial_{\lambda}(e_{\mu})e_{\nu} - e_{\mu}\partial_{\lambda}e_{\nu} \right) + O(y^2),\end{aligned}$$ $$\begin{aligned}
\nonumber%\label{}
g_{\check{\mu}\nu}(x,y)
=
g_{\check{\mu}\nu}(x) + y^{\lambda}\left(e_{\mu}\partial_{\lambda}e_{\nu}- \partial_{\lambda}(e_{\mu})e_{\nu}\right) + O(y^2),\end{aligned}$$ and $$\begin{aligned}
\nonumber%\label{}
g_{\check{\mu}\check{\nu}}(x,y)
=
g_{\check{\mu}\check{\nu}}(x) + O(y^2),\end{aligned}$$ where $
g_{\mu\nu}(x)
=
g_{\check{\mu}\check{\nu}}(x)
=
e_{\mu}e_{\nu}(x)
$ and where $
g_{\check{\mu}\nu}(x)
$ and $
g_{\mu\check{\nu}}(x)
$ can be just read off the expression of the rotated Hermitian metric in terms of vielbeins. Using again the expansions of the real and imaginary parts of the vielbeins (\[eq:vielbein expansion\]), the $\Gamma ^\rho_{\mu\nu} (x,y)$ component of the connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]) up to second order of the $y$ coordinate and its corresponding vielbein is $$\begin{aligned}
\nonumber%\label{}
\Gamma ^\rho_{\mu\nu} (x,y) = \Gamma ^\rho_{\mu\nu} (x) +
O(y^2),\end{aligned}$$ where $\Gamma ^\rho_{\mu\nu} (x)$ is just the ordinary Levi-Cività connection. With the expansions of the connection coefficients, we can now check if the theory of Hermitian gravity reduces to the theory of general relativity by plugging them into the rotated Hermitian geodesic equation, keeping only terms of linear order in the $y$ coordinate and its corresponding vielbein, yielding the ordinary geodesic equation $$\begin{aligned}
\nonumber%\label{}
\ddot{x}^\rho + \Gamma_{\mu\nu}^\rho(x){\dot{x}}^\mu {\dot{x}}^\nu
+ O(y^2) = 0\end{aligned}$$ without any first order corrections present. Though the result is not spectacular at first sight, it should be pleasing that the theory of Hermitian gravity reduces to the well tested theory of general relativity, for the Hermitian geodesic equation. In order to see if the theory predicts any interesting new physics we have to collect terms up to second order, yielding $$\begin{aligned}
\nonumber%\label{}
\ddot{x}^\rho &+& \Gamma_{\mu\nu}^\rho(x,y){\dot{x}}^\mu
{\dot{x}}^\nu + [\Gamma^\rho_{\check{\mu}\nu} (x,y)+
\Gamma^\rho_{\nu\check{\mu}}(x,y)]\dot{y}^{\check{\mu}}\dot{x}^\nu
\\\nonumber
&+&
\Gamma^\rho_{\check{\mu}\check{\nu}}(x,y) \dot{y}^{\check{\mu}}\dot{y}^{\check{\nu}}
+
O(y^3)
=
0,\end{aligned}$$ where the connection coefficients $\Gamma^\rho_{\check{\mu}\nu}
(x,y),$ $ \Gamma^\rho_{\nu\check{\mu}}(x,y)$ and $
\Gamma^\rho_{\check{\mu}\check{\nu}}$ are just the connection coefficients expanded up to linear order [@MantzThesis:2007], but where the connection coefficient $\Gamma_{\mu\nu}^\rho(x,y)$ has to be expanded up to second order since the term ${\dot{x}}^\mu {\dot{x}}^\nu$, multiplying $\Gamma_{\mu\nu}^\rho(x,y)$, is of zeroth order in the $y$ coordinate and its corresponding vielbein.
The Hermitian Einstein’s equations get corrections to the Einstein’s equations of general relativity at zeroth order. This can be seen when considering the rotated Hermitian Ricci tensor $$\begin{aligned}
\nonumber%\label{}
R_{\mu\nu}
=
R^{\lambda}_{\phantom{{\lambda}}\mu\lambda\nu} + R^{\check{\lambda}}_{\phantom{{\lambda}}\mu\check{\lambda}\nu}
=
R^{GR}_{\mu\nu} +
{R_{\text{cor}}}^{\check{\lambda}}_{\phantom{{\lambda}}\mu\check{\lambda}\nu} +
O(y^2),\end{aligned}$$ where $R^{GR}$ is the Ricci tensor according to general relativity and where $R_{\text{cor}}$ are the terms of $R^{\bar{\lambda}}_{\phantom{{\lambda}}\mu\bar{\lambda}\nu}$ of zeroth order in the expansion in the $y^{\check{\mu}}$ coordinate. Similarly the rotated Hermitian Einstein tensor $$\begin{aligned}
\nonumber%\label{}
G_{\mu\nu}
=
G^{GR}_{\mu\nu} +
{R_{\text{cor}}}^{\check{\lambda}}_{\phantom{{\lambda}}\mu\check{\lambda}\nu}
-\frac{1}{2}g_{\mu\nu}g^{\alpha\beta}{R_{\text{cor}}}^{\check{\lambda}}_{\phantom{{\lambda}}\alpha\check{\lambda}\beta}
\\\nonumber
-
\frac{1}{2}g_{\mu\nu}g^{\check{\alpha}\check{\beta}}
\big({R_{\text{cor}}}^{\lambda}_{\phantom{{\lambda}}\check{\alpha}\lambda\check{\beta}}
+
{R_{\text{cor}}}^{\check{\lambda}}_{\phantom{{\lambda}}\check{\alpha}\check{\lambda}\check{\beta}}\big)]
+
O(y^2),\end{aligned}$$ gets corrections of zeroth order in the expansion in the $y^{\check{\mu}}$ coordinate. Hence, at this point one needs to look at the solutions of Hermitian gravity in order to see if the theory contradicts experiment or not.
Hermitian Cosmology {#Hermitian Cosmology}
===================
In order to describe our Universe correctly, which is isotropic and homogeneous on large scales, our complex theory should permit solutions that possess the symmetries of isotropy and homogeneity and furthermore these solutions should correctly reduce to the Friedmann-Lema[î]{}tre-Robertson-Walker (FLRW) cosmology at low energies. Since the vielbeins of our theory are holomorphic functions, we can demand spatial isotropy and homogeneity by implementing a scale factor that is a holomorphic function of the time-like coordinate only, $$\begin{aligned}
e^{a}_{\mu}
=
a(z^0)\delta_{\mu}^{a}
\phantom{halloda}
e^{a}_{\bar{\mu}}
=
a(z^{\bar{0}})\delta_{\bar{\mu}}^{a}
\,,
\label{eq:vielbein:cosmology}\end{aligned}$$ where the complex scale factor can be specified in terms of its real and imaginary parts $$\begin{aligned}
\nonumber%\label{}
a(z^0)
=
a_{R}(z^0)
+
i a_{I}(z^0)
\phantom{hallo}
\bar{a}(z^{\bar{0}})
=
a_{R}(z^{\bar{0}})
-
i a_{I}(z^{\bar{0}}).\end{aligned}$$ Note that $z^0 = t + i\frac{G_N}{c^4} E $. The [*Ansatz*]{} (\[eq:vielbein:cosmology\]) yields a cosmology with flat spatial sections, which suffices for our purpose.[^8] With this [*Ansatz*]{} for the vielbein, the connection coefficients become holomorphic (or anti-holomorphic) functions. Consider for example the following two connection coefficients $$\begin{aligned}
\label{eq:connection H Cosmo unrotated}
\Gamma^{\rho}_{\mu\nu}
&=&
\frac{1}{2}\frac{a'}{a}(\delta^{\rho}_{\nu}\delta^{0}_{\mu}
+
\delta^{0}_{\nu}\delta^{\rho}_\mu)
\\\nonumber
\Gamma^{\rho}_{\bar{\mu}\nu}
&=&
\frac{1}{2}\frac{\bar{a}'}{\bar{a}}(\delta^{\bar{0}}_{\bar{\mu}}\delta^{\rho}_{\nu}
-
\eta^{\bar{0}\rho}\eta_{\bar{\mu}\nu}).\end{aligned}$$ These expressions for the connection coefficients can then be used to obtain the components of the Hermitian Ricci tensor. The expression for the mixed components of the Ricci tensor is then $$\begin{aligned}
\nonumber%\label{}
R_{\bar{\mu}\nu}
=
\frac{\bar{a}'a'}{\bar{a}a}
\left[\left(\frac{ d - 1}{2}\right)\delta^{\bar{0}}_{\bar{\mu}}\delta^{0}_{\nu} + (d-1)\eta_{\bar{\mu}\nu}\right],\end{aligned}$$ where $d$ is again the complex dimension of the manifold. Taking $d$ to be four, we obtain the following expressions for the independent mixed components of the Hermitian Ricci tensor[^9] $$\begin{aligned}
\nonumber%\label{}
R_{\bar{0}0}
=
-\frac{3}{2}\frac{\bar{a}'a'}{\bar{a}a}
\phantom{halloda}
R_{\bar{i}j}
=
3\frac{\bar{a}'a'}{\bar{a}a} \eta_{\bar{i}j}\end{aligned}$$ and the following expression for the Hermitian Ricci scalar $$\begin{aligned}
\nonumber%\label{}
R
=
C^{\bar{\mu}\nu}R_{\bar{\mu}\nu}
+
C^{\mu\bar{\nu}}R_{\mu\bar{\nu}}
=
21 \frac{\bar{a}'a'}{(\bar{a} a)^2}.\end{aligned}$$ The nonzero unmixed components are then $$\begin{aligned}
\nonumber%\label{}
R_{00}(z^0)
=
\frac{9}{2}\left(\frac{a'}{a}\right)^2 - 3\left(\frac{a''}{a}\right)\end{aligned}$$ and its complex conjugate. The independent components of the Hermitian Einstein tensor then become $$\begin{aligned}
\nonumber
&&G_{\bar{0}0}
=
9\frac{\bar{a}'a'}{\bar{a}a}
\,,
\phantom{halloda}
G_{\bar{i}j}
=
-\frac{15}{2}\frac{\bar{a}'a'}{\bar{a}a}\eta_{\bar{i}j}
\\\label{eq:einstein t H cosmo unrotated}
&&G_{00}
=
\frac{9}{2}\left(\frac{a'}{a}\right)^2 - 3\left(\frac{a''}{a}\right)
\,\end{aligned}$$ and their complex conjugates. If we now make use of the connection coefficients (\[eq:connection H Cosmo unrotated\]) we can easily check that the Einstein tensor of Hermitian gravity (\[eq:einstein t H cosmo unrotated\]) obeys the Bianchi identity (\[Bianchi:8d\]), implying that the Einstein tensor is divergenceless, as it should be. This represents a nontrivial check of the accuracy of our calculation.
Consider now the matter action. For definiteness (and simplicity) we shall consider the two scalar field action (\[eq:scalar action\]). Since in standard general relativity one can obtain any desired expansion $a=a(t)$ by appropriately choosing the scalar field potential, we expect that the action (\[eq:scalar action\]) does not pose any important restrictions to the Hermitian cosmology.
The underlying symmetries of a (flat) FLRW cosmology together with holomorphy then imply that the scalar fields are of the form, $\phi=\phi(z^0)$ and $\psi=\psi(z^0)$. With this observation we get that the nonvanishing components of the stress energy tensor (\[Tmn:scalars\]) are, $$\begin{aligned}
T_{\mu\bar\nu} &=& \alpha \delta_\mu^0\delta_{\bar\mu}^{\,\bar 0}
\phi' \bar\phi'
+ \eta_{\mu\bar\nu}\big(\alpha\phi' \bar\phi'- a\bar a V\big)
\nonumber
\\
T_{\mu\nu} &=& \beta \delta_\mu^0\delta_{\mu}^{\,0}
{\psi'}^2
\,,
\label{Tmn:scalars:cosmology:2}\end{aligned}$$ plus the Hermitian conjugates. Here we used $\phi'=(\partial/\partial z^0)\phi$ and $\bar\phi'=(\partial/\partial z^{\bar 0})\bar\phi$. The nonvanishing components in (\[Tmn:scalars:cosmology:2\]) are, $$\begin{aligned}
\label{Tmn:scalars:cosmology:3}
T_{0\bar 0} = a\bar a V
\phantom{ha}
T_{i\bar j} = \delta_{i\bar j}\big(\alpha\phi' \bar\phi'- a\bar a V\big)
\phantom{ha}
%\nonumber
%\\
T_{00} = \beta{\psi'}^2
.\end{aligned}$$ When combined with Eqs. (\[eq:einstein t H cosmo unrotated\]) these yield the following equations for Hermitian cosmology (\[eq:Hermitian Einstein’s equations:1\]–\[eq:Hermitian Einstein’s equations:2\]),
$$\begin{aligned}
&G_{00} \equiv
\frac{9}{2}\left(\frac{a'}{a}\right)^2 - 3\left(\frac{a''}{a}\right)
= 8\pi G_N \beta {\psi'}^2
\label{eq:Hermitian Einstein's equations:3}
\\
&G_{0\bar0}
+
C_{0\bar0}\Lambda
\equiv
9\frac{\bar{a}'a'}{\bar{a}a} - a\bar a \Lambda
=
8\pi G_N a\bar a V
\label{eq:Hermitian Einstein's equations:4}
\\\nonumber
&\frac13\delta^{i\bar j}\big(G_{i\bar j} + C_{i\bar
j}\Lambda\big)
\equiv
-\frac{15}{2}\frac{\bar{a}'a'}{\bar{a}a}
+
a\bar a\Lambda
\\
&=
8\pi G_N \big(\alpha\phi' \bar\phi'- a\bar a V\big) \,,
\label{eq:Hermitian Einstein's equations:5}\end{aligned}$$
which together with the scalar field equations of motion, $$\begin{aligned}
-3\alpha\frac{\bar a'}{a\bar a^2}\phi' - \partial_{\bar\phi}V &=& 0
\nonumber\\
-3\beta\frac{\bar a'}{a\bar a^2}\psi' - \partial_{\psi}V &=& 0
\label{eq:scalar eom:cosmology}\end{aligned}$$ represent the closed system of equations of Hermitian cosmology with scalar fields. These equations are obtained by varying the matter action (\[eq:scalar action\]) with respect to $\bar\phi$ and $\psi$, respectively. The scalar equations of motion (\[eq:scalar eom:cosmology\]) can be can be also obtained from the covariant stress-energy conservation. Inspired by the form of stress-energy in FLRW spaces, $T_{\mu\nu}=a^2\delta_\mu^{\;0}\delta_\nu^{\;0}(\rho+p)+p
g_{\mu\nu}$, the appropriate Hermitian gravity generalization is of the form, $$\begin{aligned}
\nonumber
T_{\mu\nu}
&=&
(\rho_h+p_h)a^2\delta_\mu^{\;0}\delta_\nu^{\;0}
\\
T_{\mu\bar\nu}
&=&
a^2\delta_\mu^{\;0}\delta_{\bar\nu}^{\;\bar 0}(\rho+p)
+p C_{\mu\bar\nu}
\,.
\label{stress energy:Hermitian}\end{aligned}$$ Comparing this with Eqs. (\[Tmn:scalars:cosmology:3\]) then implies $$\begin{aligned}
\label{Tmn:scalars:cosmology:p+rho}
\rho = V
\phantom{hall}
p = \alpha\dot\phi\dot{\bar\phi}- V
\phantom{hall}
\rho_h+p_h = \beta{\dot\psi}^2\end{aligned}$$ where $(1/a)\phi' = \dot \phi$ and $(1/a)\psi' = \dot \psi$. While the pressure has a standard form, note that the kinetic term does not contribute to the energy density of Hermitian gravity. [^10] The stress energy conservation $\nabla^m T_{mn}=0$ then implies, $$\dot\rho + 3 \bar H (\rho_h+p_h) + 3 H(\rho+p) = 0$$ and the complex conjugate equation, where $\dot{\rho}=(1/a)\partial_{0}\rho$ and where $H=a'/a^2 = \dot a/a$ and $\bar H = \bar a'/{\bar a}^2 = \dot {\bar a}/\bar a$. This together with Eqs. (\[Tmn:scalars:cosmology:p+rho\]) implies the scalar field equations (\[eq:scalar eom:cosmology\]), which checks the consistency of our formulation of Hermitian gravity with scalar fields.
Next, it is convenient to divide equation (\[eq:Hermitian Einstein’s equations:3\]) by $a^2$ and Eqs. (\[eq:Hermitian Einstein’s equations:4\]–\[eq:Hermitian Einstein’s equations:5\]) by $a\bar a$, respectively. Combining Eqs. (\[eq:Hermitian Einstein’s equations:4\]) and (\[eq:Hermitian Einstein’s equations:5\]) results in the constraint equations,
$$\begin{aligned}
H\bar H &=& \frac{8\pi G_N}{9}\big(V+\lambda\big)
\,,\qquad \lambda = \frac{\Lambda}{8\pi G_N}
\label{eq:herm:cosmology:1}
\\
\dot\phi\dot{\bar\phi} &=& \frac{1}{6\alpha}\big(V+\lambda\big)
\,. \label{eq:herm:cosmology:2}\end{aligned}$$
Equation (\[eq:Hermitian Einstein’s equations:3\]) can be recast as, $$\begin{aligned}
-\dot H -\frac12 H^2 &=& \frac{8\pi G_N}{3} \beta {\dot \psi}^2
\,.
\label{eq:herm:cosmology:3}\end{aligned}$$ With a help of Eq. (\[eq:herm:cosmology:1\]), the scalar equations (\[eq:scalar eom:cosmology\]) become, $$\begin{aligned}
\dot\phi &=& - \frac{3H}{8\pi G_N\alpha}
\frac{\partial_{\bar\phi}V}{V+\lambda}
\label{eq:herm:cosmology:4}
\\
\dot\psi &=& - \frac{3H}{8\pi G_N\beta}\frac{\partial_{\psi}V}{V+\lambda}
\,.
\label{eq:herm:cosmology:5}\end{aligned}$$
Equations (\[eq:herm:cosmology:1\]–\[eq:herm:cosmology:5\]) (and their hermitean conjugates) are the fundamental equations of hermitean cosmology. Note that there are 5 (2 real and 3 complex) equations for 3 complex quantities $H$, $\phi$ and $\psi$, so the system is overdetermined, and there is no guarantee that a solution exists.
We shall now show that a solution exists, and moreover we shall explicitly construct a class of solutions that gives rise to a power law expansion of the scale factor $a=a(z^0)$.
Firstly, Eq. (\[eq:herm:cosmology:1\]) implies that $\bar V = V$ and $\bar\lambda=\lambda$ are real. Secondly, Eq. (\[eq:herm:cosmology:4\]) and its complex conjugate imply that $\partial_{\bar \phi} \ln\big(V+\lambda\big)$ is a holomorphic functions of $\phi$ and that $\partial_{\phi} \ln\big(V+\lambda\big)$ is an antiholomorphic function of $\bar\phi$. Consequently the potential is determined to be of the form, $$\ln(V+\lambda) = A_1 \phi + \bar A_2 \bar \phi + A_3
\,,$$ where $A_1$, $A_2$ and $A_3$ are complex constants independent on $\phi$ (still possibly dependent on $\psi$ and $\bar\psi$). The reality of $V$ and $\lambda$ then implies that $A_1=A_2\equiv
\Omega$. Writing $A_3$ as $A_3=\ln(W)$ we have, $$V = -\lambda + W\exp\big[\Omega\big(\phi+\bar\phi\big)\big]
\,.
\label{eq:V:1}$$ Finally, since $V$ is real, $W$ and $\Omega$ must be real functions of $\psi$ and $\bar\psi$. Holomorphy is such a powerful symmetry that - even though Eq. (\[eq:V:1\]) represents the most general solution to Eqs. (\[eq:herm:cosmology:4\]) and its complex conjugate – the potentials $V$ for $\phi$ is, up to two ‘constants’ $W$ and $\Omega$, completely fixed.
Now multiplying Eq. (\[eq:herm:cosmology:4\]) with its complex conjugate, making use of Eq. (\[eq:herm:cosmology:1\]) and inserting the resulting equation into (\[eq:herm:cosmology:2\]), we obtain, $$\frac{4\pi G_N\alpha}{3} = \partial_{\bar \phi}\ln\big(V+\lambda)
\partial_{\phi}\ln\big(V+\lambda)
\equiv \Omega^2
\,,
\label{eq:omega}$$ implying that there are two allowed values for $\Omega$, $$\Omega_\pm \equiv \pm \omega = \pm \sqrt{\frac{4\pi G_N\alpha}{3}}
\,,
\label{eq:omega2}$$ fixing thus $\Omega$ completely (up to a sign). (This sign ambiguity reflects the symmetry of the theory (\[eq:scalar action\]) under the transformation, $\phi\rightarrow -\phi$, $\bar \phi \rightarrow -\bar\phi$.) With this Eq. (\[eq:herm:cosmology:4\]) becomes $$\dot \phi = \mp\frac{1}{2\omega}\frac{d}{dt}\ln(a)
\,,
\nonumber$$ which is solved by $$a=a_0 {\rm e}^{\mp 2\omega (\phi-\phi_0)}
\,.
\label{avsphi}$$ ($\phi_0$ is unphysical as it can be absorbed in the definition of $a_0$.) This means that $\phi$ is not an independent field, but a constrained field which is just a reparametrization of the scale factor $a$. This is not surprising, given the fact that $\phi$ solves the constraint equations of Hermitian gravity.
The remaining equations to be solved are (\[eq:herm:cosmology:3\]) and (\[eq:herm:cosmology:5\]), and the remaining freedom in the potential is in $W=W(\psi,\bar\psi)$. Similarly as above, we can see from Eq. (\[eq:herm:cosmology:5\]) that $\partial_\psi \ln(V+\lambda)
=
\partial_\psi\ln(W) $ must be a holomorphic function of $\psi$. We conclude that $W$ must be a product of a holomorphic function of $\psi$ and an antiholomorphic function of $\bar\psi$ (and they must be mutually equal), $$W = w(\psi)\bar w(\bar\psi)
\,.
\label{Eq:W}$$ With this observation, making use of (\[eq:herm:cosmology:5\]), Eq. (\[eq:herm:cosmology:3\]) can be recast as, $$\epsilon \equiv -\frac{\dot H}{H^2} = \frac12
+ \frac{3}{8\pi G_N\beta}\Big(\partial_\psi\ln(w(\psi))\Big)^2
\,.
\label{eq:herm:cosmology:3b}$$ Let us now consider the case when $\epsilon = {\rm const.}$ In this case Eq. (\[eq:herm:cosmology:3b\]) can be easily solved for $w(\psi)$ in terms of $\epsilon$, $$w
= w_0\exp\bigg[\sqrt{\frac{8\pi G_N\beta}{3}\Big(\epsilon-\frac12\Big)}\,\psi
\bigg]
\,,
\label{eq:w:power-law}$$ where $w_0$ is a field independent constant. The potential $V$ that yields a power law expansion is therefore given by $$\begin{aligned}
\label{eq:V:power-law}
V
&=&
- \frac{\Lambda}{8\pi G_N}
+
V_0 \exp\big[\pm\omega(\phi+\bar\phi)\big]\times
\\\nonumber
&\times&
\exp\left[\sqrt{\frac{8\pi G_N\beta}{3}}
\bigg(\Big(\epsilon-\frac12\Big)^{1/2}\,\psi
+ \Big(\bar\epsilon-\frac12\Big)^{1/2}\,\bar\psi
\bigg)\right]
\,,\end{aligned}$$ where $V_0=w_0\bar w_0$ is a (real) constant and $\omega =
\sqrt{4\pi G_N \alpha/3}$. Note that nonvanishing $\text{Im}[\epsilon]$ breaks the charge-parity (CP) symmetry of the $\psi$ field, as can be seen from the potential (\[eq:V:power-law\]). However, CPT is conserved. We will discuss below how $\text{Im}[\epsilon]$ breaks time reversal symmetry, T. Recall that a power law expansion means $$\epsilon \equiv \frac{3}{2}(1+w_f) = {\rm const.}
\,,
\label{eq:epsilon=const}$$ where $w_f=p/\rho$ is the (complex) equation of state parameter of a ‘cosmological fluid’ with a ‘pressure’ $p$ and an ‘energy density’ $\rho$. Since $\epsilon = (d/dt)(1/H)$, a constant epsilon implies a power law expansion with, $$H = \frac{1}{\epsilon z^0}
\,,\qquad
a(z^0) = a_0\, \Big(\frac{z^0}{\zeta_0}\Big)^{1/\epsilon}
\,,
\label{eq:power law expansion}$$ where $a_0$ and $\zeta_0$ are (complex) constants. In standard cosmology, $\epsilon=3/2$ ($\epsilon=2$) correspond to matter (radiation) era, while $0<\epsilon\ll 1$ corresponds to a slow roll inflation.
From Eqs. (\[eq:herm:cosmology:4\]–\[eq:herm:cosmology:5\]) and (\[eq:V:power-law\]) we find, $$\dot\phi = \mp \frac{H}{2\omega}
\,,\qquad
\dot\psi = - \frac{H}{2\omega}\sqrt{\frac{\alpha}{\beta}(2\epsilon-1)}
\,,
\nonumber$$ such that in a Universe expanding as a power law the two fields are not independent, $$\psi-\psi_0 = \pm \sqrt{\frac{\alpha}{\beta}(2\epsilon-1)}(\phi-\phi_0)
\,.
\label{eq:fields relation}$$ When this is inserted into Eq. (\[avsphi\]) we immediately get $$a=a_0\exp\bigg[-\sqrt{\frac{16\pi G_N\beta}{3(2\epsilon-1)}}(\psi-\psi_0)\bigg]
\,.
\label{eq:avspsi}$$ This means that – just like $\phi$ – in a power law expansion $\psi$ corresponds to an $\epsilon$-dependent reparametrization of the scale factor of the Universe. Note that $\epsilon = 1/2$ is a singular point of the relation (\[eq:avspsi\]). This is not unexpected, since from Eq. (\[eq:herm:cosmology:3\]) we know that $\epsilon = 1/2$ corresponds to the case when $\dot\psi=0$ and hence also $W=0$, such that $a$ is given by (\[avsphi\]) and does not dependent on $\psi$ (in fact in this case $\psi$ does not even exist).
More general cosmologies, with $\epsilon$ in Eq. (\[eq:herm:cosmology:3b\]) being a function of $z^0$ are possible, provided one chooses $w(\psi)$ in Eq. (\[Eq:W\]) of a more general (non-exponential) form. In these more general cosmologies no simple relation between $\psi$ and $a$ exists such that Eq. (\[eq:avspsi\]) must be suitably generalized.
Let us have a more careful look at the power law solution (\[eq:power law expansion\]). Recall that in the physical space an observer sees the expansion rate ${\cal H}$ that can be obtained from the rotated Hermitian Einstein tensor $G_{i\bar j}$ (\[eq:einstein t H cosmo unrotated\]) as follows, $${\cal H}^2 = -\frac{2}{45}\Re\bigg(\frac{G_{i\bar j}}{a\bar a}\bigg)
= H\bar H
\,,
\nonumber$$ from which we conclude (see Eq. (\[eq:power law expansion\])), $${\cal H} = \frac{1}{|\epsilon|\sqrt{t^2+(G_NE/c^4)^2}}
\,.
\label{eq:H:physical}$$ Note that at late times $t^2\gg G_N|E|/c^4$ the expansion rate approaches that of general relativity, $${\cal H} \rightarrow \frac{1}{|\epsilon|t}
\,,\qquad (t\rightarrow \infty)
\,,
\label{eq:H:physical:late}$$ with $\epsilon_{GR}$ given by $|\epsilon|$ of Hermitian gravity. In contrast to general relativity at early times $|t|\leq G_N
|E|/c^4$ the expansion rate does not diverge. Instead, it reaches a maximal value at $t=0$ given by $${\cal H} \rightarrow {\cal H}_{\rm max} =\frac{c^4}{|\epsilon|G_N|E|}
\,\qquad (t\rightarrow 0)
\,,
\label{eq:H:physical:early}$$ which is nonsingular as long as $E\neq 0$ (below we discuss the physical relevance of the singularity at $t=0=E$). This behavior of ${\cal H}$ corresponds to a bouncing cosmology. Indeed, since the expansion rate is symmetric under time reversal, $t\rightarrow
-t$, for $t<0$ the Universe passes through a contracting phase, followed by a mirror symmetric expanding phase for $t>0$. The time dependence of ${\cal H}$ on time $t$ (on an $E={\rm const.}$ hypersurface) is shown in figure (\[fig:HubbleMax\]).
![The observed expansion rate as a function of time. When moving backwards in time, the expansion rate of the Hermitian Hubble parameter ${\text H}_{\text{Her}}$ reaches a maximal value at $t=0$, whereas the Hubble parameter of general relativity ${\text H}_{\text{GR}}$ becomes infinite in finite time.[]{data-label="fig:HubbleMax"}](HubbleMax){width="\columnwidth"}
For completeness we now consider the scale factor of Hermitian cosmology. The observed scale factor ${\cal A}$ corresponds to the rotated metric tensor (\[eq:rotated line element:1\]–\[eq:rotated line element:3\]), $g_{\mu\nu}\equiv
{\cal A}^2(x,y)\eta_{\mu\nu} = \text{Re}[C_{\mu\bar\nu}]$. This then implies, $$\begin{aligned}
\label{eq:A}
{\cal A}
=
\sqrt{a\bar a}
=
{\cal A}_0 \left(\frac{\sqrt{t^2+(G_N E/c^4)^2}}{|\zeta_0|}
\right)^{(1/\epsilon)_R}
\times
\\\nonumber
\times \exp\bigg[\!-\!\Big(\frac{1}{\epsilon}\Big)_I
\bigg({\rm Arctan}\Big(\frac{G_N E}{c^4t}\Big)
\!-\! \frac{\pi}{2}{\rm sign}(t)\bigg)
\Bigg]
\,,\end{aligned}$$ where we chose $\zeta_0 = |\zeta_0|$ (the phase ${\rm
arg}(\zeta_0)$ can be absorbed in ${\cal A}_0$), $(1/\epsilon)_R =
\text{Re}[\epsilon]/|\epsilon|^2$, $(1/\epsilon)_I =
-\text{Im}[\epsilon]/|\epsilon|^2$, and we chose the Riemann sheet of $a=a(z^0)$ such that ${\cal A}$ is continuous at $t=0$, as required by the equations of motion for $a$. The observed scale factor reaches a minimum at $t= 0$ and expands symmetrically under time reversal, for $\text{Im}[\epsilon] = 0 $. Nonvanishing $\text{Im}[\epsilon] $, however, violates charge-parity (CP) symmetry of the $\psi$ field, as can be seen from the potential (\[eq:V:power-law\]). Since CPT is conserved, T must also be violated. A manifestation of this T violation can be seen in figure \[fig:scalefactor\], in which we can see that the contracting phase is not a mirror image of the expanding phase.
![The observed scale factor as a function of time. The expansion rate reaches a minimal value at $t=0$. Nonzero $\text{Im}[\epsilon] $ breaks time reversal symmetry, $T$.[]{data-label="fig:scalefactor"}](scalefactor){width="\columnwidth"}
The sign of $\text{Im}[\epsilon]$ determines the direction of the tilt in the scale factor function. Note that ${\cal H}$ [*cannot*]{} be obtained from ${\cal A}$ as $(\partial_t {\cal
A})/{\cal A}$. [^11] This should not surprise us, given the fact that in Hermitian gravity a derivative of a projected quantity onto a space-time hypersurface is not in general equal to the projected derivative of the same quantity. Mathematically, the difference arises because the projection procedure must be made consistent with the Cauchy-Riemann equations.
At late times $t\gg |G_N E/c^4|$ the solution (\[eq:A\]) approaches a power law expansion of general relativity, $$\begin{aligned}
\label{eq:A:2}
{\cal A}
\stackrel{t\rightarrow \infty}{\longrightarrow}
\hat {\cal A}_0
\left(\frac{t}{|\zeta_0|}
\right)^{(1/\epsilon)_R}
\phantom{h}
\hat{\cal A}_0
=
{\cal A}_0\exp\bigg[\frac{\pi}{2}
\Big(\frac{1}{\epsilon}\Big)_I \bigg].\end{aligned}$$ From this we see that at late times it is difficult to distinguish the standard FLRW cosmology and Hermitian cosmology. Indeed, the only difference is in the size of the Universe: when $(1/\epsilon)_I = -\text{Im}[\epsilon]/|\epsilon|^2>0$ ($(1/\epsilon)_I < 0$) the Universe of Hermitian cosmology appears greater (smaller) than the FLRW Universe with $\epsilon \leftrightarrow [(1/\epsilon)_R]^{-1}$. Since the absolute value of the scale factor cannot be observed (only ratios are observable), this difference cannot be used to distinguish between the standard and Hermitian cosmologies. If we had any information about the size of the early Universe, we could make the desired distinction.
One the other hand, at early times (when $t$ and $G_N E/c^4$ are comparable), the two cosmologies differ quite dramatically. Consider first the Universe which expands such that $E={\rm
const}$. In this case Eq. (\[eq:A\]) represents a bouncing universe with a minimal size given by, $${\cal A}_{\rm min} = {\cal A}(t=0) = {\cal A}_0 \left(\frac{G_N |E|}{c^4|\zeta_0|}
\right)^{(1/\epsilon)_R}
\,.
\label{eq:Amin}$$ The Universe behaves regularly ‘everywhere’ provided the cosmological singularity at $ E\rightarrow 0, t\rightarrow 0$ is never reached.
Cosmological singularity {#Cosmological singularity}
------------------------
In order to find out how accessible the singular point of Eqs. (\[eq:H:physical\]) and (\[eq:A\]) actually is, we consider a freely falling observer, which falls ‘backwards in time’ towards the singularity. In order to study how velocities and energy change in an expanding universe, we need to solve the corresponding geodesic equations.
Let us begin with general relativity. We are working in a spatially flat FLRW space. In conformal coordinates the Levi-Cività connection is of the form, $$\begin{aligned}
\nonumber
\Gamma^\mu_{\alpha\beta}
=
\frac{a'}{a}\Big(\delta^\mu_{\;\alpha}\delta_\beta^{\;0}
+ \delta^\mu_{\;\alpha}\delta_\beta^{\;0}
+ \delta^0_{\;\mu}\eta_{\alpha\beta}\Big).\end{aligned}$$ This implies the following geodesic equation and line element (for a massive observer), $$\begin{aligned}
\label{FLRW:geodesic eq}
\frac{du_c^\mu}{d\tau}
+ \frac{a'}{a}\bigg(2u_c^0u_c^\mu
- \frac{\delta^\mu_{\,0}}{a^2}\bigg)
=
0,
\phantom{ha}
\eta_{\alpha\beta}u_c^\alpha u_c^\beta
=
-\frac{1}{a^2},\end{aligned}$$ where $\tau$ is the proper time observed by a freely falling observer (in the frame in which all 3-velocities vanish): $(ds)^2=-(d\tau)^2$, and $u^\mu_c =dx_c^\mu/d\tau$ is the 4-velocity in conformal coordinates $x_c^\mu = (\eta,x^i_c)$ (here we take $c=1$). The spatial equation (\[FLRW:geodesic eq\]) is easily solved, $$\frac{d (a^2 u_c^i)}{d\eta} = 0
\,,
\label{FLRW:geodesic eq:i}$$ where we made use of the definition of conformal time, $u_c^0 d\tau = d\eta$. This means that in an expanding universe, $u_c^i\propto 1/a^2$, such that the physical momentum, $p^i_p = m a u_c^i$, scales as $p^i_p\propto 1/a$, where $m$ is observer’s mass. The time component of Eq. (\[FLRW:geodesic eq\]) implies, $$\frac{d}{d\eta}\Big[a^2\Big(a^2(u_c^0)^2-1\Big)\Big] = 0
\,,
\label{FLRW:geodesic eq:0}$$ which is consistent with the line element in Eq. (\[FLRW:geodesic eq\]) and with Eq. (\[FLRW:geodesic eq:i\]). Eqs. (\[FLRW:geodesic eq\]) and (\[FLRW:geodesic eq:i\]) can be also used to determine the scaling of the physical energy, $E_p \equiv p^0_p = m a u_c^0$: $$E_p^2 - \sum_i({p_p^i})^2 = m^2
\,,$$ from where it follows, $E_p^2-m^2 \propto 1/a^2$ (this can be also concluded from Eq. (\[FLRW:geodesic eq:0\])).
Let us now consider Hermitian gravity. The relevant connection coefficients are given in Eqs. (\[eq:connection H Cosmo unrotated\]), such that the corresponding geodesic equation (\[eq:geodesic equation arbitrary complex\]) and the line element (\[eq:Hermitian familiar form line element\]) are then, $$\begin{aligned}
\nonumber
\frac{du_c^\mu}{d\tau}
+ \frac{a'}{a}u_c^0u_c^\mu
+ \frac{\bar{a}'}{\bar{a}}\Big(u_c^{\bar 0}u_c^{\mu}
&+&
\eta^{\mu\bar 0}\frac{1}{a\bar a}\,\Big)
= 0
\\\label{FLRW:geodesic eq:her}
\eta_{\alpha\bar\beta}u_c^\alpha u_c^{\bar \beta}
&=&
-\frac{1}{a\bar a}\end{aligned}$$ where again $\tau$ is a real affine parameter defined as the proper ‘time’ of a freely falling observer (in the frame in which all 3-velocities vanish and $E=0$): $(ds)^2=-2(d\tau)^2$, and $u^\mu_c =dz_c^\mu/d\tau$ is the complex proper 4-velocity in conformal coordinates $z_c^\mu = (z^0_c,z^i_c)$. Notice that from the definition $dz_c^0/d\tau = u_c^0$, it follows that the second term in Eq. (\[FLRW:geodesic eq:her\]) can be absorbed by a simple rescaling of $u_c^\mu$, such that it simplifies to $$\begin{aligned}
\nonumber
\frac{d(au_c^\mu)}{d\tau}
+ \bar H \Big((\bar au_c^{\bar 0})(au_c^{\mu})
+
\eta^{\mu\bar 0}\,\Big)
&=& 0
\\\label{FLRW:geodesic eq:her:2}
\eta_{\alpha\bar\beta}(au_c^\alpha)(\bar a u_c^{\bar \beta})
&=& -1 ,\end{aligned}$$ where $\bar H = \bar a'/{\bar a}^2 \equiv \dot {\bar a}/\bar a$, $dz^0=adz_c^0$ and $\dot {\bar a} = (1/a)da/dz_c^0$. When split into components Eq. (\[FLRW:geodesic eq:her:2\]) yields,
$$\begin{aligned}
\frac{du^0}{d\tau}
&+& \bar H
\big(u^{\bar 0}u^{0} - 1\big)
= 0
\label{FLRW:geodesic eq:her:3a}
\\
\frac{du^i}{d\tau}
&+& \bar H u^{\bar 0}u^{i}
= 0
\,,\qquad u^0u^{\bar 0}-u^iu^{\bar i}
= 1
\,,
\label{FLRW:geodesic eq:her:3b}\end{aligned}$$
where we defined the complex ‘physical’ 4-velocities $u^\mu=au_c^\mu$ and $u^{\bar \mu}=\bar au_c^{\bar \mu}$. The corresponding complex conjugate equations must also hold. The temporal equation (\[FLRW:geodesic eq:her:3a\]) and its complex conjugate can be combined to give, $$\frac{d}{d\tau}\ln\big(u^0u^{\bar 0}-1\big) = - \big(H u^0+\bar Hu^{\bar 0}\big)
= - \frac{d\ln(a\bar a)}{d\tau}
\,.
\label{FLRW:geodesic eq:her:4}$$ The last equality follows from $H u^0 = d\ln(a)/d\tau$ and $\bar H u^{\bar 0} = d\ln(\bar a)/d\tau$. This can be straightforwardly integrated from $\tau_0$ to $\tau$ resulting in the scaling, $$\frac{u^0u^{\bar 0}-1}{(u^0u^{\bar 0})_0-1} =
\frac{(a\bar a)_0}{a\bar a}
= \frac{u^iu^{\bar i}}{(u^iu^{\bar i})_0}
\,,
\label{FLRW:geodesic eq:her:sol:1}$$ where the last equality follows from the constraint in Eq. (\[FLRW:geodesic eq:her:3a\]), or equivalently from Eq. (\[FLRW:geodesic eq:her:3b\]). In Eq. (\[FLRW:geodesic eq:her:sol:1\]) $u^\mu=u^\mu(\tau)$, $a=a(\tau)$ and we have defined $(u^0u^{\bar 0})_0 = u^0(\tau_0)u^{\bar 0}(\tau_0)$, $(u^iu^{\bar i})_0=u^i(\tau_0)u^{\bar i}(\tau_0)$ and $(a\bar a)_0=a(\tau_0)\bar a(\tau_0)$. Analogously to general relativity the spatial components of particles’ physical (complex) velocities scale as, $$u^iu^{\bar i} = \Big(\frac{dx^i}{d\tau}\Big)^2
+ G_N^2 \Big(\frac{dp^{\check i}}{d\tau}\Big)^2
\propto \frac{1}{a\bar a} = \frac{1}{{\cal A}^2}
\,.$$
Next we recall that, $$u^0 = \frac{dz^0}{d\tau}
\,,\qquad
u^{\bar 0} = \frac{dz^{\bar 0}}{d\tau}
\nonumber$$ and we define a radial and angular (time-like) coordinates, $$z^0 \equiv r{\rm e}^{i\theta} = \frac{x^0+iy^0}{\sqrt{2}}
\,.
\nonumber$$ Now, by making use of the definition $u^0=(d/d\tau)(r{\rm e}^{i\theta})$ one immediately arrives at the identity, $${\cal E}\equiv \frac12{\dot r}^2 + V(r,\theta) = 0
\,,\qquad
V = \frac12\frac{L^2}{r^2} - \frac12u^{0}u^{\bar 0}
\,,
\label{FLRW:geodesic eq:her:6}$$ where we defined an ‘angular momentum’ $$L=r^2\dot\theta
\,.
\label{eq:L:def}$$ This angular momentum (or more precisely the angular velocity $\omega_\theta=\dot\theta$) characterizes the rate of mixing between the time and energy coordinates in Hermitian gravity. Equation (\[FLRW:geodesic eq:her:6\]) represents the conserved ‘energy density’ of Hermitian cosmology. Indeed, since ${\cal
E}=0$, the energy density (\[FLRW:geodesic eq:her:6\]) is trivially conserved, $\dot {\cal E}=0$. The angular momentum (\[eq:L:def\]) is, however, not generally conserved, implying that the time and energy generally mix. This can be seen from the imaginary part of Eq. (\[FLRW:geodesic eq:her:3a\]) which – when divided by $\bar\epsilon \bar H$ and using the Hubble parameter (\[eq:power law expansion\]) of power law expansion – yields, $$\dot L \equiv \frac{d}{d\tau}(r^2\dot \theta)
= - \frac{\epsilon_I}{|\epsilon|^2}(u^0u^{\bar 0}-1)
\,.
\label{FLRW:geodesic eq:her:5}$$ Requiring that the derivative of energy integral (\[FLRW:geodesic eq:her:6\]) vanishes, one obtains the equation of motion for $r$, which corresponds to the real part of Eq. (\[FLRW:geodesic eq:her:3a\]) (divided again by $\bar\epsilon \bar H$). That means that Eq. (\[FLRW:geodesic eq:her:6\]) is an integral of motion and remarkably the dynamics of particles in Hermitian cosmology reduces to a study of motion in a (simple) potential given in Eq. (\[FLRW:geodesic eq:her:6\]).
In order to illustrate how to completely solve the geodesic equations of Hermitian cosmology, we now restrict ourselves to the simple case when $\epsilon_I=0$ (recall that in standard FLRW cosmology $\epsilon$ is by definition a [*real*]{} parameter). In this case Eq. (\[FLRW:geodesic eq:her:5\]) implies that the angular momentum $L=L_0$ is conserved, $\dot L_0=0$ and the potential (\[FLRW:geodesic eq:her:6\]) acquires the simple form, $$V = \frac12 \frac{L_0^2}{r^2} - \frac{U_0}{2r^{{2}/{\epsilon}}} - \frac12
\,,\qquad U_0 = [(u^{0}u^{\bar 0})_0-1]
(a\bar a)_0 |\zeta_0|^{{2}/{\epsilon}}
\,,
\label{eq:potential:epsReal}$$ where $U_0\geq 0$ parameterizes the time-like velocity at a time $\tau_0$. Provided $\epsilon\neq 1$ this potential has an extremum $V_e$ at the radius $r_e$ given by, $$\begin{aligned}
\nonumber
V_e &=& - \frac{U_0}{2}\frac{\epsilon-1}{\epsilon}
\left(\frac{U_0}{\epsilon L_0^2}\right)^\frac{1}{\epsilon-1}
-\frac12
\\\label{re:Ve}
r_e &=& \left(\frac{\epsilon
L_0^2}{U_0}\right)^\frac{\epsilon}{2(\epsilon-1)}
\phantom{ha}(L_0\neq 0, U_0\neq 0).\end{aligned}$$ When $\epsilon>1$ (decelerated expansion) the extremum is a minimum, as can be seen from figure \[fig:HermitianCosmology1\]. Whenever $0<\epsilon<1$ (accelerated expansion) the extremum is a maximum, as is depicted in figure \[fig:HermitianCosmology2\]. Choosing $\epsilon = 2$ corresponds to the radiation era and the value $\epsilon = 1/2$ is close to the $\epsilon$ parameter of today’s Universe. The critical value of $\epsilon= 1$ yields a curvature dominated universe, as is shown in figure \[fig:HermitianCosmology3\]. In this case V does not have an extremum (formally, an extremum $V\rightarrow -1/2$ is reached for $r\rightarrow \infty$). As $r\rightarrow 0$ the potential approaches $+\infty$ ($-\infty$) when $L_0^2>U_0$ ($L_0^2<U_0$).
\[fig:HermitianCosmology\]
Since the energy integral (\[FLRW:geodesic eq:her:6\]) is conserved, particle dynamics in the potential (\[eq:potential:epsReal\]) with $\epsilon>1$ (or when $\epsilon=1$ and $L_0^2>U_0$) is such that they reach the minimal distance $r_c$ (a turning point) given by the point where $\dot r(r_c)=0$. In other words, the Universe of Hermitian gravity generically exhibits a bounce whenever $\epsilon>1$ and $L_0\neq 0$. The critical (minimal) radius is given by $V(r_c)=0$, as can be seen from (\[FLRW:geodesic eq:her:6\]). In the radiation era ($\epsilon=2$), we can solve analytically for the critical radius $$\begin{aligned}
\nonumber
r_c
\equiv
\sqrt{t_c^2 + (G_NE)^2} = \Delta -\frac12 U_0 \,,
\\\label{eq:rc}
\Delta
=
\sqrt{(U_0/2)^2+L_0^2} \qquad ({\rm radiation}\;\;{\rm
era}) \,.\end{aligned}$$ The integral from of Eq. (\[FLRW:geodesic eq:her:6\]) is, $$\int \frac{dr}{\sqrt{-2V}}=\pm \tau
\,,
\label{eq:general integral rvstau}$$ which cannot be performed analytically for a general $\epsilon$. In radiation era ($\epsilon = 2$) integrating (\[eq:general integral rvstau\]) gives, $$\begin{aligned}
\label{rvstau}
&\pm& \tau
=
\sqrt{r^2 + U_0r-L_0^2}
\\\nonumber
&-&\frac{U_0}{2}\ln \Bigg(\frac{r+({U_0}/{2})+\sqrt{r^2 + U_0r-L_0^2}}
{\Delta}\Bigg)
\,,\end{aligned}$$ where we chose the proper time $\tau$ such that $r(\tau=0)=r_c$. This can be inverted close to the bounce, $$r \simeq r_c + \frac{\Delta}{2}\Big(\frac{\tau}{r_c}\Big)^2
\,,\qquad (r-r_c\ll \Delta,\tau\gg r_c)
\nonumber$$ At late times one gets the expected linear behavior plus a logarithmic correction which characterizes Hermitian gravity, $$r \simeq \tau + \frac{U_0}{2}\bigg[\ln\Big(\frac{\tau}{\Delta}\Big)-1\bigg]
\,,\qquad (r-r_c\gg \Delta,\tau\gg r_c)
\,.
\nonumber$$ In the limiting case when $\epsilon = 1$ (curvature domination) Eqs. (\[FLRW:geodesic eq:her:6\]) and (\[eq:potential:epsReal\]) can be integrated to give, $$\sqrt{r^2 + U_0 -L_0^2} = \pm \tau
\label{eq:eps=1}$$ such that when $U_0<L_0^2$ there is a bounce with the minimal Hubble length given by, $$r_c = \sqrt{L_0^2 -U_0} \,,\qquad (\epsilon=1,\,L_0^2>U_0)
\,.
\label{eq:rc:eps=1}$$
The existence of a minimal Hubble length $r_c$ as given by Eqs. (\[eq:rc\]) and (\[eq:rc:eps=1\]) means that even when time $t_c$ is set to zero (‘Big Bang’), the Universe reaches its maximal – but finite – expansion rate (\[eq:H:physical\]) $${\cal H}_{\rm max} = \frac{1}{\epsilon r_c}
= \frac{1}{\epsilon\sqrt{t_c^2+(G_NE_c/c^4)^2}}
\,,
\label{eq:H:physical:max}$$ with $r_c$ given in Eq. (\[eq:rc\]). (Even if $L_0$ were set to zero initially, a small nonvanishing $\epsilon_I$ would violate angular momentum conservation (\[FLRW:geodesic eq:her:5\]), such that we expect that $L\neq 0$ generically close to the bounce. Moreover, the choice $L_0=0$ represents a set of measure zero in the space of all initial conditions $\{u^\mu(\tau_0)|
\eta_{\mu\bar\nu}u^\mu(\tau_0)u^{\bar\nu}(\tau_0)=-1\}$, and in this sense the condition $L_0=0$ is ‘almost never’ realized.) Equation (\[eq:H:physical:max\]) constitutes the main result of our analysis of Hermitian cosmology, according to which Hermitian cosmology is nonsingular at the classical level.
Note that Hermitian cosmology predicts $r_c$, but at what time, $t_c$, and energy, $E_c$, $r_c$ is reached depends on the initial conditions embodied by $U_0$ and $L_0$. In other words: $\theta_c$ is not predicted since the corresponding angular velocity is associated with a conserved quantity $L_0$. To see this let us consider the evolution of the mixing angle $\theta$, $$\theta = \theta_0 + \int_{r_0}^r\frac{Ldr}{r^2\sqrt{-2V}}
\,.
\label{eq:theta:general}$$ This can be integrated for example in radiation era ($\epsilon=2$). This is the case, since $\epsilon_I=0$, $L=L_0$ is conserved, and the integral (\[eq:theta:general\]) evaluates to $$\theta = \theta_0
-{\rm Arcsin}\bigg(\frac{\frac{L_0^2}{r}-\frac{U_0}{2}}{\Delta}\bigg)
\,,\qquad \Delta^2 = L_0^2 + \Big(\frac{U_0}{2}\Big)^2
\,,
\label{eq:theta:radiation}$$ where we absorbed the value of the integral at $r_0$ into $\theta_0$. Because of the undetermined $\theta_0$, $\theta_c\equiv\theta(r_c)$ is indeed not predicted. Yet demanding $\theta\rightarrow 0$ when $r\rightarrow\infty$ gives $\theta_0 =
-{\rm Arcsin}(U_0/(2\Delta))$. At the minimal radius $r=r_c$, Eq. (\[eq:theta:radiation\]) implies $\theta_c=\theta_0 -\pi/2$, such that $\theta_c$ can be anywhere between $-\pi$ and $\pi/2$, depending on $L_0$ and $U_0$. For example, in the limit when $L_0/U_0\rightarrow 0$, $\Delta \theta \rightarrow -\pi$, while in the opposite limit when $L_0/U_0\rightarrow \infty$, $\Delta
\theta \rightarrow -\pi/2$. Note that in the latter case the Universe’s expansion rate at the minimal radius $r_c$ is completely determined by $E_c$.
To complete the analysis of the geodesic equation, one needs to integrate Eq. (\[FLRW:geodesic eq:her:3b\]). By observing that $\bar H u^{\bar 0}=d\ln(\bar a)/\tau$, one integral can be trivially performed, resulting in $$\frac{dz^i}{d\tau} = u^i(\tau_0)\frac{\bar a_0}{\bar a(\tau)}
\,.
%\label{FLRW:geodesic eq:her:}
\nonumber$$ This can be integrated to get $z^i=z^i(\tau)$ in special cases by making use of the dependence of the scale factor $a=a(r,\theta)$ in Eq. (\[eq:A\]), based on which analysis of the causal structure of Hermitian cosmology can be performed. We postpone this analysis for future work.
When $\epsilon<1$ and when $V_e<0$ in Eq. (\[re:Ve\]) (or when $\epsilon=1$ and $L_0^2<U_0$) the Universe collapses towards the Big Bang singularity $r\rightarrow 0$ in a finite time. This will be the case only when the weak energy condition is violated, that is when $\rho+3p<0$, where $\rho$ and $p$ denote the energy density and pressure of the cosmological fluid, respectively. (These statements are based on the relation, $\epsilon = (3/2)(1+w)$, where $w=p/\rho$, which holds in standard FLRW cosmology.) Notice that even when $\epsilon<1$, the Universe may exhibit a bounce, provided $V_e>0$, or equivalently if the angular momentum is large enough, $L_0^2>U_0^{2-\epsilon}\epsilon^{-\epsilon}(1-\epsilon)^{-(1+\epsilon)}$. In this case there is a finite barrier for a Universe to tunnel to smaller radii where $V(r)<0$; if that happens, the Universe hits eventually the Big Bang singularity ${\cal H}\rightarrow \infty$. This means that inflation and bounce cosmology are not mutually incompatible.
We have thus shown that Hermitian gravity solves the problem of Big Bang singularity of Einstein’s theory in a natural way.
The cosmological constant problem {#The cosmological constant problem}
=================================
Let us first recall Eqs. (\[eq:Hermitian Einstein’s equations:1\]–\[eq:Hermitian Einstein’s equations:2\]), which we now write as,
$$\begin{aligned}
G_{\mu\nu} + C_{\mu\nu} \Lambda &=& 8\pi G_N T_{\mu\nu}
\label{eq:HEeqs:1a}
\\
G_{\mu\bar\nu} + C_{\mu\bar\nu}\Lambda &=& 8\pi G_N T_{\mu\bar\nu}
\,. \label{eq:HEes:2a}\end{aligned}$$
Now imposing the reciprocity symmetry on shell implies $$C_{\mu\nu} = 0
\,,$$ which means that the geometric cosmological term cannot contribute to the holomorphic equation (\[eq:HEeqs:1a\]). Furthermore, as we have seen in section \[Hermitian Cosmology\], the reciprocity symmetry reduces the Hermitian sector (\[eq:HEes:2a\]) to the constraints (\[eq:herm:cosmology:1\]–\[eq:herm:cosmology:2\]).
A simple proof that these constraints cannot be met unless $\Lambda$ is fully compensated by a constant term in the scalar potential follows from the observation that the form of the scalar potential $V=V(\phi,\psi)$ is uniquely given by Eq. (\[eq:V:1\]), with $\Omega = \pm\sqrt{4\pi G_N\alpha/3}$, $\lambda = \Lambda/(8\pi G_N)$ and $W=W(\psi,\bar\psi)$. Note that the term $-\Lambda/(8\pi G_N)$ in the potential (\[eq:V:1\]) cancels exactly the geometric cosmological constant $\Lambda$ in Eq. (\[eq:HEes:2a\]). This proof applies only to Hermitian cosmology governed by two scalar fields $\phi$ and $\psi$ as described by Eqs. (\[eq:complex action:full\]–\[eq:scalar action:fields\]).
Since this is an important point, we shall now construct an alternative proof, which shows that the assumption that the late times Universe approaches a de Sitter phase with a constant expansion rate governed by some $\Lambda_{\rm eff}>0$ leads to contradiction, resolved by requiring $\Lambda_{\rm eff}\rightarrow 0$.
Before we proceed, let us recall the standard FLRW cosmology filled with a matter with an equation of state, $w_M=p/\rho>-1$ ($\epsilon_M = (3/2)(1+w_M)$) and a cosmological term $\Lambda$. The (classical) Hubble parameter is of the form,
$$\begin{aligned}
H_{GR} = \sqrt{\frac{\Lambda}{3}}
\coth\bigg(\epsilon_M\sqrt{\frac{\Lambda}{3}}\,t\bigg)
\\ \label{Hcl}
\dot H_{GR} = \frac{\Lambda}{3}\frac{\epsilon_M}
{\cosh^2\Big(\epsilon_M\sqrt{\frac{\Lambda}{3}}\,t\Big)}
\,,\end{aligned}$$
such that at late times $t\gg (1+w_M)^{-1}\sqrt{\Lambda/3}$, the expansion rate $H_{GR}$ approaches the de Sitter attractor, $H_{GR}\rightarrow H_{dS}=\sqrt{\Lambda/3}$, and $\dot
H_{GR}\rightarrow 0$ exponentially fast. This means that a universe filled with any matter with an equation of state with $w_M>-1$ will eventually approach the late time de Sitter attractor. This is the case, simply because the energy density in any matter fluid, with $w_M>-1$, dilutes as $\rho_M\propto
1/a^{3(1+w_M)}\propto 1/t^2$ as the Universe expands (provided $w_M$ is constant), such that at sufficiently late times the cosmological constant necessarily dominates.
To construct an alternative proof, let us assume that at late times the Universe approaches a solution with a non-zero effective cosmological constant $\Lambda_{\rm eff}$, which yields a constant expansion rate $H\rightarrow H_{dS}=\sqrt{\Lambda_{\rm eff}/3}$ and $\dot
H\rightarrow 0$. $\Lambda_{\rm eff}$ is not necessarily the original geometric cosmological constant, yet it must be strictly positive and $\partial_{z^0} \Lambda_{\rm eff}\rightarrow 0$ as $|z^0|\rightarrow \infty$. Firstly, from Eq. (\[eq:herm:cosmology:1\]) we see that as $|z^0|\rightarrow
\infty$, $$\begin{aligned}
\nonumber
H\bar H
=
\frac{1}{9}\big(8\pi G_N V + \Lambda\big)
\rightarrow \frac{\Lambda_{\rm eff}}{9} = {\rm const.}
\\\label{eq:HbarH}
(\Lambda_{\rm eff} = 8\pi G_N V_{\rm eff}) \,,\end{aligned}$$ or equivalently, $$\begin{aligned}
\nonumber
(\partial_{z^{0}}H)\bar H
\rightarrow \frac{8\pi G_N }{9} \partial_{z^0}V_{\rm eff}&\rightarrow& 0
\\ \label{eq:HbarH:dot}
H\partial_{z^{\bar 0}}{\bar H}
\rightarrow \frac{8\pi G_N}{9}\partial_{z^{\bar 0}}V_{\rm eff}
\rightarrow 0
\phantom{ha} (|z^0|&\rightarrow& \infty) \,.\end{aligned}$$ Next, we multiply Eq. (\[eq:herm:cosmology:4\]) by $\dot{\bar \phi}$ and make use of Eq. (\[eq:herm:cosmology:1\]) to arrive at, $$\dot\phi \dot{\bar \phi} \rightarrow
- \frac{1}{3\alpha}\frac{1}{\bar H}
\frac{\partial V_{\rm eff}}{\partial z^{\bar 0}}
\,, \label{eq:dotphi-dotphi}$$ where we made use of $\dot{\bar \phi}\partial_{\bar \phi} V
\rightarrow \dot{\bar \phi}\partial_{\bar \phi} V_{\rm eff}
\equiv \partial V_{\rm eff}/\partial z^{\bar 0} $, with $V_{\rm eff} = \Lambda_{\rm eff}/(8\pi G_N)$. Now combining Eqs. (\[eq:dotphi-dotphi\]) with Eq. (\[eq:herm:cosmology:2\]) yields, $$\frac{\partial \ln[V_{\rm eff}]}{\partial z^{\bar 0}} = -\frac12 \bar H
\,. \label{eq:constraint:z0 derivative}$$ The analogous complex conjugate equation also holds. But from Eq. (\[eq:HbarH:dot\]) we know that at late times $V_{\rm eff}$ must approach a constant, and thus $$\frac{\partial \ln[V_{\rm eff}]}{\partial z^{\bar 0}}
\rightarrow 0
\,,\qquad
\frac{\partial \ln[V_{\rm eff}]}{\partial z^{0}}
\rightarrow 0
\,\qquad (|z^0|\rightarrow \infty) \,,$$ implying finally that at late times $\bar H\rightarrow 0$, which together with Eq. (\[eq:HbarH\]) gives, $$H\bar H \rightarrow \frac{\Lambda_{\rm eff}}{9} \rightarrow 0
\,\qquad (|z^0|\rightarrow \infty) \,. \label{eq:HbarH:2}$$ This completes the proof that there is no late time de Sitter attractor driven by a nonvanishing effective cosmological term $\Lambda_{\rm eff}>0$ in Hermitian gravity (with the two scalar field action (\[eq:scalar action\]) used in this article).
To summarize, we have shown that the consistency of Hermitian gravity constraints requires $\Lambda_{\rm eff} = \Lambda + 8\pi
G_N V_0 \rightarrow 0$, where $V_0$ represents the time (and energy) independent part of the scalar potential $V$. In other words, any cosmological term of Hermitian gravity must be fully and precisely compensated by the corresponding scalar potential.
The question is whether this holds more generally when other types of matter fields (fermions and gauge fields) are included. And moreover, what happens when quantum corrections are included. We postpone the discussion of these (important) questions for future work.
Nevertheless, note that an appropriate choice of the potential for the second scalar field $\psi$ can lead to arbitrary (power law) expansion rate, which also includes a near exponential expansion with $\epsilon\simeq 0$. Even though this type of conformal scalar $\psi$ matter behaves similar to a cosmological term, it is not completely identical. In fact, the choice $\epsilon = 0$ in Eq. (\[eq:V:power-law\]) is very particular (it entails [*fine tuning*]{}), and thus does not comprise a cosmological constant problem. Let us now consider the limit $\epsilon\rightarrow 0$. The potential (\[eq:w:power-law\]–\[eq:V:power-law\]) $$w(\psi,\bar\psi) \rightarrow w_0
\exp\Big[ i \omega\sqrt{(\beta/\alpha)}\, (\psi-\bar \psi)\Big]$$ is oscillatory (here we used $\sqrt{-1}=i$). This potential becomes exponential if $\beta/\alpha<0$.
Note also that $\epsilon=1/2$ ($w=-2/3$) has special relevance. This power-law accelerated expansion is realized in the absence of the second field $\psi$.
Let us now rewrite Eq. (\[eq:power law expansion\]) as, $$a = a_0\left(1+h\epsilon z^0\right)^{1/\epsilon}
\label{eq:a:shifted}$$ where we shifted time $z^0\rightarrow z^0 + \zeta_0$ and we defined, $h=1/(\epsilon \zeta_0)$. Now upon taking the limit $\epsilon\rightarrow 0$, Eq. (\[eq:a:shifted\]) reduces to, $$a = a_0\exp\left(h z^0\right)
\,\qquad
(h \in \mathbb{C})
\,,
\label{eq:a:shifted:2}$$ representing an (exponentially expanding) complex de Sitter universe of Hermitian gravity with the complex Hubble parameter, $$H = h
\,. \label{eq:H:dS}$$ This holomorphic de Sitter space must be distinguished from the de Sitter space induced by a (real) cosmological term in the Hermitian sector of the theory.
In summary, we found that, as a consequence of the reciprocity symmetry, Hermitean gravity does not admit a cosmological term at the classical level neither in the holomorphic sector nor in the Hermitean sector of the theory. Yet it does admit a holomorphic de Sitter space realized by a holomorphic scalar field (with a holomorphic kinetic term and with a suitably fine tuned exponential potential).
We have thus formulated a generalized theory of gravitation which (at the classical level) allows Minkowski space, but does not admit de Sitter space realized by a positive cosmological constant.
Discussion {#Discussion}
==========
We have formulated a generalized theory of gravity on Hermitian manifolds. Given the extensive literature on complex manifolds, we summarize (and emphasize) the novel aspects of our work and compare it to existing literature:
- Our Hermitian theory of gravity lives on a Hermitian manifold of real dimension eight. There are four space-time ($x^\mu$) and four momentum-energy ($p^\mu$) coordinates. The fundamental dynamical quantity of the theory is a holomorphic tetrad, which is a function of $z^\mu = x^\mu + i(G_N/c^3)p^\mu$. The tetrad transforms by means of holomorphic coordinate transformations (\[tetrad:coord.transformations\]), which in general mix space-time and momentum-energy. This extends and generalizes both the principle of covariance and equivalence of general relativity. We identify the reciprocity symmetry with the operation of the almost complex structure operator, which transforms $\partial/\partial x^\mu$ into $\partial/\partial
y^\mu$ and $\partial/\partial y^\mu$ into $-\partial/\partial
x^\mu$. The reciprocity symmetry of a Hermitian manifold demands that the world on ‘very large scales’ ($\ell \gg l_{\rm Pl}$) (general relativistic limit) mirrors the world on ‘very small scales’ ($\ell \ll l_{\rm Pl}$) (microscopic super-Planckian world), but with the role of space-time and momentum-energy exchanged [@Born1938Reciprocity]. This symmetry leaves the commutation relation (\[eq:commutation relations\]) invariant. The reciprocity symmetry implies holomorphy of the tetrad fields. Holomorphy reduces the degrees of freedom of an eight dimensional theory to the degrees of freedom of an effectively four dimensional world, as required by all known observations.
- The eight dimensional formulation of the theory is symmetric, and yet the (four dimensional) metric contains an antisymmetric tensor, which corresponds to the imaginary part of the metric tensor $C_{\mu\bar\nu}$ (\[eq:Einstein’s decomposition of the Hermitian line element\]), which gives rise to dynamical torsion (this still awaits a rigorous proof). Our theory differs from other dynamical theories of torsion (see for example [@Moffat:1978tr]) in the leading order dynamics of torsion. It results from a theory that is projected on the space-time submanifold such that different orders in $p^\mu$ mix as a consequence of the Cauchy-Riemann equations (see remark [6]{} below). In this work we do not address the dynamics of the antisymmetric part of the metric, which may be of importance for example for the dark matter of the Universe [@Prokopec:2006kr], for spinning black holes and for the Lens-Thirring effect. Yet the fact that our Hermitian gravity theory corresponds to a ‘standard’ gravity theory of a symmetric metric field on an eight dimensional (Hermitian) manifold, is a strong indication that the theory of torsion within our Hermitian gravity does not suffer from the stability problems [@Damour:1991ru; @Janssen:2006jx] of – for example – the NGT of Ref. [@Moffat:1978tr].
- We define parallel transport by means of a metric compatible covariant derivative $\nabla_\mu$. Contrary to most (mathematical) literature on Hermitian manifolds [@Nakahara:2003nw], our covariant derivative is metric compatible, but [*not*]{} tetrad compatible (\[eq:vielbein compat\]) in the sense discussed. We consider our definition of the covariant derivative as more natural and better physically motivated, as it stems from the action principle for test particles, $S=-m\int ds$. Our covariant derivative implies ‘nonstandard’ Hermitian connection coefficients (\[eq:Hermitian connection coefficients z basis 4 d\]).
- The causal structure of the theory is changed such that in the flat space limit, the space-time-momentum-energy line element is invariant under the $U(1,3)$ group (the Hermitian line element is also invariant under complex translations and hence invariant under the Hermitian generalization of the Poincaré group). The momentum-energy coordinates can be interpreted as coordinates describing non-inertial frames. The $U(1,3)$ reduces to its subgroup, the Lorentz group $SO(1,3)$, whenever observers move inertially with respect to each other; the momentum energy part of the Hermitian flat space line element vanishes. When observers move non-inertially with respect to each other, the principle of covariance of general relativity is broken, but at the same time replaced by an extended principle of covariance (namely, the flat space metric is invariant under the U(1,3) group). Yet this breaking becomes significant only in strong gravitational fields and for large momenta and energies of observers/particles, and hence does not necessarily contradict observations.
The causal structure of the flat space limit is changed in such a way that there is a minimal time for events to be in causal contact and a maximal radius $r_{\rm max}$ for a non-local instantaneously causally related volume. The speed of light can exceed the conventional speed of light in non-inertial frames. The requirement that signals can propagate results in an upper limit on the four force squared $f^2$, which describes non-inertial transformations. Since there is no lower bound on $f^2$, there is in principle no upper limit on the group velocity, such that superluminal propagation is allowed within our theory. When the non-inertial frame of a test particle is put ‘on-shell’, such that the four momentum-energy squared is given by the particle’s mass, $p^2 = -m^2c^2$, then $r_{\rm max}\rightarrow G_Nm/c^2$ becomes one half of the Schwarzschild radius. Our analysis is based on the geodesic equation which does not take account of the self-gravity of test particles. This suggests that the above mentioned violation of causality will get hidden within the corresponding particle’s black hole radius, possibly rendering any violation of causality unobservable. In conclusion, only a more proper study of this phenomenon can fully resolve the question of causality in Hermitian gravity.
- We define an action principle for gravity and matter, where we describe the matter by two scalar fields. The pure gravity action is holomorphic in the sense that the tetrad field is a holomorphic function. The reciprocity symmetry is imposed by a constraint action, such that it is realized at the level of the equations of motion (on-shell). This assures that the Bianchi identities are satisfied. The scalar field action is covariant and built out of scalar fields that are holomorphic functions (of $z^\mu$). One scalar field has a Hermitian kinetic term, and another a holomorphic kinetic term; the potential is the product of a holomorphic function and its anti-holomorphic counterpart. Both scalars obey the covariant stress-energy conservation law, such that the Hermitian Einstein equations with scalar matter are consistent.
- We study the general relativistic limit of the theory, which is realized by projecting the dynamics onto the four dimensional space-time hypersurface. An essential element in this projection are the Cauchy-Riemann equations, which are a consequence of the reciprocity (holomorphy) symmetry of the theory. The resulting projected theory is [*holographic*]{} in the sense that, having a complete knowledge of the (complex) tetrad projected onto the four dimensional space-time manifold, allows for an unambiguous reconstruction of the full eight dimensional dynamics of Hermitian gravity (the reconstruction is essentially based on the principle of analytic extension generalized to Hermitian manifolds). We find that – to leading order in momentum-energy $p^\mu$ – the geodesic equations reduce to those of general relativity. On the other hand, the (projected) dynamical (Einstein’s) equations are not mutually identical even at zeroth order in $p^\mu$; the Cauchy-Riemann equations mix different orders of $p^\mu$. Thus in order to check the validity of our Hermitian formulation of gravity, one ought to explicitly construct and study the Hermitian analogues of [*each*]{} of the important solutions of general relativity. Only such a detailed comparison can establish the validity of Hermitian gravity, or rule it out.
- In order to investigate whether our Hermitian gravity is a viable alternative to general relativity, we study some important aspects of Hermitian cosmology. For definiteness and simplicity, we focus on flat, homogeneous and isotropic universes which expand according the power law. This class of solutions includes most of the important cosmological solutions, including the matter era, radiation era, inflation, and – as a limit – de Sitter space. As said before, our matter is described by two scalar fields. The purpose of the scalar field with a Hermitian kinetic term is to satisfy the constraints of the Hermitian sector of the theory. This field is used to ‘mark’ the scale factor of the Universe. The scalar field with a holomorphic kinetic term drives the Universe’s expansion. We show that at late times, when $t\gg (G_N/c^4)E$, Hermitian cosmology reduces to FLRW cosmology of general relativity, where $E$ denotes the relevant energy scale. At early times the two theories deviate significantly. While Einstein’s theory exhibits the well known Big Bang singularity, where the curvature invariants diverge, and the theory stops giving reliable predictions, our Hermitian gravity predicts a [*bounce*]{} Universe with a calculable minimal size and maximal space-time curvature. The contracting and expanding phases of an Hermitian gravity bounce can be asymmetric. This is a consequence of time reversal violation induced if the scalar field that drives the expansion violates CP (CPT is conserved). There is a caveat though: the observers which do not exhibit a mixing between the time-like and energy-like coordinates might still experience a Big Bang singularity. However, such observers are rare, and represent a negligible class of observers with very special initial conditions (mathematically speaking, the phase space corresponding to these observers is of measure zero). Moreover the time-energy rotation can be absent only in those universes where the mixing between time and energy is not dynamically generated. Yet there is no reason to presume that our Universe does not contain such a dynamical mixing.
- Our analysis of Hermitian cosmology confirms the expectation that, even at zeroth order, Hermitian gravity differs from Einstein’s gravity. The difference becomes significant, however, when space-time curvature is large, which is still in essence an untested sector of Einstein’s theory. In future work we hope to investigate other aspects of the theory, whenever space-time curvature is large, such that the difference between the two theories can again become significant, e.g. various types of black hole solutions.
- We consider the cosmological constant problem within our theory: the pure Hermitian gravity and two holomorphic scalar fields in a cosmological setting. Our analysis shows that any cosmological constant is forbidden at the classical level, thus solving the gravitational hierarchy problem within this framework. While this is a very welcome property of the theory, it is still to a large extent a mystery, and awaits a further and deeper understanding. In particular, we are interested in the question whether a link can be established between the reciprocity symmetry and the vanishing of cosmological constant. Moreover, we would like to find out whether the cosmological constant vanishes when other kinds of matter fields (in particular fermionic and gauge fields) are included. Furthermore, we would like to investigate whether our proof can be extended to include quantum effects.
- Finally, we are of course interested in quantizing Hermitian gravity. At this stage we stress the curious fact that the commutation relations – when imposed on the space-time and momentum-energy coordinates (\[eq:commutation relations\]) – respect the reciprocity symmetry. This is an important hint on how to quantize Hermitian gravity.
There are various other open questions which we have not addressed here. They include: (1) can violation of the principles of equivalence and covariance be observed; (2) can Hermitian gravity describe the observed inwards spiralling of the Taylor-Hulse binary pulsar; (3) does our theory meet all of the Solar system tests; (4) is the bending of light consistent with the predicted bending by general relativity; (5) can Hermitian cosmology produce cosmological perturbations consistent with observations, [*etc.*]{}
Acknowledgements {#acknowledgements .unnumbered}
================
We have enormously benefited from discussions with Willem Westra, who was actively participating at the early stages of this project. The authors acknowledge financial support by FOM grant 07PR2522 and by Utrecht University.
[99]{}
Y. Mao, M. Tegmark, A. Guth and S. Cabi, “Constraining Torsion with Gravity Probe B,” Phys. Rev. D [**76**]{}, 104029 (2007) \[arXiv:gr-qc/0608121\]. J. W. Moffat, “New Theory Of Gravitation,” Phys. Rev. D [**19**]{} (1979) 3554. T. Janssen and T. Prokopec, “Problems and hopes in nonsymmetric gravity,” J. Phys. A [**40**]{} (2007) 7067 \[arXiv:gr-qc/0611005\]. T. Janssen and T. Prokopec, “Instabilities in the nonsymmetric theory of gravitation,” Class. Quant. Grav. [**23**]{} (2006) 4967 \[arXiv:gr-qc/0604094\]. T. Damour, S. Deser and J. G. McCarthy, “Theoretical problems in nonsymmetric gravitational theory,” Phys. Rev. D [**45**]{} (1992) 3289. T. Damour, S. Deser and J. G. McCarthy, “Nonsymmetric Gravity Theories: Inconsistencies And A Cure,” Phys. Rev. D [**47**]{} (1993) 1541 \[arXiv:gr-qc/9207003\]. M. Born, “A Suggestion For Unifying Quantum Theory and Relativity,” Royal Society of London Proceedings Series A [**165**]{} (1938).
M. Born, “Reciprocity Theory of Elementary Particles,” Reviews of Modern Physics [**21**]{} (1949) 463-473.
A. Einstein, “On The Electrodynamics Of Moving Bodies,” Annalen der Physik [**XVII**]{} (1905) 891-921.
M. Nakahara, “Geometry, topology and physics,” [*Boca Raton, USA: Taylor & Francis (2003) 573 p*]{}
G. ’t Hooft, “Dimensional reduction in quantum gravity,” arXiv:gr-qc/9310026. R. Bousso, “The holographic principle,” Rev. Mod. Phys. [**74**]{} (2002) 825 \[arXiv:hep-th/0203101\]. A. Einstein, “A generalization of the relativistic theory of gravitation,” Annals Math. [**46**]{} (1945) 578. A. Einstein and E. G. Strauss, “A generalization of the relativistic theory of gravitation. 2,” Annals Math. [**47**]{} (1946) 731. A. Einstein, “A Generalized Theory of Gravitation,” Rev. Mod. Phys. [**20**]{} (1948) 35.
S. G. Low, “Reciprocal relativity of noninertial frames and the quaplectic group,” Found. Phys. [**36**]{}, 1036 (2006) \[arXiv:math-ph/0506031\]. S. G. Low, “Reciprocal relativity of noninertial frames: quantum mechanics,” J. Phys. A: Math. Theor. [**40**]{} (2007) 3999-4016 \[arXiv:math-ph/0606015\].
S. G. Low, “Relativity group for noninertial frames in Hamilton’s mechanics,” J. Math. Phys. [**48**]{} (2007) 102901 \[arXiv:0705.2030\]
J. Govaerts, P. D. Jarvis, S. O. Morgan, S. G.L ow, “World-line Quantisation of a Reciprocally Invariant System,” J. Phys. A: Math. Theor. [**40**]{} (2007) 12095-12111 \[arXiv:0706.3736\].
J. W. Moffat, “Superluminary Universe: A Possible Solution To The Initial Value Problem In Cosmology,” Int. J. Mod. Phys. D [**2**]{} (1993) 351 \[arXiv:gr-qc/9211020\]. J. D. Barrow and J. Magueijo, “Solving the flatness and quasi-flatness problems in Brans-Dicke cosmologies with a varying light speed,” Class. Quant. Grav. [**16**]{} (1999) 1435 \[arXiv:astro-ph/9901049\].
A. H. Chamseddine, “Hermitian geometry and complex space-time,” Commun. Math. Phys. [**264**]{} (2006) 291 \[arXiv:hep-th/0503048\]. A. H. Chamseddine, “Gravity in complex Hermitian space-time,” arXiv:hep-th/0610099.
C. L. M. Mantz, “Holomorphic Gravity,” Utrecht University Master’s Thesis (2007) [http://www1.phys.uu.nl/ wwwitf/Teaching/2007/Mantz.pdf.]{}
T. Prokopec and W. Valkenburg, “Antisymmetric metric field as dark matter,” arXiv:astro-ph/0606315. T. Prokopec and W. Valkenburg, “The cosmology of the nonsymmetric theory of gravitation,” Phys. Lett. B [**636**]{} (2006) 1 \[arXiv:astro-ph/0503289\].
[^1]: Here the Latin indices $a,b$ run from $0,1,...,d-1$, since they represent local indices, and $\eta = $ diag$(-1,1,1,1)$. We will the discuss the meaning of these indices later.
[^2]: Checks ( $\check{}$ ) are put on indices to denote the imaginary part of a coordinate or on indices of objects, which are projected onto its basis vector.
[^3]: This is of course quite different from ’t Hooft’s [*holographic*]{} principle [@'t; @Hooft:1993gx] for quantum gravity.
[^4]: We can prove this as follows. $
\langle\hat{w}^{\mu};\frac{\partial}{\partial\bar{z}^{\bar{\beta}}}\rangle
=
\frac{1}{2}\langle \hat{u}^{\mu}
+
i \hat{v}^{\check{\mu}} ;
\frac{\partial}{\partial x^{\beta}}
+
i\frac{\partial}{\partial y^{\check{\beta}}} \rangle
=
\frac{1}{2}\langle e^{\mu}_{\alpha}d x^{\alpha}
+
e^{\mu}_{\check{\alpha}} d y^{\check{\alpha}}
+
i(e^{\check{\mu}}_{\alpha}d x^{\alpha}
+
e^{\check{\mu}}_{\check{\alpha}}d y^{\check{\alpha}});
\frac{\partial}{\partial x^{\beta}}
+
i\frac{\partial}{\partial y^{\check{\beta}}}\rangle
=
\frac{1}{2}[ e^{\mu}_{\beta}
+
i( e^{\mu}_{\check{\beta}}
+
e^{\check{\mu}}_{\beta})
-
e^{\check{\mu}}_{\check{\beta}}].
$ Note that the whole expression is equal to zero by the definition that vielbeins are holomorphic, because this implies that $\hat{w}^{\mu} = e^\mu_\alpha (z) dz^{\alpha}$. Now both the real and imaginary part of the expression vanish, yielding $$\begin{aligned}
\nonumber
e^{\mu}_{\beta}
=
e^{\check{\mu}}_{\check{\beta}}
\phantom{hallo}
e^{\mu}_{\check{\beta}}
= -
e^{\check{\mu}}_{\beta}.\end{aligned}$$
[^5]: In principle one can pick any constant such that the units come out right. If one decides to construct the constant without introducing new constants, Newton’s constant over the speed of light cubed is a unique choice up to factors of order unity. Since we are, up to this point, constructing a classical theory, there is no room for Planck’s constant.
[^6]: In this article we define dynamical torsion to be a second order differential equations constraining the anti-symmetric part of the metric (\[eq:Einstein’s decomposition of the Hermitian line element\]). This has in principle nothing to do with the vanishing or nonvanishing of the torsion tensor.
[^7]: In order to show that $\nabla_\rho C_{\bar{\mu}\nu} = 0$, one needs to realize that $C^{\mu\bar{\epsilon}}C_{\nu\bar{\epsilon}} \equiv \beta^\mu_\nu
\neq \delta^\mu_\nu = C^{\bar{\epsilon}\mu}C_{\nu\bar{\epsilon}}$, since the Hermitian metric is nonsymmetric (it is Hermitian). The action of $\beta$ on the metric, $\beta^\alpha_\nu
C_{\bar{\mu}\alpha} = C_{\nu\bar{\mu}}$, can be derived by inserting the identity in the Hermitian line element (\[eq:Hermitian familiar form line element\]).
[^8]: To generalist the [*Ansatz*]{} (\[eq:vielbein:cosmology\]) to space-times with a constant spatial curvature, one would have to replace $a(z^0)\delta_\mu^a$ in Eq. (\[eq:vielbein:cosmology\]) by the corresponding vielbein whose spatial indices describe the geometry of a static 3-sphere (3-hyperboloid) for a space with positively (negatively) curved spatial sections. Thus for a space-time with positively curved spatial sections ($\kappa>0$) we have, $e_\mu = a(z^0)[\delta_\mu^{0}+\delta_\mu^{\chi}
+(1/\sqrt{\kappa})\sin(\sqrt{\kappa}\chi)\delta_\mu^{\theta}
+(1/\sqrt{\kappa})\sin(\sqrt{\kappa}\chi)\sin(\theta)\delta_\mu^{\varphi}]$, where $\chi\in[0,\pi/\sqrt{\kappa}]$, $\theta\in[0,\pi]$ and $\varphi \in[0,2\pi)$ are the spherical coordinates on $S^3$. For a space with a negative curvature ($\kappa<0$) the tetrad $e_\mu$ is obtained from the tetrad of the closed universe with the replacement, $(1/\sqrt{\kappa})\sin(\sqrt{\kappa}\chi)\rightarrow
(1/\sqrt{-\kappa})\sinh(\sqrt{-\kappa}\chi)$, where now $\chi\in[0,\infty)$.
[^9]: Here the Latin indices $i$ and $j$ take values $1,2,3$.
[^10]: One would arrive at a more standard expression for the scalar energy density if one would replace $\rho\rightarrow (\rho-p)/2$ in Eq. (\[stress energy:Hermitian\]).
[^11]: Indeed, integrating ${\cal H}$ would result in the scale factor proportional to $$\Bigg(\frac{c^4t}{G_NE}+\sqrt{\Big(\frac{c^4t}{G_NE}\Big)^2+1}\;\Bigg)^{1/|\epsilon|}
\,,$$ which differs from Eq. (\[eq:A\]).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Charles-Alban Deledalle, Loïc Denis, Sonia Tabti, Florence Tupin'
title: 'MuLoG, or How to apply Gaussian denoisersto multi-channel SAR speckle reduction?'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the generation of large-scale vortices in rotating turbulent convection by means of Cartesian direct numerical simulations. We find that for sufficiently rapid rotation, cyclonic structures on a scale large in comparison to that of the convective eddies, emerge, provided that the fluid Reynolds number exceeds a critical value. For slower rotation, cold cyclonic vortices are preferred, whereas for rapid rotation, warm anti-cyclonic vortices are favoured. In some runs in the intermediate regime both types of cyclones co-exist for thousands of convective turnover times. The temperature contrast between the vortices and the surrounding atmosphere is of the order of five per cent. We relate the simulation results to observations of rapidly rotating late-type stars that are known to exhibit large high-latitude spots from Doppler imaging. In many cases, cool spots are accompanied with spotted regions with temperatures higher than the average. In this paper, we investigate a scenario according to which the spots observed in the temperature maps could have a non-magnetic origin due to large-scale vortices in the convection zones of the stars.'
author:
- 'Petri J. Käpylä$^{1,2}$, Maarit J. Mantere$^{1}$ and Thomas Hackman$^{1,3}$'
bibliography:
- 'paper.bib'
title: 'Starspots due to large-scale vortices in rotating turbulent convection'
---
Introduction
============
Rotating turbulent convection is considered to play a crucial role in the generation of large-scale magnetic fields [@M78; @KR80; @RH04] and differential rotation of stars [@R89]. The interaction of rotation and inhomogeneous turbulence leads to the so-called $\alpha$-effect, which can sustain large-scale magnetic fields [e.g. @B01; @KKB09b]. However, in many astrophysically relevant cases large-scale shear flows are also present, which further facilitate dynamo action by lowering the relevant critical dynamo number. In the Sun, for example, the entire convection zone is rotating differentially [cf. @Schouea98; @Thompsonea03], and a meridional flow towards the poles is observed in the near surface layers [e.g. @ZK04]. These flows are most often attributed to rotationally influenced turbulent angular momentum and heat transport [cf. @R89; @RC01; @MBT06; @KMGBC11]. In the solar case the large-scale flows and also the magnetic activity are largely axisymmetric [e.g. @PBKT06]. This means that the sunspots, which are concentrations of strong magnetic fields, are almost uniformly distributed in longitude over the solar surface. The fact that we observe the sunspots and can attribute magnetic fields to them, has strongly influenced the interpretation of data from stars other than the Sun.
The giant planets Jupiter and Saturn are also likely to have outer convection zones [e.g. @Busse76], but they rotate much faster than the Sun. Bands of slower and faster rotation alternate in their atmospheres, reminiscent of rapidly rotating convection [e.g. @Busse94; @HA07]. However, especially in Jupiter, large spots in the form of immense storms are observed [@Marcus93]. Remarkably, the largest of these, the Great Red Spot, has persisted at least 180 years. Similar features are observed also in Saturn [e.g. @saturn1991] and other giant planets. The spots on giant planets are not of magnetic origin although dynamos are likely to be present in the interiors of the planets. Thus their explanation is probably related to hydrodynamical processes within the convectively unstable layers.
Late-type stars with higher rotation velocities in comparison to the Sun, on the other hand, often exhibit light curve variations that are usually interpreted as large spots on the stellar surface [e.g. @Chuga66; @Henry1995]. In some cases the observational data can be fitted with a model with two large spots at a 180 degree separation in longitude @BT98. There is also evidence that these ‘active longitudes’ are not equal in strength [e.g. @Jyri11; @Marjaana11], and that the relative strenght of the spots can, at least temporarily, reverse in a process dubbed ’flip-flop’ [cf. @Jetsu1993]. One interpretation of the data is that the spots are of magnetic origin and that the flip-flops are related to magnetic cycles reminescent of the solar cycle [e.g. @BBIT98]. On the other hand, it has been proposed that the flip-flops are only short-term changes related to the activity cycle, while the structure generating the temperature minima would migrate in the orbital reference frame, that could be interpreted as an azimuthal dynamo wave [e.g. @Jyri11; @Marjaana11]. Again, this interpretation relies on the magnetic nature of the cool spots.
The cool spots detected by photometry and Doppler imaging using spectroscopic observations have been taken as an indirect proxy of the magnetic field on the stellar surface, deriving from the analogy to sunspots - strong magnetic field hinders convection and causes the magnetized region to be cooler than its surrounding. Zeeman-Doppler imaging of spectropolarimetric observations [e.g. @Semel89; @Donati89; @PK02; @Carroll07] provides means to directly measure the magnetic field strength and orientation on the stellar surface. In the study of @Donati97 spectropolarimetric observations of several stars were collected during 23 nights extending over a five year interval. They report that the Zeeman signitures of the cool stars almost always exhibit a very complex shape with many successive sign reversals. This points to a rather complicated field structure with different magnetic regions of opposite polarities. Furthermore, the magnetic regions detected were mostly 500 to 1,000 K cooler than, and sometimes at the same temperature as, but never warmer than the surrounding photosphere. In the published temperature and magnetic field maps for AB Dor [@DC97], however, no clear correlation between temperature and magnetic field strength can be seen: in the temperature maps a pronounced cool polar cap with weak fringes towards lower latitudes are visible, whereas the strongest magnetic fields are seen as patchy structures at lower latitudes with a clearly different distribution than the temperature structures. Similar decorrelation of temperature minima and magnetic field strength has been reported with the same method for different objects [e.g. @Donati99; @Jeffers11], and also for the same objects with different methods [e.g. @Hussain2000; @Oleg11]. The phenomenon, therefore, seems to be wide-spread, and method-independent.
One possible explanation to the decorrelation of magnetic field and temperature structures could be that there is simply less light coming from the spotted parts than from the unspotted surface. Thus the Zeeman signatures from cool spots may be “drowned” in the signal from the unspotted surface or bright features. However, this should lead to systematic effects where the detected magnetic field strength would be correlated with the surface temperature. The least sqaures deconvolution technique [LSD, e.g. @Donati97], which is necessary for enhancing the Zeeman signal, may influence the temperature and magnetic Doppler imaging differently. The latitudes of any surface features in Doppler images are always more unrealiable than the longitudes, a fact that will not make a comparison of temperature and magnetic field maps any easier. One could thus expect, that there could be artificial discrepancies in the latitudes of magnetic and temperature features. Still, the lack of connection between even the longitudes of cool spots and magnetic features is surprising.
In this paper we consider a completely different scenario, according to which the formation of temperature anomalies on the surfaces of rapidly rotating late-type stars could occur due to a hydrodynamical instability creating large-scale vortices, analogously to the giant planets in the solar system. To manifest this mechanism in action, we simulate rotating turbulent convection in local Cartesian domains, representing parts of the stratified stellar convection zones located near the polar regions. We show that under such a setting, large-scale vortices or cyclones are indeed generated provided that the rotation is sufficiently rapid and the Reynolds number exceeds a critical value. Depending on the handedness of the vortex, which on the other hand depends on the rotation rate, the resulting spot can be cooler or warmer than the surrounding atmosphere.
We acknowledge that our model is rather primitive, lacking realistic radiation transport, spherical geometry, and relying on a polytropic setup for the stratification so detailed comparison with observations is not possible at this point. However, the main purpose of the present paper is to show a proof of concept of the existence of large-scale vortices with temperature anomalies close to those observed in rapidly rotating hydrodynamic convection. We also note that similar large-scale cyclonic structures have recently been reported from large-eddy simulations of turbulent convection [@Chan03; @Chan07]. We make comparisons to these studies when possible.
The model {#sec:model}
=========
Our model setup is similar to that used by [@KKB09b] but without magnetic fields. A rectangular portion of a star is modeled by a box situated at colatitude $\theta$. The box is divided into three layers: an upper cooling layer, a convectively unstable layer, and a stable overshoot layer (see below). We solve the following set of equations for compressible hydrodynamics: $$\frac{\mathrm{D} \ln \rho}{\mathrm{D}t} = -{\bm{\nabla} \cdot }{\bm U},$$ $$\frac{\mathrm{D} \bm U}{\mathrm{D}t} = -\frac{1}{\rho}{\bm \nabla}p + {\bm g} - 2\bm{\Omega} \times \bm{U} + \frac{1}{\rho} \bm{\nabla} \cdot 2 \nu \rho \mbox{\boldmath ${\sf S}$}, \label{equ:UU}$$ $$\frac{\mathrm{D} e}{\mathrm{D}t} = - \frac{p}{\rho}{\bm{\nabla} \cdot }{\bm U} + \frac{1}{\rho} \bm{\nabla} \cdot K \bm{\nabla}T + 2 \nu \mbox{\boldmath ${\sf S}$}^2 - \frac{e\!-\!e_0}{\tau(z)}, \label{equ:ene}$$ where $\mathrm{D}/\mathrm{D}t = {\partial}/{\partial}t + \bm{U} \cdot \bm{\nabla}$ is the advective time derivative, $\nu$ is the kinematic viscosity, $K$ is the heat conductivity, $\rho$ is the density, $\bm{U}$ is the velocity, $\bm{g} = -g\hat{\bm{z}}$ is the gravitational acceleration, and $\bm{\Omega}=\Omega_0(-\sin \theta,0,\cos \theta)$ is the rotation vector. The fluid obeys an ideal gas law $p=(\gamma-1)\rho e$, where $p$ and $e$ are pressure and internal energy, respectively, and $\gamma = c_{\rm P}/c_{\rm V} = 5/3$ is the ratio of specific heats at constant pressure and volume, respectively. The specific internal energy per unit mass is related to the temperature via $e=c_{\rm V} T$. The rate of strain tensor $\mbox{\boldmath ${\sf S}$}$ is given by $${\sf S}_{ij} = \onehalf (U_{i,j}+U_{j,i}) - \onethird \delta_{ij} {\bm{\nabla} \cdot }\bm{U}.$$ The last term of Eq. (\[equ:ene\]) describes cooling at the top of the domain. Here $\tau(z)$ is a cooling time which has a profile smoothly connecting the upper cooling layer and the convectively unstable layer below, where $\tau\to\infty$.
The positions of the bottom of the box, bottom and top of the convectively unstable layer, and the top of the box, respectively, are given by $(z_1, z_2, z_3, z_4) = (-0.85, 0, 1, 1.15)d$, where $d$ is the depth of the convectively unstable layer. Initially the stratification is piecewise polytropic with polytropic indices $(m_1, m_2, m_3) = (3, 1, 1)$, which leads to a convectively unstable layer above a stable layer at the bottom of the domain. In a system set up this way, convection transports 20 per cent of the total flux [cf. @BCNS05]. Due to the presence of the cooling term, a stably stratified isothermal layer is formed at the top. The horizontal extent of the box, $L_{\rm H}\equiv L_x=L_y$, is $4d$. All simulations with rotation are made at the North pole, corresponding to $\theta=0\degr$. The simulations were performed with the [Pencil Code]{}[^1], which is a high-order finite difference method for solving the compressible equations of magnetohydrodynamics.
Units and nondimensional parameters
-----------------------------------
Nondimensional quantities are obtained by setting $$\begin{aligned}
d = g = \rho_0 = c_{\rm P} = 1\;,\end{aligned}$$ where $\rho_0$ is the initial density at $z_2$. The units of length, time, velocity, density, and entropy are $$\begin{aligned}
&& [x] = d\;,\;\; [t] = \sqrt{d/g}\;,\;\; [U]=\sqrt{dg}\;,\;\; \nonumber \\ && [\rho]=\rho_0\;,\;\; [s]=c_{\rm P}.\end{aligned}$$ We define the Prandtl number and the Rayleigh number as $$\begin{aligned}
{{\rm Pr}}=\frac{\nu}{\chi_0}\;,\;\; {{\rm Ra}}=\frac{gd^4}{\nu \chi_0} \bigg(-\frac{1}{c_{\rm P}}\frac{{\rm d}s}{{\rm d}z
} \bigg)_0\;,\end{aligned}$$ where $\chi_0 = K/(\rho_{\rm m} c_{\rm P})$ is the thermal diffusivity, and $\rho_{\rm m}$ is the density in the middle of the unstable layer, $z_{\rm m} = \onehalf(z_3-z_2)$. The entropy gradient, measured at $z_{\rm m}$, in the nonconvecting hydrostatic state, is given by $$\begin{aligned}
\bigg(-\frac{1}{c_{\rm P}}\frac{{\rm d}s}{{\rm d}z}\bigg)_0 = \frac{\nabla-\nabla_{\rm ad}}{H_{\rm P}}\;,\end{aligned}$$ where $\nabla-\nabla_{\rm ad}$ is the superadiabatic temperature gradient with $\nabla_{\rm ad} = 1-1/\gamma$, $\nabla = ({\partial}\ln T/{\partial}\ln
p)_{z_{\rm m}}$, and where $H_{\rm P}$ is the pressure scale height. The amount of stratification is determined by the parameter $\xi_0 =(\gamma-1) e_0/(gd)$, which is the pressure scale height at the top of the domain normalized by the depth of the unstable layer. We use $\xi_0 =1/3$ in all cases, which results in a density contrast of about 23 across the domain. We define the Reynolds and Peclet numbers via $$\begin{aligned}
{\rm Re} = \frac{{u_{\rm rms}}}{\nu {k_{\rm f}}}\;,\;\; {{{\rm Pe}}} = \frac{{u_{\rm rms}}}{\chi_0 {k_{\rm f}}} = \Pr\ {\rm Re}\;,\end{aligned}$$ where ${k_{\rm f}}= 2\pi/d$ is adopted as an estimate for the wavenumber of the energy-carrying eddies, and ${u_{\rm rms}}=\sqrt{3 u_z^2}$. This definition neglects the contributions from the large-scale vortices that are generated in the rapid rotation regime. Note that with our definitions ${{\rm Re}}$ and ${{\rm Pe}}$ are smaller than the usual one by a factor $2\pi$. The amount of rotation is quantified by the Coriolis number, defined as $$\begin{aligned}
{\rm Co} = \frac{2\Omega_0}{{u_{\rm rms}}{k_{\rm f}}}\;. \label{equ:Co}\end{aligned}$$ We also quote the value of the Taylor number, $${{\rm Ta}}=\left(2\Omega_0 d^2/\nu\right)^2,$$ which is related to the Ekman number via ${\rm Ek}={{\rm Ta}}^{-1/2}$.
Boundary conditions
-------------------
The horizontal boundaries are periodic for all variables. Stress-free conditions are used for the velocity at the vertical boundaries. $$\begin{aligned}
U_{x,z}=U_{y,z}=U_z=0.\end{aligned}$$ Temperature is kept constant on the upper boundary and the temperature gradient $$\begin{aligned}
\frac{dT}{dz}=\frac{-g}{c_{\rm V}(\gamma-1)(m+1)},\end{aligned}$$ is held constant at the lower boundary, yielding a constant heat flux $F_0=-K {\partial}T/{\partial}z$ through the lower boundary.
[cccccccccccc]{} A1 & $256^2\times 128$ & $0.048$ & $0.020$ & $33$ & $8$ & $0.24$ & $2.0\cdot10^6$ & $15.5$ & $4.0\cdot10^8$ & $1.7\cdot10^{-5}$ & yes (A)\
A2 & $256^2\times 128$ & $0.018$ & $0.017$ & $13$ & $6$ & $0.48$ & $1.0\cdot10^6$ & $14.4$ & $5.6\cdot10^7$ & $1.7\cdot10^{-5}$ & no\
A3 & $256^2\times 128$ & $0.022$ & $0.019$ & $21$ & $7$ & $0.36$ & $1.3\cdot10^6$ & $12.3$ & $1.0\cdot10^8$ & $1.7\cdot10^{-5}$ & no\
A4 & $256^2\times 128$ & $(0.063)$ & $0.023$ & $37$ & $9$ & $0.24$ & $2.0\cdot10^6$ & $10.3$ & $2.3\cdot10^8$ & $1.7\cdot10^{-5}$ & yes (A)\
A5 & $256^2\times 128$ & $0.021$ & $0.020$ & $16$ & $9$ & $0.48$ & $1.0\cdot10^6$ & $7.9$ & $2.5\cdot10^7$ & $1.7\cdot10^{-5}$ & no\
A6 & $256^2\times 128$ & $0.024$ & $0.023$ & $24$ & $9$ & $0.36$ & $1.3\cdot10^6$ & $7.0$ & $4.4\cdot10^7$ & $1.7\cdot10^{-5}$ & no\
A7 & $256^2\times 128$ & $(0.093)$ & $0.026$ & $42$ & $10$ & $0.24$ & $2.0\cdot10^6$ & $6.1$ & $1.0\cdot10^8$ & $1.7\cdot10^{-5}$ & yes (A+C)\
A8 & $256^2\times 128$ & $0.028$ & $0.027$ & $28$ & $11$ & $0.36$ & $1.3\cdot10^6$ & $3.6$ & $1.6\cdot10^7$ & $1.7\cdot10^{-5}$ & no\
A9 & $256^2\times 128$ & $0.082$ & $0.028$ & $45$ & $11$ & $0.24$ & $2.0\cdot10^6$ & $3.4$ & $3.6\cdot10^7$ & $1.7\cdot10^{-5}$ & yes (C)\
A9b & $256^2\times 128$ & $(0.070)$ & $(0.031)$ & $49$ & $12$ & $0.24$ & $2.0\cdot10^6$ & $2.1$ & $1.6\cdot10^7$ & $1.7\cdot10^{-5}$ & decay\
A10 & $256^2\times 128$ & $0.032$ & $0.033$ & $53$ & $13$ & $0.24$ & $2.0\cdot10^6$ & $1.0$ & $4.0\cdot10^6$ & $1.7\cdot10^{-5}$ & no\
A11 & $256^2\times 128$ & $0.038$ & $0.038$ & $61$ & $15$ & $0.24$ & $2.0\cdot10^6$ & 0 & $0$ & $1.7\cdot10^{-5}$ & no\
B1 & $256^2\times 128$ & $0.017$ & $0.016$ & $26$ & $13$ & $0.48$ & $4.0\cdot10^6$ & $9.7$ & $1.0\cdot10^8$ & $8.6\cdot10^{-6}$ & no\
B2 & $256^2\times 128$ & $(0.021)$ & $(0.017)$ & $37$ & $13$ & $0.36$ & $5.4\cdot10^6$ & $9.1$ & $1.8\cdot10^8$ & $8.6\cdot10^{-6}$ & yes (A+C)\
B3 & $256^2\times 128$ & $(0.034)$ & $(0.020)$ & $63$ & $15$ & $0.24$ & $8.0\cdot10^6$ & $8.0$ & $4.0\cdot10^8$ & $8.6\cdot10^{-6}$ & yes (A+C)\
C1 & $256^2\times 128$ & $0.011$ & $0.011$ & $17$ & $16$ & $0.96$ & $8.0\cdot10^6$ & $14.8$ & $1.0\cdot10^8$ & $4.3\cdot10^{-6}$ & no\
C2 & $256^2\times 128$ & $(0.014)$ & $(0.012)$ & $25$ & $18$ & $0.72$ & $1.1\cdot10^7$ & $13.6$ & $1.8\cdot10^8$ & $4.3\cdot10^{-6}$ & no\
C3 & $256^2\times 128$ & $(0.022)$ & $(0.014)$ & $44$ & $21$ & $0.48$ & $1.6\cdot10^7$ & $11.6$ & $4.0\cdot10^8$ & $4.3\cdot10^{-6}$ & yes (A)\
D1 & $256^2\times 128$ & $0.013$ & $0.013$ & $42$ & $51$ & $1.20$ & $4.0\cdot10^7$ & $7.2$ & $1.4\cdot10^8$ & $1.7\cdot10^{-6}$ & no\
D2 & $512^2\times 256$ & $(0.038)$ & $(0.013)$ & $101$ & $49$ & $0.48$ & $1.0\cdot10^8$ & $7.5$ & $9.0\cdot10^8$ & $1.7\cdot10^{-6}$ & yes (A+C) \[tab:runs\]
Results {#sec:results}
=======
We perform a number of numerical experiments in order to determine the conditions under which large-scale cyclones are excited. The basic input parameters and some key diagnostic outputs of the simulations are listed in Table \[tab:runs\]. We perform a few (Set A) or a single (Sets B, C, and D) progenitor run with a given Peclet number in each Set from which the rest of the runs are obtained by continuing from a saturated snapshot and changing the value of the kinematic viscosity $\nu$ in order to change ${{\rm Re}}$. The higher resolution run D2 was remeshed from a lower resolution case D1.
Excitation of large-scale vortices
----------------------------------
We perform several sets of runs where the Peclet number and input energy flux are constant, whereas the Reynolds and Coriolis numbers are varied. We are limited to exploring a small number of cases due to the slow growth of the vortices, see Table \[tab:runs\]. Typically the time needed for the saturation of the cyclones is several thousand convective turnover times (see Fig. \[purms.eps\]). Thus many of our runs were ran until the presence or the absence of the cyclones was apparent.
![Upper panel: total velocity from Runs A5 and A7. Lower panel: velocity components $\sqrt{u_x^2}$ (black), $\sqrt{u_y^2}$ (red), and $\sqrt{u_z^2}$ (blue) from Run A7. The jump at $t {u_{\rm rms}}{k_{\rm f}}\approx 500$ is due to a lowering of $\nu$ at this point.[]{data-label="purms.eps"}](purms.eps){width="\columnwidth"}
We find that a reliable diagnostic indicating the presence of large-scale vortices is to compare the rms-value of the total velocity, ${U_{\rm rms}}$, and the volume average of the quantity ${u_{\rm rms}}=\sqrt{3u_z^2}$. The latter neglects the horizontal velocity components which significantly grow when large-scale cyclones are present (see the lower panel of Fig. \[purms.eps\]). In the cyclone-free regime, irrespective of the rotation rate, we find that ${U_{\rm rms}}\approx{u_{\rm rms}}$ suggesting that the flow is only weakly anisotropic (see Table \[tab:runs\]). In the growth phase of the vortices one of the horizontal velocity components is always stronger, but the relative strength of the components changes as a function of time (see the lower panel of Fig. \[purms.eps\]). This undulation is related to quasi-periodic changes of the large-scale pattern of the flow, although their ultimate cause is not clear.
Another quantitative diagnostic is to monitor the power spectrum of the flow from a horizontal plane within the convection zone. A typical example is shown in Fig. \[pspec\_256x128b1\] where power spectra of the velocity from the middle of the convection zone at two different times from Run B3 are shown. The snapshot from $t {u_{\rm rms}}{k_{\rm f}}= 1830$ is the initial state for Run B3, taken from Run B1, showing no cyclones. The power spectrum shows a maximum at $k/k_1=7$, indicating that most of the energy is contained in structures having a size typical of the convective eddies. However, as the run is continued further, a large-scale contribution due to the appearance of the vortices, peaking at $k/k_1=1$ grows, and ultimately dominates the power spectrum. We note that this run was not ran until saturation so the peak at $k/k_1=1$ is likely to be even higher in the final state. The presence of the vortices is also clear by visual inspection of the flow. A typical example is shown in Fig. \[psnap\_512x256a2\], where the vertical velocity component, $u_z$, is shown from the periphery of the domain for Run D2.
![Power spectra of velocity from early (dashed line) and late (solid) times from Run B3.[]{data-label="pspec_256x128b1"}](pspec_256x128b1.eps){width="\columnwidth"}
The data in Table \[tab:runs\] suggests that large-scale vorticity is excited provided that the Reynolds number exceeds a critical value, ${{\rm Re}}_{\rm c}$. For ${{\rm Pe}}\approx 10$ (Set A) we find that ${{\rm Re}}_{\rm c}$ is around 30, although the sparse coverage of the parameter range does not allow a very precise estimate to be made. We find a similar value for ${{\rm Re}}_{\rm c}$ in Sets B and C, whereas for ${{\rm Pe}}\approx50$ in Set D, the critical Reynolds number is greater than 42. In Set C, Runs C1 and C2 were started from a snapshot of Run C3 at a time when vortices were already clearly developing. In both cases we find that the cyclones decay, suggesting that their presence is not strongly dependent on the history of the run.
The critical Coriolis number in Set A is somewhere between 2.1 and 3.4. Again a very precise determination cannot be made, but continuing from a saturated snapshot of Run A9 with a somewhat lower rotation rate indicates that the vortices decay (Run A9b). We have limited the present study to the North pole ($\theta=0$), but vortices are also excited at least down to latitude $\theta=45\degr$ in the study of [@Chan07].
![Vertical velocity component $U_z$ at the periphery of the box from Run D2. See also http://www.helsinki.fi/$\sim$kapyla/movies.html. The top and bottom panels show slices near the top and bottom of the convectively unstable layer, respectively.[]{data-label="psnap_512x256a2"}](512x256a_Ur.ps){width="\columnwidth"}
{width="80.00000%"}
Thermal properties of the cyclones
----------------------------------
In order to study the possible observable and other effects of the vortices, we ran a few simulations in Set A (Runs A1, A4, A7, A9, A10, and A11) to full saturation. Figures \[puu\_slices\] and \[pT\_slices\] show the vertical velocity and temperature in the saturated regime from the six runs listed above. In the non- and slowly rotating cases (the two rightmost panels in the lower rows of Figs \[puu\_slices\] and \[pT\_slices\]), convection shows a typical cellular pattern. Vorticity is generated at small scales at the vertices of the convection cells, but no large-scale pattern arises. We note that long-lived large-scale circulation can also emerge in non-rotating convection [e.g. @Bukai09]. However, such structures are not likely to be of relevance in rapidly rotating stars.
When the rotation is increased to ${{\rm Co}}\approx3.3$, a cyclonic, i.e.rotating in the same sense as the overall rotation of the star, vortex appears (the lower left panels of Figs. \[puu\_slices\] and \[pT\_slices\]). Vertical motions are suppressed within the vortex and it appears as a cool spot in the temperature slice. Increasing rotation further to ${{\rm Co}}\approx6$, also an anti-cyclonic, i.e. rotating against the overall fluid rotation, warm vortex appears (the rightmost upper panels of Figs. \[puu\_slices\] and \[pT\_slices\]). In Run A7 the two vortices coexist for thousands of convective turnover times. In the most rapidly rotating cases A1 and A4 (the two leftmost panels in the upper rows of Figs. \[puu\_slices\] and \[pT\_slices\]) a single anti-cyclonic vortex persists in the saturated regime. A similar behaviour as a function of rotation was found by [@Chan07] from large-eddy simulations. The anti-cyclonic vortices show vigorous convection whereas in the surrounding regions convection appears suppressed. Due to the enhanced energy transport by convection, the anti-cyclones appear as warmer structures than their surroundings in the temperature slices.
Figure \[pgeos\] shows that in Runs A9 and A1 the flow is in geostrophic balance, i.e. that the flow follows the isocontours of pressure for both types of vortices. The cyclone in Run A9 shows as a low pressure area, similarly to the cyclones in the atmosphere of the Earth, whereas the anti-cyclone in Run A1 coincides with a high pressure region. A weaker high pressure region is present also in Run A9. It is not clear whether this kind of single or two spot configuration lasts if the domain is larger in the horizontal directions, or whether a greater number of spots appear. We find that the temperature contrast between the spot and the surrounding medium is of the order of five per cent (Fig. \[ptempc\]) for both types of vortices. Although the relative temperature contrast between the vortex and the surrounding vortex-free convection seems to be a robust feature in the simulations, we must remain cautious when comparing the results with observations. This is due to the rather primitive nature of the simulations that lack realistic radiation transport. Convection in our model is also fairly inefficient by design, only 20 percent of the total flux being carried by it.
Dynamo considerations and discussion
------------------------------------
Figure \[plot\_heli\] shows the horizontally averaged kinetic helicity, $\overline{\bm\omega \cdot{\bm u}}$, where $\bm{\omega}=\bm\nabla\times{\bm u}$, from Runs A9 and A1 from the initial, purely convective cyclone-free, and final fully saturated stages of the simulations. The data is averaged over a period of roughly 60 convective turnover times in each case. We find that in Run A9, where a cool cyclonic vortex appears, there is almost no change in the kinetic helicity between the initial and final stages of the simulation. In this run convection, and thus vertical motions, are largely suppressed within the vortex (see Fig. \[puu\_slices\]). Furthermore, the dominant contribution to the vorticity due to the cyclone arises via the vertical component $\omega_z={\partial}_x u_y -{\partial}_y u_x$, which is positive for a cyclonic vortex. These two effects seem to compensate each other and the helicity within the cyclone is not greatly enhanced or depressed with respect to the surroundings. This would indicate that the influence of the cyclonic vortices on the magnetic field amplification would be minor, as the helicity remains unaltered. On the other hand, the strong horizontal motions connected to the cyclone might be able to amplify the field by advecting the field lines.
{width="80.00000%"}
![Pressure (colors) and horizontal flows from the middle of the convection zone in Runs A9 (left panel) and A1 (right panel).[]{data-label="pgeos"}](pgeos.ps){width="\columnwidth"}
In Run A1, on the other hand, a more pronounced effect is seen, and the helicity is decreased up to a factor of two in the saturated stage (see the right panel of Fig. \[plot\_heli\]). This change is brought about by the different handedness of the vorticity in the anti-cyclone and by the vigorous convection within it (see the upper row of Fig. \[puu\_slices\]). The combination of these produces significantly greater helicity in the anti-cyclones, but a predominantly different sign than the surroundings and leads to an overall decrease noted in Fig. \[plot\_heli\]. The decreased amount of helicity would indicate weaker amplification of the magnetic field by anti-cyclones compared to their surroundings. Again, the strong horizontal motions might counteract by amplifying the field by advection.
The simulations presented here were performed with a setup identical to that used in [@KKB09b hereafter KKB09] to study large-scale dynamo (LSD) action in rotating convection. In KKB09 the generation of large-scale magnetic fields, given that the Coriolis and magnetic Reynolds numbers exceeded critical values, were reported. The critical Coriolis number for LSD action was found to be roughly four, which is close to the critical value for the cyclones to emerge.
The relation of the two phenomena is an interesting question, that can be only partially answered by the existing magnetohydrodynamic runs. This is because the fluid Reynolds number in the runs of KKB09 was in most cases lower than ${{\rm Re}}_{\rm c}$ required for the vortices to appear. Only two runs (A10 and D1 of KKB09) are clearly in the parameter regime exceeding the critical values found here, and another four runs (A5, A6, B5, and C1 of KKB09) where the parameters were close to marginal. The Reynolds and Coriolis numbers for these runs were calculated from the saturated state of the dynamo which in all cases lowers the turbulent velocities somewhat, decreasing the Reynolds and increasing the Coriolis numbers correspondingly. Furthermore, a different definition of the Reynolds number was used by KKB09 as in the present study. A reanalysis of the data of KKB09 suggests that early stages of cyclone formation are in progress in all of the runs listed above. However, the magnetic field grows on a significantly shorter timescale than the cyclones, and the magnetic field saturates already before thousand convective turnover times. None of the runs was continued much further than twice that, making it impossible to decide in favor or against the maintenance of vortices based on these runs.
Nonetheless, indications of growing cyclones appear in the kinematic regime, i.e. when the magnetic field is weak in comparison to kinetic energy of the turbulence, but they are far less clear, or even absent, when the magnetic field saturates. This raises two related questions: firstly, are the vortices responsible for the emergence of the large-scale magnetic fields, and secondly, can the vortices coexist in the regime where strong magnetic fields are present? The current data suggests that the presence of the vortices is not essential for the large-scale magnetic fields which persist throughout the saturated state, whereas the vortices remain less prominent or suppressed. This fact is related to the second issue. As noted above, the simulations of KKB09 are too short for the vortices to fully saturate. Thus, we cannot conclusively say whether the lack of the vortices in the dynamo regime is due to the magnetic field simply reducing the Reynolds number below the critical value, or a direct influence of the Lorentz force on the growing vortices. We will address the questions related to magnetic fields and dynamo action in more detail in a forthcoming paper.
![Temperature as a function of $x$ from a quiescent (solid lines) and cyclonic (dashed) regions for Runs A9 (left panel) and A1 (right panel). The positions of the cuts are indicated in the leftmost panels of Fig. \[pT\_slices\] with corresponding linestyles. The normalization factor $\overline{T}$ is the horizontal average of the temperature.[]{data-label="ptempc"}](ptempc.eps){width="\columnwidth"}
![Horizontally averaged kinetic helicity $\overline{\omega\cdot
u}$ as a function of $z$ from a quiescent (solid lines) and cyclonic (dashed) states for Runs A9 (left panel) and A1 (right panel). The vertical dotted lines at $z=0$ and $z=d$ indicate the bottom and top of the convectively unstable layer, respectively.[]{data-label="plot_heli"}](plot_heli.eps){width="\columnwidth"}
Observational implications
--------------------------
If large-scale cyclones such as those found in the present study occur in real stars, they will cause observational signatures on the stellar surface due to their lower or higher temperature. The temperature contrasts seen in the surface maps derived by Doppler imaging are somewhat stronger than the value of roughly five percent found in this study; for instance, on the surface of the active RS CVn binary II Peg, analysed by @Marjaana11 and [@Thomas11], the coolest spot temperatures, depending on the season, are 10-20 per cent below the mean surface temperature. Similar spot temperatures have also been obtained by analysing molecular absorption bands, but cooler stars seem to have a lower spot contrast [@Oneal98]. Taken that the numerical model is quite simple, for instance in the sense that the transport of energy by convection is underestimated, this discrepancy is not overwhelmingly large. Interestingly, Doppler images commonly also show hot surface features [cf. @Heidi07; @Marjaana11; @Thomas11]. These may be artefacts of the Doppler imaging procedure, but it is not ruled out that they could arise from the anti-cyclonic vortices seen in the present study.
It is obviously very hard to explain the active longitudes and their drift based on the vortex-instability scenario; we believe that a large-scale dynamo process is responsible for these basic features, as commonly believed [e.g. @KR80; @MBBT95; @tuominen2002starspot]. Nevertheless, it is possible that the vortex-instability contributes to the formation of starspots, and may interfere with the dynamo-instability, especially during the epochs of lower magnetic activity of the stellar cycle. Although it is very hard to predict the implications of the vortices in the magnetohydrodynamic regime, it would appear natural that spots, either cool or warm, generated by a hydrodynamic vortex-instability, could also contribute to the apparent decorrelation of magnetic field from the temperature structures.
The influence of the cyclones and anticyclones on the net helicity, important for the amplification of the magnetic field, is either close to zero (cyclones) or to decrease the net helicity (anti-cyclones). This would imply that the magnetic field amplification would be equally or even more difficult in the regions of the vortices; this picture, however, may be complicated by the presence of strong horizontal motions present in these structures, that might amplify the magnetic field simply by their capability for advecting the field lines.
Conclusions {#sec:conclusions}
===========
We report the formation of large-scale vortices in rapidly rotating turbulent convection in local f-plane simulations. The vortices appear provided the Reynolds and Coriolis numbers exceed critical values. Near the critical Coriolis number, the vortices are cyclonic and cool in comparison to the surrounding atmosphere, whereas for faster rotation warm anti-cyclonic vortices appear [see also @Chan07]. The relative temperature difference between the vortex and its surroundings is of the order of five per cent in all cases. This is of the order of the contrast deduced indirectly from photometric and spectroscopic observations of late-type stars. In our simulations the typical size of the vortices is comparable to the depth of the convectively unstable layer. However, we have not studied how the size of the structures depends e.g. on the depth of the convection zone.
We propose that the vortices studied here can be present in the atmospheres of rapidly rotating late-type stars, thus contributing to rotationally modulated variations in the brightness and spectrum of the star. Such features have generally been interpreted to be caused by magnetic spots, reminiscent to sunspots. However, our results suggest that the turbulent convection and rapid rotation of these stars can generate large-scale temperature anomalies in their atmospheres via a purely hydrodynamical process. Similar vortex-structures are observed in the atmospheres of Jupiter and Saturn. Although their definitive explanation is still debated, it is possible that they are related to rapidly rotating thermal convection in the atmosphere.
However, several issues remain to be sorted out before the reality of cyclones and anti-cyclones in the surface layers of stars can be established. The current model is highly simplified and neglects the effects of sphericity and magnetic fields. In spherical geometry more realistic large-scale flows can occur which might lead to other hydrodynamical instabilities. However, current rapidly rotating simulations in spherical coordinates have not shown evidence of large-scale vortices [e.g. @BBBMT08; @KKBMT10; @KMGBC11], although non-axsisymmetric features are seen near the equator [@BBBMT08]. It is possible that the lack of large-scale vortices in these simulations is related either to the lack of spatial resolution or too short integration time.
Magnetic fields, on the other hand, are ubiquitous in stars with convection zones. Furthermore, on the Sun they form strong flux concentrations, i.e. sunspots. At the moment, direct simulations cannot self-consistently produce sunspot-like structures in local geometry [e.g. @KBKMR11]. However, the magnetic fields in global simulations are also very different from the high-latitude active longitudes deduced from observations, namely showing more axisymmetric fields residing also near the equator [e.g. @KKBMT10; @BMBBT11]. The apparently poor correlation between magnetic fields and temperature anomalies in surface maps based on Doppler imaging also suggests that an alternative mechanism might be involved. The presence of large-scale high-latitude vortices presents such an alternative.
Currently it is not clear what happens to the vortices when magnetic fields are present. Our previous dynamo simulations in the same parameter regime [@KKB09b] did not show clear signs of vortices in the saturated regime of the dynamo although this might be explained by the too short integration time. Addressing this issue, however, is not within the scope of the present paper and we will revisit it in a future publication.
The simulations were performed using the supercomputers hosted by CSC – IT Center for Science Ltd. in Espoo, Finland, who are administered by the Finnish Ministry of Education. Financial support from the Academy of Finland grants No. 136189, 140970 (PJK) and 218159, 141017 (MJM), and the ‘Active Suns’ research project at University of Helsinki (TH) is acknowledged. The authors acknowledge the hospitality of NORDITA during their visits.
[^1]: http://code.google.com/p/pencil-code/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the evolution of hypersurfaces on the unit sphere $\mathbb{S}^{n+1}$ by their mean curvature. We prove a differential Harnack inequality for any weakly convex solution to the mean curvature flow. As an application, by applying an Aleksandrov reflection argument, we classify convex, ancient solutions of the mean curvature flow on the sphere.'
address:
- 'Department of Mathematics, University of California San Diego, La Jolla, USA'
- 'Institut für Diskrete Mathematik und Geometrie, Technische Universität Wien, Wiedner Hauptstr. 8–10, 1040 Wien, Austria'
author:
- Paul Bryan
- 'Mohammad N. Ivaki'
title: Harnack estimate for mean curvature flow on the sphere
---
We study convex hypersurfaces $M_0$, $n\ge2$, without boundary, which are smoothly embedded in $\mathbb{S}^{n+1}.$ Let $F_0:M^n\to \mathbb{S}^{n+1}$ be a smooth embedding of $M_0$. The mean curvature flow with initial data $F_0$ is a family of hypersurfaces given by embeddings $F:M^n\times[0,T)\to \mathbb{S}^{n+1}$, $~ F(\cdot,0)=F_0(\cdot)$ that moves in the direction of the unit normal vector $\nu$ with speed equal to the mean curvature $H$, the trace of the second fundamental form $A(V,V)$ over the tangent vectors $V$, e.q., $$\partial_tF(\cdot,t)=-H(\cdot,t)\nu(\cdot,t).$$ Here, $H(p,t)$ is the mean curvature of the hypersurface $M_t:=F(M^n,t)$ at the point $F(p,t)$ where the unit outer normal vector is $\nu(p,t).$ Let $D$ denote the Levi-Civita connection of $\mathbb{S}^{n+1}.$
Suppose $M_t$ is a smooth, weakly convex solution of the mean curvature flow on the time interval $[0,T)$. For $t>0$ we have $$\partial_tH+\frac{H}{2t}+2D_VH+A(V,V)-nH\geq 0$$ for all tangent vectors $V$.
Comparing this with Hamilton’s Harnack inequality [@Hamilton; @95] for the mean curvature flow in $\mathbb{R}^{n+1}$, here we gained a bonus term $-nH$ through the ambient space $\mathbb{S}^{n+1}.$ This allows us, by employing a parabolic Aleksandrov reflection argument first developed in [@Chow; @97; @Chow-Gul; @01; @Chow-Gul; @96], a rather convenient classification of ancient, smooth, weakly convex solutions of the mean curvature flow.
The only ancient, weakly convex, embedded solutions of the mean curvature flow on the sphere are equators and shrinking geodesic spheres.
Let us note that such a classification has been obtained already in [@Hu-Sin; @2014 Theorem 6.1], with a relatively short proof applying the maximum principle to the quantity $(\|A\|^2 - \tfrac{1}{n}H^2)/H^2$. As usual for such arguments, the Codazzi equation is employed; therefore, the result pertains to $n\geq 2$. The first author and Louie [@Br-Lou] used a completely different argument to obtain the classification for $n=1$. First, a Harnack inequality is obtained, which ensures that the backwards limit as $t \to -\infty$ is an equator. Then the Aleksandrov reflection argument is employed to show maximal reflection symmetry, immediately resulting in the classification. Our work here extends the techniques of [@Br-Lou] from $n=1$ to $n\geq 2$ formally following the same procedure of obtaining the Harnack inequality and then using the Aleksandrov reflection to show the maximal reflection symmetry. We believe, that unlike the methods in [@Hu-Sin; @2014], our argument is more broadly applicable: once the Harnack inequality is obtained for a curvature flow, the remainder of the argument holds with very little modification. This theme will be explored for a range of curvature flows similar to those studied in [@Andrews; @94] in a forthcoming paper. In [@Andrews; @94], the Gauss map parametrization and support function are used to great effect, considerably simplifying the computations. A genuine difficulty on the sphere is obtaining a similar useful parametrization under which we can perform the calculations.
Before moving on to the details, a few words about the broader context are in order. It is well known that ancient solutions arise as singularity models for curvature flows (e.g. [@Hu-Sin; @1999]); therefore, their study is of great importance. On the unit sphere, as opposed to the Euclidean case studied in [@Hu-Sin; @2014], where classification can be obtained only with additional assumptions such as pinching, we find a complete classification and strong rigidity. The difference is the compactness and positive curvature of the sphere. We might conjecture that for the mean curvature flow in a positively curved, compact background, there exists at most one non-trivial ancient solution emanating from each closed minimal hypersurface. However, in general, it may be too much to hope that such ancient solutions exist.
Classifications have been obtained for the curve shortening flow in the plane [@Da-Ham-Sesu; @2010] and the Ricci flow of surfaces [@Da-Ham-Sesu; @2012]. Both examples include self-similar shrinkers (round circles for the former and constant curvature spheres for the latter) and a unique Type II ancient solution with curvature blowing up faster than $|t|$ as $t\to-\infty$ (the Angenent oval \[paper clip\] in the former case and the Rosenau in the latter). Our results here rule out the possibility of such Type II singularities and show that *in any dimension*, the only weakly convex, ancient solutions are the “self-similar" shrinkers: here, self-similar means conformally equivalent to the self-similar family of shrinking spheres in Euclidean space after realizing the sphere as (locally) conformally equivalent to Euclidean space.
In higher dimensions, classification results for ancient solutions in Euclidean space are less complete and a wider range of possibilities may occur. As mentioned above, the results of [@Hu-Sin; @2014] require additional pinching assumptions to deduce that such ancient solutions are shrinking spheres. Wang [@Wang] studied translating ancient solutions and discovered the existence of non-rotationally symmetric translators (though they blow down to spheres or cylinders). Furthermore, there are solutions asymptotic to “ovals" in that the center looks like a cylinder, but the ends look like “bowls" [@Ang; @Has-Her; @Whi]. These solutions have similarities to the Angenent oval in the plane.
In the sphere, we may consider an ancient solution moving by a rotation to be a translating solution, but we quickly realize that no such convex solutions can exist because convexity and the containment principle force the flow to lie in a fixed hemisphere at all times. Our results also show that no non-self-similar ancient solutions exist. The phenomena of becoming more and more oval as $t\to -\infty$ for convex solutions appears to be ruled out by the compactness of the sphere; such a solution would have to touch an equator at the “ends" and “fatten" out in the middle, which forces it to touch the equator in the middle and thus everywhere (by convexity).
Lastly, the connection between Harnack inequalities, ancient solutions and solitons deserves mention. In Euclidean space, the classic work [@Hamilton; @95] espouses the philosophy that Harnack inequality should be an equality on solitons. We have already observed that unlike Euclidean space, no translating convex solitons exist in the sphere. The equivalent of homothetic solitons may be taken to be those solutions of the mean curvature flow following the integral curves of the conformal “position” vector field $\sin(d)\nabla d$, where $d$ denotes the distance to a fixed point (the point of collapse). This vector field generates a one-parameter subgroup of conformal transformations, under which the mean curvature flow is not invariant, though the family of shrinking geodesic spheres is such a soliton and hence these conformal solitons are of great interest. The lack of conformal invariance suggests that conformal solitons do not satisfy equality in the Harnack inequality, and this is indeed the case for the shrinking spheres. Interested readers should consult [@hun-nor; @12; @smo; @97; @smo; @01] for more details on solitons interpreted as the flow along integral curves of a vector field, particularly [@arr-sun; @13] for details on conformal solitons.
Acknowledgement {#acknowledgement .unnumbered}
===============
The first author would like to thank Bennett Chow for helpful discussions and encouragement. He would also like to thank the second author and the Institut für Diskrete Mathematik und Geometrie, TU Wien for hosting a very enjoyable and productive visit to Vienna. The coffee houses of Vienna most certainly experienced a downturn in business after his departure. The work of the second author was supported by Austrian Science Fund (FWF) Project M1716-N25.
Preliminaries
=============
Let $g=\{g_{ij}\}$, $A=\{h_{ij}\},$ and $Rm_{ijkl}$ denote, in order, the induced metric, the second fundamental form, and the curvature tensor of $M^n$. The mean curvature of $M^n$ is the trace of the second fundamental form with respect to $g$, $H=g^{ij}h_{ij}.$ Let $\overline{g}$ and $\overline{Rm}_{\alpha\beta\gamma\theta}$ denote, respectively, the metric and the curvature tensor of $\mathbb{S}^{n+1}$. Greek indices run through $\{0,\cdots,n\}$ and Latin indices belong to the set $\{1,\cdots,n\}.$
Write $\nu$ for the outer unit normal to $M_t.$ For a fixed time, we choose a local orthonormal frame $\{\partial_0=\nu,\cdots,\partial_i=\frac{\partial F(\cdot,t)}{\partial x_i},\cdots,\partial_n\}$ in $\mathbb{S}^{n+1}.$ We use the following standard notation $$h_i^j=g^{mj}h_{im}$$ $$(h^2)_i^j=g^{mj}g^{rs}h_{ir}h_{sm}$$ $$|A|^2=g^{ij}g^{kl}h_{ik}h_{lj}=h_{ij}h^{ij}$$ $$C=g^{ij}g^{kl}g^{mn}h_{ik}h_{lm}h_{nj}=h_i^kh_k^lh_l^i.$$ Here, $\{g^{ij}\}$ is the inverse matrix of $\{g_{ij}\}.$
The relations between $A$, $Rm$, and $\overline{Rm}$ are given by the Gau[ß]{} and Codazzi equations: $$Rm_{ijkl}=\overline{Rm}_{ijkl}+h_{ik}h_{jl}-h_{il}h_{jk}$$ $$\nabla_ih_{jk}=\nabla_{k}h_{ij}.$$ Moreover, $\nabla_i$ and $\Delta$ commute as follows $$(\nabla_i\Delta-\Delta\nabla_i)f=-Rc_i^j\nabla_jf$$ for all smooth functions on $M^n.$ Therefore, in view of $\bar{R}_{\alpha\beta\gamma\theta}=\lambda (\overline{g}_{\alpha\gamma}\overline{g}_{\beta\theta}-\overline{g}_{\alpha\theta}\overline{g}_{\beta\gamma})$, $\lambda =1$, and the Gau[ß]{} equation we have $$\label{eq: commute}
(\nabla_i\Delta-\Delta\nabla_i)H=((h^2)_i^m-Hh_i^m-(n-1)\delta_i^m\lambda )\nabla_mH.$$
Evolution equations
===================
In this section, we assume that $M_t$ is a strictly convex solution of the mean curvature flow.
\[lem: lem3\] The following evolution equations hold under the mean curvature flow:
1. $\partial_tg_{ij}=-2Hh_{ij}$
2. $\partial_tg^{ij}=2Hg^{im}g^{jn}h_{ij}=2Hh^{ij}$
3. $\partial_t h_i^j=\nabla^j\nabla_iH+H(h^2)_i^j+\lambda H\delta_i^j$
4. $\partial_t h_i^j=\Delta h_i^j+|A|^2h_i^j+\lambda \{2H\delta_i^j-nh_i^j\}$
5. $\partial_t h_{ij}=\Delta h_{ij}+|A|^2h_{ij}-2H(h^2)_{ij}+\lambda \{2Hg_{ij}-nh_{ij}\}$
6. $\partial_t H=\Delta H+H|A|^2+n\lambda H$
7. $(\partial_t\Delta-\Delta\partial_t)H=2Hh^{ij}\nabla_i\nabla_jH+2h^{ij}\nabla_iH\nabla_jH.$
The computations for $(1-6)$ are straightforward (see for example [@Huisken; @87]). To obtain $(7)$, we note that $$\begin{aligned}
(\partial_t\Delta-\Delta\partial_t)H=\left(\partial_i\partial_jH-\Gamma_{ij}^k\partial_kH\right)\partial_tg^{ij}-g^{ij}\partial_kH\partial_t\Gamma_{ij}^k,\end{aligned}$$ where $\Gamma_{ij}^k=g^{kl}(\partial_ig_{jl}+\partial_jg_{il}-\partial_lg_{ij}).$ Since $\partial_t\Gamma_{ij}^k$ is tensorial, using $(1)$ and $ (2)$ we may carry out our calculations in a normal frame to prove the claim.
\[lem: lem1\] Under the mean curvature flow we have $$\begin{aligned}
\partial_t(\partial_tH)=&
\Delta \partial_tH+4Hh^{ij}\nabla_i\nabla_jH+2h^{ij}\nabla_iH\nabla_jH\\
&+(|A|^2+n\lambda )(\partial_t H)+2H^2C+2\lambda H^3,\end{aligned}$$ and $$\begin{aligned}
\nabla_i\partial_tH
=\Delta\nabla_iH+\nabla_i(|A|^2H)+((h^2)_i^m-Hh_i^m)\nabla_mH+\lambda \nabla_iH.\end{aligned}$$
Using $(3), (6)$ and $(7)$ in Lemma \[lem: lem1\], we calculate $$\begin{aligned}
\partial_t(\partial_tH)=&\partial_t(\Delta H+H|A|^2+n\lambda H)\\
=&\Delta \partial_tH+2Hh^{ij}\nabla_i\nabla_jH+2h^{ij}\nabla_iH\nabla_jH\\
&+ (|A|^2+n\lambda )(\partial_t H)+H(\partial_t|A|^2)\\
=&\Delta \partial_tH+2Hh^{ij}\nabla_i\nabla_jH+2h^{ij}\nabla_iH\nabla_jH\\
&+ (|A|^2+n\lambda )(\partial_t H)+H(2h^{ij}\nabla_i\nabla_jH+2HC+2\lambda H^2)\\
=&\Delta \partial_tH+4Hh^{ij}\nabla_i\nabla_jH+2h^{ij}\nabla_iH\nabla_jH\\
&+(|A|^2+n\lambda )(\partial_t H)+2H^2C+2\lambda H^3.\end{aligned}$$ To obtain the second evolution equation, we use identity (\[eq: commute\]) and part $(6)$ of Lemma \[lem: lem1\]: $$\begin{aligned}
\nabla_i\partial_tH=&\nabla_i(\Delta H+|A|^2H+n\lambda H)\\
=&\Delta\nabla_iH+\nabla_i(|A|^2H)+n\lambda \nabla_iH\\
&+((h^2)_i^m-Hh_i^m-(n-1)\delta_i^m\lambda )\nabla_mH\\
=&\Delta\nabla_iH+\nabla_i(|A|^2H)+((h^2)_i^m-Hh_i^m)\nabla_mH+\lambda \nabla_iH.\end{aligned}$$
In the sequel, $\{b^{ij}\}$ denotes the inverse of the second fundamental form.
\[lem: lem 3\] $$\begin{aligned}
\partial_t( b^{ij}\nabla_iH\nabla_jH)\leq&-b^{im}b^{jn}\Delta h_{mn}\nabla_iH\nabla_jH+2b^{ij}\nabla_jH\Delta\nabla_iH\\
&+2h^{ij}\nabla_iH\nabla_jH
+|A|^2b^{ij}\nabla_iH\nabla_jH+2Hb^{ij}\nabla_i|A|^2\nabla_jH\\
&+n\lambda b^{ij}\nabla_iH\nabla_jH.\end{aligned}$$
Observe $$\begin{aligned}
-\lambda b^{im}b^{jn}\{2Hg_{mn}-nh_{mn}\}\nabla_iH\nabla_jH+2\lambda b^{ij}\nabla_iH\nabla_jH\leq n\lambda b^{ij}\nabla_iH\nabla_jH\end{aligned}$$ and $$\begin{aligned}
\partial_t b^{ij}=&-b^{im}b^{jn}\partial_th_{mn}\\
=&-b^{im}b^{jn}(\Delta h_{mn}+|A|^2h_{mn}-2H(h^2)_{mn})\\
&-\lambda b^{im}b^{jn}\{2Hg_{mn}-nh_{mn}\}\\
=&-b^{im}b^{jn}\Delta h_{mn}-|A|^2b^{ij}+2Hg^{ij}\\
&-\lambda b^{im}b^{jn}\{2Hg_{mn}-nh_{mn}\}.\end{aligned}$$ Therefore, using the second part of Lemma \[lem: lem1\], we obtain $$\begin{aligned}
\partial_t( b^{ij}\nabla_iH\nabla_jH)=&-b^{im}b^{jn}(\Delta h_{mn}+|A|^2h_{mn}-2H(h^2)_{mn})\nabla_iH\nabla_jH\\
&-\lambda b^{im}b^{jn}\{2Hg_{mn}-nh_{mn}\}\nabla_iH\nabla_jH\\
&+2b^{ij}(\Delta\nabla_iH+\nabla_i(|A|^2H)+((h^2)_i^m-Hh_i^m)\nabla_mH+\lambda \nabla_iH)\nabla_jH\\
\leq& -b^{im}b^{jn}\Delta h_{mn}\nabla_iH\nabla_jH+2b^{ij}\nabla_jH\Delta\nabla_iH\\
&+2b^{im}b^{jn}H(h^2)_{mn}\nabla_iH\nabla_jH\\
&+|A|^2b^{ij}\nabla_iH\nabla_jH+2Hb^{ij}\nabla_i|A|^2\nabla_jH\\
&+2b^{ij}((h^2)_i^m-Hh_i^m)\nabla_mH\nabla_jH\\
&+n\lambda b^{ij}\nabla_iH\nabla_jH\\
=&-b^{im}b^{jn}\Delta h_{mn}\nabla_iH\nabla_jH+2b^{ij}\nabla_jH\Delta\nabla_iH\\
&+2h^{ij}\nabla_iH\nabla_jH
+|A|^2b^{ij}\nabla_iH\nabla_jH+2Hb^{ij}\nabla_i|A|^2\nabla_jH\\
&+n\lambda b^{ij}\nabla_iH\nabla_jH.\end{aligned}$$
Using the identities $$\nabla_mb^{ij}=-b^{ik}b^{jl}\nabla_mh_{kl}$$ and $$\begin{aligned}
\Delta b^{ij}=&-b^{im}b^{jn}\Delta h_{mn}\\
&+\{b^{ir}b^{ks}b^{jl}+b^{ik}b^{jr}b^{ls}\}g^{pq}\nabla_ph_{rs}\nabla_qh_{kl},\end{aligned}$$ we compute $$\begin{aligned}
\Delta( b^{ij}\nabla_iH\nabla_jH)=&-b^{im}b^{jn}\nabla_iH\nabla_jH\Delta h_{mn}+2b^{ij}\nabla_jH\Delta\nabla_iH\\\
&+2\{b^{ir}b^{ks}b^{jl}\}\nabla_iH\nabla_jHg^{pq}\nabla_ph_{rs}\nabla_qh_{kl}\\
&-4g^{mn}b^{ik}b^{jl}\nabla_mh_{kl}\nabla_n\nabla_iH\nabla_jH+2b^{ij}g^{mn}(\nabla_m\nabla_iH\nabla_n\nabla_jH).\end{aligned}$$ Thus we have proved:
\[lem: lem2\] $$\begin{aligned}
\partial_t(b^{ij}\nabla_iH\nabla_jH)\leq &\Delta (b^{ij}\nabla_iH\nabla_jH)\\
&-2\{b^{ir}b^{ks}b^{jl}\}\nabla_iH\nabla_jHg^{pq}\nabla_ph_{rs}\nabla_qh_{kl}\\
&+4g^{mn}b^{ik}b^{jl}\nabla_mh_{kl}\nabla_n\nabla_iH\nabla_jH-2b^{ij}g^{mn}(\nabla_m\nabla_iH\nabla_n\nabla_jH)\\
&+2h^{ij}\nabla_iH\nabla_jH+|A|^2b^{ij}\nabla_iH\nabla_jH+2Hb^{ij}\nabla_i|A|^2\nabla_jH\\
&+n\lambda b^{ij}\nabla_iH\nabla_jH.\end{aligned}$$
\[lem: lema6\] Define $\Theta:=\partial_tH-b^{ij}\nabla_iH\nabla_jH$. Then $$\begin{aligned}
\partial_t\Theta&\geq\Delta \Theta+ \frac{2(\Theta-n\lambda H)^2}{H}+(|A|^2+n\lambda )\Theta\\
&+2\{g^{mq}b^{np}-\frac{g^{mn}g^{pq}}{H}\}\eta_{mn}\eta_{pq}+2\lambda H^3,\end{aligned}$$ where $$\eta_{mn}=\nabla_m\nabla_nH+H(h^2)_{mn}-b^{rs}\nabla_rH\nabla_sh_{mn}.$$
Note that $$\begin{aligned}
2\frac{(\Theta-n\lambda H)^2}{H}=&2\frac{(\Delta H)^2}{H}+4|A|^2\Delta H-4\frac{\Delta H}{H}b^{ij}\nabla_iH\nabla_jH\\
&+2|A|^4H-4|A|^2b^{ij}\nabla_iH\nabla_jH\\
&+\frac{2}{H}\left(b^{ij}\nabla_iH\nabla_jH\right)^2\end{aligned}$$ and $$\begin{aligned}
2\{g^{mq}b^{np}&-\frac{g^{mn}g^{pq}}{H}\}\eta_{mn}\eta_{pq}\\
=&2g^{mq}b^{np}\nabla_m\nabla_nH\nabla_p\nabla_qH-2\frac{(\Delta H)^2}{H}\\
&+4Hh^{ij}\nabla_i\nabla_jH-4|A|^2\Delta H\\
&-4g^{mq}b^{np}b^{rs}\nabla_m\nabla_nH\nabla_rH\nabla_sh_{pq}+4\frac{\Delta H}{H}b^{ij}\nabla_iH\nabla_jH\\
&+2H^2C-2|A|^4H\\
&-2Hb^{ij}\nabla_iH\nabla_j|A|^2+4|A|^2b^{ij}\nabla_iH\nabla_jH\\
&+2\{b^{ir}b^{ks}b^{jl}\}\nabla_iH\nabla_jHg^{pq}\nabla_ph_{rs}\nabla_qh_{kl}-\frac{2}{H}\left(b^{ij}\nabla_iH\nabla_jH\right)^2.\end{aligned}$$ Now the claim follows from adding up these last two identities and considering Lemmas \[lem: lem1\] and \[lem: lem2\].
Harnack estimate and Backwards Convergence
==========================================
Harnack estimate
----------------
If $M_0$ is not an equator, the strong parabolic maximum principle and the evolution equation of $h_i^j$ in Lemma \[lem: lem1\] imply that for any $t>0$, $M_t$ is strictly convex. For strictly convex hypersurfaces, $$A(V,V)+2D_VH$$ is minimized by $V=(V^1,\cdots,V^i=-b^{ij}\nabla_iH,\cdots,V^n)$ and thus to prove the main theorem, it suffices to verify that for all $t>0$ $$\partial_tH-b^{ij}\nabla_iH\nabla_jH-n\lambda H +\frac{H}{2t}\geq 0.$$ Fix $0<\varepsilon<T.$ We will apply the maximum principle to $$Q:=\frac{\Theta-n\lambda H}{H}=\frac{\Delta H+|A|^2H-b^{ij}\nabla_iH\nabla_jH}{H}$$ on the time interval $[\varepsilon, T).$ Using Lemmas \[lem: lem3\], \[lem: lema6\], and the inverse-concavity of the mean curvature, we calculate $$\begin{aligned}
\partial_tQ=\partial_t \frac{\Theta}{H}\geq&\Delta Q+2\langle \nabla Q,\frac{\nabla H}{H}\rangle+ 2Q^2+2\lambda H^2\\
&+2\frac{\{g^{mq}b^{np}-\frac{g^{mn}g^{pq}}{H}\}\eta_{mn}\eta_{pq}}{H}\\
\geq& \Delta Q+2\langle \nabla Q,\frac{\nabla H}{H}\rangle+ 2Q^2.\end{aligned}$$ An ODE comparison with $q(t)=-\frac{1}{2(t-\varepsilon)}$ which satisfies $\frac{d}{dt}q(t)=2q^2(t)$ and $\lim\limits_{t\to \varepsilon^+}q(t)=-\infty$ shows that $Q(\cdot,t)\geq q(t)$ for any $t>\varepsilon.$ Allowing $\varepsilon\to 0$ completes the proof of the Harnack estimate.
Backwards convergence
---------------------
\[cor: backward limit of principal curvatures\] Any weakly convex ancient solution of the mean curvature flow satisfies $$|A|\leq c_0\exp(nt)$$ for $t\le 0$ for some $c_0$ depending only on $M_0.$
By the strong parabolic maximum principle, any weakly convex ancient solution of the mean curvature flow must become strictly convex, unless it is a non-moving equator. By Theorem 11 for strictly convex, ancient solutions we have $$\frac{\partial_tH-b^{ij}\nabla_iH\nabla_jH-n\lambda H}{H}\geq 0.$$ Since $\lambda =1$, we get $$\partial_t\log H\geq n.$$ Integrating both sides of this inequality against $dt$ on the time interval $[t,0]$ gives $$H(\cdot,t)\leq H(\cdot,0)\exp(nt).$$
Any weakly convex ancient solution of the mean curvature flow satisfies $$|\nabla^mA|^2\leq c_m\exp(2nt)$$ for $t\le 0,$ where $c_m$ depends only on $M_0$ and $m.$
The proof follows by induction on $m.$ Using the fourth evolution equation in Lemma \[lem: lem3\], $\partial_t\Gamma_{ij}^k=A\ast\nabla A$ and that the commutator $[\nabla^k,\Delta]T $ is given by $$[\nabla^k,\Delta]T =\sum\limits_{j=0}^k\nabla^jRm\ast\nabla^{k-j}T$$ for any tensor $T$, we can compute the following evolution equations:
1. $$\partial_t |A|^2=\Delta |A|^2-2|\nabla A|^2+2|A|^4+2\lambda(2H^2-n|A|^2)$$
2. $$\begin{aligned}
\partial_t |\nabla^mA|^2=&\Delta |\nabla^mA|^2-2 |\nabla^{m+1}A|^2\\
&+\sum\limits_{i+j+k=m}\nabla^iA\ast\nabla^jA\ast\nabla^kA\ast\nabla^mA\\
&+\lambda\nabla^mA\ast\nabla^mA.
\end{aligned}$$
On the other hand, by Lemma \[cor: backward limit of principal curvatures\] we get $$\partial_t |A|^2\leq\Delta |A|^2-2|\nabla A|^2+c_1\exp(2nt),$$ and $$\partial_t |\nabla A|^2\leq\Delta |\nabla A|^2+c_2|\nabla A|^2+c_3\exp(2nt).$$ Here, $c_1,c_2,c_3\geq0$ are independent of $t.$ Therefore, $$\begin{aligned}
\partial_t ((t-s)|\nabla A|^2+b|A|^2)\leq&\Delta((t-s)|\nabla A|^2+b|A|^2)\\
&+(1-2b+(t-s)c_2)|\nabla A|^2+(c_1b+c_3)\exp(2nt).\end{aligned}$$ We want to apply the maximum principle on the time interval $[s,s+1].$ We choose $b$ large enough, independent of $t$, so that the coefficient of $|\nabla A|^2$ becomes negative. Thus the maximum principle implies that $$(t-s)|\nabla A|^2\leq c_4\exp(2n t)(1-\exp(2n(s-t) ))+bc_0^2 \exp(2ns).$$ In particular for $t=s+1$ we have $$|\nabla A|^2(\cdot, t)\leq c_5\exp(2n t).$$ This verifies the bound on $|\nabla A|.$ Higher derivative estimates follow by induction and using the test function $(t-s)|\nabla^mA|^2+b_{m-1}|\nabla^{m-1}A|^2:$ $$\partial_t |\nabla^{m-1} A|^2\leq\Delta |\nabla^{m-1} A|^2-2 |\nabla^{m}A|^2+c_1\exp(2nt).$$ $$\partial_t |\nabla^m A|^2\leq\Delta |\nabla^m A|^2+c_2|\nabla^m A|^2+c_3\exp(2nt).$$
Having established the higher derivative curvature bounds, convexity of $M_t$ implies that the backwards limit of $M_t$ is independent of subsequences. Therefore we have proved:
\[thm8\] Let $M_t$ be an ancient, embedded, weakly convex solution of the mean curvature flow. Then $M_t$ converges exponentially fast in $C^{\infty}$ to a unique equator $M_{-\infty}$ as $t\to-\infty.$
Classification of Ancient Solutions
===================================
In this section, we use Theorem \[thm8\] to classify convex, embedded ancient solution of the mean curvature flow on $\sphere^{n+1}$. The proof uses the Aleksandrov reflection as in [@Br-Lou].
We begin with some preliminaries of the Aleksandrov reflection on $\sphere^{n+1}$. First, we will work relative to the limiting equator obtained in Theorem \[thm8\]. Let $E$ be an equator that bounds the *open* hemispheres $H^{\pm}$ with centers $\pm \basepoint$, and let $\vertvec = \overrightarrow{\origin\basepoint}$ be the unit vector in $\RR^{n+2}$ that points from the origin $\origin$ to $\basepoint$. Let $\radialdistance(x) = d_{\sphere^{n+1}} (\basepoint, x)$ denote the spherical distance from $\basepoint$ to $x \in \sphere^{n+1}$. The radial projection onto $E$ is the map $x \in \sphere^{n+1} \mapsto \radialprojection(x) \in E$, where $\radialprojection$ is the nearest point on $E$ to $x$. If $x \ne \pm \basepoint$, then $\radialprojection(x)$ is a single point. If $x = \pm \basepoint$, then $\radialprojection(x) = E$. In any event, given $y \in \radialprojection(x)$, there is a unique length minimizing geodesic joining $x$ to $y$ and this geodesic must pass through $\pm \basepoint$.
It is convenient to make use of the ambient $\RR^{n+2}$ and define the height function $\height(x) = \ip{x}{\vertvec}$. The radial distance is related to the height function via $$\height(x) = \cos(\radialdistance(x))$$ which is monotonically decreasing in $\radialdistance$.
Now for the Aleksandrov reflection, let $\reflectionvector \in \RR^{n+2}$ be any unit vector that $\ip{\reflectionvector}{\vertvec} < 0$. Let $\reflectionplane = \reflectionvector^{\perp}$ be the hyperplane through the origin with the normal vector $\reflectionvector$. Let $\reflectionhalfspace^{\pm} = \{\pm \ip{x}{\reflectionvector} > 0\}$ denote the halfspaces with the boundary $\reflectionplane$. For any subset $S \subset \sphere^{n+1}$, write $\reflectionset{S}^{\pm} = S \intersect \reflectionhalfspace^{\pm}$. Lastly, let $\delta > 0$ denote the angle $\reflectionvector$ makes with $E$; therefore, $\sin \delta = \ip{\reflectionvector}{-\vertvec}$.
The Aleksandrov reflection across $\reflectionplane$ is the map defined by $$\reflectionmap: x \in \RR^{n+1} \mapsto x - 2\ip{x}{\reflectionvector} \reflectionvector.$$
This map is an (orientation reversing) isometry of $\RR^{n+2}$ fixing $\reflectionplane$ and in particular fixing the origin. Therefore, it induces an isometry of $\sphere^{n+1}$. For $x \in E$, we have $\ip{x}{\vertvec} = 0$ and $$\height(\reflectionmap(x)) = \ip{\vertvec}{x - 2 \ip{x}{\reflectionvector} \reflectionvector} = 2 \sin\delta \ip{x}{\reflectionvector}.$$ In the case $x \in E^+$, $\height(\reflectionmap(x)) > 0$, and in the case $x \in E \intersect \reflectionplane$, $\height(\reflectionmap(x)) = 0$.
In geodesic polar coordinates, $(\radialdistance, \sigma) \in (0, \pi) \times E \simeq \sphere^{n+1}\backslash \{\pm \basepoint\}$, a smooth, closed hypersurface $M \subset \reflectionhalfspace^+$ that bounds a region in $\reflectionhalfspace$ is a smooth graph $(f(\sigma), \sigma)$ over $E$ if and only if its (outer) normal $\nor$ satisfies $\ip{\nor}{\vertvec} < 0$ (e.q., $M$ has no vertical tangents). In particular, for all $\epsilon>0$ there is a $\xi>0$ such that if $M$ is a graph with $\ip{\nor}{\vertvec} < -\xi$ and $N$ is $\epsilon$-close to $M$ in $C^{\infty}$, then $N$ is a graph over $E$.
Let $M_t$ be a convex, embedded ancient solution of the mean curvature flow on $\sphere^{n+1}$. Then $M_t$ is a family of shrinking geodesic spheres emanating from an equator at $t=-\infty$.
Let $E = M_{-\infty} \simeq \sphere^n$ be the limiting equator $\lim\limits_{t\to-\infty} M_t$.
Since $M_t$ smoothly converges to $E$ as $t\to-\infty$, we may write $M_t$ as the graph of a smooth positive function over $E$ in the geodesic polar coordinates, $M_t = \{(f_t(\sigma), \sigma) \in (0,\pi) \times E$}. We have $f_t \to \pi/2$ smoothly and uniformly in $C^{k}$, for any $k$, as $t \to -\infty$. Moreover, for $\delta \in (0,\pi/4)$, $\reflectionmap(E) = \{(g_{-\infty}(\sigma), \sigma)$} is a graph over $E$. Note that $\reflectionmap(E)$ is not a graph when $\delta = \pi/4$ (it is an equator perpendicular to $E$), so let us fix a $\delta_0 \in (0,\pi/4)$ to give us a little room. Then it follows from [@Br-Lou Lemma 5.2] that both $\reflectionset{M_t}^-$ and $\reflectionmap(\reflectionset{(M_t)}^+)$ are non-empty and are graphs for all times $t\in(-\infty, S)$ with $S>-\infty$ independent of $\delta \in (0, \delta_0)$. Let us write $g_t$ for the graph of $\reflectionmap(M_t)$ over $E$.
As noted above, $\height(\reflectionmap(x)) > 0$ on $E^+$ and $\height(x) = \cos(\radialdistance(x))$. Thus, continuity implies that for $\epsilon > 0$ there exits an $\eta>0$ such that $\radialdistance(\reflectionmap(x)) < \pi/2 - \epsilon$ provided $x \in E_{\eta} = \{x \in E: d(x, E \intersect \reflectionplane) > \eta\}$. Since $M_t \to_{C^{\infty}} E$, we can choose $T_{\delta} < 0$, such that $d(M_t, E) < \epsilon/2$ for all $t < T_{\delta}$; that is, $\radialdistance (x) > \pi/2 - \epsilon/2$ for all $x \in M_t$. Now for $x \in M_t^+ \intersect \radialprojection^{-1} E_{\eta}$, since $\reflectionmap$ is an isometry, we have $d(\reflectionmap(x), \reflectionmap(\radialprojection(x))) < \epsilon/2$; therefore, $\radialdistance(\reflectionmap(x)) < \pi/2 - \epsilon/2$. Consequently, away from the strip $\{x \in E: d(x, E \intersect \reflectionplane) \leq \eta\}$, we have $\radialdistance(\reflectionmap(\reflectionset{(M_t)}^+)) < \pi/2 - \epsilon/2$ and $\radialdistance(M_t^-) > \pi/2 - \epsilon/2$. That is, $\reflectionmap(\reflectionset{(M_t)}^+) > \reflectionset{(M_t)}^-$ away from the strip.
Now consider the great circle $C$ in $\equator$ joining $\radialprojection(x)$ to $\reflectionmap(\radialprojection(x))$ for $x \in M_t^+$. From the formula $\radialprojection(x) = \tfrac{x'}{|x'|}$ where $x' = x - \ip{x}{\vertvec}\vertvec$ is orthogonal projection onto the hyperplane $e^{\perp}$ and the fact that $$V = \tfrac{1}{2\ip{x}{\reflectionvector}}(x - \reflectionmap(x)),$$ we see that $C$ lies in the plane spanned by $\reflectionvector' = \reflectionvector - \ip{\reflectionvector}{\vertvec}\vertvec$ and $\radialprojection(x)$. Letting $y$ be the closest point to $\pi(x)$ in the intersection $ C \cap \reflectionplane$, we find that in fact $C$ lies in the plane spanned by $\reflectionvector$ and $y$. In particular, since $$\ip{\reflectionvector}{\reflectionvector'} = \cos^2(\delta),$$ the angle $C$ makes with $P$ is independent of the choice of $x$. This allows us to employ the Backwards Approximate Symmetry Lemma [@Br-Lou Lemma 5.1] (which applies whenever $f_t$ converges smoothly to $C$) to conclude that (possibly by decreasing $T_{\delta}$) over $C$ $$\reflectionmap(\reflectionset{(M_t)}^+) \geq \reflectionset{(M_t)}^-$$ on the strip $\{x \in E: d(x, E \intersect \reflectionplane) \leq \eta\}$. Here, $T_{\delta}$ depends only on $\|f_t - \pi/2\|_{C^{2}}$ and $\delta$; therefore, $T_{\delta} > - \infty$ is independent of $C$. Thus we find that $$\label{eq:backwards_approximate_symmetry}
\reflectionmap(\reflectionset{(M_t)}^+) \geq \reflectionset{(M_t)}^-$$ everywhere for all $t \in (-\infty, T_{\delta})$.
Let us now define $T_{\delta}$ so that $(-\infty, T_{\delta})$ is the largest interval on which the relation holds. Also define $T = \inf\limits_{\delta \in (0,\delta_0)} T_{\delta}$. We want to show that $T > -\infty$, and hence that the relation holds on the non-empty, open interval $(-\infty, T)$. To show $T>-\infty$, we apply the maximum principle: Recall that both $\reflectionset{M_t}^-$ and $\reflectionmap(\reflectionset{(M_t)}^+)$ are non-empty and are graphs for all $t\in(-\infty, S)$ with $S>-\infty$ independent of $\delta$. Since the relation is true on $(-\infty, \min\{T_{\delta},S\})$, the maximum principle applied in the time interval $[\min\{T_{\delta},S\}/2,S)$ ensures that $$\label{eq:longtime_approximate_symmetry}
\reflectionmap(\reflectionset{(M_t)}^+) \geq \reflectionset{(M_t)}^-$$ for all $t \in (-\infty, S)$ and any $\delta \in (0,\delta_0)$. Hence $T\ge S.$
To complete the proof, we use [@Br-Lou Proposition 5.3] (which applies in any dimension) to conclude that $M_t$ is a geodesic sphere for all $t \in (-\infty, T)$ and thus for all negative times by the uniqueness of solutions.
[10]{} S. Angenent, “Formal asymptotic expansions for symmetric ancient ovals in mean curvature flow." Networks and Heterogeneous Media **8**(2013) 1–8. B Andrews, “Harnack inequalities for evolving hypersurfaces." Mathematische Zeitschrift **217**(1994): 179–197. C. Arezzo and J. Sun, “Conformal solitons to the mean curvature flow and minimal submanifolds." Mathematische Nachrichten **286**(2013): 772-790. P. Bryan and L. Janelle, “Classification of convex ancient solutions to curve shortening flow on the sphere." Journal of Geometric Analysis, (2015) DOI: 10.1007/s12220-015-9574-x. R.S. Hamilton, “Harnack estimate for the mean curvature flow." Journal of Differential Geometry **41**(1995): 215–226. B. Chow, “Geometric aspects of Aleksandrov reflection and gradient estimates for parabolic equations." Communications in Analysis and Geometry **5**(1997): 389–409. B. Chow and R. Gulliver, “Aleksandrov reflection and geometric evolution of hypersurfaces." Communications in Analysis and Geometry **9**(2001): 261–280. B. Chow and R. Gulliver, “Aleksandrov reflection and nonlinear evolution equations. I. The $n$-sphere and $n$-ball." Calculus of Variations and Partial Differential Equations **4**(1996): 249–264. P. Daskalopoulos, R. Hamilton and N. Sesum, “Classification of compact ancient solutions to the curve shortening flow." Journal of Differential Geometry **84**(2010): 455–464. P. Daskalopoulos, R. Hamilton and N. Sesum, “Classification of ancient compact solutions to the [R]{}icci flow on surfaces." Journal of Differential Geometry **91**(2012): 171–214. G. Huisken, “Deforming hypersurfaces of the sphere by their mean curvature." Mathematische Zeitschrift **195**(1987): 205–219. G. Huisken and C. Sinestrari, “Mean curvature flow singularities for mean convex surfaces." Calculus of Variations and Partial Differential Equations **8**(1999): 1-14. G. Huisken and C. Sinestrari, “Convex ancient solutions of the mean curvature flow." arXiv preprint arXiv:1405.7509 (2014). R. Haslhofer and O. Hershkovits, “Ancient solutions of the mean curvature flow." arXiv preprint arXiv:1308.4095 (2013). N. Hungerb[ü]{}hler and T. Mettler, “Soliton solutions of the mean curvature flow and minimal hypersurfaces." Proceedings of the American Mathematical Society **140**(2012): 2117–2126. K. Smoczyk, “Harnack inequalities for curvature flows depending on mean curvature." New York Journal of Mathematics **3**(1997): 103–118. K. Smoczyk, “A relation between mean curvature flow solitons and minimal submanifolds." Mathematische Nachrichten **229**(2001): 175–186. X-J.Wang, “Convex solutions to the mean curvature flow." Annals of Mathematics. Second Series **173**(2011): 1185–1239. B. White, “The nature of singularities in mean curvature flow of mean convex sets." Journal of the American Mathematical Society **16**(2003): 123–138. J.I.E. Urbas, “An expansion of convex hypersurfaces." Journal of Differential Geometry **33**(1991): 91–125.
| {
"pile_set_name": "ArXiv"
} |
---
author:
-
-
bibliography:
- 'bib/references.bib'
title: |
You Cannot Fix What You Cannot Find!\
[An Investigation of Fault Localization Bias in]{}\
---
Automated Program Repair, Spectrum-based Fault Localization, Benchmarking, Empirical Assessment, Bias.
Acknowledgements {#acknowledgements .unnumbered}
================
[ This work is supported by the Fonds National de la Recherche (FNR), Luxembourg, under projects RECOMMEND 15/IS/10449467 and FIXPATTERN C15/IS/9964569.]{}
| {
"pile_set_name": "ArXiv"
} |
**A Science White Paper in response to the 2012 Heliophysics Decadal Survey**
**“Solar & Heliospheric Physics” and “Solar Wind-Magnetospheric Interactions”**
**An Experimental Plasma Dynamo Program for Investigations\
**
**of Fundamental Processes in Heliophysics**
Plasma experiments in laboratory settings offer the opportunity to address fundamental aspects of the solar dynamo and magnetism in the solar atmosphere. Experiments are currently under construction that can investigate the self-generation of magnetic fields and related processes in large, weakly magnetized, fast flowing, and hot (conducting) plasmas. These and future experiments will probe questions that are of crucial importance to heliophysics in the solar interior, atmosphere and wind. Uniquely, laboratory plasma experiments coupled with theoretical explorations can serve to calibrate the simulation codes which are being used to understand the solar dynamo, magnetic reconnection and flares in the solar atmosphere, the nature of CMEs, and the interactions between planetary magnetospheres and the solar wind. Laboratory plasma experiments are likely to contribute new understanding complementary to the traditional observational and modeling approach normally used by space physicists.
We argue here that ground-based laboratory experiments have direct connections to NASA based missions and NSF programs, and that a small investment in laboratory heliophysics may have a high payoff. We will use the Madison Plasma Dynamo Experiment (MPDX)[^1] as an example, but advocate here for broad involvement in community-scale plasma experiments.
#### Fundamental plasma processes in heliophysics and connection to NASA missions and NSF programs
The 22-year solar cycle stands out as one of the most remarkable and enigmatic examples of magnetic self-organization in nature. The Sun’s cycles of magnetic activity profoundly affect our modern technological society. Unsurprisingly, solar magnetism is a fundamental focus of current and future NASA missions.
Magnetic fields that emerge at the solar surface as sunspots are built by dynamo action in the solar convection zone. In most solar dynamo models, organized fields are built in the tachocline, an interface layer deep in the Sun at the bottom of the solar convection zone. These magnetic structures then become buoyantly unstable and rise through the turbulent convection to emerge at the surface. After emergence, magnetic reconnection in the solar atmosphere can lead to flares and coronal mass ejections (CMEs) which have substantial impacts on the heliosphere and on Earth’s magnetosphere in particular. These processes of magnetic flux creation, emergence, and reconnection form many of the driving questions behind the current Solar Dynamics Observatory (SDO), Hinode, the Stereo mission, the very successful Solar and Heliospheric Observatory (SOHO), and the upcoming ground-based NSF Advanced Technology Solar Telescope (ATST). Understanding the origins of the heliospheric magnetic field and the evolution of CMEs in the heliosphere is fundamental to the future Solar Sentinels mission and plays important roles in the future Solar Probe and Interface Region Imaging Spectrograph (IRIS) mission as well. Indeed, the evolution of solar magnetism forms the driving questions behind the Living With a Star (LWS) program at NASA: namely, *how and why does the Sun vary, how does the Earth respond, and what are the impacts to humanity?*
As a community, we have made significant progress in understanding solar magnetism through a combination of theoretical treatments and observational techniques, but many key components of the solar dynamo remain poorly understood. In particular, the processes behind magnetic buoyancy and the generation of magnetic field through turbulent correlations occur deep in the solar interior and are difficult to constrain with either helioseismic observations or numerical simulations. Magnetic reconnection in the solar atmosphere is likewise poorly understood at present. Our understanding of interactions between the Sun and the Earth have grown tremendously and are greatly facilitated by in situ observations of the solar wind in the near-Earth environment by missions like the Advanced Composition Explorer (ACE) and the Magnetospheric Multi-scale Mission (MMM), but many details of interactions between coherent magnetized plasma structures like CMEs with the Earth’s magnetosphere remain unclear.
![\[fig:eta profiles\] $(a)$ Radial profiles of magnetic diffusivity $\eta$ used in various dynamo models. Shown in grey lines are three representative profiles from 3D convective solar dynamo simulations using the anelastic spherical harmonic (ASH) code [e.g., @Brun_et_al_2004; @Brown_et_al_2010b; @Miesch_et_al_2010]. Also shown is the double-step profiles used in 2D mean-field Babcock-Leighton models that have been used to predict the current solar cycle . The molecular diffusivity for a hydrogen plasma at solar conditions is also shown, multiplied by $10^7$ for display purposes. $(b)$ The Prandtl number $\mathrm{Pr}=\nu/\kappa$ (dashed) and magnetic Prandtl number $\mathrm{Pm}=\nu/\eta$ (solid) for a hydrogen plasma at solar conditions. Simulations use values of order unity. ](SOHO_GONG_eta_profiles_and_Prandtl.jpg){width="\linewidth"}
A fundamental limitation facing theoretical explorations of heliophysically relevant plasma processes is the vast separation in parameter space between the natural phenomena and the capabilities of even the largest super computers. As one example, cutting edge global-scale solar dynamo simulations run as part of the NASA High-End Computing (HEC) program[^2] and as part of the NSF Partnerships in Advanced Computing Infrastructure Teragrid program[^3] typically employ turbulent diffusivities which are more than ten orders of magnitude larger than the molecular diffusivities of the solar plasma (Figure \[fig:eta profiles\]$a$). This also applies to the mean-field models which have been used in attempts to predict amplitude and timing of the current solar cycle . Yet these simulations still require substantial computational resources and represent a significant investment by NASA HEC and the NSF Teragrid.
It is deeply impractical to consider simulating global-scale dynamo action in the Sun using molecular diffusivities and doing so would likely require a century of further growth in computational resources. Instead, it is vital to formulate better estimates and models of the turbulent transport processes that occur in a plasma under solar conditions. As an example, even the ratio between how quickly magnetic fields are transported and mixed by turbulence compared to the transport of momentum by the same is poorly constrained. Molecular values of this magnetic Prandtl number Pm are of order $10^{-1}$–$10^{-5}$ (Fig. \[fig:eta profiles\]$b$); turbulent values are thought to be of order unity, but this has not been adequately explored for real turbulent transport in a stratified plasma. These problems are not unique to the solar interior, but are true for simulations of photospheric convection that couple to the chromosphere and corona [e.g., @Nordlund_et_al_2009], and simulations that capture the interactions between the solar wind, CMEs and planet magnetospheres throughout the heliosphere and to the boundary with interstellar space itself.
The growth of computational resources has made it feasible to directly simulate the plasma regime explored by experiments like MPDX. A program of laboratory experimentation that is tightly coupled with significant numerical simulation efforts will lead to new insights into key plasma processes. In particular, these laboratory plasma experiments will likely achieve self-sustaining dynamo action, will explore the dynamics of magnetic buoyancy in a stratified plasma atmosphere, and will explore the interactions between an Earth-like magnetosphere and magnetotail with a solar-wind like plasma flow. In addition to revealing the fundamental underlying plasma processes at work in these systems, these experiments can anchor the computationally intensive simulations which are used to model the solar dynamo, the buoyant rise and emergence of magnetic structures in the solar interior, and interactions of the solar wind and CMEs with the Earth’s magnetosphere.
#### The importance of simulations coupled to experiments
Just as we cannot directly simulate the turbulent conditions of solar convection, we cannot construct laboratory experiments that capture the vast spectrum of scales present in the Sun. Some important physical ingredients acting in solar convection are very difficult to reproduce in experiments: in particular, it is very difficult to produce conditions that mimic the radial self-gravity of systems like the Sun, and experiments that do must generally be conducted at great cost in low-g environments [e.g., @Hart_et_al_1986].
Laboratory experiments provide great opportunities for exploring key physical processes occurring in plasmas under solar conditions, but they cannot capture the integrated global-scale picture. Global-scale couplings between convection, magnetism and rotation are likely crucial to fully understanding the solar cycle as a whole and may be important for understanding eruptive flares and CMEs as well. At present, simulations are the only option for exploring the coupled global-scale system. Simulations can offer insight into processes that are difficult or impossible to directly measure in the laboratory (e.g., the full 3-D velocity field and magnetic field at all points in a volume at all instances in time) and can suggest profitable avenues for additional experimental exploration. But this is not a one-way process: the laboratory experiments validate the codes and inform those simulations in how to correctly capture processes that currently elude simple theoretical explanation, including turbulent transport, magnetic buoyancy instabilities, and the non-linear couplings that arise between flows and magnetism.
Ultimately, theoretical explorations, numerical simulations and laboratory experiments must work together in a tightly coupled fashion to achieve optimal results. But this is feasible and can yield large and rapid payoffs. In the MPDX experiment for example, this work has begun already for modeling of the dynamo and magneto-rotational instability scenarios using the NIMROD code, which solves the MHD equations with a number of two-fluid extensions. These simulations have indicated that sustained dynamo action is possible and in which parameter regimes this is most achievable. New simulations are exploring the possibility of buoyancy instabilities, including magnetic buoyancy, and suggest that these may be experimentally achievable. Direct comparisons between the numerical simulations and pathfinder prototype experiments are completing the code [*validation*]{} loop even before construction finishes on the main experiment (MPDX). Theoretical explorations of the simulations and the experiment are in turn providing the ultimate link to the plasma astrophysics.
Tightly linked programs of laboratory experimentation and numerical simulation, backed by significant theoretical analysis to understand both systems, offer opportunities for rapid advancement in our understandings of solar plasma physics.
![The Madison Plasma Dynamo Experiment (MPDX), with human figure to scale (left). Plasma Dynamo Laboratory Overview (right). The drawing shows the new Plasma Dynamo Facility under construction at the UW Madison. The main experimental hall is two stories high, with approximately 2000 ft$^2$ of floor space and houses the 3 m diameter Plasma Dynamo Experiment. \[fig:MPDX\] ](MPDX_experiment_and_facilities.jpg){width="6.5in"}
#### Status of MPDX and future plasma experiments
Compared to space missions, ground based laboratory experiments can be quite affordable and can be built on rapid timescales. As an example, the MPDX experiment has been funded for construction through the NSF Major Research Instrumentation program with at total device cost of \$2.5M (part of which was \$750k in cost-sharing from the UW). In 2009 a Plasma Dynamo Kick-Off Meeting, sponsored by the NSF Center for Magnetic Self-Organization,[^4] was held to help formulate the scientific program and to provide feedback on the design of the experiment. Funding began in 2009, and the 3 year project is on schedule: 2010 was spent on design and lab remodeling, in 2011 the machine will be assembled, and in 2012 the experiment should be ready for initial operations. An illustration of the experimental apparatus and the Plasma Dynamo Laboratory built around it is shown in Figure \[fig:MPDX\].
There are significant barriers to running such programs on a community scale. Using MPDX as an example again, after construction there is not yet a program in place to operate this experiment. This experiment is larger than a typical single-investigator NSF experiment and might be considered a [*medium-scale*]{} basic plasma or laboratory plasma astrophysics experiment as called for in the recent Plasma 2010 report from the National Research Council. The operations of such an experiment and the diagnostic requirements are similar in scale to university-scale experiments in fusion research, yet there is no mechanism for funding the operation of a community-based facility such as this. A major challenge is adequately funding the operations of such experiments and also the significant theoretical explorations and numerical simulation efforts which are vital to maximizing the scientific returns of such facilities.
#### Heliophysically relevant plasma experiments
Future plasma experiments could directly address many plasma processes that are directly relevant to heliophysics and NASA. Most experimental plasma facilities at present are focused on other science, with substantial investments in exploring basic plasma processes and the application to fusion energy generation. Plasma experiments can now directly address several key components of the solar dynamo and of eruptive events in the solar atmosphere.
A variety of laboratory experiments are feasible in the near future that are clearly heliophysically relevant:
1. large-scale dynamo experiments (generation of the large scale solar magnetic field),
2. small-scale turbulent dynamo experiments (small scale structure of the solar dynamo and surface magnetism),
3. buoyancy driven stratified convection (solar convection zone and surface),
4. magnetic buoyancy instabilities in a stratified atmosphere (sunspot emergence and solar tachocline physics),
5. flow driven plasma turbulence in near-equipartition (solar wind),
6. flow driven plasma turbulence impacting a model magnetosphere (space weather, sun-earth connection),
7. magnetic field line stretching with potentially explosive, flow-driven reconnection (magnetotail, solar flares),
8. plasma instabilities at high $\beta$ and low collisionality (solar wind spectrum).
A program of heliophysically relevant laboratory plasma experiments can provide important insight into these problems. Such a program will surely suggest additional unexpected avenues for research, as expertise is gained in experimental techniques, and in the theoretical and numerical explorations of such experiments.
The proposed plasma experiments transcend the scope of *single user facilities* and instead require substantial *community-based facilities*. Such experiments should be supported at a larger scale than can be undertaken by individual universities and research groups. This support must extend past the mere construction of experimental facilities: instead, significant advances can be achieved by ongoing support for operations, for theoretical explorations, and for tightly coupled numerical simulations. These synergistic efforts are likely to maximize scientific output and may rapidly advance our understanding of plasma processes occurring in the solar interior, atmosphere and wind, whereas isolated efforts must continue to wrestle with intractable limitations. In summary, we advocate that heliophysics could greatly benefit from the support of community-scale, heliophysically relevant plasma experiments.
#### References
\
[10]{} url \#1[[\#1]{}]{}urlprefix\[2\]\[\][[\#2](#2)]{}
A S, [Miesch]{} M S and [Toomre]{} J 2004 [*[[ApJ]{}]{}*]{} [**614**]{} 1073–1098
B P, [Miesch]{} M S, [Browning]{} M K, [Brun]{} A S and [Toomre]{} J 2010 [*[[ApJ]{}]{}*]{} [submitted]{}
M S, [Brown]{} B P, [Browning]{} M K, [Brun]{} A S and [Toomre]{} J 2010 *Preprint*
M and [Gilman]{} P A 2006 [*[[ApJ]{}]{}*]{} [**649**]{} 498–514
P 2010 [*Living Reviews in Solar Physics*]{} [**7**]{} 3:1–91
, [Stein]{} R F and [Asplund]{} M 2009 [*Living Reviews in Solar Physics*]{} [**6**]{} 2:1–116
Hart, J E, Toomre, J, Deane, A E, Hurlburt, N E, Glatzmaier, G A, Fichtl, G H, Leslie, F, Fowlis, W W and Gilman, P A 1986, [*Science* ]{} [**234**]{} 61
[^1]: This work builds upon excitement in recent years of using liquid metals to study dynamos and will extend these studies to more astrophysically relevant parameters. Use of a plasma for such experiments allows the magnetic Reynolds number (the dimensionless parameter governing self-excitation of magnetic fields) to be approximately a factor of 10 larger than in liquid metal experiments. These experiments will be the first to investigate self-excited dynamos in a plasma, the state of matter that makes up most naturally occurring astrophysical dynamos.
[^2]: using for example the Pleiades and Columbia supercomputers
[^3]: using supercomputers at Pittsburg Supercomputing Center (PSC), San Diego Supercomputer Center (SDSC), Texas Advanced Computing Center (TACC), National Institute for Computational Science (NICS), at the National Center for Supercomputing Applications (NCSA), and others
[^4]:
| {
"pile_set_name": "ArXiv"
} |
\
------------------------------------------------------------------------
[${}^{a)}$Electronic mail: [email protected]]{}
[${}^{b)}$On leave from the Division for Optical Problems in Information Technologies, National Academy of Sciences of Belarus, Zhodinskaya St. 1/2, 220141 Minsk, Belarus]{}
Introduction
============
Multi-component (or coupled) nonlinear PDEs had been a subject of considerable interest for many years (see, for instance, Ref. 1, and references therein). Recent revival of interest in multi-component PDEs is due to new discoveries and technological advances in nonlinear optics and physics of condensed matter. An important example is the incoherent spatial optical solitons, or self-trapped spatially incoherent light beams, recently experimentally observed in nonlinear media \[2\], which are described by the multi-component nonlinear Schrödinger (NLS) equations \[3\]. Another example of possible application of the coupled NLS equations is the creation and dynamics of solitary waves in the multispecies Bose-Einstein condensates \[4\]. Similar models of coupled nonlinear PDEs appear in the wavelength division multiplexing, i.e., copropagation of pulses in an optical fibre on beams with different wavelengthes \[5-7\] and in other important applications \[8,9\].
Some of the multi-component models are integrable. Integrable multi-component PDEs have another specific feature, which makes them important for applications as zero-approximation models for analytical description of the real phenomena. It has been known for quite some time that dimensional reductions of matrix generalizations of the integrable PDEs, such as the NLS and KdV equations, can produce a variety of new integrable equations \[10\]. For instance, some of the coupled NLS equations are integrable reductions of the general matrix NLS equation. The $N$-dimensional matrix NLS equation is the simplest integrable PDE associated with the $(N+1)$-dimensional Zakharov-Shabat spectral problem \[11-13\]. Recently a variety of integrable coupled higher-order NLS equations was discovered \[14-16\], which are important in view of applications to the soliton propagation of sub-picosecond pulses in optical fibre \[17-19\]. Some of these integrable PDEs arise as dimensional reductions of the matrix complex modified KdV (cmKdV) equation, which is also associated with the Zakharov-Shabat spectral problem.
In most cases, the multi-component PDE is not integrable. However, frequently the terms destroying integrability contain small parameters and the non-integrable equation can be considered as a perturbation of the integrable one. In this case, a perturbation theory is required for analytical description of the effect of small perturbations. For instance, one is especially interested in the dynamics of solitons in nearly integrable PDEs. Soliton solutions to multi-component equations have many parameters, and their evolution may exhibit a variety of new interesting regimes. Therefore, it is necessary to have at hand a perturbation theory for multi-component nearly integrable equations. Such perturbation theory is developed in the present paper.
Perturbation theory for nearly integrable PDEs has a long history \[20-45\]. There are two basic approaches in the perturbation theory based on the IST method. The first one originated in works of Kaup \[20\] and Karpman and Maslov \[21\], where the perturbation theory was developed for nearly integrable PDEs associated in the integrable limit with the $2\times 2$ matrix Zakharov-Shabat spectral problem. Quite different approach originated from works \[22-26\]. It was found that an integro-differential operator, generating the whole hierarchy of integrable PDEs related to a given spectral problem, called the recursion operator, has a complete set of eigenfunctions, which can be used for the perturbation expansion. Several other methods, not related to the IST, were applied for description of the perturbed soliton dynamics. For instance, a method based on the Green functions was developed in Ref. 29. The IST-independent perturbation theories for solitons are usually referred to as the direct perturbation theories (see for instance, Ref. 40 and references therein). However, notwithstanding the long history of the perturbation theory, with rare exceptions, only the $2\times 2$ matrix spectral problems were considered. It was noted that construction of the perturbation theory for higher-dimensional matrix spectral problems along the standard approach becomes technically more involved.
To overcome technical difficulties of the standard approach when dealing with multi-component PDEs, the method based on the Riemann-Hilbert (RH) problem was proposed in Ref. 41, where the perturbation theory was developed for the Zakharov-Shabat spectral problem of an arbitrary matrix dimension. The RH problem was used before for construction of the perturbation theory for the Landau-Lifschitz equation \[30\], the NLS and Maxwell-Bloch equations \[34\], which are integrable by the $2\times2$ matrix spectral problem. The approach of Ref. 41 was applied to the Manakov system \[42\] (i.e., the two-component NLS equation), modified NLS equation \[43\] and massive Thirring model \[44\]. These examples demonstrate that the perturbation theory based on the RH problem always works. Recently, the RH problem was applied to nearly integrable equations on the half line, arising from the singular dispersion relations \[45\]. In Refs. 41-45 the perturbation-induced evolution equations for the spectral data were derived with the help of some matrix functional (below, the evolution functional $\Pi(x,t,k)$). It is important to emphasize that the form of the evolution functional is invariant under the gauge transformations of the considered PDE \[43\]. Thus, once constructed, the evolution functional is valid not only for the whole hierarchy of PDEs associated with a given spectral problem, but also for their images under the gauge transformations. Writing the dispersion law, generating the spectral problem, in an abstract form $\Lambda(k)$ (see the next section for details) we discover that the form of the evolution functional remains invariant under the change of the [*spectral problem*]{} as well. This invariance trivially extends to the general initial-boundary value problems. For instance, for the half-line, where one would expect a difference, we have found that the evolution functional has similar form \[45\]. Therefore, it seems that the approach based on the evolution functional is universal for construction of the perturbation theories for nearly integrable PDEs. It is also technically simple. Derivation of the perturbation-induced evolution equations for the spectral data using the evolution functional reduces to calculation of integrals.
This paper is a further development of Ref. 41. The previous results are substantially advanced. In particular, the evolution equations for the spectral data are considerably simplified with the help of some identities found for the evolution functional. We start section \[secLax\] with a brief discussion of the multi-component integrable PDEs associated with the Zakharov-Shabat spectral problem. We consider two examples, the matrix NLS and cmKdV equations, however, our approach is valid for many other multi-component PDEs. We have not made an attempt to give a complete exposition of the properties of integrable equations. We need only the Lax representation. Hence a way of deriving the Lax pair for an integrable PDE from the dispersion relation of its linearization is briefly indicated. For completeness of the exposition, a detailed derivation of the RH problem is given in section \[secRH\]. Solution of the RH problem for multi-component equations involves some technicalities, which are discussed and detailed derivations are provided in the Appendices. We derive evolution equations for the spectral data with account for perturbations in section \[pert\]. There, for an important special case of the vector nearly integrable PDEs the main result of this paper is formulated in a theorem. In the final section \[exmpl\] the equations of the first-order perturbation theory for a single vector soliton are given in explicit form.
Preliminaries: Integrable multi-component nonlinear PDEs {#secLax}
========================================================
Here we briefly discuss integrable PDEs with emphasis on the multi-component equations whose reductions are important for applications. In particular, we consider the matrix nonlinear Schrödinger and complex modified Korteweg-de Vries equations. We do not try to review this subject, for general considerations the reader can consult, for instance, Refs. 1, 10-13, 46-53 and the references therein. The purpose of this section is to remind some of the basic notions in the IST method. Though the approach below can be applied to any nonlinear PDE solvable by the RH problem, we restrict the consideration to the integrable equations associated with the $N$-dimensional Zakharov-Shabat spectral problem (\[lax1\]).
Consider the integrable PDEs which arise as the compatibility condition for the following $N\times N$ matrix linear system (Lax pair) $$\partial_x \Phi = ik[A,\Phi] + iQ(x,t)\Phi\equiv \Phi\Lambda(k)+U(x,t,k)\Phi,
\label{lax1}$$ $$\partial_t \Phi = {i\omega(k)} \Phi A + V(x,t,k)\Phi\equiv\Phi\Omega(k)+V(x,t,k)\Phi,
\label{lax2}$$ with $$A=\left(\begin{array}{cc} I_n & 0\\ 0 & -I_{N-n} \end{array}\right),\quad
Q=\left(\begin{array}{cc} 0& \bf{q}\\ \overline{\bf{q}} & 0 \end{array}\right),$$ $$\bf{q} = \left(\begin{array}{cccc} q_{11}& q_{12}&\ldots& q_{1,{N-n}} \\
q_{21}& q_{22}&\ldots& q_{2,{N-n}} \\
\vdots & \vdots & & \vdots \\
q_{n1}& q_{n2}&\ldots& q_{n,{N-n}} \end{array}\right),\quad
\overline{\bf{q}} = \left(\begin{array}{cccc} \overline{q}_{11}&
\overline{q}_{21}&\ldots& \overline{q}_{n1} \\
\overline{q}_{12}& \overline{q}_{22}&\ldots& \overline{q}_{n2} \\
\vdots & \vdots & & \vdots \\
\overline{q}_{1,{N-n}}& \overline{q}_{2n}&\ldots& \overline{q}_{n,{N-n}}
\end{array}\right),$$ where $\Lambda(k) = - ikA$ and $\Omega = i\omega(k)A$ are the dispersion laws, $Q$ is called the potential. Here the overline does not mean complex conjugation by default, e.g., in general, the functions $q_{ij}$ and $\overline{q}_{ij}$ are not considered as complex conjugate to each other. When the overline does denote complex conjugation in the text below, each time it will be specially indicated. This special case corresponds to the Hermitian potential $Q$, $Q^{\dag}=Q$ or ${\bf q}^\dagger = {\overline{\bf q}}$, and it will be referred to as the involution. The temporal evolution equation (\[lax2\]) is specified by choice of the dispersion relation $\omega(k)$ in the following manner. For simplicity, let the dispersion relation be polynomial $\omega(k)=\sum_{p=1}^{M}w_p k^p$, then $$V(k) = -{\cal P}\{\Phi\Omega\Phi^{-1}\}\equiv -\Omega(k)
+\sum_{p=0}^{M-1}V_pk^p.
\label{V}$$ Here the matrix function $\Phi(k)$ is expanded into the asymptotic series: $$\Phi(k)=I+k^{-1}\Phi^{(1)}+k^{-2}\Phi^{(2)}+..., \quad k\to \infty,$$ and the operator ${\cal P}$ takes the polynomial in $k$ part of $\Phi\Omega\Phi^{-1}$ on the asymptotics. For example, the Zakharov-Shabat spectral problem (\[lax1\]) is derived in this way $$U=-{\cal P}\{\Phi\Lambda\Phi^{-1}\}=ikA+i[\Phi^{(1)},A],$$ with the obvious relation $$Q=[\Phi^{(1)},A].
\label{Q}$$ Hence, the Lax pair satisfies the property $\text{Tr}U = -\text{Tr}\Lambda$ and $\text{Tr}V = - \text{Tr}\Omega$.
The integrable nonlinear PDE related to the Lax pair (\[lax1\])-(\[lax2\]) is given by the compatibility condition (in our case, polynomial in $k$) $$i\partial_t Q -\partial_x V + [ikA+iQ,V] = 0
\label{NLEE}$$ via setting $k=0$, while the positive powers of $k$ supply the expressions of the coefficients $V_p$ in (\[V\]) as functions of the potential $Q$ and its $x$-derivatives, $V_p=V_p(Q,Q_x,Q_{xx},...)$. For instance, choosing $\omega(k)=2k^2$ we obtain the well-known matrix nonlinear Schrödinger equation. Indeed, in this case $$V=-2ik^2A-2ikQ-AQ_x+iAQ^2$$ and the equation (\[NLEE\]) becomes $$iAQ_t + Q_{xx}+2Q^3 = 0.
\label{NLS}$$ For a complete classification of the matrix integrable NLS equations with various reductions to Hermitian symmetric spaces consult Ref. 10. A particular important case of equation (\[NLS\]) is the vector NLS equation, a generalization of the two-component vector NLS, which was shown to be integrable by Manakov \[12\]. The vector NLS equation corresponds to the Hermitian potential and the reduction $n=N-1$ (see the expression for $A$). In this case we have $$Q = \left(\begin{array}{cccc} 0 & \ldots & 0 & q_1 \\
\vdots & & \vdots & \vdots \\ 0 & \ldots & 0 & q_{n} \\
\overline{q}_{1}& \ldots & \overline{q}_{n} & 0 \end{array}\right)
\label{specialQ}$$ and matrix equation (\[NLS\]) becomes the vector NLS equation: $$i\partial_t q_l +\partial_{x}^2 q_l +2\left(\sum_{j=1}^{n}|q_j|^2\right)q_l=0,\quad l=1,...,n.
\label{vectorNLS}$$
Let us consider another important example of multi-component nonlinear integrable equations. It is given by setting $\omega(k)= 4k^3$. After simple computation we get $$V = -4ik^3A+4ik^2Q+2ikA\left(iQ_x+Q^2\right)-iQ_{xx}+[Q,Q_x]-2iQ^3,$$ which produces the matrix cmKdV equation $$Q_t + Q_{xxx} +3\left(Q^2Q_x + Q_xQ^2\right) = 0.
\label{CMKdV}$$
Combining together the considered two dispersion laws, i.e., letting $\omega(k) = 4i\epsilon k^3 +2i\beta k^2$, one can obtain $$iQ_t + \beta A \left( Q_{xx} + 2Q^3\right) + i\epsilon Q_{xxx} +
3i\epsilon \left( Q^2Q_x + Q_xQ^2\right)
= 0,$$ a special case of the (generally, non-integrable) matrix higher-order NLS equation $$iE_z + A(\alpha_1 E_{\tau\tau} + \alpha_2 E^3) + i \left\{ \alpha_3
E_{\tau\tau\tau}
+ \alpha_4 (E^3)_\tau +\alpha_5 (E^2)_\tau E \right\} = 0,
\label{HONLS}$$
Our examples are just illustrative. There is many other integrable matrix PDEs which we do not mention here. However, the perturbation theory developed in section \[pert\] applies to such PDEs also.
Riemann-Hilbert problem for multi-component PDEs {#secRH}
================================================
In this section we derive the RH problem for the multi-component integrable PDEs and discuss its solution (for more details consult Refs. 46, 54-57). We are interested in the initial-value problem for nonlinear PDEs on the whole real line with the asymptotically vanishing conditions (Cauchy problem) $q_{ij}\to 0$ as $|x|\to \infty$. The vanishing asymptotics allows us to concentrate entirely on the spectral equation (\[lax1\]) in derivation of the RH problem, while $t$-dependence enters parametrically in our approach (for the RH problem for initial-boundary value problems consult Ref. 47). Below we omit the explicit $t$-dependence for simplicity of the presentation. To begin with, let us summarize the properties of the spectral problem (\[lax1\]). Define the following $N\times N$ matrix projectors $$H_1 = \text{diag}(I_n,0),\quad H_2 = \text{diag}(0,I_{N-n}).
\label{Hs}$$ Then $A = H_1 -H_2$ and any matrix can be decomposed into the sum of two matrices, commuting and anti-commuting with $A$: $$\Phi = \Phi^{(c)} + \Phi^{(a)},\quad \Phi^{(c)} = H_1 \Phi H_1 + H_2 \Phi H_2,\quad
\Phi^{(a)} = H_1 \Phi H_2 + H_2 \Phi H_1,
\label{}$$ where $[A,\Phi^{(c)}] = 0$ and $\{A, \Phi^{(a)}\} = 0$. We will use the block-index notations for the decomposition of matrix $\Phi$ with respect to the projectors $H_1$ and $H_2$: $$\Phi = \left(\begin{array}{cc} \Phi_{I,I} & \Phi_{I,II}\\
\Phi_{II,I} & \Phi_{II,II}\end{array}\right).$$
The RH problem is a one-to-one mapping (nonlinear Fourier transform) between the set of smooth (e.g., belonging to the Schwartz space) potentials $Q(x)$ and some set of the spectral data. To identify the RH problem one must construct two solutions, one, $\Phi_+(x,k)$, to the spectral equation (\[lax1\]) and the other, $\Phi_-^{-1}(x,k)$, to the adjoint equation (the second function is inverse of some matrix function satisfying (\[lax1\]), hence the “-1" in its definition), holomorphic with respect to the spectral parameter $k$ in some complementary domains covering the whole complex $k$-plane. Such solutions can be built from the columns and rows of the Jost solutions $J_\pm$, i.e., solutions defined by the asymptotic conditions: $J_\pm(x,k)\to I$ as $x\to\pm\infty$.
As $\text{Tr} Q = 0$, letting $x\to \pm\infty$ we conclude that det$J_\pm = 1$. Hence, the columns of either of the two Jost solutions give a linear basis in the space of solutions of the spectral equation. The inverse matrices $J_\pm^{-1}$ satisfy the adjoint spectral equation: $$\partial_x \widetilde\Phi= ik [A, \widetilde\Phi] - i\widetilde\Phi Q.
\label{adj_lax1}$$ Each column of $J_\pm(x,k)$ and, respectively, row of $J_\pm^{-1}(x,k)$ is holomorphic and bounded in either upper (Im$k\ge0$) or lower (Im$k\le0$) half of the complex $k$-plane. Indeed, this can be easily seen from the Volterra integral equations for the Jost matrices written for the two blocks of columns: $$J_\pm(x,k)H_1 = H_1 + i\int\limits_{\pm\infty}^x\text{d}
\xi\, e^{-2ik(x-\xi)H_2} Q(\xi)J_\pm(\xi,k)H_1,$$ $$J_\pm(x,k)H_2 = H_2 + i\int\limits_{\pm\infty}^x\text{d}
\xi\, e^{2ik(x-\xi)H_1} Q(\xi)J_\pm(\xi,k)H_2.$$ The columns of $J_+(k)H_1$ and $ J_-(k) H_2$ are holomorphic and bounded in the upper half of the complex plane, while columns of $J_+(k)H_2$ and $J_-(k)H_1$ have the same property in the lower half plane. Similarly, the rows of $J_\pm^{-1}$ satisfy integral equations $$H_1J^{-1}_\pm(x,k) = H_1 - i\int\limits_{\pm\infty}^x\text{d}\xi\,
H_1J^{-1}_\pm(\xi,k) Q(\xi)e^{2ik(x-\xi)H_2},$$ $$H_2J^{-1}_\pm(x,k) = H_2 - i\int\limits_{\pm\infty}^x\text{d}\xi\,
H_2J^{-1}_\pm(\xi,k)Q(\xi) e^{-2ik(x-\xi)H_1},$$ from which we immediately conclude that $H_1 J^{-1}_+(k)$ and $H_2J^{-1}_-(k)$ are holomorphic and bounded in the upper half plane, while $H_2 J^{-1}_+(k)$ and $ H_1J^{-1}_-(k)$ have the same properties in the lower half plane.
On the real line, the Jost solutions are transformed into each other by the scattering matrix $S(k)$, $$J_-(x,k) = J_+(x,k) e^{ikxA}S(k) e^{-ikxA}.
\label{S}$$
For the Hermitian potential, $Q^\dagger = Q$, the matrix Jost solutions satisfy the involution (here the overline means complex conjugation) $$J_\pm^\dagger(x,k) = J^{-1}_\pm (x,\overline{k}),
\label{invol_J}$$ where the spectral parameter takes complex values in the upper or lower half plane depending on the considered column of the Jost matrix. In this case, the scattering matrix also satisfies the involution $$S^\dagger(k) = S^{-1}(k),\quad k\in \text{Re}.
\label{invol_S}$$
The holomorphic matrix functions $\Phi_+(k)$ and $\Phi^{-1}_-(k)$, satisfying equation (\[lax1\]) and (\[adj\_lax1\]), respectively, are given in terms of columns and rows of the Jost solutions: $$\Phi_+ = J_+H_1 + J_-H_2, \quad
\Phi^{-1}_- = H_1J^{-1}_+ + H_2 J^{-1}_- .
\label{phi+phi-}$$ The above defined matrix functions are holomorphic and bounded in the upper and lower half planes, respectively. They have the following asymptotics $$\Phi_\pm(x,k)\to I,\quad k\to\infty,
\label{phi_asymp_k}$$ which follow from the Volterra integral equations for the Jost solutions. For the involution (\[invol\_J\]) the matrices $\Phi_+(k)$ and $\Phi^{-1}_-(k)$ are related via $$\Phi^{\dagger}_+(k) = \Phi_-^{-1}(\overline{k}).
\label{invol_phi}$$ These matrices can be conveniently expressed in terms of only one Jost solution and elements of the scattering matrix $S(k)$. Indeed, $$\Phi_+ = J_+e^{ikxA}(H_1 + SH_2)e^{-ikxA}
\equiv J_-e^{ikxA}(H_2 + S^{-1}H_1)e^{-ikxA},$$ $$\Phi^{-1}_- = e^{ikxA}(H_1 + H_2S^{-1})e^{-ikxA}J^{-1}_+
\equiv e^{ikxA}(H_2 + H_1S)e^{-ikxA}J^{-1}_-.$$ Denote $S_+ = H_1 + SH_2$, $S_- = H_2 + S^{-1}H_1$. These matrices provide a factorization of the scattering matrix: $S_+ = SS_-$. Similarly, $\overline{S}_+ = H_1 + H_2 S^{-1}$ and $\overline{S}_- = H_2 + H_1 S$, which also factorize the scattering matrix: $\overline{S}_+ S =
\overline{S}_-$. Then
\[rep\_factor\] $$\Phi_+ = J_+e^{ikxA}S_+ e^{-ikxA} \equiv J_-e^{ikxA}S_- e^{-ikxA},
\label{rep_factor_a}$$ $$\Phi^{-1}_- = e^{ikxA}\overline{S}_+e^{-ikxA}J^{-1}_+
\equiv e^{ikxA}\overline{S}_-e^{-ikxA}J^{-1}_-.
\label{rep_factor_b}$$
The factorization matrices have the block-triangular structure. For instance, $S_+$ and $\overline{S}_+$ are upper and lower block-triangular, respectively: $$S_+ = \left( \begin{array}{cc} I_n & {\bf b} \\ 0 & {\bf a} \end{array}\right),\quad
\overline{S}_+ = \left( \begin{array}{cc} I_n & 0 \\ \overline{{\bf b}}
& \overline{{\bf a}}
\end{array}\right),
\label{factors}$$ where ${\bf b} = S_{I,II}$, ${\bf a} = S_{II,II}$, $\overline{{\bf b}} = (S^{-1})_{II,I}$, and $\overline{{\bf a}} =
(S^{-1})_{II,II}$. The following identity follows from these definitions $$\overline{{\bf b}}{\bf b} +\overline{{\bf a}}{\bf a} = I_{N-n}, \quad k\in\text{Re}.
\label{scRH}$$ For the involution (\[invol\_phi\]) the factorizations satisfy $$S^\dagger_\pm(k) = \overline{S}_\pm(k), \quad k\in \text{Re}.$$ Hence, $\overline{{\bf b}} = {\bf b}^\dagger$ and $\overline{{\bf a}} =
{\bf a}^\dagger$ in the case of involution.
Considering the product $\Phi^{-1}_-\Phi_+$ we obtain the problem of analytic factorization of a matrix $G(k)$ given on the real line, i.e., the matrix RH problem: $$\Phi^{-1}_-(x,k)\Phi_+(x,k) = e^{ikxA} G(k) e^{-ikxA}, \quad k\in \text{Re}
\label{RH}$$ and $\Phi_\pm(x,k)\to I$ for $k\to \infty$. Here the matrix $G=\overline{S}_+S_+\equiv\overline{S}_-S_-$ reads $$G = \left( \begin{array}{cc} I_n & {\bf b} \\ \overline{{\bf b}} &
I_{N-n} \end{array}\right).
\label{G}$$
As it was mentioned, the RH problem is a nonlinear mapping between the potential $Q(x)$ and the set of the spectral data, which are necessary for unique identification of the solution to (\[RH\]). For instance, given a potential, one can obtain the spectral data by solving the spectral equation and its adjoint for $\Phi_+(x,k)$ and $\Phi^{-1}_-(x,k)$. Conversely, by asymptotic expansion of $\Phi_\pm(x,k)$ as $k\to\infty$ one recovers the potential. The asymptotic expansion of $\Phi_\pm(x,k)$ can be derived via integration by parts (in the blocks with $e^{\pm
2ikx}$) in the Volterra integral equations for $J_\pm$ and $J^{-1}_\pm$. It reads $$\Phi_\pm(x,k) = I - \left( AQ(x) + i\int\limits_{-\infty}^x\text{d}\xi\,
Q^2(\xi)H_2 +i\int\limits^{\infty}_x\text{d}\xi\, Q^2(\xi)H_1
\right)\frac{1}{2k} +{\cal O}\left(\frac{1}{k^2}\right).$$ Hence, we obtain (cf. with (\[Q\])) $$Q(x) = \lim_{k\to\infty}k[\Phi_+(x,k),A] = \lim_{k\to\infty}k[A,\Phi^{-1}_-
(x,k)] . \label{QRH}$$
Solution of the RH problem: Normalization {#solution-of-the-rh-problem-normalization .unnumbered}
-----------------------------------------
Let us discuss the way of solution of the RH problem. The coordinate dependence is not important for this purpose and it is omitted below. In general, the determinants $\det\Phi_+(k)$ and $\det \Phi^{-1}_-(k)$ have zeros and the RH problem is said to be non-regular or with zeros. Note that the determinants do not depend on $x$ as it readily seen from equations (\[lax1\]) and (\[adj\_lax1\]). We consider only the RH problem with zero index, i.e., $$\int\limits_{-\infty}^\infty\text{d}\,\text{ln}\{\det G(k)\} = 0,$$ assuming that $G(k)$ is non-degenerate. Since $\Phi_\pm\to I$ as $k\to\infty$, the index is equal to the difference between the number of zeros in the upper and lower half planes. Let $k_1,...,k_M$ and $\overline{k}_1, ..., \overline{k}_M$ be zeros (where some may be equal) of $\det\Phi_+(k)$ and $\det \Phi^{-1}_-(k)$, respectively. Two conditions are imposed on the zeros. First, the geometric multiplicity of a zero must be equal to its order (which we will refer to as the algebraic multiplicity). Here the geometric multiplicity of $k_j$ is defined as the dimension $d_j$ of the null space of $\Phi_+(k_j)$, i.e., $d_j=N-\text{rank}\Phi_+(k_j)$. Trivially, the two multiplicities are equal to 1 in the case of simple zeros. In the case of involution zeros of the RH problem come in complex conjugate pairs (due to formula (\[invol\_phi\])), with coinciding algebraic (and geometric) multiplicities within each pair. (Due to the involution the index is equal to zero, in this case $\det G(k)$ is real and definite.) Second, as we are mainly interested in the Hermitian potentials $Q$, i.e., in the case of the involution, we will consider only paired zeros $k_j$ and $\overline{k}_j$, $j=1,...,s\le M$, whose algebraic multiplicities are equal, however without assuming them to be complex conjugate to each other. The algebraic multiplicity $\nu_j$ of $k_j$ always satisfies the following inequality (see Appendix A) $$\nu_j\ge N - \text{rank}\Phi_+(k_j)=d_j,$$ while $\text{rank}\Phi_+ \ge \text{max}(n,N-n)$ by construction of $\Phi_+$ (\[phi+phi-\]). For instance, if $n= N-1$, there can be not more than one vector in the null space of $\Phi_+(k)$, i.e., $d_j=1$ for all $j=1,...,s$. Hence, for this reduction, zeros of $\det\Phi_+(k)$ must be simple to satisfy the equal multiplicity condition. Similar for zeros of $\Phi^{-1}_-(k)$. (More detailed consideration of the multiplicities of zeros is placed in Appendix A.)
Let $\det \Phi_+(k)$ and $\det \Phi^{-1}_-(k)$ have zeros $k_1,...,k_s$ and $\overline{k}_1,...\overline{k}_s$, respectively, with the (algebraic and geometric) multiplicities $\nu_1,..., \nu_s$. Exactly $\nu_j$ independent vector-columns $ | p^{(j)}_l \rangle $ (vector-rows $ \langle \overline{ p}{}^{(j)}_l |$), where $l=1,...,\nu_j$, correspond to each zero $k_j$ (respectively, $\overline{k}_j$), such that $$\Phi_+(k_j) | p^{(j)}_l \rangle = 0,\quad \langle
\overline{ p}{}^{(j)}_l |\Phi^{-1}_-(\overline{k}_j) = 0. \label{vec_def}$$ To specify a unique solution to the RH problem, additionally to the continuous datum $G(k)$, the set of the discrete data $\{ k_j, |
p^{(j)}_l \rangle, \overline{k}_j, \langle \overline{ p}{}^{(j)}_l |,
l=1,...,\nu_j, j=1,...,s\}$ must be given. This becomes evident from the fact that the RH problem with zeros (i.e., non-regular) can be regularized, i.e., reduced to a modified RH problem without zeros by factoring them out with some rational matrix $\Gamma(k)$: $$\Phi_+(k) = \phi_+(k) \Gamma(k),\quad
\Phi^{-1}_-(k) = \Gamma^{-1}(k)\phi^{-1}_-(k).
\label{normphi+phi-}$$ The regularized RH problem reads $$\phi^{-1}_-(k)\phi_+(k) = \Gamma(k) e^{ikxA}G(k)e^{-ikxA}\Gamma^{-1}(k),
\quad k\in \text{Re}
\label{regRH}$$ and $\phi_\pm(k)\to I$ as $k\to\infty$.
The rational matrix functions $\Gamma(k)$ and $\Gamma^{-1}(k)$, (for details see Appendix B) $$\Gamma = I - \sum_{j,l;i,m} \frac{| p^{(i)}_m
\rangle\left(D^{-1}\right)_{im,jl}\langle \overline{ p}{}^{(j)}_l |}{k -
\overline{k}_j} ,\quad
\Gamma^{-1} = I + \sum_{j,l;i,m} \frac{
| p^{(j)}_l \rangle\left(D^{-1}\right)_{jl,im} \langle \overline{
p}{}^{(i)}_m |}{k - {k}_j}, \label{Gammas}$$ where $$D_{im,jl} = \frac{\langle \overline{ p}{}^{(i)}_m | p^{(j)}_l \rangle }{k_j
- \overline{k}_i},$$ have the same zeros as $\Phi_+(k)$ and $\Phi^{-1}_-(k)$, respectively, and the same null spaces: $$\Gamma(k_j)| p^{(j)}_l \rangle = 0,\quad \langle \overline{ p}{}^{(j)}_l |
\Gamma^{-1}(\overline{k}_j) = 0,\quad l=1,...,\nu_j, \quad
j=1,...,s.$$ Moreover, $\det \Gamma =
\prod_{j=1}^s\left(\frac{k-k_j}{k-\overline{k}_j}\right)^{\nu_j}$. Precisely these properties allow us to factor out zeros of $\Phi_+(x,k)$ and $\Phi^{-1}_-(x,k)$.
It is important to emphasize that there is a freedom in choice of the basis vectors spaning the null spaces. Indeed, it is easy to verify that the regularization matrix $\Gamma$ is invariant under the transformations: $$| p^{(j)}_l \rangle \to \sum_{m=1}^{\nu_j}| p^{(j)}_{m} \rangle M^{(j)}_{m l},
\quad
\langle \overline{ p}{}^{(j)}_l | \to \sum_{m=1}^{\nu_j} \overline{M}{}^{(j)}_{lm}
\langle \overline{ p}{}^{(j)}_{m} |,
\label{invariance}$$ where $M^{(j)}$ and $\overline{M}{}^{(j)}$ are arbitrary non-degenerate $\nu_j\times\nu_j$-matrices. For the involution, due to equation (\[invol\_phi\]) we can choose $\langle
\overline{ p}^{(j)}_l | = | p^{(j)}_l \rangle^\dagger$. This evidently leads to $$\Gamma^{\dagger}(k) = \Gamma^{-1}(\overline{k}).
\label{invol_gam}$$
The $(x,t)$-dependence of the RH data can be found in the following way. The $t$-dependence of the continuous datum (which does not depend on $x$) can be derived from equations (\[RH\]) and (\[lax2\]). We obtain $$\partial_t G = [G,\Omega].
\label{G_t}$$ This equation can be cast in a more explicit form: $$\partial_t {\bf b} = -2i\omega(k) {\bf b},\quad
\partial_t \overline{{\bf b}} = 2i\omega(k) \overline{{\bf b}}.
\label{b_t}$$
Recalling that $\text{Tr}Q=0$ and $\text{Tr}V = -\text{Tr}\Omega$, one can verify that the determinants of $\Phi_\pm$ do not depend on the co-ordinates. Therefore, the zeros are co-ordinate independent as well: $$\partial_x k_j = \partial_t k_j =0,\quad \partial_x
\overline{k}_j = \partial_t \overline{k}_j = 0,\quad j=1,...,s.
\label{k_xt}$$
Now, let us derive the co-ordinate dependence of the vector-parameters. Differentiating equation (\[vec\_def\]) we obtain $$\Phi_+(k_j)\left( \partial_x | p^{(j)}_l \rangle - ik_j A | p^{(j)}_l \rangle \right) = 0, \quad
\Phi_+(k_j)\left( \partial_t | p^{(j)}_l \rangle + \Omega(k_j) | p^{(j)}_l \rangle \right) = 0.$$ Therefore $$\partial_x | p^{(j)}_l \rangle = ik_jA | p^{(j)}_l \rangle + \sum_{m=1}^{\nu_j}F^{(j)}_{lm}
| p^{(j)}_m\rangle, \quad
\partial_t| p^{(j)}_l \rangle = -\Omega(k_j) | p^{(j)}_l \rangle + \sum_{m=1}^{\nu_j}K^{(j)}_{lm}
| p^{(j)}_m\rangle,$$ where $F$ and $K$ are some matrices. Using a suitable invariance transformation of the type (\[invariance\]) one can put $F=K=0$ without loss of generality. Hence we have $$\partial_x | p^{(j)}_l \rangle = ik_jA | p^{(j)}_l \rangle, \quad
\partial_t | p^{(j)}_l \rangle = -\Omega(k_j) | p^{(j)}_l \rangle.
\label{p_xt}$$ Similarly $$\partial_x \langle \overline{ p}^{(j)}_l | = - \langle \overline{ p}^{(j)}_l | i\overline{k}_jA,\quad
\partial_t \langle \overline{ p}^{(j)}_l | = \langle \overline{ p}^{(j)}_l | \Omega(\overline{k}_j).
\label{barp_xt}$$
Some comments are needed on the reconstruction of the potential $Q$. First of all, the soliton part of the potential is given by the rational matrix function $\Gamma(k)$ (\[Gammas\]). The pure soliton potentials are also called reflectionless, they solve the simplest RH problem with $G=I$, i.e., with zero reflection coefficients: ${\bf b}(k) = \overline{\bf b}(k) =0$. The discrete RH data have the following meaning. Zeros provide the amplitudes and velocities of the solitons, while the null vectors give their initial position, polarization and phase parameters. Second, the radiation part of the potential is given by the solution to the regularized RH problem (\[regRH\]). The continuous RH data ${\bf b}(k)$ and $\overline{\bf
b}(k)$ represent the nonlinear spectral densities of radiation. The regular RH problem is equivalent to some matrix integral equation of the Fredholm type (for instance, consult Ref. 46). Though the solution cannot be given in explicit form, its properties can be explored by the standard technique of the theory of Fredholm integral equations.
For the following, it is convenient to introduce $x$-independent null vectors $| P^{(j)}_l \rangle$ and $\langle \overline{P}^{(j)}_l | $ setting $$| p^{(j)}_l \rangle = e^{ik_j x A} | P^{(j)}_l \rangle, \quad
\langle \overline{ p}{}^{(j)}_l | = \langle \overline{P}{}^{(j)}_l |
e^{-i\overline{k}_j xA}.
\label{def_P}$$ Then, we have an $x$-independent set of the RH data, which we will call the spectral data: $\{ {\bf b}(k), \overline{{\bf b}}(k),
k_j, | P^{(j)}_l \rangle, \overline{k}_j, \langle \overline{P}{}^{(j)}_l |,
l=1,...,\nu_j, j=1,...,s\}$.
As the illustrative example, consider one-soliton solution. It is given by $\Gamma(k)$ having only one pole, say $k_1=i\eta+\xi$. In this case $$Q = (k_1 - \overline{k}_1)
[A,P_r],\quad P_r = \sum_{l,m=1}^{\nu_1}| p_l \rangle
\left(\Delta^{-1}\right)_{lm}
\langle \overline{ p}_m |,$$ with $\Delta_{lm} = \langle \overline{ p}_l | p_m \rangle$. Here $\langle \overline{ p}_l | \Phi^{-1}_-(\overline{k}_1) = 0$ and $\Phi_+(k_1)| p_l \rangle = 0$. In the particular case of $n=N-1$ and the Hermitian potential we get the vector soliton solution, for which $$P_r = \frac{|p\rangle\langle p|}{\langle p|p\rangle},\quad
|p\rangle = \exp\left\{ik_1xA\right\}|P\rangle.$$ Define complex parameters $C_l= P_l/P_N$. The $t$-dependence of $C_l$ then follows from equation (\[p\_xt\]). It can be accounted for in a convenient way by introducing real $t$-dependent parameters $\bar{x}$, and $\delta_l$, $l=1,...,n$, the soliton position and phases of its components. Let $C_l =
\theta_l e^{2(\eta-i\xi)\bar{x}}$, where $\theta_l = s_le^{i\delta_l}$. The amplitudes $s_l$, satisfying $$\sum_{l=1}^n s_l^2=1,
\label{sdeltas}$$ describe polarization of the multi-component soliton. Integrating equation (\[p\_xt\]) we obtain $$\bar{x} = \bar{x}_0 + \frac{\text{Im}\{\omega(i\eta
+\xi)\}t}{\eta},\quad
\delta_l = {\delta_l}_0+\frac{2}{\eta}\text{Im}
\left\{\left(\xi-i\eta\right)\omega(i\eta+\xi)\right\}t.
\label{tdependence}$$ The vector soliton then takes the form $$q_l = 2i\eta\theta_l e^{i(\xi/\eta )z }
\text{sech}z
\label{one-sol}$$ with $z=2\eta(x-\bar{x})$. For instance, for the vector NLS equation $\omega(k) = 2k^2$ and we arrive at the $n$-component generalization of the well-known vector soliton solution for the two-component NLS equation \[12\].
Perturbation-induced evolution of the spectral data {#pert}
===================================================
A perturbation of the integrable PDE, following from the Lax representation (\[lax1\])-(\[lax2\]), can be written in the form $$iQ_t -V_x + [ikA + iQ, V] = R,
\label{perturb_eq}$$ where $R$ (which is $k$-independent in case of the Zakharov-Shabat spectral problem) represents the terms destroying integrablity. The perturbation matrix $R$, similar to the potential $Q$, satisfies $[A,R]=0$. For the Hermitian potential $Q^\dagger = Q$ we get also $R^\dagger = - R$.
Below we derive evolution equations for the spectral data with account for arbitrary perturbation. For simplicity of the presentation we will frequently omit the explicit dependence on $x$ and $t$. To distinguish between the “integrable" and “perturbation" contributions to the evolution let us assign the variational derivatives to the latter. For instance, the perturbation-induced evolution in equation (\[perturb\_eq\]) reads $$i\frac{\delta Q}{\delta t} = R,\quad \text{or, explicitly}\quad
i\frac{\delta {\bf q}}{\delta t} = {\bf r}\quad \text{and}\quad
i\frac{\delta \overline{\bf q}}{\delta t} = -\overline{\bf r},$$ where ${\bf r} = R_{I,II}$ and $\overline{\bf r} = - R_{II,I}$. For Hermitian $Q$, $ \overline{\bf r}= {\bf r}^\dagger $.
We start with derivation of evolution equations for $\Phi_+$ and $\Phi^{-1}_-$, the solution to the RH problem (\[RH\]). In other terms, we are going to derive the generalized Lax pair (see equations (\[glax1\]) and (\[ev\_phi\]) below) for the [*perturbed*]{} nonlinear PDE (\[perturb\_eq\]). Differentiation of equation (\[lax1\]) with respect to $t$ gives $$\partial_x \left(\frac{\delta \Phi}{\delta t}\right) = ik[A,\frac{\delta
\Phi}{\delta t}]
+ iQ\frac{\delta \Phi}{\delta t} + R\Phi,$$ and consequently $$\partial_x \left(\Phi^{-1}\frac{\delta \Phi}{\delta t} \right) = ik[A,\Phi^{-
1}\frac{\delta \Phi}{\delta t}]
+ \Phi^{-1}R\Phi.$$ Let us integrate the above formula for $\Phi = J_\pm$. We get $$\frac{\delta J_\pm}{\delta t}(x,k) = J_\pm(x,k)e^{ikxA}\left(\;
\int\limits_{\pm\infty}^x\text{d}\xi\, e^{-ik\xi A}J^{-
1}_\pm(\xi,k)R(\xi)J_\pm(\xi,k)e^{ik\xi A}\right)
e^{-ikx A}.$$ Here we have used that $\delta J_\pm \to 0$ as $x\to \pm\infty$. Let us employ the relation between $J_\pm$ and $\Phi_+$ to rewrite this formula in a more convenient form. Using (\[rep\_factor\_a\]) we get $$\frac{\delta J_\pm}{\delta t}(x,k) = J_\pm(x,k)e^{ikxA} S_\pm(k)
\Upsilon(\pm\infty,x;k)
S^{-1}_\pm(k) e^{-ikxA}.
\label{varJ1}$$ In much the same way, using (\[rep\_factor\_b\]), we derive $$\frac{\delta J^{-1}_\pm}{\delta t}(x,k)=-e^{ikxA}\overline{S}{}^{-1}_\pm(k)
\overline{\Upsilon} (\pm\infty,x;k)\overline{S}_\pm(k)
e^{-ikxA}J^{-1}_\pm(x,k).
\label{varJ2}$$ Here we have introduced the notations: $$\Upsilon(\pm\infty,x;k) = \int\limits_{\pm\infty}^x\text{d}\xi\,
e^{-ik\xi A}\Phi^{-1}_+ R \Phi_+ e^{ik\xi A},
\label{Ups}$$ $$\overline{\Upsilon}(\pm\infty,x;k) = \int\limits_{\pm\infty}^x\text{d}\xi\, e^{-
ik\xi A}
\Phi^{-1}_- R \Phi_- e^{ik\xi A}.
\label{barUps}$$ These matrix functionals will enter every formula describing perturbation-induced evolution of the spectral data. For the involution, due to formula (\[invol\_phi\]) these matrix functionals satisfy $$\Upsilon^\dagger(\pm\infty,x;k) = -
\overline{\Upsilon}(\pm\infty,x,\overline{k}).
\label{invol_ups}$$ Using equations (\[varJ1\]), (\[varJ2\]), and the definition of the scattering matrix $S$ (\[S\]) we get $$\frac{\delta S}{\delta t} = e^{-ikxA}\frac{\delta }{\delta t}\left( J^{-1}_+
J_-\right) e^{ikxA}
= - \overline{S}{}^{-1}_+\overline{\Upsilon}(\infty,x)\overline{S}_+ S + S S_-
\Upsilon(- \infty,x)
S^{-1}_-$$ $$= S_+ \Upsilon(- \infty,x) S_-^{-1} + \overline{S}{}^{-1}_+
\overline{\Upsilon}(x,\infty)\overline{S}_-.$$ Setting $x\to\pm\infty$ in this formula produces two simple equivalent formulae: $$\frac{\delta S(k) }{\delta t} = S_+(k) \Upsilon(- \infty,\infty;k) S_-^{-1}(k)
\equiv
\overline{S}{}^{-1}_+(k) \overline{\Upsilon}(-\infty,\infty;k)\overline{S}_-
(k).
\label{varS}$$ Now we can easily obtain the perturbation-induced evolution of the solution to the RH problem. Taking into account the definitions of $S_\pm
$ and $\overline{S}_\pm$ and (\[rep\_factor\]) we write:
\[var\_phi\] $$\frac{\delta \Phi_+ }{\delta t}
= \frac{\delta J_+ }{\delta t}e^{ikxA}S_+e^{-ikxA} + J_+e^{ikxA}
\frac{\delta }{\delta t}
( H_1 + SH_2)e^{-ikxA}$$ $$= \Phi_+e^{ikxA}\left( - \Upsilon (x,\infty) + \Upsilon(-
\infty,\infty)H_2\right)e^{-ikxA}
= \Phi_+e^{ikxA}\Pi e^{-ikxA},
\label{dphi+}$$ $$\frac{\delta \Phi^{-1}_- }{\delta t}
= e^{ikxA}\overline{S}_+e^{-ikxA}\frac{\delta J^{-1}_+ }{\delta t}
+ e^{ikxA} \frac{\delta }{\delta t}( H_1 + H_2S^{-1}) J^{-1}_+e^{-ikxA}$$ $$= e^{ikxA}\left(\overline{\Upsilon}(x,\infty) - H_2\overline{\Upsilon}(-
\infty,\infty)\right)
\Phi^{-1}_-e^{-ikxA}
=- e^{ikxA}\overline{\Pi}\Phi^{-1}_-e^{-ikxA}.
\label{dphi-}$$
The r.h.s.’s of the above formulae contain the evolution functionals, $$\Pi(x,k) = - \Upsilon (x,\infty;k)H_1 + \Upsilon(-\infty,x;k)H_2,
\label{Pi}$$ $$\overline{\Pi}(x,k) = H_1\overline{\Upsilon}(x,\infty;k) -
H_2\overline{\Upsilon}(-\infty,x;k),
\label{barPi}$$ which account for the perturbation-induced evolution of the solution to the RH problem. As follows from formula (\[invol\_ups\]), for the case of the involution the evolution functionals satisfy $$\Pi^\dagger(k) = - \overline{\Pi}(\overline{k}).
\label{invol_pi}$$ From the definitions (\[Pi\]), (\[barPi\]) and also (\[Ups\]), (\[barUps\]) it is easy to see that the matrices $e^{ikxA}\Pi(k) e^{-ikxA}$ and $
e^{ikxA}\overline{\Pi}(k)e^{-ikxA}$ are meromorphic and bounded in the upper and lower half planes of the $k$-plane, respectively. They have simple poles respectively at zeros of $\det\Phi_+(k)$ and $\det\Phi^{-1}_-(k)$.
Let us write down the generalized Lax representation for the perturbed PDE (\[perturb\_eq\]) in terms of $\Phi_+(x,k)$: $$\partial_x \Phi_+= ik[A,\Phi_+] + iQ\Phi_+,
\label{glax1}$$ $$\partial_t \Phi_+ = \Phi_+ \Omega + V\Phi_+ + \Phi_+ e^{ikxA}\Pi e^{-ikxA}.
\label{ev_phi}$$ It is easy to check by direct calculation that the compatibility condition for the above linear system is equivalent to the perturbed equation (\[perturb\_eq\]) (indeed, $\partial_x \Pi = e^{-ikxA}\Phi^{-1}_+R\Phi_+e^{ikxA}$).
Now, let us derive the perturbation-induced evolution of the spectral data. From (\[var\_phi\]) we immediately obtain $$\frac{\delta G}{\delta t} = \frac{\delta }{\delta t} \left( e^{-ikxA}\Phi^{-1}_-
\Phi_+e^{ikxA}\right)
= G\Pi - \overline{\Pi}G.$$ Hence, the complete evolution of the continuous datum reads $$\partial_t G = [G,\Omega] + G\Pi - \overline{\Pi}G,
\label{evG}$$ where we have taken into account the integrable evolution given by (\[G\_t\]). It is important to notice that the l.h.s. of (\[evG\]) does not depend on $x$. Therefore, we can put $x\to\pm\infty$ in this equation to simplify it. Equation (\[evG\]) can be rewritten in a more explicit form. Using the definition (\[G\]), formulae (\[Pi\]) and (\[barPi\]) we get two equivalent evolution equations for [**b**]{} corresponding to the two limits $x\to\pm\infty$:
$$\partial_t {\bf b} = -2i\omega(k){\bf b} + {\bf b}H_2\Upsilon(-
\infty,\infty;k)H_2
+ H_1\Upsilon(-\infty,\infty;k)H_2,
\label{ev_b1}$$
$$\partial_t {\bf b} = -2i\omega(k){\bf b} - H_1\overline{\Upsilon}(-
\infty,\infty;k)H_1{\bf b}
- H_1\overline{\Upsilon}(-\infty,\infty;k)H_2.
\label{ev_b2}$$
Similarly,
$$\partial_t \overline{{\bf b}} = 2i\omega(k)\overline{{\bf b}}
+ \overline{{\bf b}}H_1\Upsilon(-\infty,\infty;k)H_1 + H_2\Upsilon(-
\infty,\infty;k)H_1,
\label{ev_barb1}$$
$$\partial_t \overline{{\bf b}} = 2i\omega(k) \overline{{\bf b}}
- H_2\overline{\Upsilon}(-\infty,\infty;k)H_2 \overline{{\bf b}}
- H_2\overline{\Upsilon}(-\infty,\infty;k)H_1.
\label{ev_barb2}$$
Evolution of zeros $k_j$ and $\overline{k}_j$ of $\det\Phi_+(k)$ and $\det\Phi^{-1}_-(k)$, respectively, is derived by differentiation of the determinants. In the integrable limit zeros do not depend on $t$. Then, for instance, $$\frac{\text{d} k_j}{\text{d} t} = -\left. \frac{\partial_t \det
\Phi_+(k)}{\partial_k \det \Phi_+(k)}
\right|_{k=k_j} = -\left.\frac{ \text{Tr}\{\Pi(k)\}\det\Phi_+(k)}{\partial_k
\det\Phi_+(k)}
\right|_{k=k_j}.$$ To calculate the r.h.s. in this formula recall that the evolution functional $\Pi$ has simple pole at $k=k_j$ and that $\det\Phi_+(k)$ can be written as $$\det\Phi_+(k) = \det\phi_+(k)\det\Gamma(k) = \det\phi_+(k)\prod_{i=1}^s
\left(\frac{ k - k_i}{k-\overline{k}_i}\right)^{\nu_i},$$ where $\det\phi_+(k)\ne0$. Simple calculations give $$\frac{\text{d} k_j}{\text{d} t} = - \frac{\text{Tr} \{\text{Res}
\Pi(k_j)\}}{\nu_j}.$$ Here “Res" denotes the residue of $\Pi(k)$ at $k=k_j$. Noticing that the l.h.s. does not depend on $x$, we simplify the above equation setting $x\to\pm\infty$: $$\frac{\text{d} k_j}{\text{d} t} = - \frac{\text{Tr}\{\text{Res} \Upsilon (-
\infty, \infty; k_j)H_2\}}{\nu_j}
\equiv \frac{\text{Tr}\{\text{Res} \Upsilon (-\infty, \infty;
k_j)H_1\}}{\nu_j}.
\label{ev_k}$$ Similarly, using $\text{det}\Phi^{-1}_-$, we get $$\frac{\text{d} \overline{k}_j}{\text{d} t} = -
\frac{\text{Tr}\{\text{Res}\overline{\Upsilon}
(-\infty,\infty;\overline{k}_j) H_2\}}{\nu_j}
\equiv \frac{\text{Tr}\{\text{Res} \overline{\Upsilon}(-
\infty,\infty;\overline{k}_j) H_1\}}{\nu_j}.
\label{ev_bark}$$
In derivation of the perturbation-induced evolution of the null vectors we will use the following remarkable identities (written for the $x$-independent null vectors, see (\[def\_P\])) $$\text{Res}\Pi(k_j) | P^{(j)}_l \rangle = - \frac{\delta k_j}{\delta t}|
P^{(j)}_l \rangle, \quad
\langle \overline{ P}{}^{(j)}_l | \text{Res}\overline{\Pi}(\overline{k}_j)
= \langle \overline{ P}{}^{(j)}_l | \frac{\delta \overline{k}_j}{\delta t},
\label{id_vec}$$ as well as other two identities: $$\Phi_+(k_j)e^{ik_j xA}\text{Res}\Pi(k_j) = 0,\quad
\text{Res}\overline{\Pi}(\overline{k}_j)
e^{-i\overline{k}_jxA}\Phi^{-1}_-(\overline{k}_j)= 0,
\label{id_phi}$$ which follow from (\[var\_phi\]). To verify the identities (\[id\_vec\]) one can proceed as follows. Introduce functions $F_j^{(+)}(k) = (k - k_j)\Phi^{-1}_+(k)$ and $F_j^{(-)}(k) = (k -\overline{k}_j)\Phi_-(k)$. The regularization matrices $\Gamma^{-1}(k)$ and $\Gamma(k)$ have simple poles at $k=k_j$ and $k=\overline{k}_j$, respectively. Hence the introduced matrix functions are holomorphic in some neighborhoods of these points. Now compute the product $$\left\{F^{(+)}_j(k) \frac{\delta \Phi_+(k)}{\delta t}| p^{(j)}_l
\rangle\right\}_{k=k_j}
= \left\{ (k-k_j)e^{ikxA}\Pi(k)e^{-ikxA}| p^{(j)}_l \rangle\right\}_{k=k_j}$$ $$= e^{ik_j xA}\text{Res}\Pi(k_j)e^{-ik_j xA}| p^{(j)}_l \rangle.$$ On the other hand, $$\left\{F^{(+)}_j(k) \frac{\delta \Phi_+(k)}{\delta t}\right\}_{k=k_j}
= \left\{ \frac{\delta }{\delta t}(k - k_j) I\right\}_{k=k_j}
- \left\{\frac{\delta F^{(+)}_j(k)}{\delta t}\Phi_+(k)\right\}_{k=k_j}$$ $$= -\frac{\delta k_j}{\delta t}I - \frac{\delta F^{(+)}_j(k_j)}{\delta
t}\Phi_+(k_j).$$ Multiplication by $| p^{(j)}_l \rangle $ of the latter formula and comparison with the former leads to the first identity in (\[id\_vec\]). The second one can be checked in similar way.
Now let us derive evolution equations for the $x$-independent null vectors defined in (\[def\_P\]). To this goal we simply differentiate equation (\[vec\_def\]) $$\frac{\delta }{\delta t} \left( \Phi_+(k_j)e^{ik_jxA}| P^{(j)}_l \rangle \right)
= \left\{\Phi_+(k)e^{ikxA} \Pi(k)|P^{(j)}_l\rangle\right\}_{k=k_j}
+ \frac{\delta k_j}{\delta t}\frac{\partial \Phi_+(k_j)}{\partial k}e^{ik_j
xA}|P^{(j)}_l \rangle$$ $$+\frac{\delta k_j}{\delta t}\Phi_+(k_j)ixAe^{ik_j xA} |P^{(j)}_l \rangle
+\Phi_+(k_j)e^{ik_jxA}\frac{\delta |P^{(j)}_l \rangle }{\delta t} = 0.$$ Denote $\Pi_r(k_j)$ the value of the regular part of $\Pi(k)$ at $k =k_j$, $$\Pi_r(k_j) = \left\{\Pi(k) - \frac{\text{Res}\Pi(k_j)}{k - k_j}\right\}_{k=k_j}.
\label{Pi_r}$$ Then, using (\[id\_vec\]) and (\[id\_phi\]) to cancel out secular terms, we arrive at $$\Phi_+(k_j)e^{ik_j xA}\left\{ \frac{\delta |P^{(j)}_l \rangle }{\delta t} +
\Pi_r(k_j)
|P^{(j)}_l \rangle \right\} = 0.$$ Using the same arguments as in section \[secRH\] for derivation of the integrable evolution we get, adding the latter, $$\frac{\text{d} |P^{(j)}_l \rangle }{\text{d} t} = - \Omega(k_j)|P^{(j)}_l
\rangle
- \Pi_r(k_j)|P^{(j)}_l \rangle.
\label{var_P}$$ Similarly, $$\frac{\text{d} \langle \overline{P}{}^{(j)}_l |}{\text{d} t} = \langle
\overline{P}{}^{(j)}_l |
\Omega(\overline{k}_j) + \langle \overline{P}{}^{(j)}_l |
\overline{\Pi}_r(\overline{k}_j),
\label{var_barP}$$ where $$\overline{\Pi}_r(\overline{k}_j) = \left\{\overline{\Pi}(k) -
\frac{\text{Res}\overline{\Pi}(k_j)}
{k - \overline{k}_j}\right\}_{k=\overline{k}_j}.
\label{barPi_r}$$ In the case of the involution equation (\[var\_barP\]) is Hermitian conjugate to (\[var\_P\]). Note that the l.h.s.’s of equations (\[var\_P\]) and (\[var\_barP\]) do not depend on $x$. Hence we can sent $x\to\pm\infty$ to considerably simplify these equations. Consider, for instance, (\[var\_P\]). Letting $x\to\infty$ and introducing the notations $$|P^{(j)}_l \rangle = H_1 |P^{(j)}_l \rangle + H_2 |P^{(j)}_l \rangle \equiv
|P^{(j)}_l, 1\rangle + |P^{(j)}_l , 2\rangle$$ and $\Upsilon^{(1)} = H_1\Upsilon H_2$, $\Upsilon^{(2)} = H_2\Upsilon H_2$ we get the following system
\[sys\] $$\frac{\text{d} |P^{(j)}_l, 1\rangle }{\text{d} t} = - i\omega(k_j) |P^{(j)}_l,
1\rangle
- \Upsilon^{(1)}_r(-\infty,\infty;k_j) |P^{(j)}_l , 2\rangle,
\label{P_1}$$ $$\frac{\text{d} |P^{(j)}_l , 2\rangle }{\text{d} t} = \left\{
i\omega(k_j) - \Upsilon^{(2)}_r(-\infty,\infty;k_j)\right\} |P^{(j)}_l,
2\rangle.
\label{P_2}$$
Here $\Upsilon_r (k_j)$ is value of the regular part of the matrix $\Upsilon(k)$ at $k=k_j$: $$\Upsilon_r(k_j) = \left\{\Upsilon(k) - \frac{\text{Res}\Upsilon(k_j)}{k -
k_j}\right\}_{k=k_j}.$$
When dealing with vector PDEs, i.e., for $n=N-1$, one can define the polarization-phase parameters of vector solitons as quotients of components of the null vectors (note that the zeros are simple in this particular case). Let $C^{(j)}_{l} = P^{(j)}_l/ P^{(j)}_N$, $l=1,...,n=N-1$, where $|P^{(j)}\rangle = (P^{(j)}_1, ... ,P^{(j)}_{N})^T$. Then from (\[sys\]) we obtain: $$\frac{\text{d} C^{(j)}_l}{\text{d} t} = \left\{ -2i\omega(k_j)
+ {\Upsilon_r}_{NN}(-\infty,\infty;k_j) \right\}C^{(j)}_l
- {\Upsilon_r}_{lN}(-\infty,\infty;k_j),\quad l=1,...,n.$$ This particular case ($n=N-1$) contains the vector NLS (\[vectorNLS\]) and the complex modified KdV (\[CMKdV\]) equations as examples. In view of considerable importance of such vector nonlinear PDEs, we formulate the result of this section in the following theorem.
[**Theorem:**]{} [*Let* ]{} $$iQ_t + V_0(Q,Q_x,Q_{xx},...) = \epsilon R(x,t,Q,Q_x,Q_{xx},...),
\label{evolution}$$ [*be a perturbed nonlinear PDE associated with the $N\times N$ matrix Zakharov-Shabat spectral problem,*]{} $$\partial_x \Phi = ik[A,\Phi] + iQ(x,t)\Phi, \quad
A = \text{diag} (I_n,-1),
\label{spectral}$$ [*where ($n=N-1$)*]{} $$Q= \left( \begin{array}{cc}0 & {\bf q} \\ {\bf q}^\dagger & 0
\end{array}\right),\quad
R = \left( \begin{array}{cc} 0 & {\bf r} \\ -{\bf r}^\dagger & 0
\end{array}\right),\quad {\bf q} = \left(q_1, q_2, ..., q_n\right)^T, \quad
{\bf r} = \left(r_1, r_2, ..., r_n\right)^T.$$ [*Here the matrix function*]{} $V_0$ [*represents the limiting integrable evolution given by the dispersion relation*]{} $\Omega(k) =
i\omega(k)A$, [*while*]{} $R$ [*contains the terms destroying integrability ($\epsilon$ is a small parameter). Then the perturbed evolution is equivalent to the following evolution of the spectral data:*]{} $$\frac{\text{d} k_j}{\text{d} t} = -
\epsilon\text{Res}\Upsilon_{NN}(t,k_j),
\label{equ1}$$ $$\frac{\text{d} C^{(j)}_l}{\text{d} t} = \left\{ -2i\omega(k_j) +
\epsilon{\Upsilon_r}_{NN}(t,k_j)\right\}
C^{(j)}_l - \epsilon{\Upsilon_r}_{lN}(t,k_j), \quad l=1,...,n,
\label{equ2}$$ $$\frac{\partial b_l(k) }{\partial t}=
\left\{ -2i\omega(k) + \epsilon\Upsilon_{NN}(t,k)\right\}
b_l(k) + \epsilon\Upsilon_{lN}(t,k), \quad l=1,...,n.
\label{equ3}$$ [*Here*]{} $$\Upsilon(t,k) = \int\limits_{-\infty}^\infty\text{d}\,x
e^{-ikxA}\Phi_+^{-1}(x,t,k)
R(x,t)\Phi_+(x,t,k)e^{ikxA},
\label{ups}$$ $$\Upsilon_r(t,k_j) = \left\{\Upsilon(t,k)
-\frac{\text{Res}\Upsilon(t,k_j)}{k - k_j}\right\}_{k=k_j}.$$ [*The matrix function*]{} $\Phi_+(x,t,k)$ [*solves the spectral problem (\[spectral\]) and the RH problem*]{} $$\Phi_+^{\dagger}(x,t,k)\Phi_+(x,t,k) = e^{ikxA} G(k,t) e^{-ikxA}, \quad
k\in \text{Re},$$ $$\Phi_+(k)\to I,\quad k\to\infty$$ [*of analytic factorization of matrix*]{} $G(k,t)$, $$G(k,t) = \left( \begin{array}{cc} I_n & {\bf b}(k,t) \\ {\bf
b}^\dagger(k,t)& 1 \end{array}\right),
\quad {\bf b} = \left(b_1,b_2,...,b_n\right)^T.$$ [*The Riemann-Hilbert problem has (simple) zeros*]{} $k_j(t)$, $j=1,...,s$, [*given by* ]{} $a(k_j,t) = \det\Phi_+(x,t,k_j) = 0$. [*The vector-columns*]{} $ |C^{(j)} (t)\rangle = \left(C^{(j)}_1(t),
C^{(j)}_2(t),...,C^{(j)}_n(t),1\right)$ [*are the null vectors of the matrices*]{} $\Phi_+(x,t,k_j)e^{ik_jxA}$: $$\Phi_+(x,t,k_j)e^{ik_jxA} |C^{(j)}(t)\rangle = 0,\quad j=1,...,s.$$ [*The initial spectral data are obtained via solution of the spectral equation (\[spectral\]) and represent the spectral characterization of the potential*]{} $Q(x,0)$. [*For real*]{} $k$, [*the spectral densities of radiation*]{} $b_l(k,t)$ [*and the function*]{} $a(k,t)$ [*satisfy the following identity* ]{} $$\overline{a}a =1-\sum_{l=1}^n \overline{b}_lb_l.$$ [*The potential ${\bf q}(x,t)$ is reconstructed by the formula*]{} $$q_l(x,t) = 2\lim_{k\to\infty}k(\Phi_+)_{l,N}(x,t,k),\quad l=1,...,n.$$
Some comments are necessary on the use of equations (\[equ1\])-(\[equ3\]) for the spectral data. These equations are [ *exact*]{}, i.e., they account for the perturbation exactly. As a consequence, for the non-integrable PDE (\[evolution\]), these equations are non-closed ODEs. Equations (\[equ1\])-(\[equ3\]) are non-closed because they contain explicitly the matrix $\Phi_+(x,t,k)$, solution of the RH problem, obtaining which requires knowledge of the spectral data. Therefore, equations (\[equ1\])-(\[equ3\]) serve as the generating equations for the perturbation expansion: expanding the spectral data into the asymptotic power series in $\epsilon$, one obtains the sequence of closed approximate ODEs for the spectral data. On this way, one does not need to solve the RH problem – the computations are algebraic. In the next section we consider a single multi-component soliton as an example.
Multi-component soliton under perturbations {#exmpl}
===========================================
In this section we apply the theorem for construction of the perturbation theory for a single multi-component soliton. This can be done without specifying the dispersion relation $\omega(k)$ determining the evolution of the spectral data in the unperturbed PDE. Hence our results apply to all nearly integrable vector PDEs associated with the Zakharov-Shabat spectral problem. Below we derive equations describing the evolution of the soliton parameters and give formulae for the first-order radiation. For a single vector soliton given by formula (\[one-sol\]), i.e., $q_l = 2i\eta\theta_l e^{i(\xi/\eta )z }\text{sech}z$, where $ z=2\eta(x-\bar{x})$, the regularization matrix $\Gamma$ has the form $$\Gamma = I - \frac{i\eta}{(k-\overline{k}_1)\cosh{z}}
\left( \begin{array}{cc} e^{-z}|\theta\rangle \langle\theta | &
e^{i(\xi/\eta)z}
|\theta \rangle \\ \langle \theta |e^{-i(\xi/\eta)z} &
e^{z} \end{array}\right).
\label{gamma1}$$ Here $k_1=\xi+i\eta$ and we have used the vector notations $|\theta\rangle = (\theta_1,...,\theta_n)^T$ and $\langle \theta|
= (\overline{\theta}_1,...\overline{\theta}_n)$. To simplify some of the calculations introduce the following basis (in the $n$-dimensional subspace) $$|\theta^{(1)}\rangle = |\theta\rangle,\quad |\theta^{(2)}\rangle,\quad
...,\quad |\theta^{(n)}\rangle; \quad \langle
\theta^{(l)}|\theta^{(m)}\rangle = \delta_{lm}.$$ The explicit form of the vectors $|\theta^{(l)}\rangle$ for $l=2,...,n$ will not be needed at all. Also we will use the basis $$|e_1\rangle=(1,0,...,0)^T,\quad |e_2\rangle=(0,1,0,...,0)^T,\quad
...,\quad |e_n\rangle=(0,...,0,1)^T.$$ With the help of the unitary transformation matrix $\Xi$, defined as $$\Xi = \left(\begin{array}{cc} B & 0 \\ 0 & 1\end{array}\right),\quad
B = \sum_{l=1}^n|e_l\rangle\langle\theta^{(l)}|,
\label{Xi}$$ the regularization matrix $\Gamma$ can be considerably simplified: $$\widetilde{\Gamma} = \Xi\Gamma\Xi^{-1} = I -
\frac{i\eta}{(k-\overline{k}_1)\cosh{z}}
\left( \begin{array}{cc} e^{-z}|e_1\rangle \langle e_1 | &
e^{i(\xi/\eta)z} |e_1 \rangle \\ \langle e_1
|e^{-i(\xi/\eta)z} & e^{z} \end{array}\right).
\label{gamma2}$$ Note the evident property $\widetilde{\Gamma}^{-1} = \widetilde{\Gamma}^\dagger$.
The transformation matrix $\Xi$ depends on $t$ due to the perturbation-induced time dependence of the $\theta$-parameters, but does not depend on $x$. This simple fact allows us to use the transformation with $\Xi$ inside the integrals defining $\Upsilon$ (\[ups\]). Taking into account that $\Phi_+=\Gamma$ for the pure soliton solution of the unperturbed PDE, in the first order we get $$\widetilde{\Upsilon} =
\Xi\Upsilon\Xi^{-1} = \int\limits_{-\infty}^\infty\text{d}x\, e^{-ikxA}
\widetilde{\Gamma}^{-1}\widetilde{R}\widetilde{\Gamma} e^{ikxA}.$$ Here we have defined $$\widetilde{R} = \Xi R\Xi^{-1} = \left(\begin{array}{cc}
0 & B|r\rangle \\ -\langle r|B^\dagger & 0 \end{array}\right),
\label{Rmod}$$ where, for convenience of the presentation below, we have changed the notation for the perturbation: $|r\rangle={\bf r}=(r_1,...,r_n)^T$.
In fact, we will need only one diagonal element $$\Upsilon_{NN} = \left(\Xi^{-1} \widetilde{\Upsilon}\Xi\right)_{NN}
= \int\limits_{-\infty}^\infty\text{d}x\,
\left(\widetilde{\Gamma}^\dagger\widetilde{R}\widetilde{\Gamma}\right)_{NN}$$ and the following non-diagonal matrix elements $$\Upsilon_{lN} = \left(\Xi^{-1} \widetilde{\Upsilon}\Xi\right)_{lN}
= \sum_{m=1}^n B^{-1}_{lm}\widetilde{\Upsilon}_{mN} =
\theta_l \widetilde{\Upsilon}_{1N}
+ \sum_{m=2}^n B^{-1}_{lm}\widetilde{\Upsilon}_{mN}
=\theta_l \widetilde{\Upsilon}_{1N} + F_l,$$ where $l=1,...,n$. The second term simplifies as follows $$F_l= \int\limits_{-\infty}^\infty\text{d}x\, e^{-2ikx}
\sum_{m=2}^n B^{-1}_{lm}\widetilde{R}_{mN}\widetilde{\Gamma}_{NN}
=\int\limits_{-\infty}^\infty\text{d}x\, e^{-2ikx}\widetilde{\Gamma}_{NN}
\sum_{m=2}^n \langle e_l|B^{-1}|e_m\rangle
\langle e_m|B|r\rangle$$ $$=\int\limits_{-\infty}^\infty\text{d}x\, e^{-2ikx}
\widetilde{\Gamma}_{NN}\left(r_l - \theta_l\langle \theta|r\rangle\right).
\label{Fj}$$ After simple calculations we get: $$\Upsilon_{NN} = \frac{i}{4}\int\limits_{-\infty}^\infty\text{d}z\,
\text{sech}^2z \left(\frac{e^{-z}}{k-k_1}+\frac{e^z}{k-\overline{k}_1}
\right)\left(r_0(z) +\overline{r}_0(-z)\right),
\label{gamNN}$$ $$\Upsilon_{lN} = \frac{\theta_l}
{8\eta}\int\limits_{-\infty}^\infty\text{d}z\,
\text{sech}^2z\frac{\exp\{-2ikx+i(\xi/\eta)z\}}{(k-k_1)(k-\overline{k}_1)}
\biggl\{-4\eta^2\overline{r}_0(-z)$$ $$+\left[e^{-2z}(k-\overline{k}_1)+e^{2z}(k-k_1)\right]^2
r_0(z)\biggr\} + F_l.
\label{gamjN}$$ Here we have used the notation $$r_0 = e^{-i(\xi/\eta)z}\widetilde{R}_{1N}
= e^{-i(\xi/\eta)z}\langle\theta|r\rangle
=e^{-i(\xi/\eta)z}\sum_{l=1}^n\overline{\theta}_l r_l.
\label{r0}$$
Evolution of the soliton parameters
-----------------------------------
Let us first derive evolution equations for the soliton parameters $\eta$, $\xi$, $\bar{x}$, and $\theta_l$, $l=1,...,n$. Using the identity $-2ik_1x +i(\xi/\eta)z = z +2\eta\bar{x}-2i\xi\bar{x}$ and the definition $C_l = \theta_l e^{2(\eta-i\xi)\bar{x}}$ from section \[secRH\], we obtain: $$\text{Res}\Upsilon_{NN}(k_1) = \frac{i}{4}\int\limits_{-\infty}^\infty
\text{d}z\,e^{-z}\text{sech}^2z\left(r_0(z) + \overline{r}_0(-z)\right),
\label{resups}$$ $${\Upsilon_r}_{NN}(k_1) = \frac{1}{8\eta}\int\limits_{-\infty}^\infty
\text{d}z\,e^z\text{sech}^2z\left(r_0(z) + \overline{r}_0(-z)\right),
\label{upsrNN}$$ $${\Upsilon_r}_{lN}(k_1) =
C_l\left\{-2i\bar{x}\text{Res}\{\Upsilon_{NN}(k_1)\}
+{\Upsilon_r}_{NN}(k_1) + J_0\right\} + f_l,
\label{upsrjN}$$ where $$J_0 = \frac{1}{4\eta}\int\limits_{-\infty}^\infty
\text{d}z\,\text{sech}^2z\left(\cosh{z} + ze^{-z}\right)
\left(r_0(z) - \overline{r}_0(-z)\right)$$ and $$f_l =F_l(k_1) =\frac{e^{2(\eta-i\xi)\bar{x}}}
{4\eta}\int\limits_{-\infty}^\infty
\text{d}z\,\text{sech}z \left(e^{-i(\xi/\eta)z}r_l
- \theta_l r_0\right).$$
Let us first consider the more involved derivation of equations for $\bar{x}$ and $\theta_l$. From equations (\[equ2\]), (\[upsrNN\]), and (\[upsrjN\]) we get: $$\frac{\text{d}C_l}{\text{d}t} = \left(-2i\omega(k_1)
+ 2\epsilon i\bar{x}\text{Res}\Upsilon_{NN}(k_1) - \epsilon
J_0\right)C_l - \epsilon f_l,
\label{dCjdt}$$ from which it follows that $$\frac{\text{d}|C_l|^2}{\text{d}t} =\Bigl[4\text{Im}\{\omega(k_1)\}
-4\epsilon\bar{x}\text{Im}\{\text{Res}\Upsilon_{NN}(k_1)\} -
2\epsilon\text{Re}\{J_0\}\Bigr]|C_l|^2 -
2\epsilon\text{Re}\{f_l\overline{C}_l\}.$$ Recalling that $\sum_{l=1}^n|\theta_l|^2=1$ we obtain: $$\frac{\text{d}\bar{x}}{\text{d}t} = \frac{e^{-4\eta\bar{x} }}{4\eta}
\sum_{l=1}^n \frac{\text{d}|C_l|^2}{\text{d}t} - \frac{\bar{x}}{\eta}
\frac{\text{d}\eta}{\text{d}t},$$ $$\frac{\text{d}\theta_l}{\text{d}t} = \theta_l
\left(C_l^{-1}\frac{\text{d}C_l}{\text{d}t}
-2\frac{\text{d}(\eta\bar{x})}{\text{d}t}
+2i\frac{\text{d}(\xi\bar{x})}{\text{d}t}\right).$$ The rest calculations are straightforward substitutions and using the identity $\sum_{l=1}^n f_l\overline{C}_l = 0$, which follows from the definitions of $C_l$, $r_0$, and $f_l$. After simple calculations one gets a system of equations for the soliton parameters: $$\frac{\text{d}\eta}{\text{d}t} =
-\frac{\epsilon}{2}\int\limits_{-\infty}^\infty
\text{d}z\,\text{sech}z\text{Re}\{r_0\},
\label{foreta}$$ $$\frac{\text{d}\xi}{\text{d}t} =
-\frac{\epsilon}{2}\int\limits_{-\infty}^\infty
\text{d}z\,\text{sech}z\,\text{tanh}z\text{Im}\{r_0\},
\label{detadt}$$ $$\frac{\text{d}\bar{x}}{\text{d}t} = \frac{\text{Im}\{\omega(k_1)\}}{\eta}
-\frac{\epsilon}{4\eta^2}\int\limits_{-\infty}^\infty
\text{d}z\, z\text{sech}z\text{Re}\{r_0\},
\label{forx}$$ $$\frac{\text{d}\theta_l}{\text{d}t} = i\theta_l\left\{
\frac{2\text{Im}\{\overline{k}_1\omega(k_1)\}}{\eta}
-\frac{\epsilon}{2\eta^2}\int\limits_{-\infty}^\infty
\text{d}z\,\text{sech}z\Bigl[\xi z\text{Re}\{r_0\}
+\eta(1-z\text{tanh}z)\text{Im}\{r_0\}\Bigr]\right\}$$ $$+\frac{\epsilon}{4\eta}
\int\limits_{-\infty}^\infty \text{d}z\,\text{sech}z
\left(\theta_l r_0 - e^{-i(\xi/\eta)z}r_l\right).
\label{fortheta}$$
These equations can be compared with the adiabatic equations derived by Karpman \[21\] for the single scalar soliton. First, it is convenient to introduce the average phase $\bar{\delta}$ of the soliton by setting $$\bar{\delta} = \sum_{l=1}^n|\theta_l|^2\delta_l.$$ The evolution equation for the average phase then follows from equation (\[fortheta\]): $$\frac{\text{d}\bar{\delta}}{\text{d}t} =
\frac{2\text{Im}\{\overline{k}_1\omega(k_1)\}}{\eta}
-\frac{\epsilon}{2\eta^2}\int\limits_{-\infty}^\infty
\text{d}z\,\text{sech}z\Bigl[\xi z\text{Re}\{r_0\}
+\eta(1-z\text{tanh}z)\text{Im}\{r_0\}\Bigr].
\label{fordelta}$$ Remarkably, the slow evolution of the soliton amplitude $\eta$, phase gradient $\xi$, position $\bar{x}$ and average phase $\bar{\delta}$ is given by equations similar to those derived for the single scalar soliton. The only trace of the vector nature of the soliton in equations (\[foreta\])-(\[forx\]) and (\[fordelta\]) is that the “scalar" perturbation $r_0$ obtains by averaging the original vector perturbation as follows $$r_0 = e^{-i(\xi/\eta)z}\sum_{l=1}^n\overline{\theta}_lr_l.$$
The equation for $\theta_l = s_le^{i\delta_l}$ can be cast in the form of two separate equations, one for the polarization parameters $s_l$ and the other for the phases $\delta_l$. We get: $$\frac{\text{d}s_l}{\text{d}t} = \frac{\epsilon}{4\eta}
\int\limits_{-\infty}^\infty \text{d}z\,\text{sech}z
\text{Re}\left\{s_lr_0
-e^{-i(\xi/\eta)z-i\delta_l}r_l\right\},
\label{forsj}$$ $$\frac{\text{d}\delta_l}{\text{d}t} =
\frac{2\text{Im}\{\overline{k}_1\omega(k_1)\}}{\eta}
-\frac{\epsilon}{2\eta^2}\int\limits_{-\infty}^\infty
\text{d}z\,\text{sech}z\Bigl[\xi z\text{Re}\{r_0\}
+\eta(1-z\text{tanh}z)\text{Im}\{r_0\}\Bigr]$$ $$+\frac{\epsilon}{4\eta}
\int\limits_{-\infty}^\infty \text{d}z\,\text{sech}z
\text{Im}\left\{r_0
- s_l^{-1}e^{-i(\xi/\eta)z-i\delta_l}r_l\right\}.
\label{fordeltaj}$$ Note that the equation for $\delta_l$ contains $s_l$ in the denominator as a reflection of the fact that the phase is not defined for the components which are not excited.
First-order radiation
---------------------
Now let us consider the evolution of the spectral densities $b_l(k)$ of radiation. Taking into account radiation in the first-order approximation amounts to solving the linearized regular Riemann-Hilbert problem (or the jump problem), $$\phi_+(k) -\phi_-(k) = \Gamma(k)
\left(\begin{array}{cc} 0 & e^{2ikx}|b(k)\rangle \\ \langle b(k)|e^{-2ikx}
& 0\end{array}\right) \Gamma^{-1}(k),$$ where we have used the notation $|b\rangle = {\bf b} =
(b_1,...,b_n)^T$. Solution of the above jump problem is obtained by integration and using the normalization condition $\phi_\pm(k)\to I$ as $k\to\infty$. We obtain $$\phi(k) = I +
\frac{1}{2i\pi}\int\limits_{-\infty}^\infty\frac{\text{d}\ell}{\ell-k}
\Gamma(\ell)\left(\begin{array}{cc} 0 & e^{2i\ell x}|b(\ell)\rangle \\
\langle b(\ell)|e^{-2i\ell x} & 0\end{array}\right) \Gamma^{-1}(\ell).$$
The contribution from radiation to the solution follows from the formula (\[Q\]), $$q^{(\text{rad})}_l = -2\lim_{k\to\infty}k\phi_{lN}(k)
=\frac{1}{i\pi}\int\limits_{-\infty}^\infty\text{d}k
\left\{\Gamma(k) \left(\begin{array}{cc} 0 & e^{2ikx}|b(k)\rangle \\
\langle b(k)|e^{-2ikx} & 0\end{array}\right) \Gamma^{-1}(k)\right\}_{lN}.$$ As in derivation of the equations for the soliton parameters it is convenient to use the transformation with the matrix $\Xi$. Using this transformation and formula (\[gamma2\]) one can easily simplify the formula for radiation contribution. We get $q^{(\text{rad})}_l=
q_l^{(\parallel)} + q_l^{(\perp)}$, where the “parallel" and “perpendicular" parts of radiation are defined as follows
\[qrad\] $$q_l^{(\parallel)} =
\frac{\eta\theta_l e^{i(\xi/\eta)z}}{i\pi}
\int\limits_{-\infty}^\infty\frac{\text{d}\lambda}{\lambda^2+1}
\Bigl\{(\lambda+i\tanh{z})^2 e^{i\lambda
z}\langle\theta|g(\lambda)\rangle
+ \text{sech}^2z e^{-i\lambda z}\langle
g(\lambda)|\theta\rangle\Bigr\},$$ $$q_l^{(\perp)} =
\frac{\eta e^{i(\xi/\eta)z}}{i\pi}
\int\limits_{-\infty}^\infty\frac{\text{d}\lambda}{\lambda-i}e^{i\lambda z}
(\lambda+i\tanh{z})(g_l(\lambda)- \theta_l\langle\theta|g(\lambda)\rangle).$$
Here, for convenience, we have introduced the modified spectral parameter by setting and the modified spectral densities of radiation: $$g_l(\lambda) = e^{2i(\eta\lambda+\xi)\bar{x}}
b_l(\eta \lambda +\xi).
\label{glambda}$$ The separation of radiation into parallel and perpendicular parts is due the following facts. While the perpendicular part $q_l^{(\perp)}$ satisfies the orthogonality property $$\sum_{l=1}^n \overline{\theta}_l q_l^{(\perp)} =0,
\label{property1}$$ the parallel part of the radiation is given by the same formula as the radiation of the scalar soliton (multiplied by $\theta_l$), but for the averaged spectral density $\langle\theta| g(\lambda)\rangle$.
Consider the evolution equation (\[equ3\]) for the spectral densities of radiation. In the first-order approximation the term with $\Upsilon_{NN}$ can be neglected. When deriving evolution equations for the modified spectral densities in the first-order approximation one must take into account only the fast (or “integrable") evolution of the soliton parameters involved in the definition of $g(\lambda)$. In particular, the modified spectral parameter $\lambda$ is $t$-independent. The only parameter which has the fast $t$-dependence in (\[glambda\]) is $\bar{x}$. By differentiation of (\[glambda\]) and using (\[equ3\]) and (\[forx\]) we obtain $$\frac{\partial g_l(\lambda)}{\partial t} =
i\Omega_r(\lambda)g_l(\lambda)
+\epsilon\theta_l \Upsilon^{(\parallel)}(\lambda) +
\epsilon\Upsilon^{(\perp)}_l(\lambda),
\label{forgj}$$ where $${\Omega_r}(\lambda) =
2\Bigl[\lambda\text{Im}\{\omega(k_1)\} + \text{Re}\{\omega(k_1)\}-
\omega(\eta\lambda+\xi)\Bigr].
\label{Omega}$$ The last two terms in equation (\[forgj\]) come from $\Upsilon_{lN}$ (\[gamjN\]) (the second one is the contribution of $F_l$ (\[Fj\])). They read $$\Upsilon^{(\parallel)}(\lambda) = \frac{1}{2\eta(\lambda^2+1)}
\int\limits_{-\infty}^\infty\text{d}z\, e^{-i\lambda z}\Bigl[(\lambda -
i\tanh{z})^2r_0 - \text{sech}^2z\,\overline{r}_0\Bigr],
\label{gampar}$$ $$\Upsilon^{(\perp)}_l(\lambda) = \frac{1}{2\eta(\lambda +i)}
\int\limits_{-\infty}^\infty\text{d}z\, e^{-i\lambda z}(\lambda -i\tanh{z})
(e^{-i(\xi/\eta)z}r_l - \theta_l r_0).
\label{gamperp}$$ Due to the definition of $r_0$ (\[r0\]), the perpendicular component satisfies the identity $$\sum_{l=1}^n\overline{\theta}_l \Upsilon^{(\perp)}_l = 0.
\label{propert2}$$
Integrating the equation for $g_l(\lambda)$ with $g_l(\lambda,t=0)=0$ and using the result in (\[qrad\]) we arrive at the first-order correction to initially pure soliton solution. In this case we obtain:
\[qrad1\] $$q_l^{(\parallel)} =\epsilon
\frac{\eta\theta_l e^{i(\xi/\eta)z}}{i\pi}
\int\limits_{-\infty}^\infty\frac{\text{d}\lambda}{\lambda^2+1}
\Bigl\{(\lambda+i\tanh{z})^2 e^{i\lambda
z +i{\Omega_r}(\lambda)t}\gamma^{(\parallel)}(\lambda)$$ $$+ \text{sech}^2z e^{-i\lambda
z-i{\Omega_r}(\lambda)t}\overline{\gamma}^{(\parallel)}(\lambda)\Bigr\},$$ $$q_l^{(\perp)} =\epsilon
\frac{\eta e^{i(\xi/\eta)z}}{i\pi}
\int\limits_{-\infty}^\infty\frac{\text{d}\lambda}{\lambda-i}
(\lambda+i\tanh{z})e^{i\lambda
z+i{\Omega_r}(\lambda)t}\gamma^{(\perp)}_l(\lambda),$$
where $$\gamma^{(\parallel)} = \frac{1}{2\eta(\lambda^2+1)}
\int\limits_{-\infty}^\infty\text{d}z\, e^{-i\lambda z}\Bigl[(\lambda -
i\tanh{z})^2 \hat{r}^{(+)}_0 -
\text{sech}^2z\,\hat{r}^{(-)}_0\Bigr],$$ $$\gamma^{(\perp)}_l = \frac{1}{2\eta(\lambda +i)}
\int\limits_{-\infty}^\infty\text{d}z\, e^{-i\lambda z}(\lambda -i\tanh{z})
(e^{-i(\xi/\eta)z}\hat{r}_l - \theta_l
\hat{r}^{(+)}_0),$$ with $$\hat{r}^{(+)}_0(z,t,\lambda) =
\int\limits_{0}^t\text{d}\tau\,e^{-i{\Omega_r}(\lambda)\tau}
r_0(z,\tau),\quad
\hat{r}^{(-)}_0(z,t,\lambda) =
\int\limits_{0}^t\text{d}\tau\,e^{-i{\Omega_r}(\lambda)\tau}
\overline{r}_0(z,\tau),$$ $$\hat{r}_l(z,t,\lambda) =
\int\limits_{0}^t\text{d}\tau\,e^{-i{\Omega_r}(\lambda)\tau}
r_l(z,\tau).$$
Conclusions
===========
In construction of the perturbation theory our main idea is to use the Riemann-Hilbert problem associated with the integrable PDE for the nonlinear transformation of the [*perturbed*]{} PDE to the spectral space. The evolution equations for the spectral data follow from the evolution functional, an additional object one needs to introduce into the IST theory to account for perturbations. For a single vector soliton, the equations describing evolution of the soliton parameters and first-order radiation are given in explicit form. The method is not restricted to the first order only. For instance, the second-order equations can also be derived. The perturbation theory can be applied for description of dynamics of the spatial optical solitons, soliton pulses in the multispecies Bose-Einstein condensates, soliton propagation in optical fibre with the account of the arbitrary polarization of light pulses, and for many other applications of the multi-component soliton equations.
In this paper we have restricted the consideration to the Zakharov-Shabat spectral problem. However, the approach of this paper was successfully applied to other spectral problems as well \[41-45\]. There, the evolution functional was derived and the evolution equations for the spectral data were obtained. The overall result of this and the previous works on the perturbation theory based on the Riemann-Hilbert problem is that this approach [*always works*]{}. The explicit form of the evolution functional was the same for all considered spectral problems and, moreover, it undergoes only insignificant changes in the transition from the Cauchy problem to an initial-boundary value problem \[45\].
The author is indebted to Professor E. V. Doktorov for stimulating discussions during the course of this work and his critical reading of the manuscript. This research was supported in part by the NRF of South Africa.
Some comments on multiplicity of zeros
======================================
Here we explore in more detail the algebraic and geometric multiplicities of zeros of $\det\Phi_+(k)$ and $\det\Phi^{-1}_-(k)$. For instance, consider the determinant $$\det \Phi_+ = {J_+}_1\wedge... \wedge {J_+}_n\wedge{J_-}_{n+1}\wedge ...
\wedge{J_-}_N.
\label{A1}$$ None of the columns ${J_\pm}_l(x,k)$ is equal to zero (otherwise, due to uniqueness of solution to equation (\[lax1\]) we would have ${J_\pm}_l= 0$ for all $x$). Moreover, the minors in the wedge products of columns of $J_+$ and $J_-$ satisfy linear homogeneous equations, which follow from (\[lax1\]), section \[secRH\], $$\partial_x {J_+}_1\wedge... \wedge {J_+}_n = \left\{ ik\left(
\hat{A}^{(n)} - \frac{n(n+1)}{2}
\hat{I}^{(n)}\right) + i\hat{Q}^{(n)}\right\}
{J_+}_1\wedge... \wedge {J_+}_n,$$ $$\partial_x {J_-}_{n+1}\wedge ... \wedge{J_-}_N= \left\{ ik\left( \hat{A}^{(N-
n)} - \frac{(N-n)(N-n+1)}{2} \hat{I}^{(N-n)}\right) + i\hat{Q}^{(N-n)}\right\}$$ $$\times {J_-}_{n+1} \wedge ... \wedge{J_-}_N.$$ Here $\hat{M}^{(j)}$ denotes a super-matrix, whose action on the wedge products of vector-columns is defined by the rule $$\hat{M}^{(j)}\Psi_1\wedge...\wedge\Psi_j = \sum_{p=1}^j \Psi_1\wedge...\wedge
M\Psi_p\wedge
...\wedge\Psi_j.$$ Therefore, $ \text{rank}{J_+}_1\wedge... \wedge {J_+}_n = \text{rank}\,e_1\wedge
...\wedge e_n = n$, and $\text{rank}{J_-}_{n+1} \wedge ... \wedge{J_-}_N =
\text{rank}\,e_{n+1}\wedge ...\wedge e_{N-n} = N-n$. Hence, the only possibility for $\det\Phi_+(k)=0$ is the linear dependence of the columns in ${J_+}H_1$ and ${J_-}H_2$, e.g., at least one of the columns of $J_-H_2$ is given as a linear combination of the columns of $J_+H_1$.
Let $\nu_j$ and $d_j$ be the algebraic and geometric multiplicities of the $j$-th zero $k_j$, i.e., $$\det\Phi_+(k)=(k-k_j)^{\nu_j}\psi(k), \quad \psi(k_j)\ne 0,\quad d_j = N-
\text{rank}\Phi_+(k_j).$$ Writing down the Taylor expansion of $\Phi_+(k)$ about $k=k_j$ we immediately conclude that $$\text{rank}\Phi_+(k_j) \ge N- \nu_j.
\label{A2}$$ Therefore, in general, the algebraic multiplicity is greater than the geometric one; trivially, they coincide for simple zeros. Representation (\[A1\]) gives $$\text{rank}\Phi_+(k_j)\ge \text{max} (n,N-n).$$ Hence the geometric multiplicity satisfies $$d_j \le N - \text{max}(n,N-n).
\label{A3}$$
We consider only zeros whose algebraic multiplicity is equal to the geometric one. In particular, if $n= N-1$, there can be only one vector in the null space of $\Phi_+$, i.e., $d_j=1$. Hence, in this case, zeros of $\det\Phi_+(k)$ must be simple to satisfy the equal multiplicity condition. This condition requires that the only case of a multiple zero, $\det\Phi_+(k_j)=0$, of order $\nu_j$ is that there are precisely $\nu_j$ columns of $J_-(k_j)H_2$ and $J_+(x,k)H_1$ which are linear combinations of the columns of $J_+(k_j)H_1$ and $J_-(k_j)H_2$, respectively.
The equal multiplicity condition can be guaranteed by the following constraint (imposed for some $(x,t)$) $$\text{rank} \left(\Phi_+,\; \frac{\partial \Phi_+}{\partial k}\right) = N.
\label{A4}$$ Indeed, for algebraic multiplicity $\nu_j$, by the Taylor expansion, (\[A4\]) gives $\text{rank}\Phi_+(k_j) = N - \nu_j$. Conversely, if $\text{rank}\Phi_+(k_j) = N - \nu_j$, for algebraic multiplicity $\nu_j$, then there are at least $\nu_j$ columns in $\partial \Phi_+(k_j)/\partial k $ independent from the columns of $\Phi_+(k_j)$. Hence (\[A4\]) also holds. Similar results are valid for multiplicities of zeros $\overline{k}_j$, $j=1,...,s$, of $\det \Phi^{-1}_-(k)$.
Properties of the regularization matrix
=======================================
Here we derive the regularization matrix $\Gamma(k)$ and prove its properties (in the main, we follow Refs. 13 and 58). Dependence on the co-ordinates $x$ and $t$ is not important for this purpose and omitted. Consider one pair of zeros, say, $k_s$ and $\overline{k}_s$ of $\det\Phi_+(k)$ and $\det\Phi^{-1}_-(k)$, respectively. Let the vectors $| p^{(s)}_l \rangle $ and $\langle \overline{ p}{}^{(s)}_l | $, $l=1,...,\nu_s$, satisfying $$\Phi_+(k_s)| p^{(s)}_l \rangle = 0,\quad \langle \overline{ p}{}^{(s)}_l
|\Phi^{-1}_-(\overline{k}_s) = 0,
\label{B0}$$ span the respective null spaces. Construct the following rational matrix functions $$\chi_s(k) = I - \frac{k_s - \overline{k}_s}{k - \overline{k}_s} P_s, \quad
\overline{\chi}_s(k) = I + \frac{k_s - \overline{k}_s}{k - k_s} P_s,$$ where $$P_s = \sum_{l,m = 1}^{\nu_s} | p^{(s)}_l \rangle \left( M^{-1}\right)_{lm}
\langle \overline{ p}{}^{(s)}_m |, \quad M_{lm} = \langle \overline{
p}{}^{(s)}_l | p^{(s)}_m \rangle$$ and $P_s$ is a projector: $P_s^2 = P_s$, $\text{rank}P_s = \nu_s$. It is easy to verify that $\overline{\chi}_s(k) $ is inverse to $\chi_s(k)$: $\chi_s(k)\overline{\chi}_s(k) = I$.
The determinant of $\chi_s(k)$ is easily computed in some appropriate basis, where the projector is represented by a diagonal matrix with $\nu_s$ ones and $N-\nu_s$ zeros on the diagonal. We get $$\det \chi_s(k) = \left( \frac{k - k_s}{k - \overline{k}_s}\right)^{\nu_s}.$$
Hence, with such rational matrices we can factor out the $s$-th pair of zeros. Indeed, consider the products $\Phi_+(k)\chi^{-1}_s(k)$ and $\chi_s(k)\Phi^{-
1}_-(k)$. These matrix functions are holomorphic in the upper and lower half planes, respectively (the poles are removable due to the identities (\[B0\])). On the other hand, the determinants are non-zero for $k=k_s$ and $k=\overline{k}_s$, respectively; thus one pair of zeros is factored out. By introducing a sequence of such matrices, $$\chi_j(k) = I - \frac{k_j- \overline{k}_j}{k - \overline{k}_j} P_j, \quad
\overline{\chi}_j(k) = I + \frac{k_j - \overline{k}_j}{k - k_j} P_j,\quad
j=1,...,s,
\label{B1}$$ where $\overline{\chi}_j(k) = \chi^{-1}_j(k)$, we factor out all zeros using the regularization matrix $\Gamma$ and its inverse, where $$\Gamma(k) = \chi_1(k)\chi_2(k)\cdot ... \cdot\chi_s(k).
\label{B2}$$ The projector $P_j$ is given by the following formula $$P_j = \sum_{l,m=1}^s | e^{(j)}_l \rangle \left( M^{-1} \right)_{lm} \langle
\overline{e}{}^{(j)}_m |,
\quad M_{lm} = \langle \overline{ e}{}^{(j)}_l | e^{(j)}_m \rangle.$$ Here the vectors $|e^{(j)}_l \rangle$ and $\langle \overline{e}^{(j)}_l | $ are related to the basis vectors of the null spaces of $\Phi_+(k_j)$ and $\Phi^{-1}_-(\overline{k}_j)$ by triangular equations (if scanned starting from $s$ down to $1$): $$| p^{(j)}_l \rangle = \chi^{-1}_s(k_j)\cdot\chi^{-1}_{s-1}(k_j)\cdot ...
\cdot\chi^{-1}_{j+1}(k_j)
| e^{(j)}_l \rangle,
\label{B3a}$$ $$\langle \overline{ e}{}^{(j)}_l |\chi_{j+1}(\overline{k}_j)\cdot
\chi_{j+2}(\overline{k}_j)
\cdot... \cdot\chi_s(\overline{k}_j) = \langle \overline{ p}{}^{(j)}_l |.
\label{B3b}$$ Due to $P_j^2= P_j$, these vectors satisfy the identities: $$\chi_j(k_j)| e^{(j)}_l \rangle = 0,\quad \langle \overline{e}{}^{(j)}_l
|\chi^{-1}_j(k_j) = 0,\quad
l=1,...,\nu_j.
\label{B4}$$
The regularization matrix $\Gamma(k)$ can be made parameterized entirely by the vectors from the null spaces. Indeed, let us decompose $\Gamma(k)$ and the inverse matrix into the partial fractions: $$\Gamma(k) = I - \sum_{j=1}^s\frac{\overline{B}_j}{k -
\overline{k}_j},\quad
\Gamma^{-1}(k) = I + \sum_{j=1}^s\frac{B_j}{k - k_j},
\label{B5}$$ where due to (\[B3a\]) and (\[B3b\]) we have $$B_j = \sum_{l=1}^{\nu_j}| p^{(j)}_l \rangle \langle v^{(j)}_l |, \quad
\overline{B}_j = \sum_{l=1}^{\nu_j}| \overline{v}{}^{(j)}_l \rangle \langle
\overline{p}{}^{(j)}_l |.
\label{B6}$$ From (\[B5\]) and the identity $\Gamma\Gamma^{-1} =
\Gamma^{-1}\Gamma = I$ it follows that $$\Gamma(k_j)| p^{(j)}_l \rangle = 0,\quad
\langle \overline{ p}{}^{(j)}_l | \Gamma^{-1}(\overline{k}_j) = 0,\quad
l=1,...,\nu_j,\quad j=1,...,s.
\label{B7}$$ These are the equations defining the unknown vectors $|\overline{v}{}^{(j)}_l
\rangle $ and $\langle v^{(j)}_l |$. Indeed, rewriting (\[B7\]) we have $$| p^{(j)}_l \rangle = \sum_{i=1}^s \frac{1}{k_j -
\overline{k}_i}\sum_{m=1}^{\nu_i}
| \overline{v}{}^{(i)}_m \rangle \langle \overline{ p}{}^{(i)}_m |
p^{(j)}_l \rangle,\quad
\langle \overline{ p}{}^{(j)}_l | = - \sum_{i=1}^s \frac{1}{\overline{k}_j
- k_i}\sum_{m=1}^{\nu_i}\langle \overline{ p}{}^{(j)}_l | p^{(i)}_m
\rangle\langle v^{(i)}_m |.$$ Inversion of these formulae gives $$| \overline{v}{}^{(j)}_l \rangle = \sum_{i=1}^s\sum_{m=1}^{\nu_i}| p^{(i)}_m
\rangle
\left(D^{-1}\right)_{im,jl},\quad
\langle v^{(j)}_l | = \sum_{i=1}^s\sum_{m=1}^{\nu_i}\left(D^{-
1}\right)_{jl,im}\langle
\overline{ p}{}^{(i)}_m |.
\label{B8}$$ Here the matrix $D$ is defined by $$D_{im,jl} = \frac{\langle \overline{p}_m^{(i)} |
p_l{}^{(j)} \rangle }{k_j - \overline{k}_i}.$$ Substitution of (\[B8\]) into (\[B5\]) produces the needed formulae (\[Gammas\]).
[99]{}
A. V. Mikhailov et al, Russ. Math. Surveys [**42**]{}, 1 (1987); J. Hietarinta, Phys. Rep. [**147**]{}, 87 (1987).
M. Mitchell, Z. Chen, M. Shin, and M. Segev, Phys. Rev. Lett. [**77**]{}, 490 (1996); M. Mitchel and M. Segev, Nature (London) [**387**]{}, 880 (1997).
M. Mitchell, M. Segev, T. H. Coskun, and D. N. Christodoulides, Phys. Rev. Lett. [**79**]{}, 4990 (1997); D. N. Christodoulides, T. H. Coskun, M. Mitchell and M. Segev, Phys. Rev. Lett. [**80**]{}, 2310 (1998); D. N. Christodoulides, T. H. Coskun, M. Mitchell, Z. Chen, and M. Segev Phys. Rev. Lett. [**80**]{}, 5113 (1998); N. N. Akhmediev, W. Królikowski, and A. W. Snyder, Phys. Rev. Lett. [**81**]{}, 4632 (1998); A. A. Sukhorukov and N. N. Akhmediev, Phys. Rev. Lett. [**83**]{}, 4736 (1999); O. Bang, D. Edmindon, and W. Królikowski, Phys. Rev. Lett. [**83**]{}, 5479 (1999).
C. K. Law, H. Pu, N. B. Bigelow, and J. H. Eberly, Phys. Rev Lett. [**79**]{}, 3105 (1997); T.-L. Ho, Phys. Rev. Lett. [**81**]{}, 742 (1998); S.-K. Yip, Phys. Rev. Lett. [**83**]{}, 4677 (1999).
G. P. Agrawal, [*Nonlinear Fiber Optics*]{} (Academic Press, New York, 1989); [*Optical Solitons - Theory and Experiment*]{} ed. J. R. Taylor, Cambridge Studies in Modern Optics 10 (Cambridge University Press, Cambridge, England 1992).
S. Chakravarty, M. J. Ablowitz, J. R. Sauer and R. B. Jenkins, Opt. Lett. [**20**]{}, 136 (1995).
C. Yen and L. A. Bergman, Phys Rev E [**57**]{}, 2398 (1998); [**60**]{}, 2306 (1999).
F. T. Hioe, Phys. Rev. Lett. [**82**]{}, 1152 (1999).
B. Tan and J. P. Boyd, Chaos, Solitons & Fractals [**11**]{}, 1113 (2000).
A. V. Mikhailov, Physica D [**3**]{}, 73 (1981); A. P. Fordy and P. P. Kulish, Commun. Math. Phys. [**89**]{}, 427 (1983).
V. E. Zakharov and A. B. Shabat Sov. Phys. JETP [**34**]{}, 62 (1972); Funct. Anal. Appl. [**8**]{}, 226 (1974); [**13**]{}, 13 (1979).
S. V. Manakov, Zh. Eksp. Teor. Fiz. [**65**]{}, 505 (1973) \[Sov. Phys. JETP [**38**]{}, 248 (1974)\].
S. P. Novikov, S. V. Manakov, L. P. Pitaevski, and V. E. Zakharov, [*Theory of Solitons the Inverse Scattering Method*]{} (Consultants Bureau, New York, 1984).
K. Nakkeeran, K. Porsezian, S. P. Shanmugha, and A. Mahalingam, Phys. Rev. E [**80**]{}, 1425 (1998).
T. Tsuchida and M. Wadati, J. Phys. Soc. Jpn. [**67**]{}, 1175 (1998); Phys. Lett. A [**257**]{}, 53 (1999); Inverse Problems [**15**]{}, 1363 (1999).
S. Yu. Sakovich and T. Tsuchida, J. Phys. A: Math. Gen. [**33**]{}, 7217 (2000).
F. M. Mitschke and L. F. Mollenauer, Opt. Lett. [**11**]{}, 657 (1986); Y. Kodama and A. Hasegawa, IEEE J. Quantum Electron. [**23**]{}, 510 (1987).
A. Hasegawa, [*Optical Solitons in Fibers* ]{} (Springer, Heidelberg, 1989).
A. Hasegawa and Y. Kodama, [*Solitons in Optical Communications*]{} (Clarendon Press, Oxford, 1995).
D. J. Kaup, SIAM J. Appl. Math. [**31**]{}, 121 (1976); D. J. Kaup and A. C. Newell, Proc. R. Soc. Lond. Ser. A [**361**]{}, 413 (1978).
V. I. Karpman, Phys. Scr. [**20**]{}, 462 (1979); V. I. Karpman and E. M. Maslov, Sov. Phys. JETP [**46**]{}, 281 (1977).
J. Kaup, J. Math. Anal. Appl. [**54**]{}, 849 (1976); J. Math. Phys. [**25**]{}, 2467 (1984).
V. S. Gerdjikov, M. I. Ivanov, and P. P. Kulish, Teor. Mat. Fiz. [**44**]{}, 342 (1980) (in Russian).
T. Kawata, J. Phys. Soc. Jpn. [**53**]{}, 2879 (1984).
K. Iino and Y. Ichikawa, J. Phys. Soc. Jpn. [**51**]{}, 4091 (1982); [**51**]{}, 2389 (1982).
V. S. Gerdjikov and A. B. Yanovski, Phys. Lett. A. [**103**]{}, 232 (1984); Phys. Lett. A [**110**]{}, 53 (1985); Commun. Math. Phys. [**103**]{}, 545 (1986).
V. S. Gerdjikov, Theor. Math. Phys. [**92**]{}, 374 (1992); V. S. Gerdjikov, M. I. Ivanov, Inverse Problems [**8**]{}, 831 (1992).
D. J. Kaup and T. I. Lakoba, J. Math. Phys. [**37**]{}, 308 (1996); T. I. Lakoba and D. J. Kaup, Phys. Rev. E [**56**]{}, 6147 (1997).
J. P. Keener and D. W. Mclaughlin, Phys. Rev. A [**16**]{}, 777 (1977).
Yu. S. Kivshar, Physica D [**40**]{}, 11 (1989).
Yu. S. Kivshar and B. A. Malomed, Rev. Mod. Phys. [**61**]{}, 763 (1989).
D. J. Kaup, Phys. Rev. A [**42**]{}, 5689 (1990).
R. L. Herman, J. Phys. A: Math. Gen. [**23**]{}, 2327 (1990).
E. V. Doktorov and I. N. Prokopenya, Inv. Probl. [**7**]{}, 221 (1991); Zh. Eksp. Teor. Fiz. [**100**]{}, 1129 (1991) \[Sov. Phys. JETP [**73**]{}, 623 (1991)\].
V. V. Konotop and V. E. Vekslerchik, Phys. Rev. E [**49**]{}, 2397
D. E. Pelinovsky, A. V. Buryak, and Yu. S. Kivshar, Phys. Rev. Lett. [**75**]{}, 591 (1995); D. E. Pelinovsky, Yu. S. Kivshar, and V. V. Afanasjev, Phys. Rev. E [**53**]{}, 1940 (1996).
I. V. Barashenkov, D. E. Pelinovsky, and E. V. Zemlyanaya, Phys. Rev. Lett. [**80**]{}, 5117 (1998).
N. V. Alexeeva, I. V. Barashenkov, and D. E. Pelinovsky, Nonlinearity [**12**]{}, 103 (1999).
J. Yang and D. J. Kaup, SIAM J. Appl. Math. [**60**]{}, 967 (2000).
J. Yang, J. Math. Phys. [**41**]{}, 6614 (2000).
V. S. Shchesnovich, Chaos, Solitons & Fractals, [**5**]{}, 2121 (1995).
V. S. Shchesnovich and E. V. Doktorov, Phys. Rev. E [**55**]{}, 7626 (1997).
V. S. Shchesnovich and E. V. Doktorov, Physica D [**129**]{}, 115 (1999).
V. S. Shchesnovich and I. V. Barashenkov, in preparation.
E. V. Doktorov and V. S. Shchesnovich, Inverse Problems [**17**]{}, 971 (2001).
M. J. Ablowitz and P. A. Clarkson, [*Solitons, Nonlinear Evolution Equations and Inverse Scattering*]{} (Cambridge University Press, Cambridge, 1991).
A. S. Fokas, SIAM J. Math. Anal. [**27**]{}, 738 (1996); Proc. R. Soc. Lond. A [**453**]{}, 1411 (1997); J. Math. Phys. [**41**]{}, 4188 (2000).
L. D. Faddeev and L. A. Takhtajan, [*Hamiltonian Methods in the Theory of Solitons*]{} (Springer-Verlag, Berlin-Heidelberg-New York, 1987).
J. Leon, J. Math. Phys. [**35**]{}, 3504 (1994).
A. C. Newell, [*Solitons in Mathematics and Physics*]{} (SIAM, Philadelphia 1985).
M. J. Ablowitz and H. Segur, [*Solitons and the Inverse Scattering Transform*]{} (SIAM, Philadelphia 1981).
F. Calogero and A. Degasperis, [*Spectral Transform and Solitons*]{} (North-Holland, Amsterdam, 1982).
V. S. Gerdjikov, M. I. Ivanov, P. P. Kulish, Lett. Math. Phys. [**6**]{}, 315 (1982).
X. Zhou, SIAM J. Math. Anal. [**20**]{}, 966 (1988); Commun. Pure Appl. Math. [**42**]{}, 895 (1989).
R. Beals and R. R. Coifman, Commun. Pure Appl. Math. [**37**]{}, 39 (1984); [**38**]{}, 29 (1985); Inverse Problems [**5**]{} 577 (1989).
R. Beals, P. Deift, and C. Tomei, [*Direct and Inverse Scattering on the Line*]{}, Math. Surv. Mono., [**28**]{}, American Mathematical Society, Providence, R. I., 1988.
P. J. Caudrey, Physica D [**6**]{}, 51 (1982).
T. Kawata, Riemann spectral method for the nonlinear evolution equations, p. 210, in [*Advances in Nonlinear Waves*]{}, ed. by L. Debnath, (Cambridge Univ. Press, Cambridge, 1984).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Dy$_{2}$Zr$_{2}$O$_{7}$ a disordered pyrochlore system, exhibits the spin ice freezing under the application of magnetic field. Our studies suggest the stabilization of pyrochlore phase in Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ with the substitution of nonmagnetic La, along with the biphasic mixture for the intermediate compositions. We observed that the higher La compositions (1.5 $\leq$ x $\leq$ 1.9), show spin freezing (T $\sim$ 17 K) similar to the field induced spin ice freezing for low La compositions (0 $\leq$ x $\leq$ 0.5), and the well known spin ice systems Dy$_{2}$Ti$_{2}$O$_{7}$ and Ho$_{2}$Ti$_{2}$O$_{7}$. The low temperature magnetic state for higher La compositions (1.5 $\leq$ x $\leq$ 1.9) culminates into spin glass state below 6 K. The Cole-Cole plot and Casimir-du Pr$\acute{e}$ fit shows narrow distribution of spin relaxation time in these compounds.'
author:
- 'Sheetal and C. S. Yadav\*'
bibliography:
- 'Arxiv.bib'
title: 'Evolution of spin freezing transition and structural, magnetic phase diagram of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$; x = 0 - 2.0'
---
The geometrical frustration is one of the key aspects of the cubic pyrochlore oxide A$_{2}$B$_{2}$O$_{7}$ owing to the distorted spin geometry and negligible structural disorder [@villain1979insulating; @binder1986spin]. The A and B sites of the pyrochlore form a network of corner sharing tetrahedra which is a quintessential framework for a geometrically frustrated magnet. These materials display a variety of exotic phases like spin ice (Dy$_{2}$Ti$_{2}$O$_{7}$ and Ho$_{2}$Ti$_{2}$O$_{7}$), spin liquid (Tb$_{2}$Ti$_{2}$O$_{7}$), spin glass (Y$_{2}$Mo$_{2}$O$_{7}$), order by disorder (Er$_{2}$Ti$_{2}$O$_{7}$), Kondo effect (Pr$_{2}$Ir$_{2}$O$_{7}$), unconventional anomalous Hall effect (Nd$_{2}$Mo$_{2}$O$_{7}$) and superconductivity (Cd$_{2}$Re$_{2}$O$_{7}$) [@snyder2004low; @ehlers2002dynamical; @enjalran2004spin; @greedan1986spin; @petrenko2013low; @nakatsuji2006metallic; @taguchi2003magnetic; @jin2001superconductivity]. The magnetic spin ice systems have offered an interesting physics with the signature of magnetic monopole like excitation states [@toews2018disorder; @morris2009dirac; @jaubert2011magnetic]. Although the Dy$_{2}$Ti$_{2}$O$_{7}$ and Ho$_{2}$Ti$_{2}$O$_{7}$ systems are extensively studied for spin ice properties, a complete understanding of the phenomenon is still elusive [@castelnovo2008magnetic; @fennell2009magnetic]. Among these systems, Dy$_{2}$Zr$_{2}$O$_{7}$ (DZO) has generated a quite interest, as it shows the emergence of the magnetic field induced spin freezing near 10 K and possess the magnetic entropy of R ln2 - (1/2)R ln(3/2) which is same as the spin ice Dy$_{2}$Ti$_{2}$O$_{7}$ [@devi2020emergence]. The DZO crystallizes in weak pyrochlore phase and does not exhibit any magnetic ordering down to 40 mK [@ramon2019absence]. Further partial substitution of Dy by La (up to 15 $\%$) stabilizes pyrochlore phase and spin ice behavior is observed at lower magnetic field in comparison to DZO [@devi2020emergence].
The stability of pyrochlore structure and the magnetic rare earth atom (Dy, Ho etc.) play very important role in determining the low temperature magnetic spin ice ground state of these systems. In this light we have studied the structural and magnetic phase diagram of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ system. We observed the evolution of structure from disordered pyrochlore/fluorite (0 $\leq$ x $\leq$ 0.5) to stable pyrochlore phase (1.5 $\leq$ x $\leq$ 2.0) through the biphasic mixture of these phases for 0.5 $<$ x $<$ 1.5. The compounds with weak pyrochlore structure do not show evidence of the magnetic freezing in the absence of magnetic field, whereas the stable pyrochlore compounds exhibit Dy$_{2}$Ti$_{2}$O$_{7}$ like spin freezing at T$_{f}$ $\approx$ 17.5 K. The low temperature phase of 1.5 $\leq$ x $\leq$ 1.9 compositions show spin glass state below 6 K.
DZO exhibits disordered fluorite structure (space group: Fd$\bar{3}$m) with the remnants of pyrochlore phase [@devi2020emergence; @ramon2019absence; @mandal2006preparation; @sayed2011sm2; @glerup2001structural]. The xrd patterns of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ reveal that the compositions corresponding to 0 $\leq$ x $\leq$ 0.5, adopt disordered fluorite structure. For 0.5 $<$ x $<$ 1.5, few super-structure peaks corresponding to pyrochlore structure start appearing at 2$\theta$ = 14$^{o}$ (111), 27$^{o}$ (311), 37$^{o}$ (331), 45$^{o}$ (511), etc; and the main peaks belonging to pyrochlore structure get splitted. The xrd data in this range (see supplementary figure S3) nicely fit with the biphasic mixture of disordered fluorite/pyrochlore and clean pyrochlore lattice structure. The Rietveld refinement for 1.5 $\leq$ x $\leq$ 2 confirms the clean face-centered-cubic pyrochlore structure with the additional reflections of (111), (311), (331), (511) planes [@anand2015investigations; @snyder2001spin]. This is also anticipated from the closer value of ionic radius ratio (r$_{A}$/r$_{B}$ ratio) and position of oxygen atom x(O) in the pyrochlore structure (figure 1a) [@pal2018high; @kumar2019spin]. As discussed in Ref[@devi2020emergence], for x $\leq$ 0.3, the r$_{A}$/r$_{B}$ ratio is less than the expected range of the pyrochlores and the parameter x(O) is close to $\sim$ 0.362 which indicates the formation of disordered fluorite/pyrochlore structure. Here, with x $\geq$ 1.5, the r$_{A}$/r$_{B}$ (where r$_{A}$ = 1.027/1.16 $\AA$ for Dy$^{3+}$/La$^{3+}$ and r$_{B}$ = 0.72 $\AA$ for Zr$^{4+}$) found to be in the pyrochlore regime (see supplementary table SI). The x(O) evolves continuously towards pyrochlore regime with La substitution (figure 1b), and supports the stabilization of pyrochlore structure for higher La concentrations. A systematic shift of peak position towards lower angle on La incorporation indicates the increase of unit cell parameter (figure 1a).
{width="8.5cm" height="7cm"}
Raman spectra of a pyrochlore structure consists six Raman active modes; A$_{1g}$, E$_{g}$, and four T$_{2g}$modes; corresponding to the vibrations of $<$A-O$>$ and $<$B-O$>$ bonds [@han2015electron; @hasegawa2010raman]. The DZO shows no Raman mode (figure 2) at 466 cm$^{-1}$ corresponding to the fluorite structure, but exhibits the signature of Raman mode for the pyrochlore structure [@hozoi2014longer; @devi2020emergence]. The pyrochlore Raman modes become more intense upon La substitution. The Raman spectra of the biphasic compositions (0.5 $<$ x $<$ 1.5) also do not exhibit any extra mode other than of pyrochlore structure, and thus indicate the absence of any impurity phase in the compounds. The Raman spectra of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ for the compositions x $\geq$ 1.5 shows sharper and well pronounced peaks (at 301 cm$^{-1}$ (T$_{2g}$), 395 cm$^{-1}$ (E$_{g}$), 493 cm$^{-1}$ (A$_{1g}$), 514 cm$^{-1}$ (T$_{2g}$) and 651 cm$^{-1}$ (T$_{2g}$)) than for x $\leq$ 0.5. This is possibly due to the increase in r$_{A}$/r$_{B}$ ratio for the La substituted DZO. Additionally, an extra mode at 701 cm$^{-1}$ (marked by M) is observed in all the substituted compounds which was observed around 650 cm$^{-1}$ in parent compound, was attribute to the vibrations of the ZrO$_{6}$ octahedra [@glerup2001structural]. This mode has been assigned to the distortion of octahedra and coordination from the ideal pyrochlore structure [@turner2017lanthanide].
We measured the magnetization of the compounds, at low field (at 100 Oe) using the zero-field-cooled (ZFC) and field-cooled (FC) protocols. In the previous study [@devi2020emergence], we have shown that the compounds having low La composition in Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$; x = 0.0, 0.15, 0.3 exhibit paramagnet like behavior down to 1.8 K. This behavior sustains up to the La substitution of x $\sim$ 1.0, but the dc magnetic susceptibility of x = 1.5 shows bifurcation of ZFC and FC curves at T$_{irr}$ $\approx$ 2.5 K (figure 3a). The T$_{irr}$ shifts towards higher temperature for further increase in La (x = 1.5 - 1.9), and then disappears with the complete replacement of Dy with La (x = 2.0). We further measured dc magnetization at various dc applied fields using the same measurement protocol. The dc magnetization for x = 1.9 composition presented in figure 3b shows the highest value of T$_{irr}$ at 100 Oe. The bifurcation point T$_{irr}$ is weakly dependent on applied field, and shows slight increase for the field up to 3 kOe. We have plotted the field response of $\Delta$M = M$_{ZFC}$ - M$_{FC}$ of Dy$_{0.1}$La$_{1.9}$Zr$_{2}$O$_{7}$ at various temperature in the inset of figure 3b, where it is clearly seen that $\Delta$M increases for H $\leq$ 3 kOe, and then decreases to zero for H $\approx$ 10 kOe. The splitting in the ZFC and FC curves gives an indication for the spin-glass transition. In conventional spin glass systems, the bifurcation point decreases monotonically on increasing field and the glass state quenched on applying the sufficiently strong field [@mydosh1993spin]. Furthermore, the shifting of T$_{irr}$ towards low temperature for H $\geq$ 5 kOe and reduction in the absolute value of $\chi_{dc}$ under the application of high magnetic field indicates the frozen spin glass state below T$_{irr}$. Spin glass is a metastable state which arises due to random site distribution and random exchange and leads to a multidegenerate ground state [@mydosh1993spin; @binder1986spin; @huang1985some]. In case of pyrochlores, glassiness arises possibly due to the mixed interactions, which undergoes glass/frozen state on cooling without any long range magnetic order.\
The characteristic behavior of the high temperature spin freezing at T $\sim$ 17.5 is evident (discussed later) as well as the irreversibility in the dc magnetization for T $<$ T$_{f}$ is associated with the development of spin correlation. The irreversibility of the low temperature spin state was further confirmed by examining the isothermal remnant magnetization (IRM) and thermoremnant magnetization (TRM). It is notable that both the magnetization curves merge at a field $\sim$ 10 kOe where the bifurcation between ZFC and FC magnetization also disappears. The IRM and TRM curves shown in the inset of figure 3a, are in consistence with expectations for a glassy system [@snyder2004low]. The difference between the IRM and TRM below 10 kOe points towards the retains “memory” of its aged value under the same final condition [@dupuis2002aging]. The saturation of both above 10 kOe shows that sufficiently high field can destroy this memory and is in consistence with the equivalence of ZFC and FC data taken at the same field. The Curie Weiss fitting of dc susceptibility data (measured at H = 5 kOe) in the temperature range 30 - 300 K, show the dominance of antiferromagnetic interactions in these systems. The obtained parameters are listed in Supplementary table II.
Figure 4a shows the field response of magnetization at 1.8 K. The magnetization isotherm for DZO (inset of 4a) exhibits rapid increase below 20 kOe and does not saturate up to the maximum applied field of 70 kOe. The maximum value of magnetization ($\sim$ 4.8 $\mu_{B}$/Dy) at 70 kOe for DZO is close to the M$_{s}$ = 5 $\mu_{B}$/Dy for Dy$_{2}$Ti$_{2}$O$_{7}$ but half of the theoretically expected value ($\sim$ 10.64 $\mu_{B}$/Dy), which indicates strong anisotropy in the system [@anand2015investigations; @fukazawa2002magnetic]. The substitution of La up to x $\leq$ 0.3 exhibits similar behavior as that of DZO [@devi2020emergence]. However, for higher concentration of La (x $\geq$ 1.5), magnetization isotherms show a large hysteresis below 5 kOe and get saturated at lower field in comparison to DZO. The obtained values of saturation magnetization for all the La substituted compounds corresponds to $\approx$ 5 $\mu_{B}$/Dy, which is same as for DZO. The irreversibility of magnetization seen in the dc magnetization M versus T curves (figure 3) around $\sim$ 6 K is evident from the magnetic isotherm curve. As shown in the figure 4b, the magnetic hysteresis for Dy$_{0.1}$La$_{1.9}$Zr$_{2}$O$_{7}$ get suppressed for T $>$ 6 K, similar to the dc magnetization where ZFC and FC curves overlap at this temperature. Interestingly, no appreciable hysteresis is found at H $\sim$ 10 kOe, at which the bifurcation between the ZFC-FC and IRM-TRM curves is lost. This is probably because of the breaking of the ice-rule on applying the strong field [@fukazawa2002magnetic]. The magnetic moment of the system increases to the expected value of magnetization (3.33 $\mu_{B}$/Dy) for the partially ordered two-in-two-out state at H = 5 kOe and closely reaches to the fully saturated value $\sim$ 5 $\mu_{B}$/Dy of three-in-one-out spin ordered state at $\sim$ 10 kOe [@fukazawa2002magnetic]. From these results, we conclude that the La substituted compounds for x $\geq$ 1.5 exhibits spin ice state at 17.5 K and undergoes a glassy phase below 7 K.
{width="8.5cm" height="11cm"}
The ac susceptibility ($\chi_{ac}$) of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (0 $\leq$ x $\leq$ 2) are shown in the figure 5, were measured in order to understand the effect of La substitution on the spin dynamics of the system. The DZO and other x $\leq$ 0.5 compositions exhibit paramagnet like behavior in real ($\chi^{\prime}$) and imaginary $\chi^{\prime\prime}$ parts of $\chi_{ac}$, measured at zero dc field (inset of figure 5a) [@devi2020emergence]. Figure 5a and 5b show the temperature variation of $\chi^{\prime}$ and $\chi^{\prime\prime}$ for frequency $\it{f}$ = 10-1000 Hz at zero dc applied field for DyLaZr$_{2}$O$_{7}$ (i.e. x = 1.0). As seen, $\chi_{ac}$ remains almost featureless up to this concentration, except the development of weak feature near 15 K. This feature (spin freezing transition) develops into full peak shape anomaly for 1.5 $\leq$ x $\leq$ 1.9 (figure 5c). It is worth nothing that the shape of our ac curves are very similar to the measurement in Dy$_{2}$Ti$_{2}$O$_{7}$ [@snyder2001spin]. The $\chi^{\prime}$ and $\chi^{\prime\prime}$ curves show a monotonous rise for all the excitation frequencies with decrease in temperature (see supplementary figure S4). Though the frequency dependence of spin freezing transition points to the spin glass like behavior, the obtained valued of Mydosh parameter; $\it{p}$ = $\Delta$T$_{f}$/T$_{f}$$\Delta$(ln*f*) $\sim$ 0.80 - 0.49 are much larger than the typical spin glass value of $\it{p}$ $\sim$ 0.005 - 0.01 [@mydosh1993spin], and thus rules out the possibility of spin cluster/glass state. Therefore, an alternate explanation is required to understand the spin relaxation in these clean systems. Further the frequency dependence of $\chi^{\prime}$ follows the Arrhenius Law, given by *f* = *f*$_{o}$exp(-E$_{a}$/k$_{B}$T), where E$_{a}$ is the average activation energy of the energy barrier (shown in supplementary figure S6). The obtained values of the relaxation time ($\tau$) and activation energy E$_{a}$ are $\sim$ 10$^{-5}$ sec and $\sim$ 150 ($\pm$20) K respectively. Variations of E$_{a}$ and $\tau$ with La compositions are shown in the figure 5d. The slow relaxation of spin thus can be explained on the basis of increase in energy barrier in the substituted compounds compared to DZO which has a low value of E$_{a}$ $\sim$ 10 K [@ramon2019absence]. The presence of single spin relaxation is evident from the Cole-Cole plot (see supplementary figure S6) and Casimir-du Pr$\acute{e}$ relation fit also (inset of 5c) [@anand2015investigations]. The data points on $\chi^{\prime\prime}$($\chi^{\prime}$) curves corresponding to different frequencies fall on the semi-circular arc of varying diameters for all compositions. Inset of figure 6a shows the variation in $\alpha$ for different composition. The shape of arc remains unchanged upon La substitution and indicates that the distribution of relaxation times is unaffected and follows the single spin relaxation process for Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$.
![Magnetic phase diagram of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$; x = 0 - 2 showing the evolution of spin freezing transition on La substitution. Data points marked by star are taken from Ref\[16\].](PD.eps)
In order to examine the spin relaxation process in detail, $\chi_{ac}$ data were taken over the frequency range 1 $<$ $\it{f}$ $<$ 1 kHz. The $\chi^{\prime\prime}$($\it{f}$) shows a well-defined peak at the same temperature as in $\chi_{ac}$ vs T measurements implying the single characteristic relaxation time $\tau$ (see supplementary figure S5). The change in the peak position with temperature reflects the evolution of $\tau$. Unlike Dy$_{2}$Ti$_{2}$O$_{7}$ the $\tau$ obtained from $\chi^{\prime\prime}$($\textit{f}$) plot follows the Arrhenius behavior [@snyder2004low]. The activated relaxation is responsible for the high temperature spin freezing observed at T $\sim$ 17.5 K. It is to note that $\tau(T)$ increasing at a faster rate than Dy$_{2}$Ti$_{2}$O$_{7}$ and at the lowest measuring temperature, the spins relax with a large relaxation time (10$^{-1}$ s) where in Dy$_{2}$Ti$_{2}$O$_{7}$ it is around (10$^{-3}$ s) [@snyder2004low]. Note that no relaxation time could be determined below 10 K and above 1.8 K due to the absence of $\chi^{\prime\prime}$ maxima. The obtained $\tau$ and a good fit of the data with Casimir-du Pr$\acute{e}$ relation allow us to understand the development of spin freezing transition observed in ac susceptibility at T$_{f}$ for only x $\geq$ 1.5. For x = 1, no relaxation peak observed in $\chi^{\prime\prime}$($\it{f}$) plot, however, a very weak signal of freezing transition is detected in $\chi^{\prime\prime}$(T) plot. The magnetic phase diagram of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$; x = 0 - 2.0 has been drawn in figure 6. The increase in La concentration clearly shows the stabilization of Dy$_{2}$Ti$_{2}$O$_{7}$ like high temperature freezing transition around 16 K for all the compositions along with the low temperature glassy phase for x $\geq$ 1.5.
In conclusion we have shown that the position of the oxygen atom plays a significant role in determining the crystal structure, and pyrochlore phase get stabilized towards the La end in the Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ series. The low temperature magnetic ground state of the system evolves from field induced spin ice state for La compositions of 0 $\leq$ x $\leq$ 0.3 to the Dy$_{2}$Ti$_{2}$O$_{7}$ like spin freezing near T $\sim$ 16 K for La compositions of 1.5 $\leq$ x $\leq$ 1.9. The spin dynamics of the system around this transition suggests slower spin relaxation in comparison to Dy$_{2}$Ti$_{2}$O$_{7}$. This spin freezing is followed by a spin glass like state below T $\sim$ 6 K.
We acknowledge Advanced Material Research Center, IIT Mandi for the experimental facility. Sheetal is thankful to IIT Mandi and MHRD India for the Senior Research Fellowship.
Supplementary information
=========================
Experimental Details
--------------------
The polycrystalline compounds of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (x: 0.5, 0.6, 0.8, 1.0, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0) were prepared by the standard solid-state reaction method [@devi2020emergence]. The constituent oxides: Dy$_{2}$O$_{3}$ (Sigma Aldrich, $\geq$99.99$\%$ purity), La$_{2}$O$_{3}$ (Sigma Aldrich, $\geq$99.999$\%$ purity) and ZrO$_{2}$ (Sigma Aldrich, 99$\%$ purity) were reacted in the alumina crucible at 1350$^{o}$C in ambient atmosphere for 50 hours. The reaction at this temperature was done thrice with the intermediate grindings. The obtained compounds were further pelletized and sintered at 1350$^{o}$C for 50 hours.
The powder X-ray diffraction measurement on all the compounds were performed using Rigaku x- ray diffractometer in the 2$\theta$ range of 10- 90$^{o}$ with the step size of 0.02$^{o}$. The Rietveld refinement of the x-ray diffraction (XRD) pattern was performed using Fullprof Suit software. Raman spectra of the compound was obtained at 300 K in back scattering geometry by using Horiba HR-Evolution spectrometer with 532 nm excitation laser. We study the magnetization as well as the real ($\chi^{\prime}$) and imaginary parts ($\chi^{\prime\prime}$) of the ac susceptibility ($\chi_{ac}$). The magnetic measurements were performed using Quantum Design built Magnetic Property Measurement System (MPMS) above 1.8 K.
Structural studies
------------------
**Figure S1**: Room temperature x-ray diffraction pattern of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (x = 0, 0.5, 1.0, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0). A clear shift in peak positions towards lower $\theta$ values indicates the increase in lattice constant with increase in the La concentration.
**Figure S2**: The Rietveld refined x-ray diffraction pattern of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (x = 0, 0.5, 1.5, 2.0) using Fd$\bar{3}$m space group. Peaks marked by star are the superstructure peaks belonging to pyrochlore structure, and indicates the pyrochlore type ordering in the compounds.
Figure S1 and S2 show the x-ray diffraction pattern of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (x = 0 - 2). The XRD patterns of the compounds is fitted nicely in Fd$\bar{3}$ space group and show the single phase of these compounds. Structural evolution from weak pyrochore phase (0 $\leq$ x $\leq$ 0.5) to completely stabilised pyrochlore phase (1.5 $\leq$ x $\leq$ 2) is clear from the x-ray studies. The observation of the superstructure peaks for x $\geq$ 0.5 directs the structural geometry towards stable pyrochlore phase. The refined parameters for all the compositions are listed in Table 1. In figure 3, we have shown the refined x-ray pattern of a mix phase compound DyLaZr$_{2}$O$_{7}$, where xrd data is nicely fitted with the parameters of disordered fluorite and pyrochlore phase using same space group and three different lattice parameters. Though the majority phase ($\sim$ 97$\%$) is pyrochlore (a = 10.6669 Å), there are two other $\sim$ 2$\%$ and $\sim$ 1$\%$ weak pyrochlore phases with lattice parameters a = 10.4496(3) Åand a = 10.3737(3) Å.
**Figure S3**: The Rietveld refined x-ray diffraction pattern of DyLaZr$_{2}$O$_{7}$. Inset shows the splitting of main peak and refinement with disordered fluorite and pyrochlore phase more clearly.
Magnetic studies
----------------
Figure S4 shows the temperature dependence of real and imaginary part of ac susceptibility measured at various frequencies between 10 - 1000 Hz. A clear relaxation peak start evolving on entering the biphasic region and becomes prominent in the stable pyrochlore phase, whereas parent compound Dy$_{2}$Zr$_{2}$O$_{7}$ exhibits a different behavior and paramagnetic down to 1.8 K. The magnetic phase transition is in consistence with the structural phase transition (discussed in main text). The ac susceptibility data is analyzed by Cole-Cole plot (figure S5 (left)) and Arrhenius fit (figure S6 (right)) and the obtained parameters are plotted in the main text. Further to analyzed the relaxation behavior in detail, ac susceptibility measurements were performed as a function of frequencies for T $<$ 40 K and is shown in figure S6.
**Figure S4**: Temperature dependence of real and imaginary part of ac susceptibility for Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (0 $\leq$ x $\leq$ 1.9) at various frequencies ranges between 10 - 1000 Hz. Left inset shows the temperature response of $\chi^{\prime}$ of Dy$_{2}$Zr$_{2}$O$_{7}$ and right inset shows the temperature response of $\chi^{\prime\prime}$ for DyLaZr$_{2}$O$_{7}$.
The inverse susceptibility data of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (measured at H = 5 kOe) is fitted with Curie-Weiss law in the high temperature regime (T $\geq$ 30 K) using the relation $\chi$ = C/(T - $\theta_{cw}$), where C and $\theta_{cw}$ stands for Curie constant and Curie-Weiss temperature respectively. From the Curie constant, the effective magnetic moment was calculated using $\mu_{eff}$ = $\sqrt{3k_{B}C/N_{A}}$, where $N_{A}$ is the Avogadro$^{\prime}$s number. The obtained value of $\theta_{cw}$ and $\mu_{eff}$ for all the substituted compounds are listed in Supplementary table II. The negative value of $\theta_{cw}$ signifies the dominance of AFM interactions in the system.
**Figure S5**: Frequency dependence of the imaginary part of ac susceptibility of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (1 $\leq$ x $\leq$ 1.9) below and above the freezing temperature in zero applied field. The presence of prominent single peak is the signature of single characteristic relaxation time.
La (x) $\mu_{eff}$ ($\mu_{B}$) $\theta_{cw}$ (K) Ref
-------- ------------------------- ------------------- ---------------------- --
0 16.4 -9.6(2) [@devi2020emergence]
0.3 14.12 -7.14(1) ,,
0.5 12.94 -7.02(3) This study
1.0 10.51 -5.63(2) ,,
1.5 7.56 -6.61(2) ,,
1.6 6.80 -5.64(2) ,,
1.7 5.93 -5.80(2) ,,
1.8 4.67 -6.63(7) ,,
1.9 3.21 -6.17(5) ,,
: Magnetic parameters of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$; 0 $\leq$ x $<$ 2
**Figure S6**: Cole-Cole plot of $\chi^{\prime}$ and $\chi^{\prime\prime}$ data around the freezing temperature and Arrhenius fit of Dy$_{2-x}$La$_{x}$Zr$_{2}$O$_{7}$ (1.5 $\leq$ x $\leq$ 1.9).
La (x) r$_{A}$/r$_{B}$ $\chi^{2}$ x(O) a($\AA$) Phase$\%$ Ref
-------- -- ----------------- -- ------------ -- ----------- -- ------------ -- ----------- -- ----------------------
0.0 1.43 1.46 0.3662(0) 10.4511(3) 100 [@devi2020emergence]
0.15 1.44 2.26 0.3674(0) 10.4832(4) 100 ,,
0.3 1.45 2.22 0.3692(0) 10.4972(3) 100 ,,
0.5 1.47 1.51 0.3710(2) 10.5521(6) 100 This study
0.6 1.48 2.15 0.3279(0) 10.5996(6) 60 ,,
0.3724(0) 10.4797(3) 35 ,,
0.3499(0) 10.3895(6) 5 ,,
0.8 1.49 1.95 0.3270(0) 10.6474(8) 88 ,,
0.3600(0) 10.4970(6) 8 ,,
0.3720(0) 10.3896(6) 4 ,,
1.0 1.52 2.85 0.3241(4) 10.6669(2) 97 ,,
0.3630(4) 10.4996(2) 2 ,,
0.3730(4) 10.3735(2) 1 ,,
1.5 1.56 1.70 0.3292(1) 10.7444(3) 100 ,,
1.6 1.57 1.36 0.3290(0) 10.7632(3) 100 ,,
1.7 1.58 1.43 0.3287(3) 10.7786(3) 100 ,,
1.8 1.59 1.23 0.3283(3) 10.7812(3) 100 ,,
1.9 1.60 1.49 0.3282(0) 10.7907(2) 100 ,,
2.0 1.61 1.40 0.3277(5) 10.7978(1) 100 ,,
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cross-correlation function (CCF) has become the standard tool for extraction of radial-velocity and broadening information from high resolution spectra. It permits integration of information which is common to many spectral lines into one function which is easy to calculate, visualize and interpret. However, CCF is not the best tool for many applications where it should be replaced by the proper broadening function (BF). Typical applications requiring use of the BF’s rather than CCF’s involve finding locations of star spots, studies of projected shapes of highly distorted stars such as contact binaries (as no assumptions can be made about BF symmetry or even continuity) and \[Fe/H\] metallicity determinations (good baselines and avoidance of negative lobes are essential). It is stressed that the CCF’s are not broadening functions. The note concentrates on the advantages of determining the BF’s through the process of linear inversion, preferably accomplished using the Singular Value Decomposition (SVD). Some basic examples of numerical operations are given in the IDL programming language.'
author:
- Slavek Rucinski
title: 'Determination of broadening functions using the Singular Value Decomposition (SVD) technique'
---
Convolution and cross-correlation
=================================
Convolution is an operation that the nature does for us. We seldom see “naked” functions. These could be a convolution of a spectrum with the spectrograph’s instrumental profile or with the radial component of the micro-turbulence velocity field in the stellar atmosphere or with a broadening function due to rapid rotation of a star. Thus, instead of a function $f(u)$, we observe a function $h(x)$ which is a convolution with some other broadening function (BF), $g(x)$: $$h(x) = \int_{-\infty}^{+\infty} f(u) \; g(x-u)\; du = f(x) \ast g(x)$$ This natural process can be easily simulated in numerical packages (examples in the IDL programming language are marked by a command-line prompt [IDL>]{}) either by a special operator:\
[IDL> h = convol(f,h) ]{}\
or through the Fourier-transform multiplication and its inverse:\
[IDL> h = float(fft(fft(f,-1)\*fft(g,-1),+1)) ]{}
Cross-correlation is an operation which for real functions differs from the convolution really only in the symmetry of the arguments. For complex functions things are slightly different (real and imaginary parts have different symmetries), but astronomers observe real spectra so we do not have to worry about the mathematical nuances. $$c(x) = \int_{-\infty}^{+\infty} f(u) \; g(u+x)\; du = f(x) \star g(x)$$ The cross-correlation (note the different asterisk above) function can be computed numerically using:\
[IDL> lag = findgen(201) - 100 ]{}\
[IDL> c = c\_correlate(f,g,lag) ]{}
Broadening functions
====================
Suppose we observe a sharp-line ($S(\lambda)$) and a broad-line ($P(\lambda)$) spectra and we want to determine the broadening and any other differences which make the latter spectrum more interesting than the former. The function $B(\lambda)$ can be a rotational broadening function for a single, rapidly-rotating star, or a more complex profile for two components of a binary, or for a star with spots (where they would show as indentations in the function). The sharp-line spectrum $S(\lambda)$ is not free of some broadening. This can be the thermal broadening of lines or micro-turbulence or some other mechanisms; we call them jointly $T(\lambda)$. Thus, schematically, the sharp-line spectrum can be written as: $$S(\lambda) = (\sum_i a_i \delta(\lambda_i)) \ast T(\lambda)$$ while the broad-line spectrum, broadened additionally by $B(\lambda)$, can be written as: $$P(\lambda) = S(\lambda) \ast B(\lambda) = (\sum_i a_i \delta(\lambda_i))
\ast T(\lambda) \ast B(\lambda)$$
The cross-correlation (CCF) with the sharp-line spectrum is frequently taken as an estimate of $B(\lambda)$: $$C(\lambda) = S(\lambda) \star P(\lambda) =
S(\lambda) \star (S(\lambda) \ast B(\lambda)) =
T(\lambda) \ast T(\lambda) \ast B(\lambda)$$ The new function ${\cal B}(\lambda)$, is [*not identical*]{} to $B(\lambda)$, because it inherits the common broadening components (such as thermal, micro-turbulence, instrumental) from both spectra. Tonry & Davis (1979) showed that if those common components are represented by Gaussians, the addition is quadratical, which for these functions means repeated convolutions. Thus, CCF cannot be really used to replace the broadening function. But it can give us some approximation of it and will remain a useful tool to have some preliminary estimate on the degree of the line broadening. For symmetrical broadening functions, it will remain the simplest tool to determine the radial velocities simultaneously from many spectral lines.
The differences between the BF and the CCF can be seen when an artificial broadened spectrum is created by a convolution and then the resulting spectrum is subject to the CCF operation. The result (Figure 1) is obviously different from the BF: The CCF shows negative baseline excursions and, most worryingly, it shows the “peak-pulling” effect which would lead to an under-estimate of the individual component velocities. While this last problem can be overcome by applying the TODCOR technique (Zucker & Mazeh 1994), we clearly see that the CCF is not the BF.
The Fourier transform de-convolution
====================================
Some attempts to determine the broadening functions (Anderson et al. 1983) utilized the well known property of the Fourier transforms of a correspondence between convolutions and multiplications in the two relevant domains. Thus, a convolution: $$P(\lambda) = S(\lambda) \ast B(\lambda)$$ transformed with the Fourier transform $\cal F$ changes into a product of the transforms: $${\cal F} \left\{ P(\lambda) \right\} =
{\cal F} \left\{ S(\lambda) \right\} \cdot
{\cal F} \left\{ B(\lambda) \right\}$$
Therefore, the broadening function can be restored with: $$B(\lambda) \simeq {\cal F}^{-1} \left\{ {\cal F}
\left\{ P(\lambda) \right\} /
{\cal F} \left\{ S(\lambda) \right\} \right\}$$
This can be done in a compact way with:\
[IDL> b = float(fft(fft(p,-1)/fft(s,-1),+1)) ]{}
While the mathematical background is simple and easy, the practice is just the opposite. First, the resulting $B(\lambda)$ spans the whole spectral window, so that one determines a lot of zeroes; there is no “compression” information whatsoever. But, more importantly, the division operation usually ... does not work: the high frequency noise becomes amplified and some sort of frequency filtering is needed. The result may then actually depend on the applied filter.
Some authors (see for example the large collection of works of D. F. Grey) use spectra transformed into the frequency domain without the division step. This can be done for broadening mechanisms describable by simple functions or obeying some symmetries, but fails in cases of spots or of BF’s of close binary systems.
Convolution in the formalism of linear equations
================================================
There are two main issues that re-casting convolution into a set of linear equations can resolve. These are: (1) How to channel information over the whole spectrum (say 2000 pixel long) into the BF window (say 200 pixels long)? (2) How to utilize all information contained in sharp-line spectra and remove the influence of the noise in the continuum (which carries no information)?
The convolution can be written as an over-determined system of linear equations which link a sharp-line spectrum $\vec{S}(n)$, via the broadening function $\vec{B}(m)$, with the broadened spectrum $\vec{P}(m)$. The mapping is through the “design matrix” $\widehat{Des}(m,n)$ which is formed from the sharp line spectrum $\vec{S}(n)$ by consecutive vertical shifts by one element. In IDL, this can be done with a simple routine:
[function map4,s,m\
; m - must be odd, n must be even\
n = n\_elements(s) & t = fltarr(m) \# fltarr(n-m+1)\
; t(m,n-m) = t(small,large-small) dimensions\
for j = 0,m-1 do for i = m/2,n-m/2-1 do t(j,i-m/2)=s(i-j+m/2)\
return,t\
end ]{}
An example of using this routine to create a design matrix for a 201-pixel long window would be: [IDL> des = map4(s,201)]{}. The program spectra must accordingly be trimmed to [n-m+1]{} with: [IDL> p = p(m/2:n-m/2-1)]{}. The system of equations has a familiar form of the over-determined linear set:
(50,60) (0,0)[(15,60)\[c\][$\widehat{Des}$]{}]{} (20,22)[(5,15)\[c\][$\vec{B}$]{}]{} (30,28)[=]{} (40,0)[(5,60)\[c\][$\vec{P}$]{}]{}
Singular value decomposition (SVD)
==================================
One of the traditional ways of solving the system of equations above would be to transform it into a system of the “normal” equations of the size reduced from $m \times (n-m)$ to $m \times m$ by multiplication of both sides by the transpose of the design matrix. The result would be the BF defined in the least-squares sense, and it is possible to stop at this point. However, the Singular Value Decomposition technique also gives us such an answer, but – in addition – makes it possible to remove the influence of the continuum and its noise.
The SVD technique is beautifully described in the “numerical techniques Bible” of Press et al. (1986). They present it as a somewhat magic black box and for most users it is just fine. If you want to learn how the technique really works, then the books of Golub & Van Loan (1989) or Craig & Brown (1986) are probably the best references.
The essence of the SVD is the property that one can represent any matrix by a product of 3 matrices; in our case: $\widehat{Des} = \widehat{U} \,\widehat{W}
\, \widehat{V}^T$. These matrices are: the column ortho-normal $\widehat{U}$ and $\widehat{V}$ and the diagonal matrix $\widehat{W}$ (this is really a vector containing the diagonal elements). The property of the columns in $\widehat{U}$ and $\widehat{V}^T$ is that the following products, $\widehat{U}^T \widehat{U} =
\widehat{I}$ and $\widehat{V}^T \widehat{V} = \widehat{I}$, give the unity array $\widehat{I}$ (1 on the diagonal).
(80,60) (0,0)[(15,60)\[c\][$\widehat{Des}$]{}]{} (18,28)[=]{} (25,0)[(15,60)\[c\][$\widehat{U}$]{}]{} (45,22)[(15,15)\[c\][$\widehat{W}$]{}]{} (65,22)[(15,15)\[c\][$\widehat{V}^T$]{}]{}
In IDL, the operation is represented by: [IDL> svdc,des,w,u,v,/double]{}. Here, [des]{} is the only input quantity and the remaining parameters are what the routine produces as output. The keyword [/double]{} is for higher precision and is optional.
One can check the correctness of the operations by the following commands:\
[IDL> wf = fltarr(m,m) ]{}\
[IDL> for i = 0,m-1 do wf(i,i) = w(i) ]{}\
[IDL> des\_check = u \#\# wf \#\# transpose(v) ]{}
The three new matrices are all invertible: $\widehat{U}$ and $\widehat{V}$ are ortho-normal arrays, so that their inverses are just transposes, while the diagonal array $\widehat{W}$ is replaced by a diagonal array $\widehat{W1}$, with the diagonal elements containing the inverses, $w1_{i,i} = 1/w_i$:\
[IDL> w1 = fltarr(m,m) ]{}\
[IDL> for i = 0,m-1 do w1(i,i) = 1./w(i) ]{}\
The solution is given by: $ \vec{B} = \widehat{V} \widehat{W_1} (\widehat{U}^T \vec{P}) $ or schematically:
(140,60) (0,22)[(5,15)\[c\][$\vec{B}$]{}]{} (8,28)[=]{} (16,22)[(15,15)\[c\][$\widehat{V}$]{}]{} (34,22)[(15,15)\[c\][$\widehat{W1}$]{}]{} (52,22)[(60,15)\[c\][$\widehat{U}^T$]{}]{} (118,0)[(5,60)\[c\][$\vec{P}$]{}]{}
Numerically: [IDL> b = reform(v\#\#w1\#\#(transpose(u)\#\#p))]{}. A routine makes this simpler: [IDL> b = svsol(u,w,v,p,/double)]{}. The elements of $\vec{B}$ are all independent, so that any – even strange or discontinuous – broadening functions can be restored as no condition imposed on the smoothness or symmetry of the result. Note that, if only one sharp-line template is used, the decomposition operation [svdc]{} is done only once, for possibly many broad-line spectra [p]{}, each giving a separate solution [b]{}.
Advantages and disadvantages of the SVD approach
================================================
On the positive side: (1) The problem can be treated as a linear-equations. (2) An “inverse” of the [*rectangular*]{} array $\widehat{Des}$ is possible. (3) The solution of $\vec{B}$ is defined in the least-squares sense (shortest modulus). (4) The result is the real broadening function. But there are also minuses: (5) One must solve a large system of, say, 2000 equations for 200 unknowns. (6) One must know a priori how many unknowns. (7) Initially, the results may turn out quite poor, because of the presence of plenty of linearly-dependent equations in the system (parts of spectra where the featureless continuum provides no broadening information).
The SVD approach offers a simple resolution of (7) as [*it permits removal of the effects of the continuum in an objective way*]{}. The key element here are the singular values contained in the diagonal of $\widehat{W}$. Since the solution involves $1/w_i$, small values in $w_i$ spoil the solution. These are exactly those problematic values that one wants to avoid. Thus, by rejecting of small values of $w_i$, one can (i) remove the linearly dependent equations, (ii) diminish the influence of the noise from the continuum, (iii) reduce the influence of the computer round-off errors (which enter multiplied by the order of the problem) and (iv) reduce the number of the unknowns (because the system is usually not over-determined at all). All these properties are related to the “conditioning” of the array $\widehat{Des}$. The reader is directed to the source texts on this subject for further reading.
The important factor is max($w_i$)/min($w_i$) which provides an estimate on how many of the singular should be used. In practice, one can keep on adding more $w_i$ and see the successively better solutions. The diagonal arrays $\widehat{W}$ and $\widehat{W1}$ will have then elements: $w_i = w_0, w_1, w_2, w_k, ... , w_{m-1}$ and $w1_i = 1/w_0, 1/w_1, 1/w_{k-1}, 0, ... , 0$ with $k$ (we call it the order of solution) spanning the whole range 0 to $m-1$. In IDL, this can be done by forming a square matrix of solutions:
[b = fltarr(m,m)\
for i = 0,m-1 do begin\
wb = fltarr(m)\
wb(0:i) = w(0:i) ; first i+1 singular values used, rest zero\
b(\*,i) = svsol(u,wb,v,p,/double)\
end for]{}
We note in passing, that the arrays $\widehat{U}$ and $\widehat{V}$ have very special properties as they contain the basis vectors in the spaces of the spectra and broadening functions, respectively. One can see this by analyzing a diagonalized system: $\widehat{W}\, \vec{Z} = \vec{D}$, obtained by keeping the same $\widehat{W}$, with $\vec{Z} = \widehat{V}^T \vec{B}$ and $\vec{D}
= \widehat{U}^T \vec{P}$. The solution of the diagonalized system would be then: $\vec{Z} = \vec{D}/\vec{W}$, but, in practice, the diagonal of $\widehat{W}$ is a vector, so that $z_i = d_i/w_i$. Plotting the columns of $\widehat{U}$ and $\widehat{V}$ can tell one a lot about the conditioning of the solution.
A few notes on the SVD solutions
================================
The first question is: How far in $k$ should one go and where to stop? The essential operation is to plot (usually in log units) the vector $\vec{W}$ (Figure 2). There are 3 parts of it: (1) the good, large singular values, (2) the small values usually representing the noise in the spectrum $S(\lambda)$ and (3) the numerical errors. You may want to stop no later than at the kink below the good part. But the real “quality control” is the fit to $\vec{P}$. If the error of the fit stops decreasing, you have found the right point (Figure 3). Beyond that point, you will start fitting the noise! However, it is useful to analyze the solutions for different orders $k$ and see how they first improve and then get worse. Sometimes the fit will remain poor, in spite of the leveling of the error curve; this usually means a wrong choice of the sharp-line spectrum. The standard error of the fit can be calculated from:\
[sig = fltarr(m) ; error\
pred = des \#\# transpose(b) ; predicted fits\
for i=0,m-1 do sig(i) = sqrt(total((pred(i,\*)-p)2)/m)]{}
Usually, even very low order solutions are well defined (Figure 2), but stopping early is not always advisable because this leads to a loss of resolution, as then not all basis vectors contribute. The solution which – in principle – has all elements in $\vec{B}$ independent suddenly acquires inter-element correlations. Thus, it may be advantageous to go to a highest possible order ($k=m$) and thus insure that elements of $\vec{B}$ are uncorrelated, and then decrease the noise by smoothing.
One should be aware that the errors may be under-estimated for the case of truncated ($k < m$) solutions. While the prescriptions of Rix & White (1992) and Rucinski, Lu & Shi (1993) are based on the theory of the full SVD, the error analysis for the truncated case has not yet been done. In this situation, it may be advantageous to utilize techniques of the external estimates, such as the bootstrap or Monte Carlo. This subject certainly requires more work ...
Conclusions
===========
One can position the SVD technique of linear broadening function restoration as located between the cross-correlation and the direct modeling of the spectra. The BF’s determined with it are much better defined than the CCF’s, and are true broadening functions, not their proxies. They also integrate the geometrical information from a spectrum, but give well defined baselines without the CCF’s negative fringes. They do require a bit more computer work, but several numerical packages can easily handle large linear systems of equations involved here. Obviously, they cannot replace the spectrum synthesis, if these are needed, but can be a useful tool in their preparation.
Anderson, L., Stanford, D. & Leininger, D. 1983, , 270, 200 Craig, I,J.D. & Brown, J.C. 1986, Inverse Problems in Astronomy, (Bristol and Boston: Adam Hilger Ltd) Golub, G.H. & Van Loan, C.F. 1989, Matrix Computations, 2nd ed. (Baltimore: Johns Hopkins Univ. Press) Press, W.H., Teukolsky, S.A., Vetterling, W.T. & Flannery, B.P. 1986, Numerical Recipes, Cambridge University Press (all editions) Rix, H.-W. & White, S.D.M. 1992, , 254, 384 Rucinski, S.M. 1992, , 104, 1968 Rucinski, S.M., Lu, W.-X. & Shi, J. 1993, , 106, 1174; Tonry, J. & Davis, M. 1979, , 84, 1511 Zucker, S. & Mazeh, T. 1994, ApJ, 420, 806
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using the Green function formalism we calculate a current-induced spin polarization of weakly magnetized graphene with Rashba spin-orbit interaction. In a general case, all components of the current-induced spin polarization are nonzero, contrary to the nonmagnetic limit, where the only nonvanishing component of spin polarization is that in the graphene plane and normal to electric field. When the induced spin polarization is exchange-coupled to the magnetization, it exerts a spin-orbit torque on the latter. Using the Green function method we have derived some analytical formulas for the spin polarization and also determined the corresponding spin-orbit torque components. The analytical results are compared with those obtained numerically. Vertex corrections due to scattering on randomly distributed impurities is also calculated and shown to enhance the spin polarization calculated in the [*bare bubble*]{} approximation.'
address: |
$^1$Faculty of Physics, Adam Mickiewicz University, ul. Umultowska 85, 61-614 Poznań, Poland\
$^2$ Institute of Molecular Physics, Polish Academy of Sciences, ul. M. Smoluchowskiego 17, 60-179 Poznań, Poland
author:
- 'A. Dyrdał$^{1}$ and J. Barnaś$^{1,2}$'
title: 'Current-induced spin polarization and spin-orbit torque in graphene'
---
Introduction
============
One of the main issues of the present-day spin electronics, that is of great importance for further development of high-density memory devices and magnetic random access memories, is effective manipulation of magnetization by a spin-polarized current. The general idea of switching and controlling orientation of a magnetic moment with electric current flowing through a system is based on coupling between the electron spin and magnetic moments. Two such interactions turned out to be especially useful – exchange interaction and spin-orbit coupling.
In a magnetically nonuniform system, the spin-polarized current generates a torque that is a consequence of: (i) exchange coupling between the conduction electrons and magnetization, and (ii) conservation of angular momentum in the system. The torque appears then as a result of the spin angular momentum transfer from a spin-polarized current (or pure spin current) to magnetic moments. Therefore, this torque is called spin-transfer torque. [@Manchon2007; @Ralph2008] Such a torque leads, among others, to magnetic switching in spin vales and to domain wall displacements, as observed recently in many experiments. Moreover, these phenomena give a possibility to construct low-power non-volatile memory cells (STT-MRAM, racetrack memory), integrated circuits employing a logic-in-memory architecture as well as logic schemes processing information with spins. [@Brataas2012]
Another possibility to control orientation of magnetic moments is based on a spin torque that appears due to spin-orbit interaction in the system. The corresponding torque exerted on the magnetization is usually referred to as the spin-orbit torque, and appears also in a magnetically uniform system, like a single uniform layer. Physical mechanism of the spin-orbit torque is based on a nonequilibrium spin polarization of the system, which is induced by an external electric field (current) in the presence of spin-orbit interaction. Such a spin polarization was predicted long time ago in nonmagnetic systems, where an electric current flowing through the system with spin-orbit interaction was shown to induce not only the transverse spin-current [@dyakonov71_she; @hirsch] (so-called spin Hall effect), but also a spin-polarization of conduction electrons. [@dyakonov71; @aronov89; @edelstein90; @aronov91; @Golub; @Shen] In the case of two-dimensional electron gas with Rashba spin-orbit interaction, the induced spin polarization is in the plane of the electron gas and normal to the electric field. Such a nonequilibrium spin-polarization may be treated as an effective magnetic field, which may lead to reorientation of a magnetic moment, and also can modify or induce magnetic dynamics. The spin-orbit torque was analyzed in recent few years in many papers, mainly in metallic and semiconductor heterostructures. [@Zhang2008; @Zhang2009; @Abiague2009; @Manchon; @Gambardella; @Garello2013] While the current-induced spin polarization, known also as the inverse spin-galvanic effect, [@Ganichev2002] is well known and was investigated theoretically as well as experimentally in the recent three decades, the role of geometric phase in this effect, and consequently in the spin-orbit torque, was invoked only very recently. [@Kurebayashi; @Avci; @Li2014]
In this paper we consider the current-induced spin polarization and spin-orbit torque in graphene, which is assumed to be deposited on a substrate that ensures the presence of spin-orbit interaction of Rashba type [@Dedkov2008]. We also assume that the graphene is magnetized, which may be either due to the magnetic proximity effect to a ferromagnetic substrate (or cover layer), or due to magnetic atoms (nanoparticles) on its surface. [@Yokoyama2008; @Haugen2008; @ZPNiu2011; @ZPNiu2014; @KawakamiZutic2014] Coexistence of the Rashba spin-orbit interaction and proximity-induced magnetism in graphene was predicted theoretically and also observed experimentally.[@KawakamiZutic2014; @Qiao2010; @Qiao2014; @Wang2015; @Gong2011] As the spin transfer torque in ferromagnetic graphene junctions was already considered theoretically (see e.g. Yokoyama and Linder [@Yokoyama]), the problem of spin torques induced by spin-orbit interaction in graphene is rather unexplored.
It has been shown that the current-induced spin polarization in a defect-free nonmagnetic graphene with Rashba spin-orbit interaction is oriented in the graphene plane and is also normal to the current orientation. Moreover, sign of the spin-polarization depends on the chemical potential and also on the sign of the Rashba spin-orbit coupling parameter. [@Dyrdal2014] When the Fermi level passes through the Dirac points, the spin polarization becomes reversed. In this paper we show that the current-induced spin polarization in magnetic graphene has generally all three components. In the approximation linear with respect to the magnetization, one of these components is equal to that in the case of a nonmagnetic graphene, i.e. it is proportional to the relaxation time. The leading terms in the other two components are independent of the relaxation time.
The paper is organized as follows. In Sec. 2 we describe the model and present a general formula describing current-induced spin polarization. Analytical formulas as well as numerical results for the current-induced spin polarization are presented in Sec. 3. Vertex correction is calculated in section 4, while the spin-orbit torque is described and discussed in Sec. 5. Summary and final conclusions are in Sec. 6.
Model and method
================
Transport properties of graphene close to the charge neutrality point are determined mainly by electrons in the vicinity of Dirac points. The corresponding effective-mass Hamiltonian, $H^0_K$, which describes the low-energy electronic states in graphene around the $K$ point of the Brillouin zone, can be written as a sum of three terms, [@kane] $$H^{0}_{K} = H_{0} + H_{R} + H_{\bf{M}} .$$ The first term, $H_{0}$, describes the low energy electronic states of pristine graphene, and can be written as a matrix in the pseudospin (sublattice) space, $$\begin{aligned}
\label{hk0}
H_{0} = v \left(
\begin{array}{cc}
0 & (k_{x} - i k_{y}) \sigma_{0} \\
(k_{x} + i k_{y}) \sigma_{0} & 0 \\
\end{array}
\right),\end{aligned}$$ where $v = \hbar v_{F}$, with $v_{F}$ denoting the electron velocity in graphene, which is constant. The second term in Eq. (1) describes the Rashba spin-orbit interaction due to a substrate, $$\begin{aligned}
\label{hkR}
H_{R} = \lambda \left(
\begin{array}{cc}
0 & \sigma_{y} + i \sigma_{x} \\
\sigma_{y} - i \sigma_{x} & 0 \\
\end{array}
\right),\end{aligned}$$ with $\lambda$ being the Rashba spin-orbit coupling parameter. The last term of the Hamiltonian (1) represents the influence of an effective exchange field $\tilde{\mathbf{M}}$ created by a nonzero magnetization. Such a magnetization can appear in graphene, for instance, due to the proximity effect to a magnetic substrate. This term can be written in the form, $$\begin{aligned}
\label{hkM}
H_{\bf{M}} = - \tilde{\mathbf{M}} \cdot \left(
\begin{array}{cc}
\bsig & 0 \\
0 & \bsig\\
\end{array}
\right),\end{aligned}$$ where the exchange field $\tilde{\mathbf{M}}$ is measured in energy units. This field can be related to the magnetization $\mathbf{M}$ and the local exchange interaction between the conduction electrons and magnetization in the two-dimensional graphene, $J_{\rm ex}({\bf r}-{\bf r}^\prime )= J_{\rm ex}\,\delta ({\bf r}-{\bf r}^\prime )$, [*via*]{} the formula $\tilde{\mathbf{M}} = (J_{\rm ex}/2g\mu_B){\mathbf M}$, Here, $g$ is the Lande factor ($g=2$), $\mu_B$ is the Bohr magneton, while positive and negative $J_{\rm ex}$ (measured in the units of Jm$^2$) correspond to antiferromagnetic and ferromagnetic coupling, respectively. In the above equations, $\bsig$ is the vector of Pauli matrices, $\bsig =(\sigma_x,\sigma_y, \sigma_z )$, while the matrix $\sigma_{0}$ denotes the unit matrix in the spin space. Note that the so-called intrinsic spin-orbit interaction in graphene is very small and therefore it is neglected in our consideration. In a general case, the magnetization vector $\mathbf{M}$ may be oriented arbitrarily in space, and its orientation will be described by two spherical angles, $\theta$ and $\xi$, as indicated in Fig. 1. Moreover, the absolute magnitude of $\bf M$ is assumed to be constant, $|{\bf M} |\equiv M={\rm const}$. Hamiltonian for the second non-equivalent Dirac point, $K^\prime$, can be obtained from $H_K$ by reversing sign of the wavevector component $k_x$ and substitution $\sigma_{y} \rightarrow - \sigma_{y}$ in $H_{R}$.
![(Color online) Schematic of the system under consideration. Graphene is on a substrate which assures a nonzero magnetization and also a spin-orbit interaction of Rashba type. Orientation of the magnetic moment $\bf M$ is described by the angles $\theta$ and $\xi$. An external electric field is oriented along the axis $y$.[]{data-label="schematic"}](Fig1.eps "fig:"){width="220pt"}\
In the lowest order with respect to the exchange field $\tilde{\mathbf{M}}$, the casual Green function corresponding to the Hamiltonian (1), $G_{k}^{0} = \{ [\varepsilon + \mu + i \delta\, {\rm sign}\,(\varepsilon)] - H^{0}_{K}\}^{-1}$ has poles at $\varepsilon = E_{n} - \mu - i \delta\,{\rm sign}\,(\varepsilon)$, where $E_{n}$ ($n = 1-4$) are eigenvalues of the Hamiltonian (1) without the term $H_{\mathbf{M}}$. These eigenvalues have the following form: $$\begin{aligned}
E_{1,2} = \mp \lambda - \sqrt{k^{2}v^{2} + \lambda^{2}}\\
E_{3,4} = \mp \lambda + \sqrt{k^{2}v^{2} + \lambda^{2}},\end{aligned}$$ where $E_{1,2}$ correspond to the valance bands, while $E_{3,4}$ describe the conduction bands. Note, the bands corresponding to $n=2$ and $n=3$ touch each other at the Dirac point ($k=0$), while a gap equal to $4\lambda$ appears between the bands $n=1$ and $n=4$.
In the presence of a dynamical external electric field applied along the axis $y$, the total Hamiltonian for electrons near the $K$ point takes the form $$H = H^{0}_{K} + H_{K}^{\mathbf{A}},$$ where the second term, $$H_{K}^{\mathbf{A}} = - e \hat{v}_{y} A_{y}(t) = - i e \frac{v}{\hbar} \left( \begin{array}{cc}
0 & -\sigma_{0} \\
\sigma_{0} & 0 \\
\end{array}\right) A_{y}(t),$$ is the perturbation due to interaction with the time-dependent electromagnetic field represented by the vector potential $A_{y}(t) = A_{y}e^{-i \omega t}$. Here, $e$ is the electron charge, $\hat{v}_y$ is the $y$-component of the electron velocity operator, $\hat{\mathbf v}=\partial H^0_K/\partial \mathbf{k}$, whereas $\omega$ is the frequency of the dynamical field (later we will take the limit of $\omega \to 0$).
When an electric current flows in the system due to the electric field, electron spins become polarized as a result of the co-operation of the current and Rashba spin-orbit coupling. This nonequilibrium spin polarization of conduction electrons can be calculated (in the zero-temperature limit) using the following formula: $$\label{S_alpha}
S_{\alpha}(t) = - i \mathrm{Tr}\int \frac{d^{2} \mathbf{k}}{(2 \pi)^{2}} \hat{S}_{\alpha} G_{\mathbf{k}}(t, t^\prime)|_{t^\prime =t+0},$$ where $G_{\mathbf{k}}(t, t^\prime)$ is the zero-temperature causal Green function corresponding to the total Hamiltonian $H$ (see Eq. 7), and $\hat{S}_{\alpha}$ is the spin vertex function defined as $$\label{sa}
\hat{S}_{\alpha} = \frac{\hbar}{2} \left( \begin{array}{cc}
\sigma_{\alpha} & 0 \\
0 & \sigma_{\alpha} \\
\end{array}\right).$$ Upon Fourier transformation with respect to the time variables and expansion in a series with respect to the vector potential $A_{y} = - i E_{y}/\omega$, the expression (\[S\_alpha\]) for the induced nonequilibrium spin density takes the form $$\begin{aligned}
\label{S_op}
S_{\alpha} (\omega) = \frac{e E_y}{ \omega} {\rm{Tr}}\int\frac{d^{2}\mathbf{k}}{(2\pi)^{2}} \int\frac{d\varepsilon}{2 \pi}
\hat{S}_{\alpha}
\;G^0_{\mathbf{k}}(\varepsilon + \hbar\omega) \hat{v}_{y} G^0_{\mathbf{k}}(\varepsilon).\hspace{0.6cm}\end{aligned}$$ In the dc limit, $\omega \rightarrow 0$, the above formula leads to the following expression for the spin polarization: $$\begin{aligned}
\label{S_alpha1}
S_{\alpha} = \frac{e}{2 \pi} E_{y} \hbar {\rm{Tr}}\int\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}\hat{S}_{\alpha} G^{0R}_{\mathbf{k}} \hat{v}_{y} G^{0A}_{\mathbf{k}},\end{aligned}$$ where $G^{0R(A)}_{\mathbf{k}}$ is the retarded (advanced) Green function corresponding to the unperturbed Hamiltonian (1), taken at the Fermi level ($\varepsilon = 0$). Upon taking into account Eqs (\[sa\]) and (\[S\_alpha1\]), and also including the contribution from the second Dirac point, the expression for the induced spin polarization acquires the form $$\begin{aligned}
\label{S_alpha_main}
S_{\alpha} = \frac{e\hbar^{2}}{2 \pi} E_{y} {\rm{Tr}}\int\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}\left(
\begin{array}{cc}
\sigma_{\alpha} & 0\\
0 & \sigma_{\alpha} \end{array} \right) G^{0R}_{\mathbf{k}} \hat{v}_{y} G^{0A}_{\mathbf{k}}.\end{aligned}$$ Based on this formula we calculate analytically as well as numerically the current-induced spin polarization, as described and discussed in the subsequent section.
Current-induced spin polarization
=================================
From the general formula (\[S\_alpha\_main\]) one finds the following expression for the $\alpha$-th component of the spin polarization: $$\label{12}
S_{\alpha} = \frac{e \hbar}{2 \pi}E_{y} \int \frac{dk k}{(2\pi)^{2}} \frac{T_{\alpha}}{\Pi_{n=1}^{4}(\mu - E_{n} + i \Gamma)(\mu - E_{n} - i \Gamma)},$$ where $T_{\alpha}$ is defined as $$\label{11}
T_{\alpha} = \hbar \int_{0}^{2 \pi} d\phi {\rm{Tr}} \left[ \left(
\begin{array}{cc}
\sigma_{\alpha} & 0\\
0 & \sigma_{\alpha} \end{array} \right) g^{0R}_{\mathbf{k}} \hat{v}_{y} g^{0A}_{\mathbf{k}}\right].$$ Here, $g^{0R(A)}_{\mathbf{k}}$ is the nominator of the retarded (advanced) Green function, $\phi$ stands for the angle between the axis $x$ and the wavevector $\bf k$, while $\Gamma = \hbar/2 \tau$, where $\tau$ is the momentum relaxation time. The parameter $\Gamma$ (or equivalently relaxation time $\tau$) will be treated here as a phenomenological parameter, which effectively includes contributions due to momentum relaxation from various scattering processes (scattering on impurities, other structural defects, phonons, or electron-electron scattering). Note that $\Gamma$ depends in general on the chemical potential $\mu$ and may be also different in the two Rashba subbands. However, when the Fermi level is in the two subbands, we assume for simplicity the same $\Gamma$ for both of them.
Up to the terms linear in the exchange field $\tilde{M}$, the functions $T_{\alpha}$ ($\alpha =x, y, z$), see Eq. (15), can be written as follows:
$$\begin{aligned}
\label{Ta}
T_{x} = 16 \lambda \pi v \mu (k^{4} v^{4} - \mu^{4} + 4 \lambda^2 \mu^{2}),\\
T_{y} = 64 \pi v \lambda \tilde{M} \mu (k^{2} v^{2} - 2 \lambda^{2}) \Gamma \sin\theta, \label{Tb}\\
T_{z} = - 64 \pi \lambda \tilde{M} v^{3} k^{2} \mu \Gamma \cos\theta \sin\xi .\label{Tc}\end{aligned}$$
Note, the dependence on the orientation of $\bf M$ is contained in the above expressions for $T_y$ and $T_z$, while $T_x$ is independent of $\bf M$. Equations (14) and (16) allow finding spin polarization in a general case, i.e. for an arbitrary relaxation time. However, some analytical expressions for all components of the spin polarization can be obtained in the limit of low impurity concentration, i.e. for long relaxation times ($\tau \rightarrow \infty $).
Consider first the $x$-component of the spin polarization. Combining Eq.(\[Ta\]) with Eq.(\[12\]) and making the substitution $\sqrt{k^2 v^2 + \lambda^2} = \gamma$, one obtains $$\begin{aligned}
\label{Sx_0a}
S_{x} = 8 e \hbar E_{y} \lambda \mu \hspace{6cm}\nonumber\\ \times \int_\lambda^\infty \frac{d \gamma \gamma}{v (2\pi)^2} \frac{ (\gamma^2 - \lambda^2)^2 - \mu^4 + 4 \lambda^2 \mu^2}{[(\mu + \lambda + \gamma)^2 + \Gamma^2] [(\mu - \lambda + \gamma)^2 + \Gamma^2]}\nonumber\\
\times \frac{1}{[(\mu + \lambda - \gamma)^2 + \Gamma^2][(\mu - \lambda - \gamma)^2 + \Gamma^2]}.\hspace{0.5cm}\end{aligned}$$ From this formula follows that $S_{x}$ is independent of $\tilde{M}$ in the linear approximation with respect to the exchange field. For long relaxation times we get the same analytical formulas as those in the case of nonmagnetic graphene, [@Dyrdal2014] i.e., $$\label{sx1}
S_{x} = \frac{e}{4 \pi} \frac{2 \lambda \pm \mu}{v(\lambda \pm \mu)} \mu E_{y} \tau$$ for the Fermi level lying in the range $-2 \lambda < \mu < 2 \lambda$, and $$\label{sx2}
S_{x} = \pm \frac{e}{4 \pi} \frac{2 \lambda}{v (\mu^{2} - \lambda^{2})} \mu^{2} E_{y} \tau$$ for $|\mu| > 2 \lambda$. In both above equations (as well as below), the upper and lower signs correspond to $\mu >0$ and $\mu <0$, respectively.
{width="90.00000%"}
The spin polarization given by Eqs (\[sx1\]) and (\[sx2\]) is proportional to $\tau$. However, one should bear in mind that these formulas were derived on the assumption of long $\tau$. Therefore, one may expect some deviations from this formula when $\tau$ is finite and not too long. In Fig. \[Fig:sx\](a) we show variation of the $S_x$ component of spin polarization with the chemical potential $\mu$ and relaxation time $\tau$, obtained by numerical integration of the formula (17). Figures \[Fig:sx\]b and \[Fig:sx\]c, in turn, present cross-sections of the density plots shown in Fig. \[Fig:sx\](a) for constant values of $\tau$ and $\mu$, respectively. The results obtained from the analytical formulas are compared in Figs \[Fig:sx\](b) with those obtained by numerical integration of the formula (17). From this comparison follows that for $\tau$ of the order of $10^{-11}$s or smaller, there are some deviations from the results given by the analytical formulas, though these deviations are not large. For $\tau$ of the order of $10^{-10}$s or longer, numerical results match quite well those obtained from the analytical formulas. Since the $S_x$ component is the same in magnetic and nonmagnetic limits (within the approximations used here), and in the nonmagnetic limit it was considered and analyzed in Ref.\[\], we will not discuss this component in more detail.
From Eqs (\[12\]) and (\[Tb\]) one finds the $y$-component of the spin polarization in the following form: $$\begin{aligned}
\label{Sy_0}
S_{y} = 32 e E_{y} \hbar \lambda \mu \tilde{M} \hspace{6cm}\nonumber\\ \times \int_\lambda^\infty \frac{d \gamma \gamma}{v (2\pi)^2} \frac{(\gamma^2 - 3 \lambda^2) \Gamma \sin\theta }{[(\mu + \lambda + \gamma)^2 + \Gamma^2] [(\mu - \lambda + \gamma)^2 + \Gamma^2]}\nonumber\\
\times \frac{1}{[(\mu + \lambda - \gamma)^2 + \Gamma^2][(\mu - \lambda - \gamma)^2 + \Gamma^2]}\, .\hspace{0.8cm}\end{aligned}$$ In the limit of slow relaxation, $\Gamma \to 0$, the above formula leads to the following analytical results: $$\label{sy1}
S_{y} = \pm \frac{e \hbar}{4\pi} \frac{\tilde{M} }{\lambda} \sin\theta \frac{\mu (\mu \pm 2\lambda) - 2 \lambda^{2}}{2v\mu(\mu \pm \lambda)} E_{y}$$ for $|\mu| < 2 \lambda$, and $$\label{sy2}
S_{y} = \pm \frac{e \hbar}{4\pi} \frac{\tilde{M} }{\lambda} \sin\theta \frac{\mu^{2} - 4 \lambda^{2}}{v(\mu^{2} - \lambda^{2})} E_{y}$$ for $|\mu| > 2 \lambda$.
Numerical results for the $y$-component of the current-induced spin polarization, obtained by numerical integration of the formula (20) are shown in Fig.2(d) as a function of chemical potential $\mu$ and relaxation time $\tau$. Figures 2(e,f) present cross-sections of Fig.2(d). Figure 2(e) additionally shows the results obtained from analytical formulas, see the curves for $\tau\to\infty$. Right parts of Fig.2(d,e,f) present in more detail the corresponding shaded regions. Similarly as the $x$-component, $S_y$ is antisymmetric with respect to reversal of the sign of Fermi energy, and its dependence on $\mu$ also reveals some steps at $\mu = \pm 2\lambda$. These steps are associated with the edges of the bands $E_1$ and $E_4$. Moreover, when the Fermi level is at the Dirac point ($\mu = 0$), the analytical solution (\[sy1\]) for $S_{y}$ becomes divergent. To understand origin of the divergency in the analytical solution for $\tau\to \infty$, one should note that the solution for the $x$-component is also infinite for $\tau\to\infty$, independently of $\mu$. This clearly shows that the limit of $\tau\to \infty$ is not physical as the dissipation processes are necessary in order to stabilize a finite current-induced deviation of the system from equilibrium, and thus also a finite current density and spin polarization. Therefore, in Fig.2(e) we compare the numerical results based on the corresponding analytical formulas with those obtained by numerical integration. This comparison clearly shows that the results obtained from the analytical formulas are roughly in agreement with those obtained from numerical integration, except the vicinity of $\mu=0$, where the analytical solution diverges for $\mu\to 0$, while the numerical results based on Eq.(\[Sy\_0\]) are then finite. Moreover, some discrepancy also occurs around $\mu=\pm 2\lambda$, but now the difference is finite and rather small. Thus, one should bear in mind that the analytical results (\[sy1\]) and (\[sy2\]) for the $y$-component have limited applicability range, and are not applicable for $\mu$ in the vicinity of the Dirac points.
The $S_{z}$ component can by found from Eqs (\[12\]) and (\[Tc\]) and acquires the form $$\begin{aligned}
\label{Sz_0}
S_{z}= - 32 e E_{y} \hbar \lambda \mu \tilde{M} \cos\theta \sin\xi \hspace{4cm}\nonumber\\ \times \int_\lambda^\infty \frac{d \gamma \gamma}{v (2\pi)^2} \frac{(\gamma^2 - \lambda^2) \Gamma }{[(\mu + \lambda + \gamma)^2 + \Gamma^2] [(\mu - \lambda + \gamma)^2 + \Gamma^2]}\nonumber\\
\times \frac{1}{[(\mu + \lambda - \gamma)^2 + \Gamma^2][(\mu - \lambda - \gamma)^2 + \Gamma^2]}\, .\hspace{0.8cm}\end{aligned}$$ Similar calculations as those done for the $y$-component lead to the following analytical expressions in the limit of long relaxation time: $$\begin{aligned}
\label{sz1}
S_{z} = \mp \frac{e \hbar}{4\pi} {\tilde{M}}{\lambda} \cos\theta \sin\xi \frac{\mu \pm 2 \lambda}{2 v (\mu \pm \lambda)} E_{y}\end{aligned}$$ for $\mu < 2\lambda$, and $$\begin{aligned}
\label{sz2}
S_{z} = \mp \frac{e \hbar}{4\pi} \frac{\tilde{M}}{\lambda} \cos\theta \sin\xi \frac{\mu^{2} - 2 \lambda^{2}}{v (\mu^{2} - \lambda^{2})} E_{y}\end{aligned}$$ for $\mu > 2\lambda$.
In Fig.2(g) we present the $z$-component of the current-induced spin polarization, calculated as a function of the chemical potential and relaxation time by numerical integration of the formula (25). In turn, Figs.2(h,i) show cross-section of Fig.2(g). In Fig.2(h) we additionally compare the numerical results with those obtained from analytical solution. Now, the analytical solution is not divergent, see the curve for $\tau\to\infty$. When the relaxation time is sufficiently small, the numerical results obtained from Eq. (\[Sz\_0\]) deviate from the results obtained on the basis of the analytical formulas. These deviations are rather small for $\tau\gtrsim 10^{-11}$s, except the region near the zero chemical potential. However, the difference between the analytical and numerical results around $\mu =0$ is now much less pronounced than it was in the case of the $y$-component (compare Fig.2(e) and Fig.2(h). In turn, for $\tau\lesssim 10^{-11}$s the deviations become remarkable in the whole range of the chemical potentials shown in Fig.2(h).
All the components of the spin polarization ($S_x$, $S_{y}$ and $S_{z}$) vanish at $\mu = 0$ and are antisymmetric with respect to the sign reversal of the chemical potential. Numerical results presented above show that the spin polarization strongly depends on the Fermi level position. In the close vicinity of the Dirac points, the $y$-component of the spin polarization has pronounced peaks (positive above and negative below $\mu =0$). The other two components behave more regularly in this region. All three components exhibit some cusps (or dips) when $\mu$ is in the vicinity of $\mu = \pm 2\lambda$, i.e., when the Fermi level approaches the top edge of the band $E_{1}$ or bottom edge of the band $E_{4}$. The spin polarization also remarkably depends on the Rashba parameter $\lambda$. This dependence reveals peculiarities of the corresponding electronic structure, and remarkably depends on the Rashba parameter. In numerical calculations we assumed the Rashba spin-orbit coupling parameter $\lambda =2$ meV. Generally, this parameter depends on the substrate (or cover layer), and in real systems varies from a few to a few tens of meV, see eg. Refs \[\].
Vertex correction
=================
In the preceding section we have calculated spin polarization induced by electric field assuming effective relaxation time $\tau$ (or equivalently relaxation rate $\Gamma$). Both, $\tau$ and chemical potential were treated there as independent parameters. When considering a specific relaxation mechanism, these parameters usually are not independent. Since the dominant scattering processes are on impurities, we consider now this problem in more details. Assume the scattering potential created by randomly distributed weak short-range scatterers, which may be written as $V(\mathbf{r}) s_{0} \sigma_{0} $ with Gaussian correlations $\langle V(\mathbf{r}) V(\mathbf{r}'))\rangle = n_{i} V^{2}\delta(\mathbf{r} - \mathbf{r}')$ (where $s_{0}$ and $\sigma_{0}$ and denote unit matrix in the pseudo-spin and spin subspace respectively).
Detailed calculation of the self energy due to scattering on the point-like impurities gives $\Gamma_{1,4} = \frac{n_{i}V^{2}}{2v^{2}}(|\mu| - \lambda)$ and $\Gamma_{2,3} = \frac{n_{i}V^{2}}{2v^{2}}(|\mu| + \lambda)$, where $n_i$ is the impurity concentration while $V$ is the impurity scattering potential. When $|\mu |\gg \lambda$, then indeed $\Gamma_{1,4} \simeq \Gamma_{2,3}\equiv \Gamma$. Otherwise, we take $\Gamma$ as the average of $\Gamma_{1,4}$ and $\Gamma_{2,3}$, i.e. $\Gamma = \frac{n_{i}V^{2}}{2v^{2}}|\mu|$.
When calculating the impurity averaged conductivity, it is well known that non-crossing diagrams give an important contribution and renormalize the results obtained in the [*bare bubble*]{} approximation. Such a vertex renormalization is known to have a significant influence on the spin current induced [*via*]{} the spin Hall effect. In the case of two-dimensional electron gas with Rashba spin-orbit interaction it totally cancels the spin Hall conductivity obtained in the [*bare bubble*]{} approximation.[@Inoue04; @Raimondi05; @Dimitrova05; @Chalaev05] However, this is not a general property and in other systems the vertex corrections can only reduce partly the spin Hall effect.[@Murakami04; @Malshukov05; @Korotkov06]
The problem of disorder in graphene was discussed in many papers. [@McCann2006; @Ostrovsky2006; @Pachoud2014] However, there is still a lack of information on the influence of disorder and impurities on spin-orbit driven phenomena in graphene. This problem was raised by Sinitsyn *et al*. [@Sinitsyn] and Gusynin *et al.* [@Gusynin2014] in the context of spin Hall and spin Nernst effect in the presence of intrinsic spin-orbit interaction in graphene and in the case of spin-independent random potential. In this case problem becomes simpler because one can reduce the model to $2\times2$ space. Such a simplification, however, is not possible in the presence of Rashba spin-orbit interaction.
In the weak scattering limit, the localization corrections are vanishingly small and therefore only noncrossing ladder diagrams are important. The summation over the ladder diagrams can be represented by the vertex corrections to the current-induced spin polarization. The renormalized spin vertex function is then given by the following equation:[@mahan] $$\tilde{S}_{\alpha} = \hat{S}_{\alpha} + n_{i} V^{2} \int \frac{d^{2} \mathbf{k}}{(2\pi)^{2}} G_{\mathbf{k}}^{A} \tilde{S}_{\alpha} G_{\mathbf{k}}^{R},$$ where $\hat{S}_{\alpha}$ is defined by Eq. (10). For the point-like scattering potential one can postulate the vertex function $\tilde{S}_{\alpha}$ in the form $$\begin{aligned}
\tilde{S}_{\alpha} = a_{\alpha}\, \frac{\hbar}{2} \left(
\begin{array}{cc}
\sigma_{x} & 0 \\
0 & \sigma_{x} \\
\end{array}
\right) + b_{\alpha}\, \frac{\hbar}{2} \left(
\begin{array}{cc}
\sigma_{y} & 0 \\
0 & \sigma_{y} \\
\end{array}
\right)\nonumber\\ + c_{\alpha}\, \frac{\hbar}{2} \left(
\begin{array}{cc}
\sigma_{z} & 0 \\
0 & \sigma_{z} \\
\end{array}
\right) + d_{\alpha}\, \frac{\hbar}{2} \left(
\begin{array}{cc}
\sigma_{0} & 0 \\
0 & \sigma_{0} \\
\end{array}
\right),\end{aligned}$$ for $\alpha =x,y,z$, where $a_\alpha$, $b_\alpha$, $c_\alpha$, and $d_\alpha$ are certain parameters to be determined. To find these parameters we multiply Eq. (26) by the matrix as specified below and take the trace, $$\begin{aligned}
{\mathrm{Tr}}\left\{\left(
\begin{array}{cc}
\sigma_{i} & 0 \\
0 & \sigma_{i} \\
\end{array}
\right) \tilde{S}_{\alpha}\right\} = {\mathrm{Tr}}\left\{\left(
\begin{array}{cc}
\sigma_{i} & 0 \\
0 & \sigma_{i} \\
\end{array}
\right) \hat{S}_{\alpha}\right\}\nonumber\\ + n_{i} V^{2} \int \frac{d^{2} \mathbf{k}}{(2\pi)^{2}} {\mathrm{Tr}}\left\{ \left(
\begin{array}{cc}
\sigma_{i} & 0 \\
0 & \sigma_{i} \\
\end{array}
\right) G_{\mathbf{k}}^{A} \tilde{S}_{\alpha} G_{\mathbf{k}}^{R} \right\},\end{aligned}$$ for $i = 0,x,y,z$. Taking into account Eq. (27), one finds then a set of equations for the coefficients $a_{\alpha}, b_{\alpha}, c_{\alpha}, d_{\alpha}$.
We recall that in this paper the exchange field due to proximity effect is assumed to be small, so the current-induced spin polarization is limited to the terms linear in the exchange field. Consequently, the vertex correction is also calculated in the lowest order appropriate to have spin polarization linear in M.
For $\alpha = x$ we find that: $$\begin{aligned}
b_{x} = c_{x} = d_{x} = 0, \\
a_{x} = \frac{1}{1 - n_{i} V^{2} \mathcal{I}_{x}},\end{aligned}$$ where $$\mathcal{I}_{x} = \int \frac{dk k}{2\pi}\frac{\chi_{x}(\mu, \Gamma)}{\prod_{n = 1}^{4} (\mu - E_{n} + i \Gamma)(\mu - E_{n} - i \Gamma)}$$ and $$\begin{aligned}
\chi_{x}(\mu, \Gamma) = k^{6} v^{6} + k^{4} v^{4} (3 \Gamma^{2}-\mu^{2})+(\Gamma^{2}+\mu^{2})^3 \nonumber \\
+4 (\Gamma^{4}-\mu^{4}) \lambda^{2}+k^{2} v^{2} (\Gamma^{2}+\mu^{2}) (3 \Gamma^{2}-\mu^{2}+4 \lambda^{2})\nonumber \\
\approx (k^{2} v^{2} -\mu^{2} ) \left(k^{4} v^{4}-\mu^{4}+4 \mu^{2} \lambda^{2}\right)\nonumber\\
+\left(3 k^{4} v^{4}+3 \mu^{4} +2 k^{2} v^{2} \left(\mu^{2}+2 \lambda^{2}\right)\right)\Gamma^{2}.\end{aligned}$$ In the above equation, only terms up to the second order in $\Gamma$ have been retained, while terms of higher order have been omitted.
For $\alpha = y$ we find the following coefficients: $$\begin{aligned}
a_{y} = c_{y} = d_{y} = 0, \\
b_{y} = a_{x} = \eta.\end{aligned}$$
In turn, for $\alpha = z$ we find $$\begin{aligned}
a_{z} = b_{z} = d_{z} = 0, \\
c_{z} = \frac{1}{1 - n_{i} V^{2} \mathcal{I}_{z}} = \zeta ,\end{aligned}$$ where $$\mathcal{I}_{z} = \int \frac{dk k}{2\pi} \frac{\chi_{z}(\mu, \Gamma)}{\prod_{n = 1}^{4} (\mu - E_{n} + i \Gamma)(\mu - E_{n} - i \Gamma)}$$ and $$\begin{aligned}
\chi_{z}(\mu, \Gamma) = k^{6} v^{6} +k^{2} v^{2} \left(3 \Gamma^{2}-\mu^{2}\right) \left(\Gamma^{2} +
\mu^{2} \right)\nonumber\\
+ \left(\Gamma^{2}+ \mu^{2}\right)^{3} +k^{4} v^{4} \left(3 \Gamma^{2}-\mu^{2} - 2 \lambda^{2}\right)\nonumber\\+\left(\Gamma^{2}+ \mu^{2}\right) \left(2 \left(\Gamma^{2} -3 \mu^{2}\right) \lambda^{2} + 8 \lambda^{4} \right)\nonumber \\
\approx \left((k^{2} v^{2}-\mu^{2})^{2} - 4 \mu^{2} \lambda^{2} )\right) \left(k^{2} v^{2} + \mu^{2} - 2 \lambda^{2} \right) \nonumber\\
+ \left(3 k^{4} v^{4} + 2 k^{2} v^{2} \mu^{2} + 3 \mu^{4} - 4 \mu^{2} \lambda^{2} + 8 \lambda^{4} \right) \Gamma^{2}.\end{aligned}$$
Finally the renormalized spin-vertex functions are: $$\tilde{S}_{x} = \frac{\hbar}{2} \eta \left(
\begin{array}{cc}
\sigma_{x} & 0 \\
0 & \sigma_{x} \\
\end{array}
\right)$$ $$\tilde{S}_{y} = \frac{\hbar}{2} \eta \left(
\begin{array}{cc}
\sigma_{y} & 0 \\
0 & \sigma_{y} \\
\end{array}
\right)$$ $$\tilde{S}_{z} = \frac{\hbar}{2} \zeta \left(
\begin{array}{cc}
\sigma_{z} & 0 \\
0 & \sigma_{z} \\
\end{array}
\right)$$
This means that the results obtained in the [*bare bubble*]{} approximation should be multiplied only by a numerical factor to take into account the vertex corrections due to disorder. More specifically, the results for $S_{x}$ and $S_{y}$ should be multiplied by the factor $\eta$ while those for $S_{z}$ should be multiplied by $\zeta$. The situation is significantly different from that found in the case of spin Hall effect. This is because transport phenomena and spin polarization are affected by scattering on impurities in remarkably different ways.
{width="90.00000%"}
In Fig.3(a) we show the renormalization parameter $\eta$ as a function of chemical potential and relaxation time. Now the relaxation time is connected with the chemical potential through the relation $\frac{\hbar}{\tau} = \frac{n_{i}V^{2}}{v^{2}}|\mu|$. A single point in the $\tau ,\mu$ space corresponds to a well defined value of $n_{i}V^{2}$. However, possible values of $n_{i}V^{2}$ have been limited in Fig.3 to $n_{i}V^{2}<({n_{i}V^{2}})_{\rm max}$, where $({n_{i}V^{2}})_{\rm max}$ is a certain maximum value which is physically reasonable. The central white region is bounded by the condition $\hbar /\tau = (n_{i}V^{2})_{\rm max}|\mu|/v^{2}$ and is excluded for the considered parameters. In Fig.3c, in turn, we show the parameter $\eta$ as a function of the relaxation time and the ratio ${n_{i}V^{2}}/({n_{i}V^{2}})_{\rm max}$. As one might expect, this figure shows that the normalization parameter $\eta$ becomes reduced with decreasing ${n_{i}V^{2}}$. Figures 3(b) and 3(d) present cross-sections of Fig.3(a) and 3(c), respectively. The above described results for $\eta$ show that the $S_x$ and $S_y$ components are remarkably renormalized by the vertex correction and are enhanced by a factor of the order of 2 (between 1 and 3). This enhancement of the spin polarization is comparable to that found in the case of two-dimensional electron gas with Rashbe interaction.[@edelstein90] The parameter $\zeta$, in turn, is shown in Fig.3e-f. It is of the same order of magnitude as the parameter $\eta$ and depends on the chemical potential and relaxation time in a similar way, so we will not discuss it in more detail.
Spin-orbit torque
=================
The current-induced spin polarization is exchange-coupled to the local magnetization $\bf M$ and thus exerts a torque on $\bf M$. According to Eq.(\[hkM\]), energy of this interaction per unit area can be written as $E_{\rm ex} = -(2/\hbar )\tilde{\bf M}\cdot \bf{S}$, where $\bf{S}$ is the induced spin polarization. Taking into account the relation between $\tilde{\bf M}$ and $\bf M$, one finds the spin-orbit torque per unit area, $\btau$, exerted on the magnetization (more precisely on the corresponding equilibrium spin polarization of the system) in the form $$\label{sot}
\btau = \frac{2}{\hbar} \tilde{\mathbf{M}} \times \mathbf{S}= \frac{J_{\rm ex}}{g\mu_B\hbar} \mathbf{M} \times \mathbf{S}.$$
Let us consider in more detail some specific situations as concerns relative orientation of the magnetization and electric field (current). Let us start with the situation when the magnetization $\mathbf{M}$ is in the plane of the system and perpendicular to the current. This corresponds to $\theta = 0$ and $\xi = 0$ ($M_{x} = M \neq 0$ and $M_{y} = M_{z} = 0$). From the above general equation follows that the spin-orbit torque can be then written in a general form as $$\label{sot1}
\btau =A( - \hat{j} M_{x} S_{z} + \hat{k} M_{x} S_{y}),$$ where $\hat{i}$, $\hat{j}$, and $\hat{k}$ are unit vectors along the axes $x$, $y$ and $z$, respectively, and we introduced the following abbreviation: $A= J_{\rm ex}/g\mu_B\hbar$. Taking into account Eqs (\[sy1\]), (\[sy2\]), (\[sz1\]) and (\[sz2\]), one finds immediately that the spin-orbit torque in this geometry disappears because both $S_{y}$ and $S_{z}$ component of the spin polarization vanish.
Consider now the situation corresponding to $\theta = 0$ and $\xi = \pi/2$ ($M_{y} = M \neq 0$ and $M_{x} = M_{z} = 0$), i.e. the case when the magnetization is parallel to the electric current. From Eq.(\[sot\]) follows that the spin-orbit torque has the following general form: $$\label{sot2}
\btau =A( \hat{i} M_{y} S_{z} - \hat{k} M_{y} S_{x}).$$ The $S_{z}$ component is now nonzero, and thus both, $S_z$ and $S_x$ contribute to the torque in this geometry.
When $\theta = \pi/2$, namely $M_{z}=M \neq 0$ and $M_{x} = M_{y} = 0$, the magnetization is perpendicular to the graphene plane. The spin-orbit torque takes then the general form, $$\label{sot3}
\btau =A( - \hat{i} M_{z} S_{y} + \hat{j} M_{z} S_{x}).$$ Similarly as in the preceding situation, both $S_y$ and $S_x$ are nonzero and determine the torque.
In the last two cases the spin-orbit torque contains two components: linear term with respect to $J_{\rm{ex}}$ (proportional to $S_{x}$) and quadratic term in $J_{\rm{ex}}$ (proportional to $S_{z}$ and $S_{y}$). The spin orbit torque contains one component proportional to the relaxation time and another component whose the dominant part is independent on the relaxation time.
In a general case of arbitrary orientation of the magnetic moment, magnitude and character of the spin-orbit torque varies with the orientation of the magnetic moment. This is because two components of the current-induced spin polarization depend on the magnetization, while the third one is independent of $\bf M$. As a result the spin torque may have field-like and (anti)damping terms.
Summary
=======
We have calculated current-induced spin polarization in graphene deposited on a ferromagnetic substrate, that ensures not only Rashba spin-orbit interaction but also a ferromagnetic moment in the graphene layer. To describe electronic spectrum of graphene we have used Kane Hamiltonian that describes low-energy states around the Dirac points. Using the zero-temperature Green functions formalism and linear response theory, we have derived analytical formulas for the spin polarization, up to the terms linear in $M$. Numerical results based on the analytical formulas have been compared with those obtained by numerical integration procedure. From this comparison we have formulated applicability conditions of the analytical results. Significant deviations of the analytical results from those based on numerical integration have been found for relaxation times smaller than $10^{-10}$s.
The nonequilibrium (current-induced) spin polarization exerts a torque on the magnetization [*via*]{} the exchange interaction. This torque contains a term which is proportional to the $x$-component of the induced spin polarization and therefore is proportional to the momentum relaxation time. The torque also includes a component whose main part is independent of the relaxation time.
The spin-orbit torque due to the interplay of external electric field and Rashba coupling at the interface between graphene and a magnetic layer can be used for instance to trigger magnetic dynamics and/or magnetic switching. Indeed, such a switching was observed experimentally in a recent paper by Wang [*et al*]{}.[@wangpreprint] However, instead of graphene they used MoS$_2$ – another two-dimensional honeycomb crystal.
This work has been partially supported by the National Science Center in Poland as research project No. DEC-2013/10/M/ST3/00488 and by the Polish Ministry of Science and Higher Education through a research project ’Iuventus Plus’ in years 2015-2017 (project No. 0083/IP3/2015/73).
A. Manchon, N. Ryzhanova, N. Strelkov, A. Vedyayev, and B. Dieny, J. Phys.: Condens. Matter **19**, 165212 (2007).
D. C. Ralph, M. D. Stiles, JMMM **320**, 1190 (2008).
A. Brataas, A. D. Kent and H. Ohno, Nature Materials **11**, 372 (2012).
M. I. Dyakonov, V. I. Perel, Pisma Z. Eksp. Teor. Fiz. 13, 657 (1971); JETP Lett. 13, 467 (1971).
J. E. Hirsch, Phys. Rev. Lett. 83, 1834 (1999).
M. I. Dyakonov and V. I. Perel, Phys. Letters A [**35**]{}, 459 (1971).
A. G. Aronov and Yu. B. Lynda-Geller, JETP Lett [**50**]{}, 431 (1989).
V. M. Edelstein, Sol. State Communs. [**73**]{}, 233 (1990).
A.G. Aronov, Y.B. Lyanda-Geller, G.E. Pikus, Sov. Phys. JETP [**73**]{}, 537 (1991).
E. L. Golub, E. I. Ivchenko, Phys. Rev B **84**, 115303 (2011).
K. Shen, G. Vignale, R. Raimondi, Phys. Rev. Lett. **112**, 096601 (2014).
A. Manchon and S. Zhang, Phys. Rev. B **78**, 212405 (2008).
A. Manchon and S. Zhang, Phys. Rev. B **79**, 094422 (2009).
A. Matos-Abiague, R. L. Rodriguez-Suarez, Phys. Rev. B **80**, 094424 (2009).
A. Manchon, arXiv:1204.4869v1 (2012).
P. Gambardella and I. M. Miron, Phil. Trans. R. Soc. A **369**, 3175 (2011).
K. Garello, I. M. Miron, C. O. Avci, F. Freimuth, Y. Mokrousov, S. Blugel, S. Auffret, O. Boulle, G. Gaudin, P. Gambardella, Nature Nanotechnology 8, 587-593 (2013).
S. D. Ganichev, E. L. Ivchenko, V. V. Belkov,S. A. Tarasenko, M. Sollinger, D. Weiss, W. Wegschelder, W. Prettl, Nature **417**, 153 (2002).
H. Kurebayashi, J. Sinova, D. Fang, A.C. Irvine, J. Wunderlich, V. Novak, R.P. Campion, B.L. Gallagher, E.K. Vehstedt,L.P. Zarbo, K. Vyborny, A.J. Ferguson, and T. Jungwirth, Nature Nanotechnology **9**, 211 (2014).
C. O. Avci, K. Garello, C. Nistor, S. Godey, B. Ballesteros, A. Mugarza, A. Barla, M. Valvidares, E. Pellegrin, A. Ghosh, I. M. Miron, O. Boulle, S. Auffret, G. Gaudin, P. Gambardella, Phys. Rev. B **89**, 214419 (2014).
H. Li, H. Gao, L. P. Zarbo, K. Vyborny, X. Wang, I. Garate, F. Dogan, A. Cejchan, J. Sinova, T. Jungwirth, A. Manchon, Phys. Rev. B **91**, 134402(2015).
Yu. S. Dedkov, M. Fonin, U. Rudiger, and C. Laubschat, Phys. Rev. Lett. **100**, 107602 (2008).
T. Yokoyama, Phys. Rev. B **77**, 073413 (2008).
H. Haugen, D. Huertas-Hernando, and A. Brataas, Phys. Rev. B **77**, 115406 (2008).
Z. P. Niu, J. Phys.: Condens. Matter **23**, 435302 (2011).
Z. P. Niu, Phys. Lett. A **378**, 73 (2014).
P. Lazic, G. M. Sipahi, R. K. Kawakami, and Igor Zutic, Phys. Rev. B **90**, 085429 (2014).
Z. Qiao, S. A. Yang, W. Feng, W.-K. Tse, J. Ding, Y. Yao, J. Wang, and Q. Niu, Phys. Rev. B 82, 161414(R) (2010).
Z. Qiao, W. Ren, H. Chen, L. Bellaiche, Z. Zhang, A. H. MacDonald, and Q. Niu, Phys. Rev. Letters 112, 116404 (2014).
Z. Wang, C. Tang, R. Sachs, Y. Barlas, and J. Shi, Phys. Rev. Letters 114, 016603 (2015).
S. J. Gong, Z. Y. Li, Z. Q. Yang, Ch. Gong, Ch.-G. Duan and J. H. Chu J. Appl. Phys. [**110**]{}, 043704 (2011).
T. Yokoyama and J. Linder, Phys. Rev. B **83**, 081418(R) (2011).
A. Dyrda[ł]{} ‚ J. Barnaś, V. K. Dugaev, Phys. Rev. B 89, 075422 (2014).
C. L. Kane and E. J. Mele, Phys. Rev. Lett. **95** 226801 (2005).
A. Varykhalov, D. Marchenko, M. R. Scholz, E. D. L. Rienks, T.K. Kim, G. Bihlmayer, J. Sanchez-Barriga, and O. Rader, Phys. Rev. Lett. [**108**]{}, 066804 (2012).
D. Marchenko, A. Varykhalov, M.R. Scholz, G. Bihlmayer, E.I. Rashba, A. Rybkin, A.M. Shikin and O. Rader, Nature Commun. [**3**]{}, 1232 (2012).
D. Marchenko, J. Sánchez-Barriga, M. R. Scholz, O. Rader, and A. Varykhalov, Phys. Rev. B [**87**]{}, 115426 (2013).
J. Balakrishnan, G. K. W. Koon, M. Jaiswal, A. H. Castro Neto, and B. Özyilmaz, Nature Physics [**9**]{}, 284 (2013).
J. Balakrishnan, G. K. W. Koon, A. Avsar, Y. Ho, J.H. Lee, M. Jaiswal, S.-J. Baeck, J-H. Ahn, A. Ferreira, M. A. Cazalilla, A. H. Castro Neto, and B. Özyilmaz, Nature Communications [**5**]{} 4748 (2014).
A. Avsar, J. Y. Tan, T. Taychatanapat, J. Balakrishnan, G. K. W. Koon, Y. Yeo, J. Lahiri, A. Carvalho, A. S. Rodin, E. C. T. O’Farrell, G. Eda, A. H. Castro Neto, and B. Özyilmaz, Nature Communications [**5**]{}, 4875 (2014).
J. Inoue, G. E. W. Bauer ,L. W. Molenkamp , Phys. Rev. B [**70**]{}, 041303 (2004).
R. Raimondi and P. Schwab, Phys. Rev. B [**71**]{}, 033311 (2005).
O. V. Dimitrova, Phys. Rev. B [**71**]{}, 245327 (2005).
O. Chalaev and D. Loss, Phys. Rev. B [**71**]{}, 245318 (2005)
S. Murakami, Phys. Rev. B [**69**]{}, 241202(R) (2004).
A. G. Mal’shukov and K. A. Chao, Phys. Rev. B [**71**]{}, 121308(R) (2005).
P. L. Krotkov and S. Das Sarma, Phys. Rev. B [**73**]{}, 195307 (2006).
E. McCann, K. Kechedzhi, V. Fal’ko, H. Suzuura, T. Ando, and B.L. Altshuler, Phys. Rev. Lett. [**97**]{}, 146805 (2006).
P. M. Ostrovsky, I.V. Gornyi, A. D. Mirlin, Phys. Rev. B [**74**]{}, 235443 (2006).
A. Pachoud, A. Ferreira, B. Ozyimaz, A.H. Castro Neto, Phys. Rev. B [**90**]{}, 035444 (2014).
N.A. Sinitsyn, J.E. Hill, H. Min, J. Sinova, and A. H. MacDonald, Phys. Rev. Lett. [**97**]{}, 106804 (2006).
V. P. Gusynin, S. G. Sharapov, and A. A. Varlamov, Phys. Rev. B 90, 155107 (2014).
G.D. Mahan, *Many Particle Physics*, (Kluwer Academic/Plenum Publishers, New York, 2000).
W. Wang, A. Narayan, L. Tang, K. Dolui, Y. Liu, X. Yuan, Y. Jin, Y. Wu, I. Rungger, S. Sanvito, and F. Xiu, preprint arXiv:1502.06154.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We review the role of fundamental spin-0 bosons as bosonic coherent motion (BCM) in the Universe. The fundamental spin-0 bosons have the potential to account for the baryon number generation, cold dark matter (CDM) via BCM, inflation, and dark energy. Among these, we pay particular attention to the CDM possibility because it can be experimentally tested with the current experimental techniques. We also comment on the panoply of the other roles of spin-0 bosons–such as those for cosmic accelerations at early and late times.'
address:
- |
Department of Physics, Seoul National University, Seoul 151-747, Korea, and\
Department of Physics, Kyung Hee University, Seoul 130-701, Korea
- |
Center for Axion and Precision Physics Research (IBS), Daejeon 305-701, Korea, and\
Department of Physics, KAIST, Daejeon 305-701, Korea
- 'Department of Physics, Faculty of Science, Tokyo University of Science, 1-3 Kagurazaka, Shinjuku, Tokyo 162-8601, Japan'
author:
- 'Jihn E. Kim'
- Yannis Semertzidis
- Shinji Tsujikawa
title: ' **Bosonic Coherent Motions in the Universe**'
---
Introduction
============
Recent cosmological observations [@Planck13; @WMAP9] confirm the eight decade old Zwicky’s proposal [@ZwickyF33] that the Universe contains a large amount of dark matter (DM). The DM profile has been measured accurately enough to pinpoint DM to “it is cold dark matter (CDM) [@Planck13]." The bosonic coherent motion (BCM) can be CDM [@Preskill83; @Abbott83; @Dine83] if the coherent-boson lifetime is long enough to have survived until now [@KSVZ; @Shifman80; @DFSZ; @Dine81]. The axion proposed to solve the strong CP problem [@KimPRP; @Cheng88; @Pecceirev; @KimRMP] is fitting to this BCM scenario [@RevAxionsCDM]. The BCM is one of many possibilities of CDM scenarios [@RevAxionsCDM] which accounts for only 27% in the energy pie of the Universe. The dominant portion, 68%, in the energy pie is the homogeneous energy density, at least up to the $10^3$ Mpc scale, which is usually referred to dark energy. Dark energy (DE) being homogeneous cannot be accounted for by corpuscular particles but may be accountable by the cosmological constant or by some vacuum expectation value (VEV) of spin-0 boson(s) [@Sahni; @Pratra; @Paddy; @DErev; @Silvestri09; @Caldwell09; @Tsuji10]. The visible particles (mostly atoms in the energy count) constitute only 5% in the energy pie.
If we accept the Big Bang cosmology from the earliest possible time, $10^{-43}$ s, the success of the Standard Model of particle physics is based on the assumption of very tiny DE of order less than $10^{-46}\,{\,\textrm{GeV}}^4$ because the age of the Universe is very long $\simeq 13.8\,{\rm Gy}$ [@Planck13; @WMAP9]. So, the DE problem or the theoretical cosmological constant problem [@WeinbergCC] is not only the problem in cosmology but also a problem in particle physics. Out of despair, many adopt the anthropic scenario for the cosmological constant problem [@WeinbergAnth; @Tegmark06; @ZeeEgr13]. For the anthropic solution to work, the cosmological constant must be a free undetermined parameter in particle physics, as integration constants of Refs. [@Hawking84; @KKL00; @KimCC]. In a deterministic theory such as in string theory, possible cosmological constants must be allowed near 0 for our Universe to have adopted one of these, which is the reason trying to have as many as $10^{120}$ models, to pack the vacua with separation between them by $(10^{-3\,}{\,\mathrm{eV}})^4$, from string theory [@ZeeEgr13]. But, all those $10^{120}$ vacua must allow three family SMs, and satisfy the known SM phenomena such as the GUT scale weak mixing angle $\sin^2\theta_W=3/8$ [@Kim81; @Luo91; @Kim03], etc. But, we have only a handful of minimal supersymmetric Standard Models from string theory satisfying the requirements [@stringSM; @KimKyae07; @HuhKK09; @Lebedev07; @Nilles06]. Or, a Standard Model solution with DE$\simeq 10^{-47\,}{\rm GeV}^4$ has to be found so that the anthropic argument chooses it. This search seems more difficult than finding a vanishing cosmological constant solution theoretically. At present, we can say that the anthopic solution in string theory has not worked out yet. Therefore, in the Standard Model and in its supergravity extension, it is fair to say that the cosmological constant is assumed to be zero.
By observing the luminosities of Type-I supernovae [@Perlmutter99; @Riess98], the recent acceleration of the Universe has been established. So, explaining the DE scale of $10^{-47}\,{\,\textrm{GeV}}^4- 10^{-46}\,{\,\textrm{GeV}}^4$ became an important topic [@DErev; @Silvestri09; @Caldwell09]. In **Table\[Tab:01\]**, we list several ideas proposed to account for this recent acceleration of the Universe. Both the high scale inflation [@inflationold; @Infnew1; @InfNew2] and the recent acceleration [@Perlmutter99; @Riess98] in the Universe are based on the assumption of vanishing cosmological constant.
To determine the VEV of a scalar field, say $\phi$, one must consider all the allowed effective terms at low energy. At each interaction point, suitable symmetry requirements must be satisfied. A typical mass scale of $\phi$ is given by the effective mass term $m^2|\phi|^2$. In Fig. \[DilBreaking\], we consider only two diagrams with the dimension 4 ($d=4$) couplings. If each $d=4$ vertex of Fig. \[DilBreaking\] satisfies the global phase symmetry, the two-loop and one-loop mass terms do not break the global symmetry. On the other hand, each $d=4$ vertex satisfies the dilaton symmetry (requiring just $d=4$ couplings) but the diagrams of Fig. \[DilBreaking\] are $d=2$ terms which of course break the dilaton symmetry. One well-known model breaking the dilaton symmetry at the one-loop quantum level, including the $d\ge 6$ terms, is the Coleman-Weinberg model [@ColemanW73]. Therefore, it is not likely that a consistent calculation of a small DE scale can be performed by introducing the dilaton symmetry. However, some global phase symmetry may be suitable for this.
In Sec. \[sec:spin0\], we present the focus points of this review: the BCM scenarios and the axion detection experiments. In Sec. \[sec:CC\], we point out the difficulty of obtaining zero cosmological constant theoretically. In Sec. \[sec:inflation\], we mini-review the inflationary cosmology, in particular in view of the recent BICEP2 data. In Sec. \[sec:discussion\], we discuss the subject of this review: why the role of [*fundamental*]{} spin-0 particles are important in cosmology.
Spin-0 boson filling the Universe {#sec:spin0}
=================================
After the discovery of a fundamental spin-0 scalar particle (the Brout-Englert-Higgs boson) at the LHC, it is timely to study the roles of fundamental spin-0 bosons in the Universe. It is very interesting to note that fundamental spin-0 bosons have been employed to account for the mothers of atoms ([*i.e.*]{} baryon number generation via the Affleck-Dine mechanism [@AffDine]), CDM via BCM [@Preskill83], DE via a transient cosmological constant [@QuintAx03; @KimQ99; @Kimq00; @KimNill09; @Choi00], and even the vacuum energy needed for the high scale inflation [@inflationold; @Infnew1; @InfNew2]. Among these, we focus on CDM via BCM in this review because similar ideas can be applicable to DE and inflation models. Another attractive point discussing CDM via BCM is that it can be experimentally proved in the near future [@Yannis14].
Ideas Description \[scalarS or pseudoscalarP\] Disc. sym. Fine-tune From string
------------------------------------- --------------------------------------------------- -------------- --------------- ---------------
MOND$^{\,\rm a}$ Change Newtonian gravity. \[No boson\] Irrelevant Yes Not yet
\[0.2em\] Anthropic principle Out of many possible vacua, only those Irrelevant Irrelevant Not yet$^{b}$
suitable for age$>t{\rm _U}$ survived. \[S or P\]
\[0.2em\] Quintessence With a runaway $V\propto 1/\phi^n$ ($n>0$). \[S\] No Yes$^{\rm c}$ Not yet
\[0.2em\] Dilaton P-Gold. boson from dilaton sym. \[S\] No Yes$^{\rm d}$ Not yet
\[0.2em\] U(1)$_{\rm DE}$ Goldstone P-Gold. boson from U(1)$_{\rm DE}$ sym. \[P\] Yes No Yes$^{\rm e}$
: Typical DE models with a few pseudo-Goldstone bosons originating from global symmetries. $^{\rm a}$ Refs. [@MOND; @Becken], $^{\rm b}$ Ref. [@ZeeEgr13], $^{\rm c}$ Ref. [@Stenhardt98], $^{\rm d}$ Ref. [@Wetterich], $^{\rm e}$ Ref. [@KimNilles14]. []{data-label="Tab:01"}
We are familiar with the ether idea of the late 19th Century, filling out the Universe. The VEV idea of spin-0 particles used for breaking global symmetries [@Goldstone61] and gauge symmetries [@Higgs64] is a kind of ether. If a scalar field $\phi$ has a universal value over the entire Universe, any operation of the type ‘Poincare transformation’ does not notice a change. Thus, the VEV of a scalar field, $\langle\phi\rangle$, respects the Poincare symmetry. But, if $\phi$ is a complex field, then the VEV breaks the phase transformation symmetry, [*i.e.*]{} breaks a global phase symmetry [@Goldstone61]. Even though the Brout-Englert-Higgs mechanism [@BE64; @Higgs64; @GHK64] for breaking gauge symmetries is not a monopoly of spin-0 particles [@Techni; @WeinTechni], now the role of spin-0 particles becomes more important, especially after a hint of large tensor-to-scalar ratio $r$, based on the BICEP2 observation [@Larger].
---------------------------------------------------------------------------------------------------------------------------------------------------------
![Diagrams leading to dimension 2 interactions with dimension 4 coupling at each vertex. []{data-label="DilBreaking"}](figFeyn01 "fig:"){width="8.5cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------
Let us denote scalar and pseudoscalar particles as $s$ and $a$, respectively. Scalar particles transform under the parity operation as $P: s({\bf x}) \to +s(-{\bf x})$, and pseudoscalar particles transform as $P: a({\bf x}) \to -a(-{\bf x}) $. If they are components of a complex field, it is usually represented as the radial and phase fields, respectively, $\phi=s e^{ia/f}$ where $f$ is a mass parameter. Thus, the complex field transforms under parity as $P: \phi({\bf x}) \to\phi^*(-{\bf x}) $. Any pseudoscalar field represented as a phase can be represented by an angle field with the angle defined in the range $[0,2N\pi)$, where $N$ is the domain-wall number. A Goldstone boson arising from breaking a global phase symmetry by the VEV $v$ is a pseudoscalar field $a$ defined as [$$\begin{split}
\langle \phi\rangle=\frac{v+s}{\sqrt2}\, e^{ia/f},~~~
\langle s\rangle=0,~~~\langle a\rangle=[0, 2N\pi f).
\end{split}$$]{}
Cosmology with BCM {#subsec:spin0cosm}
------------------
On the flat Friedmann-Lemaitre-Robertson-Walker cosmological background space described by the line element $ds^2=-dt^2+a^2(t)\delta_{ij}dx^i dx^j$, the evolution of the classical scalar field $\phi$, ([[*i.e. *]{}]{}the evolution of the VEV of $\phi$), is given by [$$\begin{split}
\frac{d^2}{dt^2}\langle \phi\rangle+
3 H\frac{d}{dt}\langle \phi\rangle+V'(\langle \phi\rangle)=0\,,\label{eq:vev}
\end{split}$$]{} where $H=\dot{a}/a$ is the Hubble parameter and $V'=(d/d\langle \phi\rangle)V$ is a derivative of the potential $V$ (a dot represents a derivative with respect to the cosmic time $t$). With a discrete symmetry $\phi \to -\phi$, the leading term of $V'$ is the mass term $m^2 \langle \phi \rangle$. When $\langle \phi \rangle$ moves very slowly, we can neglect the second derivative term, and the evolution equation gives $3H \dot\phi \simeq -m^2 \phi$. $\langle \phi \rangle$ starts to change rapidly when $H$ becomes small enough to satisfy $3H\simeq m$. After this condition is met, $\langle \phi \rangle$ oscillates rapidly, as shown in Fig. \[BCMcosmology\], which is interpreted as the BCM of $\phi$.
As mentioned above, the VEV $ \langle \phi \rangle$ is assumed to be the same over the whole Universe for the Poincare invariance, otherwise the invariance is broken. In the Universe, this homogeneity is subtly broken. The inflation manages different scales of quantum fluctuations enter the horizon at different scales, basically breaking the homogeneity. A given scale condenses gravitationally. The VEV in that scale evolves according to Eq. (\[eq:vev\]), and describes the BCM of $\langle\phi
\rangle$.
From the Friedmann equation we have $3H^2{M_{\rm P}}^2=\rho$, where $\rho$ is the energy density of the Universe and ${M_{\rm P}}$ is the reduced Planck mass (${M_{\rm P}}=2.4 \times 10^{18}$ GeV). Denoting the time at the onset of oscillations of $\langle\phi\rangle$ as $t_1$, the condition for determining $t_1$ is [$$\begin{split}
\sqrt{\frac{3\rho(t_1)}{M_{\rm P}^2}}=m(t_1).
\end{split}$$]{} These oscillations are equivalent to a gas of $\phi$ particles of low-momentum. This kind of spin-0 particle coherent motion was first discussed in Ref. [@Preskill83; @Abbott83; @Dine83] for the case $\phi={\rm axion}$. It is known that the BCM behaves like CDM because of the low-momentum. Thus, the number and energy densities are given by [$$\begin{split}
n=m\langle \phi\rangle^2\,,\quad
\rho= m^2\langle \phi\rangle^2\,.
\end{split}$$]{}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ After $t_1$, the BCM potential of $\langle\phi\rangle$ at the red bullet oscillates around the minimum.[]{data-label="BCMcosmology"}](figBCM02 "fig:"){width="7cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
We denote the current age of the Universe as [$t_{\rm U}$]{}. Depending on $t_1$ and [$t_{\rm U}$]{}, we can classify BCMs as
- : If $t_1<{t_{\rm U}}$, the currently oscillating vacuum $\langle\phi\rangle$ is BCM. The BCM can be classified into the following two sub-categories.
- $\star$ [**BCM1**]{}: The lifetime of $\phi$ is long enough, $\tau_{\phi}>{t_{\rm U}}$. Then, the oscillating BCM contributes to the CDM amount. The QCD axion belongs here.
- $\star$ [**BCM2**]{}: The lifetime of $\phi$ is short, $\tau_{\phi}<{t_{\rm U}}$. Then, all $\phi$ quanta decayed already, producing SM particles. The inflaton with $\tau_\phi\sim 10^{-36}$s belongs here and reheating after inflation gives the beginning of the radiation-dominated Universe.
- : Temporary cosmological constant. On the other hand, if $\langle\phi\rangle$ has not oscillated yet, then $t_1>{t_{\rm U}}$ and $\langle\phi\rangle$ stays there now, behaves like a cosmological constant, but it is a temporary phenomenon and will eventually become [**BCM1**]{} after $t_1$. For this to be satisfied, the mass is around $10^{-33\,}$eV with a trans-Planckian decay constant [@Carrol98]. If $V(\langle\phi\rangle)$ describes [**CCtmp**]{}, the equation of state $w_\phi$, characterized by the field energy density $\rho_{\phi}=\frac12 \dot{\phi}^2 +V(\phi)$ and the pressure $P_{\phi}=\frac12 \dot{\phi}^2 -V(\phi)$, is a useful parameter, [$$\begin{split}
w_{\phi} \equiv \frac{P_{\phi}}{\rho_{\phi}}
=\frac{\frac12 \dot{\phi}^2 -V(\phi)}{\frac12 \dot{\phi}^2 +V(\phi)}\,.\label{eq:ofstate}
\end{split}$$]{} Provided that $\frac12 \dot{\phi}^2 \ll V(\phi)$, $w_{\phi}$ is close to $-1$, behaving like the cosmological constant. In order to realize the recent acceleration, we require the condition $w_{\phi}<-\frac13$.
Scalar particles
----------------
The Brout-Englert-Higgs boson is the only known fundamental scalar field. The other scalar most widely used in particle theory is dilaton, the scalar Goldstone boson arising from breaking the dilatonic symmetry. The effect of dilatonic symmetry on the cosmological constant problem has been discussed extensively in Ref. [@Wetterich]. For the solution, however, a fine-tuning is needed. The obvious effect of a VEV of a scalar field $s$ in cosmology is the universal constant on the right-hand-side of the Einstein equation.
The scalar-field cosmology in the presence of a barotropic perfect fluid was studied in 1980–90s [@Fujii82; @Fujii90; @Ford87; @Wetterich; @Peebles88; @Ratra88; @Chiba97; @Ferreira97; @Paul98; @CLW98], even before the discovery of the recent cosmic acceleration. This was chiefly motivated by the “missing matter problem” in 1980s. In 1990, Fukugita [[*et al.*]{}]{}[@Fukugita90] tested cosmological models against observations of the number count of faint galaxies and showed that these data favor the Universe with low matter density ([*i.e.*]{}, matter is missing). In the abstract of their paper they stated that “Furthermore, it is shown that the best agreement with the data is obtained with a sizable cosmological constant, including the case of zero curvature model as predicted by inflation.” In addition, it was already known in the early 1990s that the presence of a cosmological constant can make the age of the Universe longer such that it is consistent with the age of oldest globular clusters.
If the cosmological constant originates from a vacuum energy appearing in particle physics, it is vastly larger than the today’s average cosmological density [@WeinbergCC]. Because of this problem, people tried to construct dynamical cosmological constant models in which the energy density of cosmological constant varies in time, basically belonging to a kind of [**CCtmp**]{}. For example, if we consider a dilaton field $\phi$, the cosmological constant depends on $\phi$ by transforming the dilatonic action to the so-called Einstein-frame action (in which the dilaton does not have a direct coupling with the Ricci scalar) [@Fujii90; @Wetterich].
Exponential potentials often arise from the curvature of internal spaces associated with the geometry of extra dimensions (so called “modulus” fields) [@GSW; @Olive90]. Inspired by this, the exponential potential $V(s)=V_0 e^{-\lambda\, s{M_{\rm P}}}$ has been used, with constant parameters $V_0$ and $\lambda$. There are two distinct fixed points on the flat Friedmann-Lemaitre-Robertson-Walker cosmological background space [@CLW98; @DErev]: (a) the scaling solution, and (b) the scalar-field dominated solution.
For $\lambda^2>3(1+w_m)$, where $w_m$ is the equation of state for a background fluid, the solutions approach the scaling fixed point (a), characterized by the field density parameter $\Omega_{s}=3(1+w_m)/\lambda^2$ and the field equation of state $w_{s}=w_m$ [@Peebles88; @Ferreira97; @CLW98]. Even for the initial conditions where $\rho_{s}$ is larger than $\rho_m$ in the early radiation-dominated era, the field eventually enters the scaling regime in which $\rho_{s}$ is proportional to $\rho_m$ with $\rho_{s}/\rho_m={\rm constant}<1$. The field energy density of the scaling solution contributes to the total energy density of the Universe, but it does not lead to the cosmic acceleration. For $\lambda^2<3(1+w_m)$, there exists the scalar-field dominated fixed point (b), characterized by $\Omega_{s}=1$ and $w_{s}=-1+\lambda^2/3$. The late-time cosmic acceleration can be realized for $\lambda^2<2$. Since in this case the point (b) is also stable, the scalar field can be the source of DE. For $\lambda^2<2$, the scalar potential is quite shallow, so the field density in the early Universe needs to be much smaller than the background energy density (unlike the scaling solution discussed above).
After the discovery of the recent cosmic acceleration in 1998, the cosmological dynamics of “quintessence” (a canonical scalar field responsible for DE) were studied in detail for several different potentials [@Paul2; @Paul3]. One example is the inverse power-law potential $V(s)=M^{4+n}s^{-n}$, where $M$ and $n$ are positive constants. This potential can arise in globally supersymmetric QCD theories [@Bine][^1]. The Universe enters the stage of cosmic acceleration for the field value larger than $s_0 \approx {M_{\rm P}}$. Since $V(s_0)$ is of the order of $H_0^2{M_{\rm P}}^2$, one can estimate the mass scale $M$ as $M \approx 10^{-(46-19n)/(4+n)}$ GeV. For $n=O(1)$, this energy scale can be compatible with that appearing in particle physics.
In the presence of a perfect fluid with the equation of state $w_m$, there exists a so-called tracker solution for the potential $V(s)=M^{4+n}s^{-n}$. The tracker is characterized by a common, cosmic evolutionary trajectory that attracts solutions with a wide range of initial conditions [@Paul3]. The field equation of state along the tracker is given by $w_{s}=(w_mn-2)/(n+2)$, which corresponds to $w_{s}=-2/(n+2)>-1$ during the matter era. The slope of the potential $\lambda=-{M_{\rm P}}V_{,s}/V=n{M_{\rm P}}/s$ gets smaller with the growth of $s$, so $w_{s}$ approaches $-1$ in the future. The inverse power-law potential belongs to a class of freezing quintessence models [@CaLi] in which the evolution of the field gradually slows down.
There is another class of quintessence models, dubbed thawing models [@CaLi], in which the field has been frozen by Hubble friction and then it starts to evolve after the Hubble parameter drops below the field mass $m$. In this case the field equation of state $w_{s}$ is close to $-1$ at the initial stage, but it starts to grow at the late cosmological epoch. The field mass $m_s$ responsible for dark energy corresponds to $m_s \simeq 10^{-33}$ eV [@Carrol98]. The representative potential of thawing models is that of a pseudo-scalar field arising from breaking the global U(1) symmetry (which we will explain more details in Sec. \[pseudosec\]).
If we consider a scalar field $\phi$ non-minimally coupled to the Ricci scalar $R$ (like dilaton), this gives rise to a coupling with non-relativistic matter in the Einstein frame [@Amen]. The fifth force induced by such a matter coupling needs to be suppressed in the solar system. There are several ways to suppress the propagation of the fifth force in local regions of the Universe.
One is the so-called chameleon mechanism [@Khoury], under which the mass of a scalar degree of freedom is different depending on the matter densities in the surrounding environment. If the effective mass is sufficiently large in the regions of high density, the coupling between the field and non-relativistic matter can be suppressed by having a thin shell inside a spherically symmetric body. In Brans-Dicke theory (including $f(R)$ gravity) [@Brans] it is possible to suppress the propagation of the fifth force by designing the field potential $V(\phi)$ appropriately [@Khoury2; @Sawi; @Capo; @Mota; @fRgravity1; @fRgravity2].
Another is the so-called Vainshtein mechanism [@Vain], under which nonlinear scalar-field self interactions can suppress the fifth force at short distances even in the absence of the field potential. The self interactions of the form $(\partial \phi)^2 \square \phi$, which correspond to the Lagrangian of covariant Galileons [@Galileons], can lead to the decoupling of the field $\phi$ from matter within a radius much larger than the solar-system scale [@Nicolis; @Kimura; @Kase; @Kase2].
Pseudoscalar particles {#pseudosec}
----------------------
Most pseudoscalar particles observed so far are pseudo-Goldstone bosons. Let $a, \Lambda$ and $f$, respectively, be a Goldstone boson from a spontaneously-broken global U(1) symmetry, the dominant explicit symmetry breaking mass parameter, and the decay constant. Then, the mass of $a$ is [$$\begin{split}
m_a=c_a \frac{\Lambda^2}{f}\,,
\end{split}$$]{} where $c_a$ is the number given by the explicit symmetry breaking terms. For the QCD axion, the breaking of the U(1) symmetry is given by the QCD anomaly and we have $c_a \Lambda^2=[Z^{1/2}/(1+Z)] f_\pi m_\pi$ with $Z=m_u/m_d$ where $f_\pi, m_\pi, m_u,m_d$ are neutral-pion decay constant, its mass, and $u$ and $d$ quark masses [@KimRMP]. If the explicit breaking term is given by $V_{\rm br}=-(\Lambda^{4-n}\phi^n+{\rm h.c.})/2$, then we have $m_a= (f/\Lambda)^{n/2} (n\Lambda^2/f) $. As shown in Fig. \[DilBreaking\], the pseudo-Goldstone boson arising from a global symmetry [U(1)$_{\rm gl}$]{} does not appear in the loops if each vertex satisfies [U(1)$_{\rm gl}$]{}. But, it is known that all global symmetries are approximate [@KimNilles14; @Kim13worm; @Diaz14; @KraussW89; @KimJKPS14]. Most strong explicit breaking may be from the anomaly of the type [U(1)$_{\rm gl}$]{}-$G$-$G$, where $G$ is a non-Abelian gauge group.
The most waited-for pseudoscalar particle is the very light axion in the axion window because its discovery will confirm at least three: (1) a physical confirmation of instanton solutions of non-Abelian gauge theories [@Belavin75], (2) 't Hooft solution [@HooftSol] of the U(1) problem of QCD [@Weinberg75], and (3) at least some portion of CDM in the Universe. The particle [*axion*]{} was first appreciated by Weinberg and Wilczek in the Ben Lee Memorial Conference in October, 1977 [@Cline77], using the Peccei-Quinn (PQ) symmetry [@PQ77]. If $G$ is QCD, the symmetry [U(1)$_{\rm gl}$]{} is called the PQ symmetry [U(1)$_{\rm PQ}$]{} and the pseudo-Goldstone boson $a$ related to [U(1)$_{\rm PQ}$]{} is called the [*QCD axion*]{}. The axion is needed to understand the strong CP problem of “Why is the neutron electric diplole moment so small even though the gluon interactions (in the presence of instanton solutions of QCD) allow a neutron-size diplole moment?” In early days, three kinds of solutions to the strong CP problem were admitted [@KimPRP]: the calculable solution, the massless up quark case, and the axion solution. The calculable solutions have not provided yet an acceptable model with sufficiently small neutron electric diplole moment. The massless up quark case is not favored in the global fit [@KimRMP]. The remaining axion solution is checked in various cases as discussed in the next Subsect.
Field theory examples on axions with renormalizable couplings corresponding to [**BCM1**]{} are usually classified to the KSVZ and DFSZ models [@KSVZ; @Shifman80; @DFSZ; @Dine81]. But, this classification is too simple. There can be many KSVZ and DFSZ type models with one type of quark representations [@Kim98]. One may introduce many different types of quarks also for axion phenomenology.
Therefore, it is better to have a theory predicting definite PQ charges of the quarks in a full theory. The most attractive proposal along this line is the string compactification. Here, the PQ global symmetry is determined once the compactification scheme is presented. Standard models obtained from string compactification include many quarks beyond the Standard Model spectrum, in particular numerous singlet fields. Along this line, several years ago a QCD axion including non-renormalizable terms was studied and the axion-photon-photon coupling has been calculated with an approximate [U(1)$_{\rm PQ}$]{} symmetry [@ChoiKimKim]. Recently, an exact [U(1)$_{\rm PQ}$]{} symmetry has been studied in a string compactification where the axion-photon-photon coupling has been calculated below the PQ symmetry breaking scale [@Kimagg14], [$$\begin{split}
c_{a\gamma\gamma} =\frac{1123}{388}-1.98\simeq 0.91\,.
\end{split}$$]{} We expect that more calculations of $c_{a\gamma\gamma} $ will be performed in string models with the property of successful SM phenomenologies, which will guide us where to look for the QCD axion [@Yannis14].
Dark energy can be the case of [**CCtmp**]{} in the above classification. Pseudoscalar [**CCtmp**]{} have been discussed already more than a decade ago in Refs. [@QuintAx03; @Choi00; @Nomura]. But, a more plausible analysis, looking into the detail of string compactification, has been presented recently [@KimNilles14; @KimJKPS14; @KamionAxi14].
The field mass $m_a$ responsible for dark energy corresponds to $m_a \simeq 10^{-33}$ eV [@Carrol98]. Meanwhile, if the axion field is responsible for CDM, the typical mass scale is between $10^{-5}$ eV and $10^{-2}$ eV [@KimRMP]. In string theory there are many ultralight axions possibly down to the Hubble scale $H_0=10^{-33}$ eV [@axiverse]. Axions in the mass range between $10^{-28}$ eV and $10^{-18}$ eV become non-relativistic at a later cosmological epoch relative to the standard CDM. Such a light scalar field leads to the suppression of the CDM power spectrum on small scales [@axiverse; @Barbi; @Wayne] (like light massive neutrinos), so there is an observational signature for ultralight axions if the axion potential is of the form $[1-\cos(a/f_a)]^3$ [@KamionAxi14].[^2]
Axion detection
---------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The (constant) magnetic and (oscillating) electric field directions used in the axion detection experiments. The electric field follows the oscillation of the classical axion field.[]{data-label="figBandE"}](figAxDet03 "fig:"){width="7cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Figure \[figBandE\] captures the idea behind the main experimental axion dark matter detection effort. There are two equivalent pictures describing the axion to photon conversion in the presence of a Direct-Current (DC) magnetic field ${\bf B}$ [@Sikivie89; @Krauss85; @Jihnekim14], briefly described here: The axion decays to two photons through the triangle anomaly. Its lifetime, for an axion mass in the $\mu$eV range, is of order $10^{50}$ s, much larger than the lifetime of our Universe. This decay rate can be significantly enhanced in the presence of a DC ${\bf B}$-field via the inverse Primakov effect [@Sikivie89]. This decay rate is additionally enhanced by the density of the final states, e.g., the quality factor $Q$ of a resonant microwave cavity when its resonant frequency coincides with the axion field oscillation frequency.
In the second picture the axion couples to the product of ${\bf E} \cdot {\bf B}$, where ${\bf E}$ is an electric field. In the presence of a DC magnetic field there is an oscillating electric field appearing with the same frequency as the axion field. If the DC magnetic field is of finite extent, then the oscillating ${\bf E}$-field induces an azimuthal oscillating magnetic field due to Maxwell’s equations. If there is a resonant microwave cavity at the same boundary and with the same resonant frequency, it then provides feedback that enhances the oscillating ${\bf E}$-field by the quality factor of the cavity [@Krauss85; @Jihnekim14]. The power conversion of the axion DM to microwave photons estimated by the two methods is the same and it is given by $$P_{a \to \gamma} = g_{a \gamma \gamma} ^2
\left( \frac {\rho_a}{m_a} \right)
B_0^2 V C_j Q_L \,, \label{eq:axpower}$$ where $Q_L$ is the cavity loaded quality factor, $C_j$ is the mode filling factor, $g_{a \gamma \gamma}$ is the axion-photon-photon coupling constant, $\rho_a$ is the axion DM local density, $m_a$ is the axion mass, $B_0$ is the strength of the DC magnetic field and $V$ the volume of the cavity.
The expected power conversion $P_{a \to \gamma}$ is extra-ordinarily small, but nonetheless it can be within the present experimental capabilities for an axion mass in the 1–20 $\mu $eV range. Have we known the axion mass with a 1 part per million (ppm) accuracy, it would take less than a day to detect it if axions were more than 10% of the DM. The main issue is that, barring the BICEP2 results [@Marsh14; @Gondolo14], we have no such information. The best-suited axion DM mass is below about 1 meV all the way to about 1 $\mu$eV, spanning three orders of magnitude with a potential line width of about 1 ppm. Clearly, scanning the whole axion mass range will require too many steps, and therefore the sensitivity needs to be very high at each step.
Furthermore, in some theoretical scenarios, the axion DM mass is not constrained from below and can be very light, well below 1 $\mu$eV. In addition to the microwave cavity method, which is mostly applicable between 1-20 $\mu$eV, other methods include looking for axions emitted by Sun’s core, and astrophysical limits, as axions can provide another channel of energy loss, significantly altering the star lifetime. An overview of the present experimental/astrophysical limits of the axion coupling constant vs. the axion mass are given in Fig. \[figExp\].
Looking at Eq. (\[eq:axpower\]), it is clear there is a number of possible improvements one can make in this method: (i) Increase the magnetic field value, (ii) Increase the magnetic field volume, and (iii) Increase the cavity quality factor. The pioneering axion DM experiments that started in the late 1980’s [@RBF; @Florida] probed an axion DM candidate in a limited mass region, assuming a stronger axion to photon coupling than is required by theory by roughly two orders of magnitude.
Over a period of more than 15 years, the dominant axion dark matter experiment (ADMX), currently located at the University of Washington and ADMX-HF located at Yale University, have made several conceptual improvements and have improved on those limits. The second generation ADMX experiment, owing to the development of very low noise SQUID amplifiers just below 1 GHz [@Clarke10] and a number of additional smaller developments, has reached the boundaries of a plausible axion DM candidates. Currently implementing a dilution refrigerator to their system is expected to allow them to either detect or exclude an axion comprising 100% of the DM for masses in the range 1–20 $\mu$eV.
The new Center for Axion and Precision Physics (CAPP) [@capp_site], established by the Institute for Basic Science in South Korea [@ibs_site], plans to either detect or exclude an axion DM component down to the 10% level for a similar axion mass range. This will be achieved by [@Yannis14] (a) Development of a 25T and then a 35T solenoidal magnet compared to the currently used 8–9T solenoidal magnets, (b) Substantially improving, roughly by an order of magnitude, the quality factor of the microwave cavities in the presence of strong magnetic fields, and (c) Constructing and running a toroidal cavity with a large volume and a reasonable ${\bf B}$-field value so that the overall product $B^2 V$ is an order of magnitude larger than present values.
The commonly used NbTi superconducting cable has a critical current that falls very rapidly as the magnetic field increases above 10T, making it unsuitable to obtain higher B-field strengths. However, recent developments with high $T_c$ cables makes possible achieving much higher current densities at large B-field values, when they are cooled at low temperatures around 4K. This is an experimental method fuelled by the energy-storage field and prototype magnets are already under development. CAPP is collaborating with the Magnet Division of Brookhaven National Laboratory to develop a 10cm inner bore diameter capable of producing around 25T of magnetic field. Preliminary tests on different high $T_c$ cables are providing encouraging results that the goal can be met. The expected time period for this development is of order of five years, after which we develop a separate magnet with a goal of achieving 35T peak magnetic field, albeit with smaller inner bore diameter. The next step would be to configure a toroidal magnetic field, optimising the use of the magnetic field as the fringe field is minimized in that geometry. Preliminary cable testing results also point to this geometry for achieving the highest possible magnetic field values. The time scale for this development is of order ten years.
The presently used cavities have a quality factor between 50K–100K. It has been reported by ADMX that they are developing cavities with thin-film superconducting coatings on the vertical side walls with the goal of increasing the cavity quality factor by roughly a factor of five. This is possible when the B-field is shaped to be aligned with the vertical wall, minimising the transverse B-field below about 100 Gauss. Further increases of the quality factor are hindered by the top/bottom surfaces of the right-cylindrical cavity as the magnetic field angle traversing the surface is very close to 90 degrees. Our plan to further improve upon this achievement is two-fold: First, develop a toroidal cavity where the B-field can be shaped along the cavity walls reducing the transverse B-field below the required level. If that is possible, the quality factor can be increased by several orders of magnitude. Second, the top/bottom plates are going to be treated in a way that the B-field can penetrate it without affecting the superconducting layer on the inside of the cavity. Again, the quality factor can potentially increase by several orders of magnitude.
The expected axion width is of order 1ppm, [*i.e.*]{} the axion quality factor is a bit better than $10^6$. Therefore, the best one can do is to produce a cavity with the same quality factor, so the best one can expect is to gain a factor of 10 to 20 in the axion to photon power conversion. The scanning speed goes as the square root of the quality factor since there are more steps required in order to cover all possible frequencies, [*i.e.*]{} the best one can expect to do is a scanning speed improvement factor between three and five.
BICEP2 results favour axion masses in the meV range, albeit with only 1–10% of DM composed of axions. This fact makes it particularly difficult to detect it as the volume of microwave cavities are particularly small and not of much practical use at those frequencies, plus the axion DM density is very weak. If the BICEP2 results turn out to be confirmed,[^3] one could follow a different strategy in detecting axions [@Asimina14]. If the axion mass were to be found, then one could launch a dedicated axion DM experiment within a very small axion mass range having much higher chances of success.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Experimental/astrophysical limits of $g_{a\gamma}$ vs. $m_a$ [@RevAxionsCDM]. The experiments giving the limits are shown. The KSVZ and DFSZ lines are from Ref. [@Kim98], and the string calculation is the green line [@Kimagg14]. The limit of the yellow region is the largest one from Ref. [@Kim98]. The excluded gray region is from [@Marsh14; @Gondolo14], allowing a factor 5 generosity due to the domain wall annihilation problem [@BarrKim14]. The astrophysics lines in the bigger box represent that the regions above those lines are excluded.[]{data-label="figExp"}](figAxData04 "fig:"){width="7.5cm"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The cosmological constant problem and string theory {#sec:CC}
===================================================
In order to realize the present-day cosmic acceleration with the cosmological constant $\Lambda$, we require that $\Lambda$ is of the order of $H_0^2$, i.e., $\Lambda \approx H_0^2=(2.1332h \times 10^{-42}~{\rm GeV})^2$, where $h \approx 0.7$. This corresponds to the energy density $\rho_{\Lambda} \approx \Lambda {M_{\rm P}}^2 \approx 10^{-120}{M_{\rm P}}^4$. Even before the discovery of the present-day cosmic acceleration, Weinberg [@WeinbergAnth] put the bound on $\rho_{\Lambda}$, as $$-2\times 10^{-120}{M_{\rm P}}^{4} \lesssim \rho_{\Lambda}\lesssim
6\times10^{-118}{M_{\rm P}}^{4}\,.
\label{rhoregion}$$ The lower bound comes from the fact that the negative cosmological constant does not lead to the collapse of the Universe today. The upper bound corresponds to the requirement that the vacuum energy does not dominate over the matter density for redshifts $z$ larger than 1 to realize the successful structure formation.
There have been attempts to explain the very low values of $\rho_{\Lambda}$ ranging the Weinberg bound (\[rhoregion\]). For example, Bousso and Polchinski [@Bousso] employed the 4-form field $F_{\mu \nu \lambda \sigma}$ with the energy density $F_{\mu \nu \lambda \sigma}F^{\mu \nu \lambda \sigma}/48=c^2/2$, where $c$ is a constant. In the context of string theory, there are “electric charges” (membranes) sourcing the 4-form field dual to “magnetic charges” (5-branes). The constant $c$ can be quantized in integer ($n$) multiples of the membrane charge $q$, such that $c=nq$.
Bousso and Polchinski introduced $J$ 4-form fields together with $J$ membrane species with charges $q_{1},q_{2},\cdots,q_{J}$. The number $J$ can be as large as 100 in string theory. Since the flux energy density of each charge is given by $n_{i}^{2}q_{i}^{2}/2$, the effective cosmological constant reads $$\Lambda=\Lambda_{b}+
\sum_{i=1}^{J}n_{i}^{2}q_{i}^{2}/2\,, \label{qless2}$$ where $\Lambda_b$ is the bare cosmological constant. For the anti de Sitter minimum with $\Lambda_b<0$, there exist integers $n_i$ satisfying $$2|\Lambda_{b}|<\sum_{i=1}^{J}n_{i}^{2}q_{i}^{2}
<2(|\Lambda_{b}|+\Delta\Lambda)\,,\label{nicon}$$ where $\Delta\Lambda\simeq10^{-123}$ in the unit ${M_{\rm P}}=1$.
If we consider a $J$-dimensional grid with axes corresponding to $n_{i}q_{i}$, the displacement of the 4-form field is given by discrete grid points with integers $n_{i}$. The region (\[nicon\]) corresponds to a thin-shell characterized by the radius $r=\sqrt{2|\Lambda_{b}|}$ and the width $\Delta r=\Delta\Lambda/\sqrt{2|\Lambda_{b}|}$. The volume of the thin-shell is $$V_{S}=\Omega_{J-1}r^{J-1}\Delta r=
\Omega_{J-1}|2\Lambda_{b}|^{J/2-1}\Delta\Lambda\,,$$ where $\Omega_{J-1}=2\pi^{J/2}/\Gamma(J/2)$ is the area of a unit ($J-1$)-dimensional sphere. A grid cell has a volume $V_{C}=\prod_{i=1}^{J}q_{i}$. There is at least one value of $\Lambda$ for $V_{C}<V_{S}$, i.e., $$\prod_{i=1}^{J}q_{i}<\frac{2\pi^{J/2}}{\Gamma(J/2)}
|2\Lambda_{b}|^{J/2-1}\Delta\Lambda\,.\label{qicon}$$ When $J=100$, $|\Lambda_{b}|=1$ and $\Delta V=10^{-123}$ with equal charges ($q_{i}=q$, for $i=1,2,\cdots,J$), the condition (\[qicon\]) is satisfied for $q<0.035$. Since the charge $\sqrt{q}$ has the dimension of mass from Eq. (\[qless2\]), this condition translates to $\sqrt{q}<0.19$ in units of [$M_{\rm P}$]{}. Thus, the presence of many 4-form fields allows the possibility of realizing a small effective cosmological constant.
The idea of Bousso and Polchinski is based on the flux energy density originating from multiple 4-form fields. This idea was extended to the so-called flux compactification on a Calabi-Yau manifold in type II string theory. In the presence of fluxes, Kachru, Kallosh, Linde and Trivedi [@KKLT] first set up a supersymmetric anti de Sitter (AdS) vacuum with all moduli fields fixed. Then, they obtained a de Sitter vacuum by adding an anti D3-brane in a warped geometry to lift up the AdS state.
There are hundreds of different 3-cycles on the Calabi-Yau manifold in the flux compactification. A macroscopic observer can view a 5-brane wrapping a 3-cycle as a 2-brane (membrane). The 5-brane can wrap any of these 3-cycles, which gives rise to hundreds of different membranes in four-dimensional space-time. The number of vacua appearing in string theory can be extremely large. For 500 three-cycles with each cycle wrapped by up to 10 fluxes, we have $10^{500}$ vacua.
The possible presence of such a large amount of vacua led to the notion of so-called string landscape [@Susskind]. This landscape includes so many possible configurations of local minima, among which our Universe may correspond to one of them. Each vacuum in the string landscape has different matter and coupling constant. The SM is not predicted uniquely in this picture. The argument is that we may be able to find a vacuum with an extremely small energy density among $10^{500}$ vacua. However, this anthropic argument depends on “[Those packed near $\Lambda=0$ out of $10^{500}$ vacua describe particle phenomenology correctly, in particular with three chiral families and $\sin^2\theta_W=3/8$]{},” otherwise the landscpe vacua differring by $\Delta\Lambda$ describe unacceptable universes. From this reasoning, the string landscape is commented in **Table\[Tab:01\]** as ‘not yet’ established.
A general problem with the anthropic arguments is that they are often applied to a single parameter while fixing all the others. A parameter value that is ruled out in one case may be acceptable if something else is changed at the same time. In this sense, it is not clear that the anthropic arguments of $\Lambda$ provide a satisfactory answer to the cosmological constant problem.
As commented before, the DE scale may be accountable from highly suppressed non-renormalizable terms in string-allowed discrete symmetries [@KimNilles14; @KimJKPS14] if the true vacuum has zero cosmological constant. In this sense, the theoretical solution toward the vanishing cosmological constant is more difficult to solve than obtaining a tiny DE scale on top of the vanishing cosmological constant [@WeinbergCC; @ZeeEgr13].
Inflation {#sec:inflation}
=========
The possibility of an exponential expansion of the Universe was known [@LeMatre; @Sato75; @Sato81; @Kazanas80; @Star80] even before the influential paper of Guth [@inflationold] which advocates diluting away the GUT scale monopoles [@Preskill79]. For example, in the abstract of the Kazanas’s paper [@Kazanas80], it is stated that “...In particular it is shown that under certain conditions this expansion law is exponential. It is further argued that under reasonable assumptions for the mass of the associated Higgs boson this expansion stage could last long enough to potentially account for the observed isotropy of the universe.” In the papers of Sato [@Sato81; @Sato75], diluting away topological defects such as monopoles and domain walls was stressed after the advent of the modern GUT model [@GG74; @PS73]. In the Guth’s paper [@inflationold] it was clearly emphasized that the inflationary paradigm can address the solutions for the homogeneous, horizon and flatness problems.
The scalar field responsible for inflation is called ‘inflaton’. The inflaton field is a superposition of quanta of all possible wave lengths. A quantum fluctuating scale inflates exponentially and after passing the horizon, it is stretched exponentially with an almost scale-invariant form [@Mukhanov81; @Hawking82; @Starobinsky82; @GuthPi82; @Bardeen83] and the frozen-scale still inflates exponentially (see Ref. [@BTW] for a review). Different fluctuating scales go out of the horizon at different cosmic times and their exponentially stretched scales are correlated.
After the end of inflation, the quantum fluctuations enter the horizon again and become the sources of density perturbations. The prediction of nearly scale-invariant primordial perturbations generated during inflation was consistent with the temperature anisotropies of Cosmic Microwave Background (CMB) observed by the COBE satellite [@COBE92]. The recent WMAP and the Planck data of CMB refined the temperature anisotropies to very high accuracy [@WMAP9; @Planck13][^4].
The observables and the constraints implied by inflation are
A sufficient inflation, requiring the large e-fold number, $N_e>70$, for addressing horizon and flatness problems.
The amplitude of temperature anisotropies $\delta T/T\simeq 10^{-5}$, for galaxy formation with CDM.
The spectral index of scalar perturbations $n_s\simeq 0.96$, from WMAP and Planck data.
The tensor-to-scalar ratio $r \lesssim 0.2$, from WMAP and Planck data.
The non-linear estimator of scalar non-Gaussianities for the local shape $f_{\rm NL}^{\rm local}=2.7 \pm 5.8$ (68% CL), from Planck data.
As long as the slow-roll conditions are satisfied, the single-field inflationary scenario generally gives rise to local non-Gaussianities with $|f_{\rm NL}^{\rm local}|$ much smaller than 1 even for most general scalar-tensor theories with second order equations of motion [@Cremi; @Chen; @DeTsu]. Hence the slow-roll single-field models are consistent with the Planck bound of non-Gaussianities. Using the observational bounds of $n_s$ and $r$, we can distinguish between many single-field inflationary models [@Planckinf; @Kuro; @PTP]. For example, the self-coupling potential $V(\phi)=\lambda \phi^4/4$ [@Linde83] and hybrid inflation [@Hybrid] with $n_s>1$ are disfavored from the data.
The amplitude of tensor perturbations is given by ${\it P}_h=2H^2/(\pi^2 {M_{\rm P}}^2)$, so the detection of gravitational waves in CMB observations implies that the energy scale of inflation is directly known [@GWRus75; @GWStar79; @GWRuba82; @GWFabri83; @GWAbbott84]. Since the B-mode polarization of CMB is generated by tensor perturbations but not by scalar perturbations, the B-mode detection is a smoking gun for the existence of primordial gravitational waves.
If the tensor-to-scalar ratio $r$ is smaller than the order of 0.01, it is not easy to detect the CMB B-mode polarization.[^5] If $r$ is detected in the range $r>0.05$, then the energy scale during inflation corresponds to the GUT scale. The great interest in the announcement of $r\sim 0.16$ from the BICEP2 group [@Larger]is because of the implication that the Universe once passed the vacuum energy scale of $10^{16}\,{\,\textrm{GeV}}$. Even though the GUT scale [$M_{\rm GUT}$]{} is humongous from our TeV scale Standard Model, it is tiny from the point of gravity scale, the Planck mass [$M_{\rm P}$]{}. Because of the micro density perturbation, the vacuum energy at the scale $(10^{16}\,{\,\textrm{GeV}})^4$ leads to $r\sim {\rm O}(0.1)$. This phenomenon of the GUT scale energy density during inflation is usually parametrized by chaotic inflation with the potential $V(\phi)=\frac12 m^2\phi^2$ [@Linde83].
![Chaotic inflation with the potential $V(\phi)=m^2\phi^2/2$ [@Linde83]. The red bullet is the Lyth bound which is far above $M_{\rm P}$.[]{data-label="figChaotic2"}](figChaoT05){width="7cm"}
If a large $r$ of order 0.2 is detected, the field value in the $\phi^2$ chaotic inflation is bounded from below, [*i.e.*]{} $\langle\phi\rangle>15{M_{\rm P}}$, which is known as the ‘Lyth bound’ [@Lyth97]. This situation is shown in Fig. \[figChaotic2\], where the energy density at the inflationary epoch is the GUT scale. The field value $\langle\phi\rangle> 15{M_{\rm P}}$ is trans-Planckian and the energy density at [$M_{\rm P}$]{} is tiny. So, one needs a fine-tuning in the $\phi^2$ chaotic inflation. Introducing a confining force at a GUT scale, a heavy axion for the inflaton with a potential bounded from above was proposed [@Freese90], which is called [*natural inflation*]{}. In this scenario, the energy density has the upper bound of order [$M_{\rm GUT}$]{}$^4$ as shown in Fig. \[figNatKNP\](a). One period of the inflaton in this case is of order [$M_{\rm GUT}$]{}, and hence the Lyth bound is violated. To remedy this, two confining forces are introduced with two heavy axions with the resulting potential [@KNP05], [$$\begin{split}
V= &-\Lambda_1^4\cos\left(\alpha\frac{a_1}{F_1}
+\beta\frac{a_2}{F_2} \right)\\& - \Lambda_2^4
\cos\left(\gamma\frac{a_1}{F_1} +\delta\frac{a_2}{F_2} \right) +{\rm constant}\,,
\end{split}$$]{} where $\alpha,\beta,\gamma$, and $\delta$ are determined by the corresponding PQ symmetries of two heavy axions $a_1$ and $a_2$. Even though we allow O(1) couplings, the GUT mass scales can lead to [$M_{\rm P}$]{} with the probability of $\sim$1%. With mass parameters of 50[$M_{\rm GUT}$]{}, we would obtain 50[$M_{\rm P}$]{}with the probability of $\sim$1%. This is the Kim-Nilles-Peloso 2-flation model [@KNP05]. The probability of the 2-flation model with a large decay constant, [[*i.e. *]{}]{}$f_\phi >15{M_{\rm P}}$ to occur as shown in Fig. \[figNatKNP\](b), is about 1%. The green-potential in Fig. \[figNatKNP\](b) is the other heavy axion potential. It can be generalized to N-flation [@Nflation].
The axionic topological defects in the anthropic window [@Pi84; @Tegmark06] can be diluted away if inflation occurs below the anthropic window scale. With the GUT scale energy density during inflation, however, this dilution mechanism does not work. With the GUT energy scale inflation as implied by the BICEP2 [@Larger], it could have pinned down to $f_a\sim 10^{11\,}{\,\textrm{GeV}}$ [@Marsh14; @Gondolo14], using the numerical calculation of radiating axions from axionic string-wall system [@Kawasaki12]. In the numerical calculation, the Vilenkin-Everett mechanism [@Vilenkin82] of erasing the horizon scale string has not been taken into account. In addition, the hidden-sector confining force can erase horizon scale axionic strings such that the QCD axion domain wall is not a serious cosmological problem [@BarrKim14]. The hidden-sector solution needs the hidden-sector domain-wall number of $N_h=1$, which is possible in string compactification with an anomalous U(1) [@KimDW14].
![ (a) The natural inflation of [@Freese90]; (b) The KNP inflation [@KNP05]. The red bullet in the 2-flation is an O(1) value of $\langle a\rangle/f_a$.[]{data-label="figNatKNP"}](figNatKNP){width="7.5cm"}
In addition to pinning down the upper bound on $f_a$, the GUT scale inflation provokes a question, “What is the symmetry which naturally satisfies the Lyth bound [@Lyth97]?" Lyth considered this problem with respect to the $\eta$ parameter [@Lyth14bicep]. But, there exists a more fundamental question. In an ultra-violet completed theory such as string theory, every parameter is calculable. If we consider the $\phi^2$ chaotic inflation of Fig. \[figChaotic2\], there is a question, “Why do we neglect other terms?" In string theory, only discrete symmetries are permitted by the compactification process. For example, a term $\phi^{104}/{M_{\rm P}}^{100}$ can be possible if the discrete symmetries allow it. But with the trans-Planckian value, for example $\langle \phi \rangle \sim 31$, the coefficient must be tuned to 1 out of $10^{127}$, which is as bad as the cosmological constant problem.
Fortunately, there is another way for inflation to occur. We must choose the hilltop inflation, but sacrificing the single-field inflaton. It is not so bad in view of the fact that the 2-flation model already introduced two axions in the inflaton sector. Then, the inflating region is near origin such that the minimum at $f_{\rm DE}$ is far away from the origin. In the region $[0,f_{\rm DE}]$ the vacuum energy is of order [$M_{\rm GUT}$]{}$^4$. This can be obtained from the condition on the quantum numbers of the assumed discrete symmetry [@KimHilltop14]. The inflaton rolls in the yellow region in Fig. \[figHilltop\] where the inflaton takes a green curve in the two-inflatons space.
![The Hilltop inflation [@Freese90]. The green bullet is the point whose effects are observed by the BICEP2 group.[]{data-label="fig:NatKNP"}](figHilltop){width="7.5cm"}
Discussion {#sec:discussion}
==========
After the discovery of the fundamental Brout-Englert-Higgs boson, which is represented as $H_u$ and $H_d\,(=H_u^\dagger\,\rm ~in~ non-SUSY~case)$, we reviewed the cosmological role of spin-0 bosons. This finding hints the possibility of numerous spin-0 bosons ($\phi$) at the GUT scale. Spin-0 bosons at the GUT scale of the canonical dimension 1 can have more important effects to low energy physics compared to those of spin-$\frac12$ fermions of the canonical dimension $\frac32$ (Dirac fermions $\psi,\overline{\psi}$ for example) at the GUT scale. For example, the spin-0 contribution [$$\begin{split}
\frac{\phi^{2n}}{{M_{\rm P}}^{2n+2m-4}}(H_d H_u)^m \label{eq:bosonHH}
\end{split}$$]{} dominates the fermion contribution $\frac{\psi^n\overline{\psi}\,^n}{{M_{\rm P}}^{3n +2m-4 }}(H_d H_u)^m$ for $n,m\ge 1$. In addition, the existence of fundamental spin-0 bosons at the GUT scale may be extended to a larger symmetry: supersymmetric GUTs, or minimal supersymmetric Standard Models from string compactification. The interactions of the singlet fields only can take a SUSY superpotential, for example with GUT scale singlets $\phi$ and trans-Planckian singlets $\Phi$ for simplicity [@KimHilltop14] [$$\begin{split}
W=\sum_i \frac{\phi^{a_i}}{{M_{\rm P}}^{a_i+\ell_i-3}} \Phi^{\ell_i}.\label{eq:Winf}
\end{split}$$]{}
The rationale leading to the forms of Eqs. (\[eq:bosonHH\]) and (\[eq:Winf\]) are the discrete symmetries obtained from string compactification [@KimNilles14],[^6] which guarantees the absence of gravity obstruction of discrete symmetries, for example via wormholes [@Kim13worm]. The form of the interactionis (\[eq:Winf\]) can lead to inflation with trans-Planckian decay constant with a multi-field hilltop potential, [*i.e.*]{} [**BCM2**]{}. The form of the interactionis (\[eq:bosonHH\]) can lead to QCD axion via [**BCM1**]{}, and the DE scale via [**CCtmp**]{}. The fundamental scalars at the TeV, GUT and trans-Planckian scales allow all scenarios presented in Subsec. \[subsec:spin0cosm\]. These are worked out on top of vanishing cosmological constant, which is assumed in any particle physics models. At present, we do not have any persuasive hint toward a theoretical solution of the vanishing cosmological constant. Any theory for the vanishing cosmological constant must satisfy the requirements of particle phenomenology we used in this review.
The fundamental scalars may be detectable if their couplings to gluons are appreciable. The front runner in the search of fundamental scalars hinting high energy (GUT or intermediate) scales is the QCD axion which couples to the gluon anomaly.
Acknowledgement {#acknowledgement .unnumbered}
===============
JEK is supported in part by the National Research Foundation (NRF) grant funded by the Korean Government (MEST) (No. 2005-0093841) and by the IBS (IBS-R017-D1-2014-a00), YKS is supported by the IBS (IBS-R017-D1-2014-a00), and ST is supported by the Grant-in-Aid for Scientific Research from JSPS (No. 24540286).
Disclosure/Conflict-of-Interest Statement {#disclosureconflict-of-interest-statement .unnumbered}
=========================================
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
[99]{}
\#1\#2\#3[[*Phys.Rep.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Rev. Mod. Phys.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Nucl.Phys.B*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Phys.Lett.B*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Phys.Rev.D*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Phys.Rev.Lett.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[ [**\#1**]{} (\#3) \#2(E)]{} \#1\#2\#3[[*JHEP*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*JCAP*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Z.Phys.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Euro.Phys.J.C*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*J.Phys.G*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Int.J.Mod.Phys.D*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Mod.Phys.Lett.A*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Astrophys.J.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Nature*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Sov.J.Nucl.Phys.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Astrophys.J.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Mon.Not.Roy.Astron.Soc.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*JETPLett.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Prog.Theor.Phys.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*J.KoreanPhys.Soc.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[[*Front.Phys.*]{} (\#3) [**\#1**]{}: \#2]{} \#1\#2\#3[(\#3) [**\#1**]{}: \#2]{}
\#1\#2\#3[[*ibid.*]{} [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[ [**\#1**]{} (\#3) \#2(E)]{} \#1\#2\#3[ [**\#1**]{} (\#3) \#2(E)]{}
Ade PAR [[*et al.*]{}]{}(Planck Collaboration). Planck 2013 results XXII: Constraints on inflation. \[arXiv:1303.5082 \[astro-ph.CO\]\].
Hinshaw G [[*et al.*]{}]{}(WMAP Collaboration). Nine-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Cosmological Parameter Results. [*Astrophys. J. Suppl.*]{} (2013) [**208**]{}: 19. doi: 10.1088/0067-0049/208/2/19
Zwicky F. [*Helvetica Physica Acta*]{} (1933) [**6**]{}:110.
Preskill J, Wise MB, Wilczek F. Cosmology of the Invisible Axion. . doi: 10.1016/0370-2693(83)90637-8
Abbott LF, Sikivie P. A. Cosmological Bound on the Invisible Axion. . doi:10.1016/0370-2693(83)90638-X
Dine M, Fischler W. The Not So Harmless Axion. . doi: 10.1016/0370-2693(83)90639-1.
Kim JE. Weak Interaction Singlet and Strong CP Invariance. . doi: 10.1103/PhysRevLett.43.103
Shifman MA, Vainstein VI, Zakharov VI. Can Confinement Ensure Natural CP Invariance of Strong Interactions?. . doi: 10.1016/0550-3213(80)90209-6
Zhitnitsky AP. On Possible Suppression of the Axion Hadron Interactions. (In Russian) [*Yad.Fiz.*]{} (1980) [**31**]{}: 497-504 \[[*Sov.J.Nucl.Phys.*]{} (1980) [**31**]{}: 260\]
Dine M, Fischler W, Srednicki M. A Simple Solution to the Strong CP Problem with a Harmless Axion. . doi: 10.1016/0370-2693(81)90590-6
Kim JE. Light Pseudoscalars, Particle Physics and Cosmology. . doi: 10.1016/0370-1573(87)90017-2
Cheng H-Y. The Strong CP Problem Revisited. . doi: 10.1016/0370-1573(88)90135-4
Peccei RD. in [*CP Violation*]{}, ed. C. Jarlskog \[[*Adv. Ser. Direct. High Energy Phys.*]{} (World Scientific, Singapore, 1989), pp. 503–551\]
Kim JE, Carosi G. Axions and the Strong CP Problem. . doi: 10.1103/RevModPhys.82.557
Baer H, Choi K-Y, Kim JE, Roszkowski L. Dark matter production in the early Universe: beyond the thermal WIMP paragigm. \[arXiv:1407.0017 \[hep-ph\]\].
Sanhi V, Starobinsky A. The case for a positive cosmological $\Lambda$ term. . doi: 10.1142/S0218271800000542
Peebles P, Ratra B. The cosmological constant and dark energy. . doi: 10.1103/RevModPhys.75.559
Padmanabhan T. Cosmological constant: The weight of the vacuum. . doi: 10.1016/S0370-1573(03)00120-0
Copeland EJ, Sami M, Tsujikawa S. Dynamics of dark energy. . doi: 10.1142/S021827180600942X
Silvestri A, Trodden M. Approaches to Understanding Cosmic Acceleration. [*Rept.Prog.Phys.Part.Sci.*]{} (2009) [**72**]{}: 096901. doi: 10.1088/0034-4885/72/9/096901
Caldwell RR, Kamionkowski M. The Physics of Cosmic Acceleration. [*Ann.Rev.Nucl.Part.Sci.*]{} (2009) [**59**]{}: 397–429. doi: 10.1146/annurev-nucl-010709-151330
Tsujikawa S. Dark energy: investigation and modeling. [*Astrophysics and Space Science Library*]{} (2011) [**370**]{}: 331–402. doi: 10.1007/978-90-481-8685-3-8
Weinberg S. The Cosmological Constant Problem. . doi: 10.1103/RevModPhys.61.1
Weinberg S. Anthropic Bound on the Cosmological Constant. . doi: 10.1103/PhysRevLett.59.2607
Tegmark M, Aguirre A, Rees M, Wilczek F. Dimensionless constants, cosmology and other dark matters. . doi: 10.1103/PhysRevD.73.023505
A. Zee, [*Einstein gravity in a knutshell*]{} (Princeton Univ. Press, 2013), Sec. X.7, pp. 752
Hawking SW. The cosmological constant is probably zero. . doi: 10.1016/0370-2693(84)91370-4
Kim JE, Kyae B, Lee HM. Randall-Sundrum model for selftuning the cosmological constant. . doi: 10.1103/PhysRevLett.86.4223
Kim JE. Cosmological constant is probably adjustable in brane worlds. . doi: 10.1103/PhysRevD.81.123018
Kim JE, Langacker P, Levine M, Williams HH. A Theoretical and Experimental Review of the Weak Neutral Current: A Determination of Its Structure and Limits on Deviations from the Minimal SU(2)$_L\times$U(1) Electroweak Theory. . doi: 10.1103/RevModPhys.53.211
Langacker P, Luo M-X. Implications of precision electroweak experiments for $M_t, \rho_0, \sin^2\theta_W$ and grand unification. . doi: 10.1103/PhysRevD.44.817
Kim JE. $Z_3$ orbifold construction of SU(3)$^3$ GUT with $\sin^2\theta_W^0=\frac38$. . doi: 10.1016/S0370-2693(03)00567-7
Kobayashi T, Raby S, Zhang R. Searching for realistic 4d string models with a Pati-Salam symmetry – Orbifold grand unified theories from heterotic string compactification on a $Z_6$ orbifold. . doi: 10.1016/j.nuclphysb.2004.10.035
Kim JE, Kyae B. Flipped SU(5) from $Z_{12-I}$ orbifold with Wilson line. . doi: 10.1016/j.nuclphysb.2007.02.008
Huh J-H, Kim JE, Kyae B. SU(5)$_{\rm flip}\times$SU(5)$'$ from $Z_{12-I}$. . doi: 10.1103/PhysRevD.80.115012
Lebedev O, Nilles HP, Ramo-Sanches S, Raby S, Ratz M, Vaudrevange PKS, Wingerter A. A Mini-Landscape of Exact MSSM Spectra in Heterotic Orbifolds. . doi: 10.1016/j.physletb.2006.12.012
Nilles HP, Ramo-Sanches S, Ratz M, Vaudrevange PKS. From strings to the MSSM. . doi: 10.1140/epjc/s10052-008-0740-1
Perlmutter S [*et al.*]{} \[Supernova Cosmology Project\]. Measurements of $\Omega$ and $\Lambda$ from 42 High-Redshift Supernovae. . doi: 10.1086/307221
Riess AG [*et al.*]{} \[Supernova Search Team\]. Observational Evidence from Supernovae for an Accelerating Universe and a Cosmological Constant. . doi: 10.1086/300499
Milgrom M. A. Modification of the Newtonian dynamics as a possible alternative to the hidden mass hypothesis. . doi: 10.1086/161130
Beckenstein J, Milgrom M. Does the missing mass problem signal the breakdown of Newtonian gravity?. . doi: 10.1086/162570
Zlatev I, Wang L, Stenhardt PJ. Quintessence, cosmic coincidence, and the cosmological constant. . doi: 10.1103/PhysRevLett.82.896
Wetterich C. Cosmology and the fate of dilatation symmetry. . doi: 10.1016/0550-3213(88)90193-9
Kim JE, Nilles HP. Dark energy from approximate U(1)$_{\rm de}$ symmetry. . doi: 10.1016/j.physletb.2014.01.031
Guth A. The Inflationary Universe: A Possible Solution to the Horizon and Flatness Problems. . doi: 10.1103/PhysRevD.23.347
Linde AD. A New Inflationary Universe Scenario: A Possible Solution of the Horizon, Flatness, Homogeneity, Isotropy and Primordial Monopole Problems. . doi: 10.1016/0370-2693(82)91219-9
Albrecht A, Steinhardt PJ. Cosmology for Grand Unified Theories with Radiatively Induced Symmetry Breaking. . doi: 10.1103/PhysRevLett.48.1220
Coleman SR, Weinberg EJ. Radiative Corrections as the Origin of Spontaneous Symmetry Breaking. . doi: 10.1103/PhysRevD.7.1888
Affleck I, Dine M. A New Mechanism for Baryogenesis. . doi: 10.1016/0550-3213(85)90021-5
Kim JE, Nilles HP. A Quintessential Axion. . doi: 10.1016/S0370-2693(02)03148-9
Kim JE. Axion and almost massless quark as ingredients of quintessence. . doi: 10.1088/1126-6708/1999/05/022
Kim JE. Model dependent axion as quintessence with almost massless hidden sector quarks. . doi: 10.1088/1126-6708/2000/06/016
Kim JE, H. P. Nilles HP. Axionic dark energy and a composite QCD axion. . doi: 10.1088/1475-7516/2009/05/010
Choi K. String or M theory axion as a quintessence. . doi: 10.1103/PhysRevD.62.043509
Semertzidis Y. Talk presented at “10th Patras Workshop on Axions, WIMPs, and WISPs", CERN, Geneva, 29 June–4 July 2014.
Goldstone J. Field Theories with Superconductor Solutions. Nuovo Cim. (1961) [**19**]{}: 154–164. doi: 10.1007/BF02812722
Higgs PW. Broken symmetries, massless particles and gauge fields. . doi: 10.1016/0031-9163(64)91136-9; Broken Symmetries and the Masses of Gauge Bosons. . doi: 10.1103/PhysRevLett.13.508
Brout R, Englert F. Broken Symmetry and the Mass of Gauge Vector Mesons. . doi: 10.1103/PhysRevLett.13.321
Guralnik GS, Hagen CR, Kibble TWB. Global Conservation Laws and Massless Particles. . doi: 10.1103/PhysRevLett.13.585
Susskind L. Mass without scalars. . doi: 10.1016/0550-3213(79)90364-X
Weinberg S. Implications of dynamical symmetry breaking: An addendum. . doi: 10.1103/PhysRevD.19.1277, an addendum to . doi: 10.1103/PhysRevD.13.974
Ade PAR [[*et al.*]{}]{}(BICEP2 Collaboration). BICEP2 I: Detection Of B-mode Polarization at Degree Angular Scales. . doi: 10.1103/PhysRevLett.112.241101
Carroll SM. Quintessence and the Rest of the World. . doi: 10.1103/PhysRevLett.81.3067
Fujii Y. Origin of the Gravitational Constant and Particle Masses in Scale Invariant Scalar–Tensor Theory. . doi: 10.1103/PhysRevD.26.2580
Fujii Y, Nishioka T. Model of a Decaying Cosmological Constant. . doi: 10.1103/PhysRevD.42.361
Ford LH. Cosmological Constant Damping by Unstable Scalar Fields. . doi: 10.1103/PhysRevD.35.2339
Peebles PJE, Ratra B. Cosmology with a Time Variable Cosmological Constant. . doi: 10.1086/185100
Ratra B, Peebles PJE. Cosmological Consequences of a Rolling Homogeneous Scalar Field. . doi: 10.1103/PhysRevD.37.3406
Chiba T, Sugiyama N, Nakamura T. Cosmology with x matter. . doi: 10.1093/mnras/289.2.L5; Observational tests of x matter models. . doi: 10.1046/j.1365-8711.1998.02012.x
Ferreira PG, Joyce M. Structure formation with a selftuning scalar field. . doi: 10.1103/PhysRevLett.79.4740; Cosmology with a primordial scaling field. . doi: 10.1103/PhysRevD.58.023503
Caldwell R R, Dave R, Steinhardt PJ. Cosmological imprint of an energy component with general equation of state. . doi: 10.1103/PhysRevLett.80.1582
Copeland EJ, Liddle AR, Wands D. Exponential potentials and cosmological scaling solutions. . doi: 10.1103/PhysRevD.57.4686
Fukugita M, Takahara F, Yamashita K, Yoshii Y. Test for the cosmological constant with the number count of faint galaxies. . doi: 10.1086/185813
Green MB, Schwarz JH, Witten E. [*Superstring Theory*]{} (Cambridge University Press, 1987)
Olive KA. Inflation. . doi: 10.1016/0370-1573(90)90144-Q
Zlatev I, Wang L-M, Steinhardt PJ. Quintessence, cosmic coincidence, and the cosmological constant. . doi: 10.1103/PhysRevLett.82.896
Steinhardt PJ, Wang L-M, Zlatev I. Cosmological tracking solutions. . doi: 10.1103/PhysRevD.59.123504
Binetruy P. Models of dynamical supersymmetry breaking and quintessence. . doi: 10.1103/PhysRevD.60.063502
Caldwell RR, Linder EV. The limits of quintessence. . doi: 10.1103/PhysRevLett.95.141301
Amendola L. Scaling solutions in general nonminimal coupling theories. . doi: 10.1103/PhysRevD.60.043501
Khoury J, Weltman A. Chameleon fields: Awaiting surprises for tests of gravity in space. . doi: 10.1103/PhysRevLett.93.171104
Brans C, Dicke R. Mach’s principle and a relativistic theory of gravitation. . doi: 10.1103/PhysRev.124.925
Khoury J, Weltman A. Chameleon cosmology. . doi: 10.1103/PhysRevD.69.044026
Hu W, Sawicki I. Models of $f(R)$ Cosmic Acceleration that Evade Solar-System Tests. . doi: 10.1103/PhysRevD.76.064004
Capozziello S, Tsujikawa S. Solar system and equivalence principle constraints on $f(R)$ gravity by chameleon approach. . doi: 10.1103/PhysRevD.77.107501
Gannouji R, Moraes B, Mota D, Polarski D, Tsujikawa S, Winther H. Chameleon dark energy models with characteristic signatures. . doi: 10.1103/PhysRevD.82.124006
Sotiriou TP, Faraoni V. $f(R)$ theories of gravity. . doi: 10.1103/RevModPhys.82.451
De Felice A, Tsujikawa S. $f(R)$ theories. [*LivingRev.Rel.*]{} (2010) [**13**]{}: 3. doi: 10.12942/lrr-2010-3
Vainshtein A. To the problem of nonvanishing gravitation mass. . doi: 10.1016/0370-2693(72)90147-5
Deffayet C, Esposito-Farese G, Vikman A. Covariant Galileon. . doi: 10.1103/PhysRevD.79.084003
Nicolis A, Rattazzi R, Trincherini E. The Galileon as a local modification of gravity. . doi: 10.1103/PhysRevD.79.064036
Kimura R, Kobayashi T, Yamamoto K. Vainshtein screening in a cosmological background in the most general second-order scalar-tensor theory. . doi: 10.1103/PhysRevD.85.024023
De Felice A, Kase R, Tsujikawa S, Vainshtein mechanism in second-order scalar-tensor theories. . doi: 10.1103/PhysRevD.85.044059
Kase R, Tsujikawa S, Screening the fifth force in the Horndeski’s most general scalar-tensor theories. . doi: 10.1088/1475-7516/2013/08/054
Kim JE. Abelian discrete symmetries $Z_{N}$ and $Z_{nR}$ from string orbifolds. . doi: 10.1016/j.physletb.2013.08.039
Dias AG, Machado ACB, Nishi CC, Ringwald A, Vaudrevange PKS. The Quest for an Intermediate-Scale Accidental Axion and Further ALPs. . doi: 10.1007/JHEP06(2014)037
Krauss LM, Wilczek F. Discrete Gauge Symmetry in Continuum Theories. . doi: 10.1103/PhysRevLett.62.1221
Kim JE. Modeling the small dark energy scale with a quintessential pseudoscalar boson. . doi: 10.3938/jkps.64.795
Belavin AA, Polyakov A, Schwartz A, Tyupkin Y. Pseudoparticle Solutions of the Yang-Mills Equations. . doi: 10.1016/0370-2693(75)90163-X
't Hooft G. How Instantons Solve the U(1) Problem. . doi: 10.1016/0370-1573(86)90117-1
Weinberg S. The U(1) problem. . doi: 10.1103/PhysRevD.11.3583
Weinberg S, and also Wilczek F, in [*Unification of Elementary Forces and Gauge Theories*]{} edited by Cline DB, Mills FE (Harwood Academic Publishers, London, 1978)
Peccei RD, Quinn HR. CP Conservation in the Presence of Instantons. . doi: 10.1103/PhysRevLett.38.1440
Kim JE. Constraints on Very Light Axions from Cavity Experiments. doi:10.1103/PhysRevD.58.055006
Choi KS, Kim IW, Kim J E. String compactification, QCD axion and axion-photon-photon coupling. . doi: 10.1088/1126-6708/2007/03/116
Kim JE. Calculation of axion-photon-photon coupling in string theory. . doi: 10.1016/j.physletb.2014.06.021
Nomura Y, Watari T, Yanagida T. Quintessence axion potential induced by electroweak instanton effects. . doi: 10.1016/S0370-2693(00)00605-5
Kamionkowski M, Pradler J, Walker DGE. Dark energy from string axiverse. \[arXiv: 1409.0549\[hep-ph\]\].
Arvanitaki A, Dimopoulos S, Dubovsky S, Kaloper N, March-Russell J. String axiverse. . doi: 10.1103/PhysRevD.81.123530
Amendola L, and Barbieri R. Dark matter from an ultra-light pseudo-Goldsone-boson. . doi: 10.1016/j.physletb.2006.08.069
Hu W, Barkana R, Gruzinov A. Cold and fuzzy dark matter. . doi: 10.1103/PhysRevLett.85.1158
Sikivie P. Experimental Tests of the Invisible Axion. , Erratum-ibid. 52, 695 (1984). doi: 10.1103/PhysRevLett.51.1415
Krauss L, Moody J, Wilczek F, Morris EB. Calculations for Cosmic Axion Detection. . doi: 10.1103/PhysRevLett.55.1797
Hong J, Kim JE, Nam S, Semertzidis KY. Calculations of resonance enhancement factor in axion-search tube-experiments. \[arXiv:1403.1576 \[hep-ph\]\].
Marsh DJE, Grin D, Hlozek R, Ferreira PG. Tensor Detection Severely Constrains Axion Dark Matter. . doi: 10.1103/PhysRevLett.113.011801
Vissineli L, Gondolo P. Axion cold dark matter in view of BICEP2 results. . doi: 10.1103/PhysRevLett.113.011802
DePanfilis S [[*et al.*]{}]{}(Rochester-Brookhaven-Fermilab Collaboration). Limits on the Abundance and Coupling of Cosmic Axions at $4.5 < m_a < 5.0\,\mu{\rm eV}$. . doi: 10.1103/PhysRevLett.59.839
Hagmann C [[*et al.*]{}]{}(Florida Group). Results from a search for cosmic axions. . doi: 10.1103/PhysRevD.42.1297
Clarke J. Then and Now. . doi: 10.1142/S0217979210056438
“<http://capp.ibs.re.kr/html/capp_en/>”
“<http://www.ibs.re.kr/eng.do/>”
Arvanitaki A, Geraci AA. Resonant detection of axion mediated forces with Nuclear Magnetic Resonance. \[arXiv: 1403.1290 \[hep-ph\]\].
Adam R. [[*et al.*]{}]{}(Planck Collaboration). Planck intermediate results, XXX: The angular power spectrum of polarized dust emission at intermediate and high Galactic latitudes. arXiv: 1409.5738 \[astro-ph.CO\]
Bousso R, Polchinski J. Quantization of four form fluxes and dynamical neutralization of the cosmological constant. . doi: 10.1088/1126-6708/2000/06/006
Kachru S, Kallosh R, Linde A, Trivedi S. De Sitter vacua in string theory. . doi:10.1103/PhysRevD.68.046005
Susskind L. The anthropic landscape of string theory. In [*Universe or multiverse?*]{}, pp. 247–266, ed. Carr B J. \[arXiv: hep-th/0302219\].
Lemaitre G. [*Annales de la Society Scientifique de Bruxelles*]{} (1933) [**47A**]{}: 49 \[English in [*Gen. Rel. Grav.*]{} (1997) [**29**]{}: 641
Sato K. Cosmological Baryon Number Domain Structure and the First Order Phase Transition of a Vacuum. . doi: 10.1016/0370-2693(81)90805-4
Sato K. First Order Phase Transition of a Vacuum and Expansion of the Universe. . doi: 1981MNRAS.195.467S
Kazanas D. Dynamics of the Universe and Spontaneous Symmetry Breaking. . doi: 10.1086/183361
Starobinsky AA. A New Type of Isotropic Cosmological Models Without Singularity. . doi: 10.1016/0370-2693(80)90670-X
Preskill J. Cosmological Production of Superheavy Magnetic Monopoles. . doi: 10.1103/PhysRevLett.43.1365
Georgi H, Glashow SL. Unity of All Elementary Particle Forces. . doi: 10.1103/PhysRevLett.32.438
Pati J, Salam A. Unified Lepton-Hadron Symmetry and a Gauge Theory of the Basic Interactions. . doi: 10.1103/PhysRevD.8.1240
Mukhanov VF, Chibisov GV. Quantum Fluctuation and Nonsingular Universe (In Russian). [*Pis’ma Zh. Eksp. Teor. Fiz.*]{} (1981) [**33**]{}:549–553 \[\].
Hawking SW. The Development of Irregularities in a Single Bubble Inflationary Universe. . doi: 10.1016/0370-2693(82)90373-2
Starobinsky AA. Dynamics of Phase Transition in the New Inflationary Universe Scenario and Generation of Perturbations. . doi: 10.1016/0370-2693(82)90541-X
Guth AH, Pi S-Y. Fluctuations in the New Inflationary Universe. . doi: 10.1103/PhysRevLett.49.1110
Bardeen J, Steinhardt P, Turner M. Spontaneous Creation of Almost Scale - Free Density Perturbations in an Inflationary Universe. . doi:10.1103/PhysRevD.28.679
Bassett B, Tsujikawa S, Wands D. Inflation dynamics and reheating. . doi: 10.1103/RevModPhys.78.537
Smoot GF [*et al.*]{}. Structure in the COBE differential microwave radiometer first year maps. . doi: 10.1086/186504
Turner MS. Talk presented at “The 5 Year Celebration of Bethe Center”, Univ. Bonn, 11 April 2014
Creminelli P, Zaldarriaga M. Single field consistency relation for the 3-point function. . doi: 10.1088/1475-7516/2004/10/006
Chen X, Huang M, Kachru S, Shiu G. Observational signatures and non-Gaussianities of general single field inflation. . doi: 10.1088/1475-7516/2007/01/002
De Felice A, Tsujikawa S. Shapes of primordial non-Gaussianities in the Horndeski’s most general scalar-tensor theories. . doi: 10.1088/1475-7516/2013/03/030
Ade PAR [*et al.*]{}. Planck 2013 results. XXII. Constraints on inflation. \[arXiv:1303.5082 \[astro-ph.CO\]\].
Tsujikawa S, Ohashi J, Kuroyanagi S, De Felice A. Planck constraints on single-field inflation. . doi: 10.1103/PhysRevD.88.023529
Tsujikawa S. Distinguishing between inflationary models from cosmic microwave background. [*PTEP*]{} (2014) [**6**]{}: 06B104. doi: 10.1093/ptep/ptu047
Linde AD. Chaotic inflation. . doi: 10.1016/0370-2693(83)90837-7
Linde AD. Hybrid inflation. . doi: 10.1103/PhysRevD.49.748
Grishchuk LP. Amplification of gravitational waves in an istropic universe. [*Zh. Eksp. Teor. Fiz.*]{} (1974) [**67**]{}: 825-838 \[[*Sov. Phys. J. Exp. Theor. Phys.*]{} (1975) [**40**]{}: [409–415]{}\]
Starobinsky AA. Spectrum of relict gravitational radiation and the early state of the universe. [*Zh. Eksp. Teor. Fiz. Pisma Red.*]{} \[\]
Rubakov VA, Sazhin MV, Veryaskin AV. Graviton Creation in the Inflationary Universe and the Grand Unification Scale. . doi: 10.1016/0370-2693(82)90641-4
Fabbri R, Pollock MD. The Effect of Primordially Produced Gravitons upon the Anisotropy of the Cosmological Microwave Background Radiation. . doi: 10.1016/0370-2693(83)91322-9
Abbott LF, Wise MB. Constraints on Generalized Inflationary Cosmologies. . doi: 10.1016/0550-3213(84)90329-8
Lyth DH. What would we learn by detecting a gravitational wave signal in the cosmic microwave background anisotropy?. . doi: 10.1103/PhysRevLett.78.1861
Freese K, Frieman JA, Orlinto AV. Natural inflation with pseudo Nambu–Goldstone bosons. . doi: 10.1103/PhysRevLett.65.3233
Kim JE, Nilles HP, Peloso M. Completing Natural Inflation. . doi: 10.1088/1475-7516/2005/01/005
Dimopoulos S, Kachru S, McGreevy J, Wacker JG. N-flation. . doi: 10.1088/1475-7516/2008/08/003
Pi S-Y. Inflation Without Tears. . doi: 10.1103/PhysRevLett.52.1725
Hiramatsu T, Kawasaki M, Saikawa K, Sekiguchi T. Axion cosmology with long-lived domain walls. . doi: 10.1088/1475-7516/2013/01/001
Vilenkin A, Everett AE. Cosmic Strings and Domain Walls in Models with Goldstone and PseudoGoldstone Bosons. . doi: 10.1103/PhysRevLett.48.1867
Barr SM, Kim JE. New confining force solution of QCD domain wall problem. \[arXiv:1407.4311 \[hep-ph\]\].
Kim JE. High scale inflation, model-independent string axion, and QCD axion with domain wall number one. . doi: 10.1016/j.physletb.2014.05.033
Lyth DH. BICEP2, the curvature perturbation and supersymmetry. \[arXiv:1403.7323 \[hep-ph\]\]
Kim JE. The inflation point in U(1)$_{\rm de}$ hilltop potential assisted by chaoton, BICEP2 data, and trans-Planckian decay constant. . doi: 10.1016/j.physletb.2014.08.025
[^1]: However, the scalar in this case is composite.
[^2]: For this specific form, one needs fine-tunings between domain wall number one, two, and three terms in the potential.
[^3]: We note that the recent Planck-dust report extrapolated to the BICEP2 field gives the dust contribution similar to $r\approx 2$ without the dust contribution [@PlunckDust14].
[^4]: From the Planck data the existence of CDM was also confirmed (by 7$\sigma$ [@Turner14Talk]) better than any other data.
[^5]: However, the future observations like LiteBIRD may reach this range.
[^6]: See, also, the discrete gauge symmetries in the field theory language [@KraussW89].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove bounds on operator growth and infinite temperature out-of-time-ordered correlators in many-body systems with $N$ spin-$\frac{1}{2}$ degrees of freedom which interact via two-body all-to-all interactions. Our results parametrically improve previous bounds, and sharply constrain when and how quantum simulators, including trapped ion crystals and cavity quantum electrodynamics, can study quantum gravity.'
author:
- Chao Yin
- Andrew Lucas
bibliography:
- 'scram\_all2all\_spin.bib'
title: 'Bound on quantum scrambling with all-to-all interactions'
---
Introduction
============
In the past few years, the holographic correspondence between quantum many-body systems and quantum gravity in one more spacetime dimension [@Maldacena:1997re] has attracted intense interest. In particular, the realization that microscopic models including the Sachdev-Ye-Kitaev model [@Sachdev:2015efa; @Maldacena:2016hyu; @Kitaev:2017awl] might realize (aspects of) quantum gravity has set off a hunt for microscopic models that mimic quantum gravity, and might also be studied experimentally [@Chew_2017; @Chen_2018; @Marino_2019; @Lewis_Swan_2019; @Alavirad_2019; @Bentsen_2019prl; @Bentsen_2019prx].
A key property of quantum black holes (and thus a theory of quantum gravity) is that they are *fast scramblers* [@Sekino:2008he]. For our purposes, fast scrambling means that out-of-time-ordered correlators (OTOCs) exhibit exponential growth [@Shenker:2013pqa]. In a theory of $N$ spin-$\frac{1}{2}$ degrees of freedom, we expect that for any $1\le i,j\le N$, $$\left\langle [X_i(t),X_j]^2 \right\rangle \sim -\frac{1}{N}\mathrm{e}^{\lambda t}. \label{eq:toyotoc}$$ Here $X_i$, $Y_i$ and $Z_i$ denote the three Pauli matrices acting on spin $i$. The key feature of (\[eq:toyotoc\]) is that $\lambda$ is independent of $N$, and so the OTOC (as we defined it) becomes of order 1 in a *scrambling time* $$t_{\mathrm{s}} \sim \log N. \label{eq:tslogN}$$ (\[eq:tslogN\]) is believed to hold in all theories of quantum gravity, for (almost?) every pair of $i$ and $j$. Note that (\[eq:tslogN\]) serves as our (informal) definition of scrambling time.
The canonical Lieb-Robinson theorem [@Lieb1972], which says that quantum information spreads ballistically in a $d$-dimensional lattice, forbids fast scrambling in conventional lattice models. However, generalizations of the Lieb-Robinson bounds to spin systems defined on more abstract interaction graphs, including those with all-to-all interactions (each spin couples to each other spin), do suggest fast scrambling is permitted [@Lashkari_2013; @Bentsen_2019; @guo2019signaling].
Happily, it is experimentally possible to realize the spatially non-local interactions required of a fast scrambler. As a simple example, we can realize the Hamiltonian $$H = \frac{1}{N^\alpha} \sum_{i,j=1}^N J_{ij}(t)Z_i Z_j + \sum_{i=1}^N \sum_{a=1}^3 h^a_i(t) X^a_i \label{eq:H}$$ in a trapped ion crystal [@Britton_2012] (with expected exponent $\alpha=1$). Hamiltonians with similar simple all-to-all interactions can be achieved in cavity quantum electrodynamics [@Leroux_2010; @thompson2020] (with $\alpha=0$). Remarkably, in these platforms it is possible to measure certain kinds of OTOCs [@Garttner_2017; @Li_2017; @wei2019OTOC]. Here $X^a_i = \lbrace X_i,Y_i,Z_i\rbrace$ is shorthand for the three Pauli matrices, and $\alpha$ is a free parameter governing the strength of the all-to-all interactions; we take $J_{ij}$ to scale independently of $N$. In the simplest experiments, all $J_{ij}=J$ are the same. If such a system can model a fast scrambler, it would allow for near-term experimental tests of aspects of quantum gravity.
Our goal is to rigorously address the extent to which (\[eq:H\]), along with many generalizations, could realize a fast scrambler in an experiment. We will show that at infinite temperature, in this family of models (\[eq:H\]), $$t_{\mathrm{s}} \gtrsim N^{\alpha - \frac{1}{2}}. \label{eq:main}$$ (We postpone the precise statement and its proof.) Hence it is impossible to have both fast scrambling (which requires $\alpha\le \frac{1}{2}$) and extensivity of the energy spectrum ($\alpha\ge 1$), at least in the model with $J_{ij}=J$.
Implications
============
Our main result (\[eq:main\]) is complementary to recent works [@li2020fast; @belyansky2020minimal] which have proposed studying fast scrambling in models of a similar form to (\[eq:H\]). Our bound (\[eq:main\]) is not incompatible with their key results, so long as $\alpha\le \frac{1}{2}$ is taken. Whether such a small value of $\alpha$ has further interesting consequences or constraints on the many-body dynamics is an interesting open question. It is unclear whether this constraint is irrelevant for the faithful simulation of quantum gravity in an experiment, and whether or not such models can realize in a non-trivial way the temperature-dependent bound on OTOC growth conjectured in [@Maldacena:2015waa].
Indeed, (\[eq:main\]) has clear implications for how a quantum simulator, such as a trapped ion crystal, could be used for the experimental study of information scrambling and quantum gravity. Certainly we must take $\alpha\le \frac{1}{2}$ to realize fast scrambling. The only way for such a model to be thermodynamically extensive is for $J_{ij}$ to be a matrix with order 1 entries and maximal eigenvalue $N^{1/2}$. Heuristically this means that $J_{ij}$ is a random matrix [@mehta]. Unfortunately, such a regime is not yet realized in a coherent quantum simulation with hundreds of qubits. For example, focusing on trapped ion platforms, such a regime would require detuning the driving laser very far from the vibrational modes of the ion crystal [@Britton_2012], leading to very weak collective interactions.
The more practical alternative for experiments is to use the single-site fields in (\[eq:H\]) to dephase the many-body wave function, leading to genuine quantum dynamics and scrambling. After all, since all Pauli $Z$s commute in (\[eq:H\]), having single-site $X$ and $Y$ fields is mandatory to realizing chaos. The results of [@belyansky2020minimal] suggest this approach may be feasible. However, the Hamiltonian must then be strongly time-dependent, meaning that no finite temperature physics may be realized. As the emergence of a semiclassical bulk geometry out of quantum dynamics critically relies on a low temperature compared to microscopic energy scales, many questions about quantum gravity may be inaccessible.
Our bound (\[eq:main\]) is parametrically stronger than existing Lieb-Robinson bounds [@Lashkari_2013; @Bentsen_2019; @guo2019signaling]. We derived it using a more general operator growth formalism developed in [@chen2019operator; @chen2019finite; @lucas2019nonperturbative; @tran2020hierarchy], which relies on the simple relation between OTOCs and operator size at infinite temperature [@nahum_operator_2018; @von_keyserlingk_operator_2018; @Roberts:2018mnp]. The Lieb-Robinson bounds of [@Lashkari_2013; @Bentsen_2019; @guo2019signaling] might be saturated by studying OTOCs prepared in finely tuned initial states (the infinite temperature ensemble measures the value of the correlator in a typical state). Indeed, [@tran2020hierarchy] recently discovered that there are two separate notions of locality that arise in systems with power law interactions; it would not be surprising if a similar phenomenon arose in models with all-to-all interactions.
It is worth keeping in mind that “fast scrambling" is not necessarily the “fastest scrambling" in nature [@Bentsen_2019; @lucas2019star]. It is possible to find models with $N$-independent scrambling times, which are certainly not holographically dual to quantum gravity. In the future, we hope to find further-refined probes of holography and quantum gravity to better discriminate between holographic and non-holographic models with non-local interactions.
Formal discussion
=================
The remainder of this paper consists of the proof of (\[eq:main\]). First, we make precise our assumptions and state a theorem; we conclude with its proof. We study quantum many-body systems consisting of $N$ spin-$\frac{1}{2}$ degrees of freedom. The spins are labeled by vertices $v$ in the set $V$. The Hilbert space $\mathcal{H}$ is (isomorphic to) $(\mathbb{C}^2)^{\otimes N}$. As above, $X_i$, $Y_i$ and $Z_i$ denote the Pauli matrices (normalized as $X_i^2=1$) on spin $i$ ($i\in V$).
Let $\mathcal{B}$ denote the set of Hermitian operators acting on $\mathcal{H}$. It is spanned by products of Pauli matrices on every qubit, along with the identity: $$\mathcal{B} = \bigotimes_{i=1}^N \mathcal{B}_i=\bigotimes_{i=1}^N \lbrace 1, X,Y,Z\rbrace_i. \label{eq:basisB}$$ We denote elements of $\mathcal{B}$ by $|\mathcal{O})$ – these are just like Dirac kets, but with a parentheses to emphasize the vector space is $\mathcal{B}$, not $\mathcal{H}$. The appropriate inner product on $\mathcal{B}$ for studying infinite temperature chaos is $$(A|B) = 2^{-N}\mathrm{tr}(A^\dagger B).$$ The basis vectors of (\[eq:basisB\]) are orthonormal. Time translation on $\mathcal{B}$ is generated by the Liouvillian $$\mathcal{L}(t) = \mathrm{i}[H(t),\cdot]. \label{eq:liouvillian}$$ $\mathcal{L}(t)$ is an antisymmetric linear operator on $\mathcal{B}$, and $$\frac{\mathrm{d}}{\mathrm{d}t} |\mathcal{O}(t)) = \mathcal{L}(t)|\mathcal{O}(t)).$$
Define the projection operation $$\mathbb{P}_i |\mathcal{O}) = |\mathcal{O}) - \frac{1}{2}\left| 1_i \otimes \mathrm{tr}_i \mathcal{O}\right)$$ where $\mathrm{tr}_i$ denotes partial trace over qubit $i$. This operation removes all products of Pauli matrices which include the identity on site $i$. Clearly, the infinite temperature OTOC obeys $$\begin{aligned}
\left|2^{-N}\mathrm{tr}\left([X_i(t),X_j]^2\right)\right| &= \left|2^{-N}\mathrm{tr}\left([\mathbb{P}_jX_i(t),X_j]^2\right)\right| \notag \\
&\le 4 (X_i(t)|\mathbb{P}_j|X_i(t)). \label{eq:pjotoc}\end{aligned}$$ This conclusion generalizes to allow for $X_i$ and $X_j$ to be any linear superposition of Paulis. For any subset $S\subseteq V$ similarly define $\mathbb{P}_S$ to be the projection onto all operators which have at least one non-identity Pauli on at least one vertex $i\in S$.
Let $0<a<1$ be an $N$-independent constant. We *define* the scrambling time as the smallest possible time $t_{\mathrm{s}}>0$ at which the projection in (\[eq:pjotoc\]) is large: $$t_{\mathrm{s}} = \inf_{t\in\mathbb{R}^+} \left\lbrace \sup_{\mathcal{O}_i\in\mathcal{B}_i} \sum_{j=1}^N \frac{(\mathcal{O}_i(t)|\mathbb{P}_j|\mathcal{O}_i(t))}{(\mathcal{O}_i|\mathcal{O}_i)} > aN\right\rbrace. \label{eq:tsdef}$$ Our key conclusions do not depend on $a$. The quantity in the sum above is called the *average operator size* [@nahum_operator_2018; @von_keyserlingk_operator_2018; @Roberts:2018mnp].
Formally, we say that a graph $\Lambda=(V,E)$ exists in $d$ spatial dimensions ($\Lambda$ is $d$-dimensional) when the following properties hold. Let $N=|V|$ be the number of vertices. Pick any vertex $v\in V$. Define $S_D$ to be the set of all vertices that are a distance $\le D$ away from $v$: namely, for any $x\in S_D$ a path of at most $D$ edges exists from $v$ to $x$. We say $\Lambda$ is $d$-dimensional when there exist finite constants $0<c_1,c_2<\infty$ that are independent of $N$, such that for every $v$,
$$\begin{aligned}
|S_D| &\le c_1 D^d, \label{eq:SDbound}\\
|S_D| - |S_{D-1}| &\le c_2 D^{d-1}.\end{aligned}$$
We are now ready to state our main result:
Let $\Lambda=(V,E)$ be a $d$-dimensional lattice graph with $|V|=N$ vertices, such that each vertex in the graph has at most $k$ vertices, with $k$ finite and independent of $N$. Consider quantum dynamics on $\mathcal{H}=(\mathbb{C}^2)^N$ generated by $$\begin{aligned}
H(t) &= \sum_{i,j \in V} \frac{J_{ij}^{ab}(t)}{N^\alpha} X^a_iX^b_j + \sum_{\lbrace i,j\rbrace \in E} K^{ab}_{ij}(t) X^a_i X^b_j \notag \\
&+ \sum_{i\in V} h^a_i(t) X^a_i \label{eq:thmH}\end{aligned}$$ where $|J_{ij}(t)|\le 1$ and $|K^{ab}_{ij}(t)| \le 1$. We employ Einstein summation convention on $a,b$ indices. Then there exists $0<C<\infty$ such that if $\alpha < 1+\frac{1}{d}$, $$t_{\mathrm{s}} > C N^{(2\alpha-1)/(d+2)} . \label{eq:thm}$$ In other words, for any $\alpha>\frac{1}{2}$, the Hamiltonian (\[eq:thmH\]) is not a fast scrambler. For $\alpha \ge 1+\frac{1}{d}$, $t_{\mathrm{s}} > CN^{1/d}$ is not affected by the non-local interactions.
Proof of the Theorem
====================
For notational simplicity, we assume below that $H(t)$ does not depend on time. However, the proof below immediately generalizes to the time-dependent case, which we “leave as an exercise to the reader".
Choose any vertex $v\in V$, $D \in \mathbb{Z}^+$, and let $\bar S_D$ denote the complement of $S_D$. Now let us define
$$\begin{aligned}
H_{<D} &:= \sum_{\lbrace i,j\rbrace \subset S_D} K^{ab}_{ij} X^a_i X^b_j + \sum_{i\in S_D} h^a_i X^a_i, \\
H_D &:= \sum_{i \in S_D, j\notin S_D} K^{ab}_{ij}X^a_i X^b_j, \\
H_{>D} &:= \sum_{\lbrace i,j\rbrace \subset \bar S_D} K^{ab}_{ij} X^a_i X^b_j + \sum_{i\in \bar S_D} h^a_i X^a_i, \\
H_{\mathrm{<NL}} &:= \left(\sum_{i\in S_D, j\notin S_D} + \sum_{\lbrace i,j\rbrace \subset S_D}\right) \frac{J_{ij}^{ab}}{N^\alpha} X^a_iX^b_j, \\
H_{\mathrm{>NL}} &:= \sum_{\lbrace i,j\rbrace \subset \bar S_D} \frac{J_{ij}^{ab}}{N^\alpha} X^a_iX^b_j.\end{aligned}$$
$H_{>D}$ and $H_{\mathrm{>NL}}$ are the terms in the Hamiltonian that do not act on vertices in $S_D$; $H_{<D}$ acts entirely within $S_D$; $H_D$ and $H_{<\mathrm{NL}}$ denote terms which connect $S_D$ and $\bar S_D$. We define $\mathcal{L}_{<D}(t)$, etc., in the obvious way, using (\[eq:liouvillian\]). Note that $$H = H_{<D} + H_D + H_{>D} + H_{\mathrm{<NL}}+H_{\mathrm{>NL}}.$$
Following [@chen2019operator; @chen2019finite; @lucas2019nonperturbative; @tran2020hierarchy], we invoke the Duhamel identity $$\mathrm{e}^{\mathcal{L}t} = \mathrm{e}^{\mathcal{L}_{<D}t} + \int\limits_0^t\mathrm{d}s \; \mathrm{e}^{\mathcal{L}(t-s)} (\mathcal{L}-\mathcal{L}_{<D}). \mathrm{e}^{\mathcal{L}_{<D}s} \label{eq:duhamel}$$ Let $|\mathcal{O}_v)$ denote a (linear combination of) Paulis on vertex $v$ with $(\mathcal{O}_v|\mathcal{O}_v) = 1$. Since $|\mathcal{O}_v(t)) = \mathrm{e}^{\mathcal{L}t}|\mathcal{O}_v)$, we can apply (\[eq:duhamel\]). Now, how much of $|\mathcal{O}_v(t))$ has support in $\bar S_D$? Observe that $$\mathbb{P}_{\bar S_D} |\mathcal{O}_v(t)) = \mathbb{P}_{\bar S_D} \int\limits_0^t\mathrm{d}s \; \mathrm{e}^{\mathcal{L}(t-s)} (\mathcal{L}_D+\mathcal{L}_{\mathrm{<NL}}) \mathrm{e}^{\mathcal{L}_{<D}s} | \mathcal{O}_v).$$ Since $\mathrm{e}^{\mathcal{L}t}$ and $\mathrm{e}^{\mathcal{L}_{<D}t}$ are unitary transformations, they do not change the length of $|\mathcal{O}_v)$ as measured by our inner product. Thus, we use the triangle inequality to obtain $$\begin{aligned}
\left\lVert\mathbb{P}_{\bar S_D} |\mathcal{O}_v(t))\right\rVert_2 &\le \int\limits_0^t \mathrm{d}s \left\lVert \mathcal{L}_D \mathrm{e}^{\mathcal{L}_{<D}s} | \mathcal{O}_v) \right\rVert_2 \notag \\
&\;\; + \int\limits_0^t \mathrm{d}s \left\lVert \mathcal{L}_{\mathrm{<NL}} \mathrm{e}^{\mathcal{L}_{<D}s} | \mathcal{O}_v) \right\rVert_2\label{eq:2terms}\end{aligned}$$ where we have defined $\lVert \mathcal{O}\rVert_2^2 = (\mathcal{O}|\mathcal{O})$. The left hand side bounds the OTOC which controls the scrambling time.
We first bound the top line of (\[eq:2terms\]):
Let $Q_D$ denote the set of vertices exactly distance $D$ from $v$: $Q_D= S_D-S_{D-1}$. Then there exists $0<\mu<\infty$ such that $$\lVert \mathbb{P}_{Q_D} \mathrm{e}^{\mathcal{L}_{<D}t}|\mathcal{O}_v) \rVert_2 \le \mathrm{e}^{\mu t-D }. \label{eq:LR}$$
This is a well-known result [@Lieb1972]; the reader should feel free to skip. Still, we present an elegant proof of this lemma, of interest to specialists, using quantum walks [@lucas2019nonperturbative; @tran2020hierarchy]. In one dimension a slightly improved version of this proof leads to stronger bounds than the provably optimal Lieb-Robinson-style bounds of [@chen2019operator; @wang2019tightening].
For notational convenience, we denote for the proof of this lemma $\mathrm{e}^{\mathcal{L}_{<D}t}|\mathcal{O}) = |\mathcal{O}(t))$. Define $$\mathcal{F} := \sum_{x\in S_D} b^{d_x} \mathbb{P}_x$$ where $d_x$ denotes the distance from $v$ to $x$. Observe that $$\frac{\mathrm{d}}{\mathrm{d}t} (\mathcal{O}_v(t)|\mathcal{F}|\mathcal{O}_v(t)) = (\mathcal{O}_v(t)|[\mathcal{F},\mathcal{L}_{<D}]|\mathcal{O}_v(t))$$ and that it is easy to (crudely) bound the right hand side: denoting $\varphi_x(t) := \lVert \mathbb{P}_x |\mathcal{O}_v(t)) \rVert_2$, we find that (using $|K_{ij}^{ab}|\le 1$) $$\begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t} (\mathcal{O}_v|\mathcal{F}|\mathcal{O}_v) &\le \sum_{\substack{\lbrace x,y\rbrace \in S_D \\ \text{ and } \lbrace x,y\rbrace \in E}} (\mathcal{O}_v| [b^{d_x}\mathbb{P}_x + b^{d_y}\mathbb{P}_y, \mathcal{L}_{xy}]|\mathcal{O}_v) \notag \\
&\le 36 \sum_{\substack{\lbrace x,y\rbrace \in S_D \\ \text{ and } \lbrace x,y\rbrace \in E}} \varphi_x\varphi_y \left(b^{d_x} + b^{d_y}\right) \notag \\
&\le 18 \sum_{\substack{\lbrace x,y\rbrace \in S_D \\ \text{ and } \lbrace x,y\rbrace \in E}} \left(b^{d_x} + b^{d_y}\right)\left(\varphi_x^2 + \varphi_y^2\right) \notag \\
&\le (\mathcal{O}_v|\mathcal{F}|\mathcal{O}_v) \times 18(1+b)k.\end{aligned}$$ In the first line we have defined $\mathcal{L}_{xy} = \mathrm{i}[K^{ab}_{xy}X^a_x X^b_y,\cdot]$; in the third line we have used that $2\varphi_x\varphi_y \le \varphi_x^2+\varphi_y^2$; in the fourth line we have used that only nearest neighbor interactions on $\Lambda$ are allowed by the local terms: if $\lbrace x,y\rbrace \in E$, $|d_x-d_y|\le 1$. Therefore letting $\mu=9(1+b)k$, $$(\mathcal{O}_v(t)|\mathcal{F}|\mathcal{O}_v(t)) \le \mathrm{e}^{2\mu t}. \label{eq:expgrowth}$$
The final observation is that $$(\mathcal{O}_v|\mathbb{P}_{Q_D}|\mathcal{O}_v) b^{D} \le \sum_{x\in Q_D} (\mathcal{O}_v|\mathbb{P}_{x}|\mathcal{O}_v)b^D \le (\mathcal{O}_v|\mathcal{F}|\mathcal{O}_v). \label{eq:markov}$$ Combining (\[eq:expgrowth\]) and (\[eq:markov\]) and setting $b=\mathrm{e}^2$, we obtain (\[eq:LR\]).
This lemma then allows us to crudely (but easily!) bound the first line of (\[eq:2terms\]) as follows: $$\begin{aligned}
\int\limits_0^t \mathrm{d}s \left\lVert \mathcal{L}_D \mathrm{e}^{\mathcal{L}_{<D}s} | \mathcal{O}_v) \right\rVert_2 &\le 2t \lVert H_D\rVert \sup_{s\in [0,t]} \left\lVert \mathbb{P}_{Q_D} \mathrm{e}^{\mathcal{L}_{<D}s}|\mathcal{O}_v)\right\rVert_2 \notag \\
&\le Mt D^{d-1} \mathrm{e}^{\mu t - D}. \label{eq:last1}\end{aligned}$$ where $\lVert H_D\rVert$ denotes the conventional operator norm (in this case, maximal eigenvalue) of $H_D$ and $M$ is an order 1 constant related to the degree of $\Lambda$. Then, the second line of (\[eq:2terms\]) is bounded simply: $$\begin{aligned}
\int\limits_0^t \mathrm{d}s \left\lVert \mathcal{L}_{\mathrm{<NL}} \mathrm{e}^{\mathcal{L}_{<D}s} | \mathcal{O}_v) \right\rVert_2 &\le t \sup_{s\in[0,t]}\lVert [H_{<\mathrm{NL}},\mathcal{O}_v(s)]\rVert_2 \notag \\ &\le 2t \lVert H_{\mathrm{<NL}}\rVert_2 \label{eq:last2}\end{aligned}$$ where we have used the fact that $\mathcal{O}_v(s)$ has maximal eigenvalue 1 to simplify the calculation above. Then we observe that $$\begin{aligned}
\lVert H_{\mathrm{<NL}}\rVert_2^2 &= \left(\sum_{i\in S_D, j\notin S_D} + \sum_{\lbrace i,j\rbrace \in S_D}\right) \frac{(J_{ij}^{ab})^2}{N^{2\alpha}} \lVert X^a_i X^b_j \rVert_2^2 \notag \\
&\le 9|S_D|N^{1-2\alpha}. \label{eq:last3}\end{aligned}$$ Now let us combine (\[eq:last1\]), (\[eq:last2\]) and (\[eq:last3\]), evaluated at a value of $D$ obeying $$D \ge 2 \mu t + D_0. \label{eq:D0}$$where $D_0$ will be chosen below. Using (\[eq:SDbound\]), and when $t$ is large, we conclude $$\begin{aligned}
\left\lVert\mathbb{P}_{\bar S_D} |\mathcal{O}_v(t))\right\rVert_2 &\le \frac{Z^\prime
\sqrt{c_1 \max(2\mu t, D_0)^d }t}{ N^{\alpha-1/2}} \notag \\
&\;\;+ \frac{M^\prime}{2\mu} (2\mu t + D_0)^d \mathrm{e}^{-\mu t-D_0} \label{eq:last4}\end{aligned}$$ for finite constants $Z^\prime$ and $M^\prime$ independent of $N$.
We now choose $D_0$ such that $$\sqrt{\frac{a}{8}} > \mathrm{e}^{-D_0/2} \sup_{t\in\mathbb{R}^+} \frac{M^\prime}{2\mu} (2\mu t + D_0)^d \mathrm{e}^{-\mu t-D_0/2} \label{eq:sqrta2}$$ Note that $D_0$ can be chosen independent of $N$. To understand why we make this choice, we return to our definition of scrambling time. Suppose that we choose $$|S_D|<\frac{aN}{2}. \label{eq:DaN2}$$ At the scrambling time $t=t_{\mathrm{s}}$, by (\[eq:tsdef\]) and (\[eq:DaN2\]), $$\frac{a}{2} < (\mathcal{O}_v(t_{\mathrm{s}})|\mathbb{P}_{\bar S_D} |\mathcal{O}_v(t_{\mathrm{s}})). \label{eq:sqrta}$$ Now, let us assume that at the scrambling time, $t / N^{1/d} \rightarrow 0$. In this case, we can always choose a $D$ compatible with (\[eq:D0\]) and (\[eq:DaN2\]). Combining (\[eq:last4\]), (\[eq:sqrta2\]), and (\[eq:sqrta\]), we obtain $$t_{\mathrm{s}}^{1+d/2} > \sqrt{\frac{a}{8}} \frac{N^{\alpha-\frac{1}{2}}}{Z^\prime \sqrt{c_1(2\mu)^d}}, \label{eq:finalthm}$$ which leads to (\[eq:thm\]) so long as $\alpha<1+\frac{1}{d}$. If instead $\alpha\ge1+\frac{1}{d}$, (\[eq:finalthm\]) implies that $t_{\mathrm{s}}$ scales faster $N^{1/d}$, which violates our assumption that we could choose a $D$ such that $S_D\subset V$ while (\[eq:finalthm\]) holds.
If the assumptions of Theorem 1 hold, but in addition $K^{ab}_{ij}(t)=0$, then for some $0<C<\infty$, $$t_{\mathrm{s}}(N) \ge C N^{\alpha-1/2}. \label{eq:cor3}$$\[cor3\]
This is a simple extension of the proof of the main theorem. If $K^{ab}_{ij}=0$, then in (\[eq:2terms\]) we may consider $D=0$ (i.e. $S_D$ contains only the starting vertex $v$). (\[eq:last4\]) reduces to its first term with $|S_D|=1$. This implies (\[eq:cor3\]).
Tightness of bounds
===================
We conclude by showing that the simplest of our bounds, Corollary \[cor3\], cannot be algebraically improved. While this is probably not surprising due to the existence of other fast scramblers such as [@belyansky2020minimal], we present an illustrative and simple (not many-body chaotic) protocol that saturates (\[eq:cor3\]).
For $N$ sufficiently large, there exists a time-dependent $H(t)$, satisfying the assumptions of Corollary \[cor3\] with $\alpha>\frac{1}{2}$, along with $$J_{ij}^{ab}(t) = J_{ij}(t) \delta^{aZ}\delta^{bZ}$$ such that for any $\epsilon>0$, $$t_{\mathrm{s}}(N) \le K N^{\alpha + \epsilon - 1/2} \label{eq:cor4}$$ where $K$ is a finite constant that can depend on $\epsilon$.
We describe a simple protocol to grow large operators. We choose a $g=\mathrm{O}(1)$ and $M=\mathrm{O}(N)$ (to be specified more carefully later) such that $$N \ge 1 + gM. \label{eq:NgM}$$ Let $R_1,\ldots,R_g$ denote $g$ disjoint sets of vertices with $M$ elements. By (\[eq:NgM\]) there exists another vertex (let’s call it 0) not in any of these sets.
For simplicity, we work in a basis of Pauli matrices $\lbrace 1, X^+, X^-, Z\rbrace$ on every site, where we define $\sqrt{2}X^\pm = X\pm \mathrm{i}Y$ (note the slightly unusual normalization). Our goal is to build a protocol that starts with $X^+_0$ and time evolves it into an operator with average size of order $M$. The protocol will work by first expanding the operator into set $R_1$ using two-body $ZZ$ couplings, then applying a rotation to convert all $Z$ in $R_1$ into $X$, then expanding into $R_2$ using $ZZ$ couplings, and so on. After $l$ steps, the size of the operator will scale as $C^l$ with high probability. To be precise, we say that an operator $\mathcal{O}$ has size $s$ with probability $P_s$, where $$P_s = (\mathcal{O}|\mathbb{Q}_s|\mathcal{O})$$ where $\mathbb{Q}_s$ is a projection superoperator onto Pauli strings with exactly $s$ non-identity components. We will then show how to choose $g$ so that $C^g=M$ and (\[eq:cor4\]) are both obeyed.
Let us now show how to achieve the goals outlined above, starting with the first step of the protocol. Let $$U_{Z,1} = \mathrm{e}^{-\mathrm{i} \frac{\tau}{2} Z_0 Z_{R_1}}$$ where we have defined $$X^\alpha_{R_i} := \sum_{v\in R_i} X^\alpha_v.$$ A straightforward calculation (see e.g. [@lucas2019star]) shows that $$U_{Z,1}^\dagger X^+_0 U_{Z,1} = X^+_0 \mathrm{e}^{\mathrm{i}\tau Z_{R_1}}. \label{eq:UZX0}$$ The probability that $U_{Z,1}^\dagger X^+_0 U_{Z,1}$ has size $1\le s\le M+1$ after step 1 of the protocol is then $$P_{s,1} = \left(\begin{array}{c} M \\ s-1 \end{array}\right) \left(\cos^2\tau\right)^{M-s-1} \left(\sin^2 \tau\right)^{s+1}.$$ The average size of the operator is $$\sum sP_{s,1} = 1 + M\sin^2\tau .$$ Since the distribution is binomial, fluctuations about the mean are of order $\sqrt{M}|\sin\tau|$ and if $\tau$ is sufficiently large, these fluctuations will be small. Let us define an O(1) constants $c_{1,2}$ obeying $c_1<1<c_2$ such that $$p_0 := 2^{-1/2g} < \sum_{s= 1 + s_1^* }^{1+\lceil c_2s_1^* \rceil} P_{s,1}.$$ where $$s_1^* := \lceil c_1 M\sin^2\tau \rceil.$$ Most of the operator has size at least $s_1^*+1$. So, if we define the projection $$\mathbb{R}_1 := \sum_{s= 1 + \lceil c_1 M\sin^2\tau \rceil }^{1+\lceil c_2s_1^* \rceil} \mathbb{Q}_s,$$ then it suffices to keep track of only the operator $\mathbb{R}_1 U^\dagger_Z X^+_0 U_{Z,1}$ . Note also that there exists a finite constant $c_3$ such that the time it takes to run the unitary $U_{Z,1}$ by turning on only Ising couplings as given in (\[eq:thmH\]) is given by $$t_Z = c_3 N^{\alpha}\tau.$$
The next step of the protocol corresponds to rotating the $Z$s in (\[eq:UZX0\]) into $X = (X^+ + X^-)/\sqrt{2}$. We can do this in an O(1) time $t_X$ using $H(t)$ in (\[eq:thmH\]) by applying the unitary transformation $$U_{Y,1} = \mathrm{e}^{-\mathrm{i}\frac{\pi}{4}Y_{R_1}}.$$ Define the final operator to be $$\mathcal{O}_1 = U^\dagger_{Y,1}U^\dagger_{Z,1}X^+_0U_{Z,1}U_{Y,1}$$ After this step, the probability that a Pauli string has $l_\pm$ $X^\pm$ Pauli strings in $R_1$ is given as follows. Let$$j = l_+ - l_-$$ denote the difference between the number of $X^+$ and $X^-$, and let $\mathbb{J}^1_j$ project on to Pauli strings with this imbalance $j$. Then $$(\mathcal{O}_1| \mathbb{Q}_s \mathbb{J}^1_j \mathbb{Q}_s|\mathcal{O}_1) = (\mathcal{O}_1| \mathbb{Q}_s |\mathcal{O}_1) \left(\begin{array}{c} s \\ \frac{1}{2}(s+j) \end{array}\right) \frac{1}{2^s} \label{eq:jbinomial}$$ We define an O(1) constant $c_4$ such that for any value of $s\ge \lceil c_1 M\sin^2\tau \rceil$, $$p_0 < \sum_{j : |j|>c_4 \sqrt{s}} \left(\begin{array}{c} s \\ \frac{1}{2}(s+j) \end{array}\right) \frac{1}{2^s}.$$ Define the projector $\mathbb{K}_1$ onto all Pauli strings with $s$ $X^\pm$ in $R_1$, such that $s>\lceil c_1 M\sin^2\tau \rceil$ and with $|j|>c_4 \sqrt{s}$. Then clearly, $$\lVert \mathbb{K}_1 |\mathcal{O}_1) \rVert_2 \ge p_0^2.$$
For the remaining steps of the protocol, we choose the unitaries
$$\begin{aligned}
U_{Z,l} &= \mathrm{e}^{-\mathrm{i}\frac{\tau}{2} Z_{R_{l-1}}Z_{R_l}}, \\
U_{Y,l} &= \mathrm{e}^{-\mathrm{i}\frac{\pi}{4} Y_{R_l}}.
\end{aligned}$$
We define $$\mathcal{O}_l := U^\dagger_{Y,l}U^\dagger_{Z,l}\mathcal{O}_{l-1}U_{Z,l}U_{Y,l}.$$ The key observation is that $\mathcal{O}_l$ consists entirely of Pauli strings of identity, $X^+$ and $X^-$, *and* that every single Pauli string in $\mathcal{O}_{l-1}$, in our $\pm$ basis, evolves in an orthogonal subspace of operator space $\mathcal{B}$ relative to every other string during steps $l,l+1,\ldots g$. Therefore, we can easily recursively analyze the operator growth after every step of the protocol.
For example, when $l=2$, we can analyze the evolution of the operator $\mathbb{J}^1_j\mathbb{Q}_s \mathcal{O}_1$ separately for each $j$ and each $s$. Upon doing so, we find that the probability for having size $s_2$ in domain $R_2$, with imbalance $j_2$ between $X^+$ and $X^-$ in $R_2$, is given by $$P_{s_2,2}(j) = \left(\begin{array}{c} M \\ s_2 \end{array}\right) \left(\cos^2(j\tau)\right)^{M-s_2} \left(\sin^2 (j\tau)\right)^{s_2}$$ The answer only depends on $j$, since for any operator $A$ $$[Z_{R_1}, \mathbb{J}^1_j A] = 2\mathrm{i}j\mathbb{J}^1_j A.$$ Now, we define the projector $\mathbb{R}_2$ onto all operators with at least $s_2^*$ $Z$s in $R_2$ (and not more than $c_2 s_2^*$), where $$s_2^* := M\sin^2\left(c_4 \sqrt{s_1^*} \tau\right) > \frac{4c_4^2 }{\pi^2} \left(s_1^*\right)^2.$$ In deriving this formula, it was important that we could ignore zeros of $\sin^2(j\tau)$ away from $\tau=0$; we will confirm at the end of the proof that this is so. The distribution of $j_2$ after applying $U_{Y,2}$ is given by the binomial formula, similar to (\[eq:jbinomial\]). A straightforward generalization of the logic at step 1 tells us that $$\lVert \mathbb{K}_2|\mathcal{O}_2)\rVert_2 \ge p_0^4$$ where $\mathbb{K}_2$ projects onto all Pauli strings with at least $s_2^*$ $X^\pm$ in $R_2$ and with imbalance of at least $c_4 \sqrt{s_2^*}$.
Clearly this procedure extends to all $l$. The minimal size after each step is $$s_l^* > \left(\frac{4c_4^2 }{\pi^2}\right)^{l-1} \left(s_1^*\right)^l.$$ After $g$ steps, we conclude that $$\begin{aligned}
\sum_{i=1}^N (\mathcal{O}_g|\mathbb{P}_i|\mathcal{O}_g) &\ge \sum_{i=1}^N (\mathcal{O}_g|\mathbb{K}_g \mathbb{P}_i \mathbb{K}_g |\mathcal{O}_g) \ge s_g^* (\mathcal{O}_g | \mathbb{K}_g|\mathcal{O}_g) \notag \\
&\ge s_g^* p_0^{2g}= \frac{s_g^*}{2}.
\end{aligned}$$ The total runtime of the protocol is $$t_{\mathrm{s}} = g(t_Z+t_X).$$
It remains to fix $g$ and $\tau$ such that (\[eq:cor4\]) holds. Since $t_X$ is O(1) it suffices to choose $g$ and $\tau$ such that for some $0<c_5<\infty$
\[eq:cor4end\]$$\begin{aligned}
gc_3N^\alpha\tau &\le KN^{\alpha+\epsilon-1/2}, \\
c_5 M &\le s_g^*.
\end{aligned}$$
These inequalities hold if$$s_1^* = \frac{\pi^2}{4c_4^2} \left(\frac{4c_4^2c_5N}{\pi^2 g}\right)^{1/g} > \frac{4}{\pi^2} c_1\frac{N}{g}\tau^2 \label{eq:s1starend}$$ or, for a suitable $0<c_6<\infty$ that depends on $g$, $$\tau = c_6 N^{-(g-1)/2g}. \label{eq:tauchoice}$$ Choosing $g>1/2\epsilon$, we satisfy (\[eq:cor4end\]).
The final thing to confirm is that the “imbalance" of $X^+$ and $X^-$ is always so small that $j\tau \ll 1$ holds, except at the final step. Since $j\sim \sqrt{s}$ with high probability, (\[eq:s1starend\]) and (\[eq:tauchoice\]) confirm that this is so. In the final step, due to the binomial distribution of $j$, there will be negligible concentration around $j\tau/\pi \in \mathbb{Z}$, and the probability that the operator has size $\mathrm{O}(M)$ is finite.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Alexey Gorshkov, Ana Maria Rey and Brian Swingle for useful discussions.
| {
"pile_set_name": "ArXiv"
} |
0.2cm
[**Local Electronic Structure of PbVO$_{3}$, a New Member of PbTiO$_{3}$ Family, studied by XANES/ELNES** ]{}
Sher Alam$^{1}$, Alexei A. Belik$^{2}$, Y. Matsui$^{1}$
$^{1}$[*CSAAG, NIMS, Tsukuba 305-0044, Ibaraki, Japan*]{}\
$^{2}$[*ICYS, NIMS, Tsukuba 305-0044, Ibaraki, Japan*]{}\
Abstract
Recently an interesting multi-ferroic system PbVO$_{3}$ \[Chem. Mater. 2004\] was successfully prepared using a high-pressure and high-temperature technique. The crystallographic features were reported. In this note we concentrate on the theoretical XANES spectra by considering the K-edge of Vanadium. The tetragonality \[c/a=1.229 at 300 K\] of PbVO$_{3}$ is the largest in the PbTiO$_{3}$ family of compounds. Thus one is led naturally to examine the effect of the change of tetragonality and the axial oxygen position on the electronic structure \[i.e. XANES spectrum\]. We study this effect in two ways. At a given temperature we vary the tetragonality and the axial oxygen position and quantify it in terms of XANES difference spectrum. Secondly, we compute the XANES spectra at three different temperatures, 90 K, 300 K, and 530 K and quantify the change in terms of the difference spectrum. We note that in this compound the tetragonality increases almost monotonically with temperature from 12 K to 570 K without transition to the cubic phase under ambient pressure. A key objective of the current investigation is to gain an understanding of various absorption features in the vicinity of K-edge of V, in terms of valence, local site symmetry, local coordination geometry, local bond distances, charge transfer, and local projected density of states. We consider both the polarized and the unpolarized XANES spectra. In short we have performed a local electronic study, which nicely complements the crystallographic features reported recently in PbVO$_3$.
PACS numbers: 78.20.-e, 78.30.-j, 74.76.Bz Key words: XANES, ELNES, SCF, FEFF8, Tetragonality
Introduction
============
Using high-pressure and high-temperature techniques, PbVO$_{3}$ has been synthesized by Belik et al. [@ale04], and Shpanchenko et al. [@shp04].
One of the most interesting feature of PbVO$_{3}$ is its large tetragonality \[c/a=1.229 at 300 K\] in comparison to that of PbTiO$_{3}$ \[c/a=1.064 at 300 K\]. The large tetragonal distortion implies that it will have also a sizable polarization, P$_{s}$. Indeed a rough calculation [@ale04] employing the ionic model gives a value close to 100 $\mu$C/cm$^{2}$. This is to be compared to P$_{s}$=81 $\mu$C/cm$^{2}$ at 300 K for PbTiO$_{3}$, P$_{s}$=26 $\mu$C/cm$^{2}$ at 300 K for BaTiO$_{3}$, and P$_{s}$=37 $\mu$C/cm$^{2}$ at 543 K for KNbO$_{3}$, in the same family of compounds. It is well-known and reasonable to expect that [*ferroelectricity*]{} in ionic crystals is correlated with crystal structure distortion. The basic reason is simple, polarization is directly caused by the atomic displacement. A very well-known example is PbTiO$_{3}$, a perovskite-type tetragonal crystal structure, space group P4mm at 300 K and a simple cubic structure above 763 K. More interesting is the phenomenon, when a material exhibits two or all three of the properties of (anti)ferroelectricity, (anti)ferromagnetism, and (anti)ferroelasticity. These materials are called multiferroics, a recent example is BiFeO$_{3}$ which exhibits a PbTiO$_{3}$-type structure. Other examples are YMnO$_{3}$, BiMnO$_{3}$, TbMnO$_{3}$, and TbMn$_{2}$O$_{5}$. It is clear that these materials contain a magnetic transition metal ion \[such as Mn, Fe\] in conjunction with Bi$^{3+}$ and Pb$^{2+}$.
It is well-known that as the region close to the x-ray near edge is scanned in energy, the ejected photolectron sequentially probes the unoccupied electronic levels of materials. This results in a fine x-ray absorption near edge structure \[XANES\], within roughly 30-50 eV of the threshold. XANES thus contains useful chemical and structural information.
The main point is that this region is dominated by [*strong*]{} photoelectron scattering. Thus it is highly non-trivial to develop a code to take into account the multiple scattering which dominate the XANES region. Perhaps one of the most elegant XANES code is the FEFF8 series[@ank98; @ank02]. This [*ab initio*]{} based on the [*real-space multiple scattering*]{} \[RSMS\] and one of its several advantages is that it applies to both [*periodic*]{} and [*aperiodic*]{} systems.
Thus the purpose of this note is to concentrate on the theoretical calculation using the updated version of FEFF8, i.e. FEFF8.20 [@ank02], for the PbVO$_{3}$, using the crystallographic data reported recently by one of us [@ale04]. In particular we calculate the near edge spectra of this system for several cases of interest. In addition we also study the near-edge structure for the in-plane and out of plane cases.
Here we report on our results of PbVO$_{3}$, for the P4mm phase. The calculated XANES spectra are remarkably self-consistent. We have performed several cross-checks in addition to the Self-Consistent-Field \[SCF\] cycles used in FEFF8 to obtain high level accuracy. It is to be noted that the calculations are based on SCF one-electron Green’s function approach, where many-body effects are taken into account in terms of final-state potentials and a complex energy-dependent self-energy. In this scattering theoretic approach the structure in XANES is correlated with projected density of states \[pDOS\].
Another aim is to provide a concrete example of the application of XANES and DOS calculations using FEFF. Indeed, although FEFF is a powerful code, its self-consistency must be demonstrated by applications to real systems systematically. Moreover by actual calculations one can illustrate advantages and disadvantages, and thus find ways of improving the code.
This paper is organized as follows, in the next section we outline some basic points about the multiple scattering calculations based on the FEFF8 code. In section three the results and discussion of our study of PbVO$_{3}$ are given using the FEFF8 code. The final section contains the conclusions.
Multiple Scattering via FEFF8
=============================
It is well-known that XANES refers to roughly the 40-50 eV region near the edge of X-ray absorption spectroscopy \[XAS\]. The main point is that this region is dominated by [*strong*]{} photoelectron scattering. Thus it is highly non-trivial to develop a code to take into account the multiple scattering which dominate the XANES region. Perhaps one of the most elegant XANES code is the FEFF8 series[@ank98; @ank02]. This [*ab initio*]{} based on the [*real-space multiple scattering*]{} \[RSMS\] and one of its several advantages is that it applies to both [*periodic*]{} and [*aperiodic*]{} systems. In this approach the self-consistent field \[SCF\] calculations for both the local electronic structure and x-ray absorption spectra are implemented. It is important to note that the full-multiple scattering \[FMS\] is taken into account for a “small cluster” of atoms plus the higher-order multiple scattering from scatterers outside the said cluster. Some of its main advantages are: SCF estimate of Fermi energy, orbital occupancy and charge transfer.
In this note we concentrate mainly on giving the $\mu(E)$ and the difference $\mu(E)$ spectra. The value $\mu(E)$ is the main quantity in XANES. Here we look at the basic definitions restricting ourselves to brief comments. As is well-known the primary quantities with which XANES calculations are concerned are $\mu$,$\mu_{0}$, $\chi$, and $\rho_{li}(E)$[@ank98] $$\begin{aligned}
\mu_{li}(E)=\mu_{li}^{0'}(E)[1+\chi_{li}^{'}(E)].
\label{f1}\end{aligned}$$ We note that the prime is here for clarity since it reminds us that it denotes final state quantities in the [*in the presence of screened hole*]{}. The central quantity in X-ray Absorption Spectroscopy \[XAS\] is the absorption coefficient $\mu(E)$. As is known there is a close and deep connection between XAS and [*electronic structure*]{}, which is indicated and implied by the resemblance of the contribution from a site, $i$ and orbital angular momentum $l$ and the local $l$-projected electronic density of states \[LDOS\] at site $i$ $$\begin{aligned}
\rho_{li}(E)=\rho_{li}^{0}(E)[1+\chi_{li}(E)].
\label{f2}\end{aligned}$$ We provide, a concrete example of this inter-relationship between absorption and LDOS in Fig. \[fig4\]. It is clear that the main features or peaks of the two quantities, absorption and LDOS strongly resemble each other, as they should. Incidentally, this provides an independent check on our calculation.
However, it is important to bear in mind that the since core hole plays a significant role in calculation of XAS, the similarity between XAS and LDOS cannot be regarded as an absolute.
It is important to keep in mind that FEFF method starts from the most fundamental quantity i.e. the Real Space Green’s Function and constructs the physical quantities of interest from it [@ank98; @ank02]. This is one of the code’s main attraction since unlike band calculations it does not depend on symmetry. In this sense it is ideal for cluster physics[^1]. In the shorthand notation the MS expansion can be written as [@ank98] $$\begin{aligned}
G^{SC}=G^{0}tG^{0}+G^{0}tG^{0}tG^{0}+G^{0}tG^{0}tG^{0}tG^{0}+...,
\label{f3}\end{aligned}$$ where $G^{SC}$ represents the self-consistent Green’s Function and t the scattering t matrix. It is clear that in FMS we can implicitly sum Eq.\[f3\] to all orders by using matrix inversion, i.e. $$\begin{aligned}
G^{SC}=G^{0}(1-tG^{0})^{-1},
\label{f4}\end{aligned}$$ as it follows simply from the form of Eq.\[f3\].
Results and Discussion
======================
Let us now give the results and analysis of our calculated theoretical results using FEFF8.20.
The major input we use is the crystallographic and positional information refinements recently given by one of us [@ale04]. This information is available at 90 K, 300 K and 570 K and for convenience is listed in Table I.
For the purposes of this note we consider the XANES results, where we consider all the atoms upto the distance of 7 Å, from the central or core atom, i.e. V. The exact enumeration is given in Table II for the case, when the temperature is 90 K. The distances from V are calculated with Atoms \[version 3.0beta10\] program [@new01]. A total of 110 atoms including the central atom are included in the current XANES calculation, Table II. The shell and atom numeration for temperatures 300 K and 530 K are respectively given in Tables III and IV. From Tables II-IV, several points are immediately clear. We can clearly see that V atom is five-fold coordinated, with axial oxygen atoms asymmetrically situated. In other words, there is a very strong octahedral distortion with one very short Vanadyl V-O distance. Thus the axial oxygen distances in PbVO$_{3}$ are much further apart compared to PbTiO$_{3}$. This information is summarized as a comparison of PbVO$_{3}$ and PbTiO$_{3}$ in Table V. The oxidation state of the five-fold coordinated V can be readily be estimated by using the relation $Z=25.99-11.11 l_{V-OP}$[@sch00] and the value of $V-OP=1.98583$ Åfrom Table V, this gives Z=3.9274. Thus V has an average valence of approximately 4+. [*We expect this strong five-fold coordination to effect the electronic properties*]{}, and this is shown to be the case as we show below.
The main input file for FEFF8.2 is generated by Atoms and analysis is done with Athena[@new01] \[version 0.8.037\].
Typical XANES spectra for the three different temperatures 90 K, 300 K and 530 K are given, respectively in Figs. \[fig1\]-\[fig3\]. These are labelled by $\mu_{u}$ for the unpolarized, by $\mu_{ab}$ for the polarized case representing the ab-plane and by $\mu_{c}$, when the E vector is parallel to the c-axis.
In Fig. \[fig4\] the XANES spectrum is compared with the corresponding projected density of states. The agreement is quite good, showing the correctness of our calculations.
In Fig. \[fig5\] the XANES spectra for the two cases, I and II are given using the 90 K input data. Intuitively, one expects that if the axial oxygen position is changed, one would see the consequence of this change in the local electronic structure, i.e. the XANES spectra. In the example given in Fig. \[fig5\] we took the apical oxygen fractional z-coordinate as $\delta z_{_{V-OA}}=0.1612$ and labelled this as case I. For comparison we note that the corresponding values of axial oxygen are $\delta z_{_{V-OA}}=0.1138$ at room temperature for PbTiO$_{3}$ and $\delta z_{_{V-OA}}=0.2087$ and $\delta z_{_{V-OA}}=0.2102$, respectively at 90 K and 300 K for PbVO$_{3}$, Table I. Thus, for illustration we took an intermediate value between that of PbTiO$_{3}$ and PbVO$_{3}$ for this quantity, to see how it would effect the electronic structure. Incidentally, one would also expect roughly this value for the fractional z-coordinate for the system PbV$_{0.5}$Ti$_{0.5}$O$_{3}$, which has not been fabricated yet. It would be useful to fabricate PbV$_{0.5}$Ti$_{0.5}$O$_{3}$, since it could provide further insight between distortion and multi-ferroic behaviour in these and related systems.
We also expect that varying the tetragonality \[c/a\] at a given temperature will also effect the XANES spectrum. Indeed this turns out to be the case. This case is labelled as case II. Once again, for demonstration we took an intermediate between that of PbTiO$_{3}$ \[a=b=3.905 Å ,c=4.156 Å , c/a=1.06428 at 300 K\] and PbVO$_{3}$ \[a=b=3.8033 Å ,c=4.6499 Å , c/a=1.2226 at 90 K\]. For case II we chose a=b=3.905 Å ,c=4.4700 Å , c/a=1.1447 at 90 K.
The positional information \[in energy units eV\] of the features or peaks for Figs. \[fig1\]-\[fig3\] and \[fig5\] are given in Table VI.
In order to quantify the change in XANES spectra, as a result of the temperature change , varying the axial oxygen position and the tetragonality at a given temperature, we calculate the difference spectra of the cases with respect to the spectra at 90 K.
The results for difference spectra for the temperatures 300 K and 500 K data with respect to 90 K data are shown in Fig. \[fig6\].
The difference spectra which result when the axial oxygen position and tetragonality is varied are illustrated in Fig. \[fig7\].
Let us consider first some general remarks as a guide to the reader on Vanadium K-edge [@wong84] in the context of Vanadium oxides, in order to motivate the spectra shown in Figs. \[fig1\]-\[fig3\] and \[fig5\]. It is known that V forms a series of oxides VO , V$_{2}$ O$_{3}$, V$_{4}$O$_{7}$, V$_{2}$O$_{4}$ and V$_{2}$O$_{5}$ over a range of formal oxidation states. In V K-edge XANES/ELNES in these oxides, one observes a pre-edge absorption feature which is strongest in V$_{2}$O$_{5}$, followed by weak feature/shoulder on a rising absorption edge, which culminates in a strong peak approximately 20 eV from the Fermi energy. It is known that the strong peak is assigned to dipole-allowed transition 1s $\longrightarrow$ 4p, the weak shoulder/feature as a shakedown transition representing the 1s $\longrightarrow$ 4p and the pre-edge or near-edge feature near the threshold as a forbidden 1s $\longrightarrow$ 3d. The features/peaks arising at higher energies than the main 1s $\longrightarrow$ 4p transition are more difficult to be specific about. These features may arise from transition to higher np states, multiple scattering and/or shape resonances.
From Figs. \[fig1\]-\[fig3\] and \[fig5\] we can clearly see five features or peaks. The first peak P1 can be referred to as the edge peak, since it just above the Fermi energy. This feature is present in all the cases. However, there are some interesting points to note, regarding this peak. First of all, the dependence of this peak on the distortion of the local structure around V atom, is evident from our calculation of the different cases. It does not disappear with temperature, since the distortion is present at higher temperatures, in contrast to what would happen in case of centrosymmetric structures. The position of P1 also remains close to approximately 5469.1 eV, Table VI, for all the cases considered here, with the exception of case II, where it is slightly differently located at roughly 5469.4 eV. In order to show and quantify that the distortion and ultimately multi-ferrocity is caused by the asymmetric positions of the axial oxygens, we calculated the polarized XANES data along the c-axis and ab-plane. It is clear from Figs. \[fig1\]-\[fig3\] that the main contribution to distortion comes from the c-axis spectra. Thus the unpolarized data $\mu_{u}$, which is an average of all the polarizations, receives most of the contribution for P1 from the c-axis. In contrast the ab-plane spectra exhibits P1 to a much lessor degree. These points are cleared further by examining P1 in the 530 K spectra, here P1 is quite weak \[WF\] for the ab-plane data, which effects the polarization average data, so that after averaging P1 appears as a weak \[W\] feature in the unpolarized spectra, in contrast P1 remains a strong feature in the c-axis spectra.
The strong dependence of the XANES/ELNES on the local distortion is clear not just from P1, but the entire spectrum. For example, P4 \[the dominant post edge peak at roughly 5487 eV, is nearly absent for the most non-centrosymmetric case, i.e. the c-axis spectra, Table VI and Figs. \[fig1\]-\[fig3\], where we have labelled it as very weak feature \[VWF\]. P4 would be very large for centrosymmetric case. In short, it anti-correlates with distortion, in contrast to P1 which correlates with deviation from centrosymmetry. On the basis of our data, we can also conclude that P5 also correlates with distortion. The ”shakedown” peaks P2 and P3 are quite robust and only P2 slightly weakens at 530 K, Table VI.
Yet another interesting result of our calculation is that P1 arises, due to the hybridization between the p and d orbitals of the absorber, an effect which is forbidden for centrosymmetric structure but increased by local distortions. This effect is known for PbTiO$_{3}$ [@ank98], but for PbVO$_{3}$ it seems to be very pronounced. This is expected due the much bigger tetragonality of the PbVO$_{3}$ compared to PbTiO$_{3}$. But our calculations explicitly demonstrate it and further quantifies it in terms of the electronic near edge structure.
The results for the two cases I and II, are displayed in Fig. \[fig5\], the five peaks/features are apparent again. The exact positions of the features are listed in Table VI. It is interesting to note that for case I the peak P4 is dominant compared to P1, whereas it is reversed in case II, Fig. \[fig5\]. Thus once again we see the anti-correlation of the two peaks P1 and P4, as discussed above. Once again the ”shakedown” peaks P2 and P3 are present.
In order to further quantify the near edge electronic structure we have taken the difference spectra, of all the unpolarized cases discussed here, relative to the 90 K spectra. The results are displayed in Figs. \[fig6\]-\[fig7\]. The percentage change in tetragonality in going from 90 K to 300 K and 530 K, Table I, is approximately 0.6% and 0.9% respectively, with respect to the former. This translates into approximately 4% for the 300 K case and about 10.5% for the 530 K, as can be seen from the difference spectra, Fig. \[fig6\], by considering the maximum peak to peak variation. This clearly shows that we can correlate small changes in tetragonality to the local electronic behaviour. Let us turn to Fig. \[fig7\], where for case II, the tetragonality was chosen as c/a=1.1447, which is 6.4 % smaller than that of PbVO$_{3}$ relative to the latter, at the same temperature of 90 K. The change in the corresponding difference spectra is large, as expected, average variation being on the order of 20%. This shows that by taking tetragonality roughly between that of PbTiO$_{3}$ and that of PbVO$_{3}$, we can account for the large difference between their electric polarizations, which is incidentally also on the order of 20 %, albeit at different temperature of 300 K. For case I, where we attempted to simulate a behaviour between PbTiO$_{3}$ and PbVO$_{3}$, by taking the z-coordinate of the axial oxygen roughly between these two systems, keeping the tetragonality at the same value as PbVO$_{3}$, we see an average percentage change of 30 % relative to the PbVO$_{3}$ at the same temperature. In particular variation is large in region of peak P1, as expected in lieu of earlier comments, since the more the deviation from perfect octahedral symmetry, the bigger the effect, clearly the PbVO$_{3}$ at 90 K has $\delta z_{_{V-OA}}=0.2087$ to be compared with $\delta z_{_{V-OA}}=0.1612$ for case I, at the same temperature. This represents a percentage change of roughly 23 % in the $\delta z_{_{V-OA}}$ value relative to the measured PbVO$_{3}$ at 90 K. In order to see how this effects the difference spectra, we can examine the P1 region in Fig. \[fig7\] for case I. Incidentally the change is around 23 %, which is consistent with percentage change of $\delta z_{_{V-OA}}$.
The charge transfer and occupational orbital numbers are given in Tables VII and VIII. These results can be used to analyze the varying influence of the oxygen environment on the electronic structure of different transition-metal atoms in general, and in particular in the perovskite structure of our interest. Here we only concentrate on a comparison between PbVO$_{3}$ and PbTiO$_{3}$, leaving a more detailed comparison between several related materials for future work. The case of PbTiO$_{3}$ has been calculated in some detail for the purposes of comparison, since it possesses the largest tetragonality among the previous known members of this family. We see that the charge transfer for the oxygen atom in the PbVO$_{3}$ changes by approximately by 28.9% relative to the PbTiO$_{3}$ case at 300 K. For the Pb atom one see a change of roughly 16.3 % compared to that in the PbTiO$_{3}$ system. We note that for each case, the net charge transfer cancels to within $\pm 0.001$, which provides a useful and a simple test for our calculations.
The exact location of all the peaks is displayed in Table IX. Both the positive and negative features are presented for obvious reasons. We can immediately see that difference spectra shows the largest variations, and peaks roughly at the positions P1, P2, P3, P4 and P5, discussed above. This is not surprising, but clarifies the differential behaviour of each peak region for the unpolarized spectra given in Figs. \[fig1\]-\[fig3\] and \[fig5\].
Conclusions
===========
We have for the first time given the XANES study of the newly fabricated PbVO$_{3}$.
- The spectra for three important cases and the difference spectra have been calculated and presented here. It is demonstrated that electronic structure shows a calculable change in the temperature range 90 - 530 K. This is expected since the c-lattice parameter increases almost linearly with temperature, whereas the a and b lattice constants remain unchanged.
- Roughly speaking all the spectra show five peaks or features in the XANES region.
- We found that by taking polarization into account, we can clearly see the origin and the dependence of peaks and features and their sensitivities, on ab-plane or c-axis. This also allows us to trace back the origin of peaks/features in the unpolarized spectra to the ab-plane and/or c-axis.
- Charge transfers and electronic orbital occupation numbers have been calculated for several cases of interest. We have also given a comparison of charge transfer and orbital occupation numbers between PbVO$_{3}$ and PbTiO$_{3}$. For example, we find a change of approximately 28.9% in the charge transfer of the oxygen atom in going from the PbVO$_{3}$ to the PbTiO$_{3}$ with respected to the latter, for the 300 K data.
- The difference spectra confirms the previous observations and results, and further correlates the percentage changes in tetragonality and $\delta z_{_{V-OA}}$ directly to the variations observed in the difference spectra. We can also account roughly for the change in electric polarization between PbVO$_{3}$ and PbTiO$_{3}$, by looking at how changes in tetragonality and axial-oxygen deviation from ideal position correlate to the variation in the near edge electronic structure. The difference spectra allows us approximately to translate these into percentage changes.
From these calculation and analysis we can conclude that is possible to quantify the electronic structure of PbVO$_{3}$ by using the near edge structure.
Acknowledgments {#acknowledgments .unnumbered}
===============
The Sher Alam’s work is supported by Y.Matsui’s Crystal Structure and Analysis Group at the Advanced Material Laboratory \[NIMS\] and MONBUSHO via the JSPS invitation program.
[99]{} Alexei A. Belik et al.,Chem. Mater. [**17**]{}, (2005) 269-273. R.V.Shpanchenko et al.,Chem. Mater. [**16**]{}, (2004) 3267. A. L. Ankudinov et al., Phys. Rev. [**B58**]{}, (1998) 7565. A. L. Ankudinov et al., Phys. Rev. [**B65**]{}, (2002) 104107. Matt. Newille et al., J. Synchrotron Radiation [**8**]{}, (2001) 322. W. Schindler et al.,Chem. Mater. [**12**]{}, (2000) 1248. J. Wong, PRB. [**30**]{}, (1984) 5596.
Table I: Crystallographic Data at 90 K, 300 K and 530 K for PbVO$_{3}$.\
\
[|l|c|c|c|c|c|c|c|r|]{}Temperature & Atom & a Å& c Å& $\frac{c}{a}$ & x & y & z\
90 K & Pb & 3.80329(6) & 4.64989(10)& 1.2226 &0 &0 &0\
& V & & & &1/2 &1/2 &0.5677(3)\
& O1 & & & &1/2 &1/2 &0.2087(15)\
& O2 & & & &1/2 &0 &0.6919(9)\
300 K & Pb & 3.80391(5)& 4.67680(8) &1.2295 &0 &0 &0\
& V & & & &1/2 &1/2 &0.5668(4)\
& O1 & & & &1/2 &1/2 &0.2102(16)\
& O2 & & & &1/2 &0 &0.6889(10)\
530 K & Pb & 3.80721(5) & 4.69819(9) &1.2340& 0 &0 &0\
& V & & & &1/2 &1/2 &0.5674(4)\
& O1 & & & &1/2 &1/2 &0.2102(15)\
& O2 & & & &1/2 &0 &0.6915(10)\
\
\
\
Table II: Shell and atom enumeration in the P4mm, with V as a core atom at 90 K.\
\
[|l|c|c|r|]{}Shell & Atom & No.of Atoms & distance from V\
1 & O1 & 1 & 1.66931\
2 & O2 & 4 & 1.98741\
3 & O1 & 1 & 2.98058\
4 & Pb & 4,4 & 3.35757,3.76839\
5 & V & 4 & 3.80329\
6 & O1 & 4 & 4.15351\
7 & O2 & 8,4 & 4.29125,4.49449\
8 & V & 2 & 4.64989\
9 & O1 & 4 & 4.83207\
10& V & 4 & 5.37866\
11 & O2 & 4 & 5.56256\
12 & O1 & 4 & 5.63175\
13 & O2 & 4,8 & 5.73409,5.88774\
14 & V & 8 & 6.00720\
15 & O1 & 4,1 & 6.14930,6.31920\
16 & Pb & 8,8 & 6.34060,6.56740\
17 & O2 & 8,8 & 6.73848,6.88075\
\
\
\
Table III: Shell and atom enumeration in the P4mm, with V as a core atom at 300 K.\
\
[|l|c|c|r|]{}Shell & Atom & No.of Atoms & distance from V\
1 & O1 & 1 & 1.66775\
2 & O2 & 4 & 1.98583\
3 & O1 & 1 & 3.00905\
4 & Pb & 4,4 & 3.36742,3.77647\
5 & V & 4 & 3.80391\
6 & O1 & 4 & 4.15345\
7 & O2 & 8,4 & 4.29107,4.52490\
8 & V & 2 & 4.67680\
9 & O1 & 4 & 4.85017\
10& V & 4 & 5.37954\
11 & O2 & 4 & 5.58187\
12 & O1 & 4 & 5.63213\
13 & O2 & 4,8 & 5.73437,5.91138\
14 & V & 8 & 6.02845\
15 & O1 & 4,1 & 6.16391,6.34455\
16 & Pb & 8,8 & 6.34658,6.57276\
17 & O2 & 8,8 & 6.75478,6.88133\
\
Table IV: Shell and atom enumeration in the P4mm, with V as a core atom at 530 K.\
\
[|l|c|c|r|]{}Shell & Atom & No.of Atoms & distance from V\
1 & O1 & 1 & 1.67819\
2 & O2 & 4 & 1.99089\
3 & O1 & 1 & 3.02000\
4 & Pb & 4,4 & 3.37316,3.78862\
5 & V & 4 & 3.80721\
6 & O1 & 4 & 4.16067\
7 & O2 & 8,4 & 4.29633,4.53410\
8 & V & 2 & 4.69819\
9 & O1 & 4 & 4.85955\
10& V & 4 & 5.38421\
11 & O2 & 4 & 5.61384\
12 & O1 & 4 & 5.63968\
13 & O2 & 4,8 & 5.74051,5.92055\
14 & V & 8 & 6.04713\
15 & O1 & 4 & 6.17334\
16 & Pb & 8 & 6.35358\
17 & O1 & 1 & 6.37638\
18 & Pb & 8 & 6.58357\
19 & O2 & 8,8 & 6.78307,6.88827\
\
\
\
Table V: Typical oxygen distances in PbVO$_{3}$ versus PbTiO$_{3}$ at 300 K, from Pb and M \[V,Ti\] atoms. The oxygen distances for the two cases I and II at 90 K considered here are also given, in addition to the cases of PbVO$_{3}$ for 90 K and 530 K. All distances are in Å units.\
\
[|l|c|c|c||c|c|c|r|]{}Material &Pb-OA & Pb-OP &M-OA\[short\] Å& M-OP Å& M-OA\[long\] Å& Z\
PbVO$_{3}$ &2.86378 & 2.38614 & 1.66775 & 1.98583 & 3.00905 & 3.9274\
PbTiO$_{3}$ &2.80146 & 2.51937 & 1.76713 & 1.97916 & 2.38887 & 4.0015\
Case I &2.79184 &2.38090&1.89018 & 1.98741 & 2.75972 & 3.9099\
Case II &2.91458 &2.38934&1.60473 & 2.02989 & 2.86527 & 3.4379\
PbVO$_{3}$ \[90 K\]&2.85907 &2.38090&1.66931 & 1.98741 & 2.98059 & 3.9099\
PbVO$_{3}$ \[530 K\]&2.86752 &2.39258&1.67820 & 1.99089 & 3.0200 & 3.8712\
\
\
Table VI: Features/Peaks for Figs. \[fig1\]-\[fig3\] and Fig. \[fig5\]. All the positions are given in eV.\
[|l|c|c|c|c|r|]{}Case & P1 & P2 & P3 & P4 & P5\
90 K,$\mu_{u}$ & 5469.087 & 5475.655 &5480.532 & 5487.475 & 5493.943\
& & 5477.341 & & &\
90 K,$\mu_{ab}$ & 5469.087 & 5475.655 &5480.532 & 5487.475 & VWF\
90 K,$\mu_{c}$ & 5469.087 & 5475.655 &5480.532 & VWF & 5493.943\
300 K,$\mu_{u}$ & 5469.084 & 5475.652 &5480.529 & 5487.472 & VWF\
& & 5477.338 & & &\
300 K,$\mu_{ab}$ & 5469.084 & 5475.652 &5480.529 & 5487.472 & VWF\
300 K,$\mu_{c}$ & 5469.084 & 5476.757 &5480.529 & VWF & 5493.94\
530 K,$\mu_{u}$ & 5469.076 &VW(5476.749) &5481.216 & 5488.010 & VWF\
530 K,$\mu_{ab}$ & 5469.076 & W(5476.442) &5481.216 & 5487.464 & VWF\
530 K,$\mu_{c}$ & 5469.076 & 5476.749 & 5480.521 &VWF & 5492.951\
case I & 5469.035 & 5479.146 &5481.175 & 5487.424 & 5496.948\
case II& 5469.387 & 5475.131 &5479.855 & 5486.627 & 5495.962\
\
\
Table VII: Charge Transfer and Occupation orbital numbers for the polycrystalline PbVO$_{3}$ at 90 K, 300 K and 530 K. For comparison the case of polycrystalline PbTiO$_{3}$ at 300 K is given.\
\
[|l|c|c|c|c|r|]{} Atom Type & l character & 90 K & 300 K & 530 K & PbTiO$_{3}$ \[300 K\]\
O & s & 1.879 & 1.879 & 1.881 & 1.882, 1.888\
& p & 4.357 & 4.355 & 4.361 & 4.448, 4.423\
& d & 0.000 & 0.000 & 0.000 & 0.000, 0.000\
Charge Transfer& & -0.236 & -0.234 & -0.241&-0.329, -0.312\
Atoms in cluster& & 57 & 57 & 57 & 53, 15\
V, Ti & s & 0.452 & 0.451 & 0.453 & 0.429, 0.439\
& p & 6.653 & 6.653 & 6.655 & 6.663, 6.706\
& d & 3.798 & 3.795 & 3.787 & 2.643, 3.792\
Charge Transfer& & 0.099 & 0.098 & 0.106 & 0.264, 0.081\
Atoms in cluster& & 49 & 45 & 45 & 37, 15\
V$_{core}$, Ti$_{core}$ & s & 0.454 & 0.454 & 0.455 & 0.435, 0.439\
& p & 6.674 & 6.673 & 6.675 & 6.690, 6.706\
& d & 4.902 & 4.922 & 4.894 & 3.778, 3.792\
Charge Transfer& & -0.037 & -0.037 & -0.031& 0.095, 0.081\
Atoms in cluster& & 65 & 65 & 57 & 45, 15\
Pb & s & 1.756 & 1.765 & 1.763 & 1.835, 1.834\
& p & 1.300 & 1.298 & 1.285 & 1.141, 1.136\
& d & 10.334 & 10.333 & 10.322 & 10.301, 10.331\
Charge Transfer& & +0.610 & +0.605 & 0.617 & 0.723, 0.699\
Atoms in cluster& & 46 & 46 & 46 & 42, 21\
\
Table VIII: Charge Transfer and Occupation orbital numbers for the polycrystalline PbVO$_{3}$ at for case I and II. For comparison the case of polycrystalline PbTiO$_{3}$ at 300 K is given.\
\
[|l|c|c|c|r|]{} Atom Type & l character & case I at 90 K & case II at 90 K & PbTiO$_{3}$ \[300 K\]\
O & s & 1.894 & 1.870 & 1.882, 1.888\
& p & 4.442 & 4.315 & 4.448, 4.423\
& d & 0.000 & 0.000 & 0.000, 0.000\
Charge Transfer& & -0.335 & -0.185 &-0.329, -0.312\
Atoms in cluster& & 57 & 57 & 53, 15\
V, Ti & s & 0.451 & 0.439 & 0.429, 0.439\
& p & 6.685 & 6.638 & 6.663, 6.706\
& d & 3.612 & 3.869 & 2.643, 3.792\
Charge Transfer& & 0.249 & 0.052 & 0.264, 0.081\
Atoms in cluster& & 49 & 45 & 37, 15\
V$_{core}$, Ti$_{core}$ & s & 0.452 & 0.443 & 0.435, 0.439\
& p & 6.693 & 6.662 & 6.690, 6.706\
& d & 4.779 & 4.971 & 3.778, 3.792\
Charge Transfer & & 0.088 & -0.069 & 0.095, 0.081\
Atoms in cluster & & 65 & 65 & 45, 15\
Pb & s & 1.767 & 1.752 & 1.835, 1.834\
& p & 1.145 & 1.408 & 1.141, 1.136\
& d & 10.334 & 10.339 & 10.301, 10.331\
Charge Transfer& & +0.756 & +0.502 & 0.723, 0.699\
Atoms in cluster& & 46 & 43 & 42, 21\
\
Table IX: Peak positions in eV, of the difference spectra displayed in Figs. \[fig6\]-\[fig7\].\
\
[|l|c|c|c|r|]{} Peak Type & $\delta \mu_{90-300}$ & $\delta \mu_{90-530}$ & $\delta \mu_{90-case1}$ & $\delta \mu_{90-case2}$\
+ve & & 5468.787 & & 5468.787\
+ve & & & 5469.087 &\
+ve & 5469.687 & & &\
+ve & 5470.626 & & &\
+ve & & 5475.131 & 5475.131 &\
+ve & 5477.341 & 5477.341 & & 5477.341\
+ve & & & 5480.532 &\
+ve & 5481.227 & & & 5481.227\
+ve & & & 5484.989 &\
+ve & & 5486.627 & &\
+ve & 5488.342 & & &\
+ve & & & 5492.962 &\
-ve & 5468.687 & 5468.187 & &\
-ve & & 5469.387 & & 5469.687\
-ve & 5474.626 & & & 5474.626\
-ve & & 5476.198 & &\
-ve & & & 5478.56 &\
-ve & & & & 5479.198\
-ve &5480.532 & & &\
-ve & & 5481.941 & 5481.941 &\
-ve & & & & 5484.199\
-ve & & &5488.342 &\
-ve & & 5491.057 & &\
-ve & 5492 & & &\
-ve & & & & 5497\
\
[^1]: It is known that XANES and EXAFS signals are sensitive to local structure. Indeed just as XRD is indicative of long-range order, XANES and EXAFS carry signatures of short range order or disorder.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A theorem about asymptotic estimation of multiple integral of a special type is proved for the case when the integrand peaks at the integration domain bound, but not at a point of extremum. Using this theorem the asymptotic expansion of the electromagnetic deuteron form factors at high momentum transfers is obtained in the framework of two-nucleon model in both nonrelativistic and relativistic impulse approximations. It is found that relativistic effects slow down the decrease of deuteron form factors and result in agreement between the relativistic asymptotics and experimental data at high momentum transfers.'
---
[**Asymptotic estimation of some multiple integrals and the electromagnetic deuteron form factors at high momentum transfer**]{}
A.F. Krutov [^1]
*Department of General and Theoretical Physics,*
*Samara State University,*
*Samara,*
*443011 Russia*
V.E. Troitsky [^2]
*D.V. Skobeltsyn Institute of Nuclear Physics,*
*Moscow State University,*
*Moscow,*
*119992 Russia*
N.A. Tsirova [^3]
*Department of General and Theoretical Physics,*
*Samara State University,*
*Samara,*
*443011 Russia*
(Dated: 10 September 2007)
PACS numbers: 11.10.Jj, 13.40Gp, 13.75.Cs
Introduction
============
Recent advances in experimental investigations of hadron structure arouse the interest in the theoretical study of the hadron electromagnetic form factors at high momentum transfers (see e.g. review [@GiG02] and references therein). In this connection the JLab program of investigations on elastic electron-deuteron scattering experiments at $Q^2\;\simeq\;$ 10 (GeV/c)$^2$ ($Q^2=-q^2\;,\;q$ is transferred momentum) [@ArH05] attracts considerable attention. There exists a hope that these JLab experiments will help to determine the limits of application for the two-nucleon model and to clarify the interplay between nucleon-nucleon and quark approaches to the deuteron.
Using the asymptotic expansion presented in this paper we show [@KrT07-1] that the momentum transfer region in the JLab experiments is asymptotical for the deuteron considered as a nucleon–nucleon system. That is why the study of the electromagnetic deuteron form factors is interesting at $Q^2\;\to\;\infty$.
The present work is devoted to the theoretical investigation of deuteron form factors asymptotic behavior at high momentum transfer in the framework of the two-nucleon model. The form factors asymptotics is studied in accordance with the next points.
1\. As a rule, calculation of the form factors asymptotic behavior in the relativistic approaches reduces to the asymptotic estimation of $n$-tuple integrals. In the relativistic approach used in our work the deuteron form factors are expressed in terms of double integrals where integrands peak at the integration domain bound, and corresponding point is not a point of extremum. In this connection the theorem defining asymptotic expansion of $n$-tuple integrals with such integrand is proven in our paper.
2\. In general, high momentum transfers require relativistic consideration. However we begin the consideration of the asymptotic estimation of electromagnetic deuteron form factors with the nonrelativistic case and nonrelativistic impulse approximation at $Q^2\;\to\;\infty$ . This is due to the facts that, at first, the nonrelativistic calculation is a less complicated and, second, this calculation is important for establishment of the role of the relativistic effects.
3\. The asymptotic expansion of the relativistic deuteron form factors is calculated in the relativistic invariant impulse approximation in a variant of instant form of Poincare-invariant quantum mechanics (PIQM) developed in our papers previously [@BaK95; @KrT02; @KrT03; @KrT05; @KrT07]. The relativistic calculations are performed by analogy with nonrelativistic case. It is shown that relativistic effects essentially slow down the asymptotical decrease of the form factors.
4\. It is found that obtained in the framework of the two-nucleon model relativistic asymptotics coincides with the experimental data.
This paper is organized as follows. Section 2 is devoted to the proving of the central for this work theorem defining asymptotics of multiple integrals of some special type. In Sec. 3 a brief review of the formulas for the deuteron form factors in the nonrelativistic and relativistic invariant impulse approximation is given. The deuteron form factors asymptotics is calculated in nonrelativistic and relativistic impulse approximation with the help of the proven theorem in Sec. 4. In Sec. 5 asymptotics of the form factors is obtained for the deuteron wave functions in the conventional representation as a discrete superposition of Yukawa-type terms [@Mac01]. Sec. 6 contains the conclusions of this paper.
Theorem on the asymptotic expansion of some multiple integrals in the case when the maximal value of the integrand belongs to region boundary
=============================================================================================================================================
In the following we will consider integrals of the kind: $$F(\lambda)=\int\limits_{\Omega}f(\lambda,x)\exp[S(\lambda,x)]dx\;,
\label{intlap}$$ where $\Omega$ is a domain in [**R**]{}$^n$, $x=(x_1,...,x_n)$, $\lambda$ is a large positive parameter. We will use following definitions: $\partial\Omega$ is a bound of the domain $\Omega$, $[\Omega]=\Omega\cup\partial\Omega$, the bound $\partial\Omega\in
C^{\infty}$ if in the vicinity of any point $x^0\in\partial\Omega$ it can be specified by equation $x_j=\varphi(x'),\ x'\in U',\
x'=(x_1,...,x_{j-1},x_{j+1},...,x_n)$, $U'$ is a neighborhood of a point $x^{'\,0} $, and the function $\varphi (x') \in
C^{\infty}$ in $U'$.
Let us prove now a lemma and a theorem on the asymptotic estimation of integrals in Eq.(\[intlap\]).
L e m m a. [*Let $S(\lambda,x)$ be a smooth function in the $[\Omega]$, $f(\lambda,x)$ be a continuous function in the $[\Omega]$, and $M(\lambda) \in C^1$, $$M(\lambda)=\sup\limits_{x\in[\Omega]}S(\lambda,x)<\infty\;,$$ at some $\lambda_0>0$ the integral*]{} (\[intlap\])
*be absolutely convergent: $$\int\limits_{\Omega}|f(\lambda_0,x)|\exp[S(\lambda_0,x)]dx<\infty\;,$$ and the following conditions be fulfilled at $\lambda\geq\lambda_0$: $$\frac{\partial S(\lambda,x)}{\partial\lambda}\leq\frac{d
M(\lambda)}{d\lambda}\;, \label{uslovie1}$$*
$$|f(\lambda,x)|\leq C_1|f(\lambda_0,x)|\;. \label{uslovie2}$$
Then at $\lambda\geq\lambda_0$ the following estimation is valid:
$$|F(\lambda)|\leq C_2 e^{M(\lambda)}\;. \label{lemma}$$
P r o o f. At $\lambda\geq\lambda_0$ the following estimations are true: $$|F(\lambda)|\leq e^{M(\lambda)}\int\limits_{\Omega}
e^{S(\lambda_0,x)-M(\lambda_0)} e^{S(\lambda,x)
-S(\lambda_0,x)-M(\lambda)+M(\lambda_0)}|f(\lambda,x)|dx\leq$$ $$\leq e^{M(\lambda)-M(\lambda_0)} \int\limits_{\Omega}
e^{S(\lambda_0,x)+S(\lambda,x)
-S(\lambda_0,x)-M(\lambda)+M(\lambda_0)}|f(\lambda,x)|dx.$$ From conditions (\[uslovie1\]),(\[uslovie2\]) we obtain the inequality: $$|F(\lambda)|\leq C_1
e^{M(\lambda)-M(\lambda_0)}\int\limits_{\Omega} e^{S(\lambda_0,x)}
|f(\lambda_0,x)|dx\leq C_2 e^{M(\lambda)}\;.$$ Thus the statement (\[lemma\]) of the lemma is proved.
Later we will consider function $S(\lambda,x)$ described in lemma which has the maximal value in the point $x^0\in\partial\Omega$, and $S(\lambda,x),\
\partial\Omega\in C^\infty$ in the vicinity of $x^0$. This point is not the point of extremum, that means the validity of the following conditions: $$\frac{\partial S(\lambda,x^0)}{\partial n}\neq 0\;, \label{cond1}$$ and matrix of coefficients $B$: $$\left\|\frac{\partial^2S(\lambda,x^0)}{\partial\xi_i\partial\xi_j}
\right\|_{i,j=1}^{n-1}=B\;, \label{cond2}$$ gives the negative determined quadratic form. In Eqs. (\[cond1\]), (\[cond2\]) $\partial/\partial n$ specifies the internal normal derivative $\vec{n}$ to the $\partial \Omega$, and $\xi_1,...,\xi_{n-1}$ is an orthonormal basis in the tangential to the $\partial\Omega$ plane $T\,\partial\Omega_{x^0}$ at the $x^0$ point.
For convenience let us choose in the vicinity of point $x^0$ a frame $y=(y_1,...,y_n)$, so that $x^0$ is the origin of coordinate and the internal normal to $\partial\Omega$ coincides with the last basis vector of the new coordinate system. Functions $f,\ S$ in these coordinates we denote as $f^*,\ S^*$, and $U^*$ is an image of $U$ (that is an image of a half-vicinity of the point $x^0$). The equation for $\partial U^*$ in the vicinity of the point $y=0$ can be written in the following way: $$y_n=\varphi(y'), \hspace{10mm} y'\in U',\hspace{10mm}
y'=(y_1,...,y_{n-1}),$$ with $U'$ is a vicinity of the point $y'=0$, $\varphi(y')\in
C^\infty(U'),$ and at $y'\to 0,\varphi(y')=O(\mid y'\mid ^2)$.
T h e o r e m. [*Let the following conditions be fulfilled:*]{}
1$^\circ.\ f, S \in C([\Omega]).$
2$^\circ.$ $S(\lambda,x^0)$ [*is maximal value of function*]{} $S(\lambda,x)\;,\;$ $x^0 \in \partial\Omega$, [*and*]{} $x^0$ [*is not point of extremum.*]{}
3$^\circ.\ f,\ S,\ \partial\Omega \in C^{\infty}$ [*in the vicinity of the point*]{} $x^0$.
4$^\circ.$ [*The Taylor expansion of functions $S^*$ and $f^*$ in the vicinity of point $x^0$ satisfy the following relations:*]{} $$f^*(\lambda, y)=f^*(\lambda, 0)[1+o(1)]\;, \label{f*ly}$$ $$S^*(\lambda, y',\varphi(y'))-S^*(\lambda, 0)=\frac{1}{2}\langle Ay',\
y'\rangle+O(\mid y'\mid ^3)\;, \label{S*ly}$$ [*the matrix $A=\left\|\frac{\partial^2S^*(\lambda,0)}{\partial y_i\partial
y_j} \right\|_{i,j=1}^{n-1},$ angle brackets denote bilinear form: $\langle x,\ y\rangle=x_1y_1+x_2y_2+...+x_ny_n$*]{}.
[*Then at*]{} $\lambda\to\infty$ [*the following asymptotic expansion is valid:*]{} $$F(\lambda)\sim\exp[S(\lambda,x^0)]\sum\limits_{k=0}^{\infty}a_k(\lambda)\;.
\label{ocenka}$$ [*The way to calculate coefficients $a_k(\lambda)$ will be determined later.*]{}
P r o o f. Let us divide the integral (\[intlap\]) into two integrals. Integration domain of the first one is the half-vicinity $U$ of the point $x^0$, and integration domain of the second one is a remainder of integration domain of the original integral. It is easily shown by the proven lemma that the second integral is exponentially small as compared with $\exp[S(\lambda,x^0)]$. So we will estimate asymptotically the first integral only.
In the expansion of the function $S^*(\lambda,y',\varphi(y'))$ in line with the condition (\[S\*ly\]) linear components are absent, because the point $y'=0$ is a point of maximum of the function $S^*(\lambda,y',\varphi(y'))$ in the region $U'$.
Let us choose $U$ in accordance with inequalities $\varphi(y')\leq y_n\leq\delta,\ \delta>0$ at $y \in U^*$. Then we can represent the integral (\[intlap\]) within exponentially decreasing terms:
$$F(\lambda)=\int\limits_{U^*}f^*(\lambda,
y)\exp[S^*(\lambda,y)]dy\;. \label{idok}$$
Let us rewrite integral in Eq. (\[idok\]) in the following way: $$F(\lambda)=\int\limits_{U'}\Phi(\lambda, y')dy'\;,$$ with $$\Phi(\lambda, y')=\int\limits_{\varphi(y')}^{\delta}
\exp[S^*(\lambda,y)]f^*(y)dy_n\;. \label{intfi}$$ The integral (\[intfi\]) is one-dimensional, and the function $S^*(\lambda, y)$ reaches extremum on the boundary $y_n=\varphi(y')$. Asymptotic expansion of this integral can be found through integration by parts. After $N+1$ integration we obtain the sequence: $$\Phi(\lambda,y')=\sum\limits_{k=0}^{N}M^k\left[\frac{f^*(\lambda,y)}{S^{*'}(\lambda,y)}\right]
\left.\exp[S^*(\lambda,y)]\right|_{\varphi(y')}^{\delta}-\int\limits_{\varphi(y')}^{\delta}
{M^N\left[\frac{f^*(\lambda,y)}{S^{*'}(\lambda,y)}\right]'}\exp[S^*(\lambda,y)]dy_n\;,$$ with $M^0$ is a unit operator and $$M^k=-\frac{1}{S^{*'}(\lambda,y)}\frac{d^k}{dy^k_n}\;.$$ The substitution of $y_n=\varphi(y')$ provides the main contribution to the asymptotics, the value of $y_n=\delta$ is exponentially small as compared with the previous. Further integration under these conditions gives the following expansion for the function (\[intfi\]): $$\Phi(\lambda,y')=-\exp[S^*(\lambda,y',\varphi(y'))]\sum\limits_{k=0}^\infty
M^k\left[\frac{f^*(\lambda,y',\varphi(y'))}{S^{*'}(\lambda,y',\varphi(y'))}\right]\;.$$ So $$F(\lambda)=-\sum\limits_{k=0}^\infty
\int\limits_{U'}\exp[S^*(\lambda,y',\varphi(y'))]
M^k\left[\frac{f^*(\lambda,y',\varphi(y'))}{S^{*'}(\lambda,y',\varphi(y'))}\right]dy'
\label{F}$$ The point $y'=0$ is an internal point of maximum of the integrand in the expression (\[F\]). Functions $S^*(\lambda,y',\varphi(y'))$ and $f^*(\lambda,y',\varphi(y'))$ satisfy the conditions of lemma (\[uslovie1\]) and theorem (\[f\*ly\]), (\[S\*ly\]), therefore we can apply a formula for asymptotic estimation of the n-tuple Laplas integrals [@Fed77] and obtain corresponding asymptotic expansion (\[ocenka\]) of the integral in Eq. (\[F\]). Thus the theorem is proven.
Note, that in the general case it is rather difficult to write a compact formula for the coefficients $a_k(\lambda)$ in Eq. (\[ocenka\]). They can be obtained such kind of way in any specific cases. As an example, these coefficients will be obtained and written out explicitly for double integrals in the consideration of the asymptotic estimation of electromagnetic deuteron form factors in Sec. 4. Here we write out only the first asymptotic term from Eq. (\[ocenka\]) in the $x$ variables:
$$F(\lambda)\; \sim\;
-(2\pi)^{\frac{n-1}{2}}\exp[S(\lambda,x^0)]\left(\frac{\partial
S(\lambda,x^0)}{\partial n}\right)^{-1}|\det
B|^{-\frac{1}{2}}f(\lambda,x^0)\;, \label{gch}$$
where $\vec n\;,\;B$ are defined by conditions (\[cond1\]) and (\[cond2\]).
Electromagnetic deuteron form factors in the\
nonrelativistic and relativistic impulse approximation
======================================================
In the nonrelativistic impulse approximation known formulas for electromagnetic deuteron form factors can be rewritten in the following way [@JaM72]: $$G^{NR}_C(Q^2) = \sum_{l,l'}\int\,k^2\,dk\,k'\,^2\,dk'\, u_l(k)\,
\tilde g^{ll'}_{0C}(k\,,Q^2\,,k')\, u_{l'}(k')\;,$$ $$G^{NR}_Q(Q^2) =
\frac{2\,M_d^2}{Q^2}\,\sum_{l,l'}\int\,k^2\,dk\,k'\,^2\,dk'\,
u_l(k)\,\tilde g^{ll'}_{0Q}(k\,,Q^2\,,k')\,u_{l'}(k')\;,$$ $$\label{GqGNIP}
G^{NR}_M(Q^2) =-\,M_d\,\sum_{l,l'}\int\,k^2\,dk\,k'\,^2\,dk'\,
u_l(k)\,\tilde g^{ll'}_{0M}(k\,,Q^2\,,k')\, u_{l'}(k')\;.$$ Here $u_l(k)$ are the deuteron wave functions in momentum representation, $l\;,\;l'=$ 0,2 are orbital angular momenta, $\tilde g^{ll'}_{0i}(k\,,Q^2\,,k')\;,\;i=C,Q,M$ are nonrelativistic free two-particles charge, quadrupole and magnetic dipole form factors, $M_d$ is the deuteron mass. Formulas for $\tilde g^{ll'}_{0i}$ are given in [@KrT07].
Let us discuss briefly possible types of the model deuteron wave functions. There are several classes of the deuteron wave functions: obtained with microscopic model Hamiltonians of the $NN$-interaction in the non-relativistic nuclear physics (for example, see [@Mac01]), deduced from scattering amplitudes in the Bethe-Salpeter approach and its various quasipotential reductions (see [@StG97]), wave functions of the Poincare-invariant quantum mechanics (as an example see wave functions in the instant form of PIQM [@BaK95; @KrT02; @KrT03; @KrT05; @KrT07]), and also wave functions calculated in the various statements of inverse scattering problems [@Tro94; @KrM97; @ShM03]. But independently of the method any wave function can be represented as the following Laguerre polynomial expansion [@ShM03]: $$u_l(k)=\sum\limits_{m=0}^{\infty} a_{lm}\sqrt{\frac{2m!}{\Gamma
(m+l+3/2)}}\; r_0^{l+\frac{3}{2}}\;k^l\;
L_{m}^{l+\frac{1}{2}}(r_0^2 k^2)\; e^{-\frac{r_0^2 k^2}{2}}\;
\label{funcu}$$ or in the coordinate representation: $$u_{l}(r) = \sum\limits_{m=0}^{\infty} (-1)^m
a_{lm}\sqrt{\frac{2m!}{r_0\;\Gamma (m+l+3/2)}}\;
\left(\frac{r}{r_0}\right)^{l+1}\;
L_{m}^{l+\frac{1}{2}}\left(\frac{r^2}{r_0^2}\right)\;
e^{-\frac{r^2}{2r_0^2}}\;, \label{funcukoord}$$ here $L^{l+1/2}_{m}(x)$ are generalized Laguerre polynomials, $\Gamma (x)$ is an Euler gamma function, the dimensional parameter $r_0$ can be related to the deuteron matter radius (see Sec. 5).
The wave function representation as a Laguerre polynomial expansion (\[funcu\]) is very useful for the calculation of the asymptotic behavior of the form factors. However, one can avoid such representation and obtain the asymptotic expansion directly for the initial wave function.
Generally, at high transferred momentum it is necessary to take into account relativistic corrections in the electromagnetic deuteron structure. In our paper relativistic description of the deuteron is constructed in the framework of instant form of Poincare-invariant quantum mechanics (PIQM), developed by authors previously [@BaK95; @KrT02; @KrT03; @KrT05; @KrT07]. In this approach we present electromagnetic deuteron form factors by analogy with nonrelativistic case (\[GqGNIP\]). Corresponding formulas in the relativistic impulse approximation were obtained in our paper [@KrT03]: $$G^R_C(Q^2) = \sum_{l,l'}\int\,d\sqrt{s}\,d\sqrt{s'}\,
\varphi_l(s)\, g^{ll'}_{0C}(s\,,Q^2\,,s')\, \varphi_{l'}(s')\;,$$ $$G^R_Q(Q^2) =
\frac{2\,M_d^2}{Q^2}\,\sum_{l,l'}\int\,d\sqrt{s}\,d\sqrt{s'}\,
\varphi_l(s)\,g^{ll'}_{0Q}(s\,,Q^2\,,s')\,\varphi_{l'}(s')\;,$$ $$\label{GqGRIP}
G^R_M(Q^2)=-\,M_d\,\sum_{l,l'}\int\,d\sqrt{s}\,d\sqrt{s'}\,
\varphi_l(s)\,g^{ll'}_{0M}(s\,,Q^2\,,s')\, \varphi_{l'}(s')\;,$$ where $\varphi_l(s)$ are the deuteron wave functions in sense of PIQM, $g^{ll'}_{0i}((s\,,Q^2\,,s'),$ $i=C,Q,M$ are relativistic free two-particles charge, quadrupole and magnetic dipole form factors. Formulas for free form factors are given in [@KrT07].
The deuteron wave functions in sense of PIQM are solutions of eigenvalue problem for a mass squared operator for the deuteron (see, e.g. [@BaK95]): $ \hat M^2_d\, |\psi\rangle =
M^2_d\,|\psi\rangle. $ An eigenvalue problem for this operator is coincident with the nonrelativistic Schrödinger equation within a second order on deuteron binding energy ${\varepsilon_d^2}/({4M})$, the value of which is small ($M$ is an averaged nucleon mass). So the deuteron wave functions in sense of PIQM differ from nonrelativistic wave functions by conditions of normalization only. In the relativistic case the wave functions are normalized with relativistic density of states: $$\sum_{l=0,2}\,\int_0^\infty\varphi^2_l(k)\,\frac{dk}{2\sqrt{k^2 +
M^2}}=1\;,\quad
\varphi_l(k) = \sqrt[4\ ]{s}\,k\,u_l(k)\;,\quad
s=4(k^{2}+M^{2})\;.\label{relnorm}$$
Nonrelativistic formulas Eq. (\[GqGNIP\]) can be obtained from relativistic ones (\[GqGRIP\]) in the nonrelativistic limit. This limiting procedure can be performed in the most natural way in the instant form of PIQM. The reason is that in papers [@BaK95; @KrT02; @KrT03; @KrT05; @KrT07] we have constructed the successful formalism of the instant form of PIQM. In the case of other forms of PIQM (point and front forms) the obtaining of nonrelativistic limit is much more difficult.
For obtaining the asymptotic form factors behavior at high transferred momentum in the nonrelativistic and relativistic cases it is necessary to estimate asymptotically double integrals (\[GqGNIP\]) and (\[GqGRIP\]) at $Q^2\to\infty$. Notice that integrands reach its maximum value at the integration domain bound, and this point is not a point of extremum. In the previous Section the theorem defining asymptotics of $n$-tuple integrals of such kind was proven.
Asymptotic expansion of the deuteron form factors
=================================================
We start asymptotic expansion of the deuteron form factors from the nonrelativistic case. It is caused by the simplicity of the nonrelativistic formulas, so the calculation of the asymptotics is more clear. In what follows the relativistic calculation will be presented analogous to the nonrelativistic one, although more combersome. Moreover nonrelativistic calculation is interesting because nonrelativistic formulas for the form factors (\[GqGNIP\]) are conventional, that is why its correct asymptotic expansion has universal significance. Let us emphasize also that the relativistic expressions for form factors and, therefore, their asymptotic estimations depend on the choice of the method of relativisation of the two-nucleon model. Nonrelativistic calculation is also of interest because it helps to clarify the role of relativistic effects in the electromagnetic structure of the deuteron at the asymptotical momentum transfers.
As we have seen in Sec. 3, the deuteron form factors in the nonrelativistic impulse approximation can be represented by double integrals (\[GqGNIP\]). We will find its asymptotic expansion using the theorem of Section 2 and use as an example the asymptotics of the charge form factor. We shall estimate only the $l = l' = 0$ term in the sum (\[GqGNIP\]) because the asymptotics of the other terms of form factors (\[GqGNIP\]) can be derived analogously.
Let us rewrite the corresponding $l = l' = 0$ term of the charge form factor (\[GqGNIP\]) using Eq. (\[funcu\]): $$\int\,\tilde g^{00}_{0C}(k\,,Q^2\,,k')\,\exp\left[S(k,k')\right]
\,k^2\,dk\,k'\,^2\,dk'\;\times$$ $$\times\left(\sum\limits_{m} a_{0m}\sqrt{\frac{2m!}{\Gamma
(m+3/2)}}\; r_0^{\frac{3}{2}}\; L_{m}^{\frac{1}{2}}(r_0^2
k^2)\right)\left(\sum\limits_{m} a_{0m}\sqrt{\frac{2m!}{\Gamma
(m+3/2)}}\; r_0^{\frac{3}{2}}\; L_{m}^{\frac{1}{2}}(r_0^2
k'\,^2)\right)\;. \label{Gcnew}$$ We have denoted in (\[Gcnew\]): $$S(k,k') = -\frac{r_0^2}{2} \;(k^2+k'^2)\;.
\label{S}$$ The expression for $\tilde g^{00}_{0C}(k\,,Q^2\,,k')$ is commonly accepted (see, e.g., [@KrT07]): $$\tilde g^{00}_{0C}(k, Q^2, k') = \frac{1}{k\,k'\,Q}
\left[\theta\left(k' - \left|k - \frac{Q}{2}\right|\right) -
\theta\left(k' - k - \frac{Q}{2}\right)\right]
\left(G^p_E(Q^2)+G^{n}_E(Q^2)\right),\label{tg00}$$ $G^{p,n}_E(Q^2)$ are electric form factors of proton and neuteron respectively, $\theta(x)$ is a step function.
In the case under consideration the space dimension $n=2,\
(x_1,x_2)=(k,k'),\ \lambda = Q^2$ is a large positive parameter. Integration domain is determined by $\theta$-functions in Eq. (\[tg00\]) and shown in Fig.1. The location of the point of maximal value of the function $S$ can be obtained by analysis of (\[S\]) and (\[tg00\]): $(k^0,k'\,^0)=(\frac{Q}{4},\frac{Q}{4})$.
![The integration domain, location of the point of maximal value, and transition to the new variables for the nonrelativistic case[]{data-label="dff"}](nerel1)
Let us perform the transition to the new basis as we have descripted before. We perform the shift of the origin of coordinates to the point of maximal value of the function $S$. Then we rotate the obtained frame for the internal normal to the boundary in the new origin to be coincident with the last basis vector of the new frame. This procedure is illustrated in Fig.1.
In the other words we perform the transition to the new variables in Eq. (\[Gcnew\]): $$k=\frac{1}{\sqrt{2}}\left(t' + t\right) + \frac{Q}{4}\;, \quad
k'=\frac{1}{\sqrt{2}}\left(t' - t\right) + \frac{Q}{4}\;.
\label{tt'}$$ At this transformation function $S(k,k')$ gets dependence on large parameter $Q^2$: $$S^*(Q^2,t,t') =
-\frac{r_0^2}{2}\left(t^2+t'\;^2+\frac{Q}{\sqrt{2}}\;t'+\frac{Q^2}{8}\right)
\label{s}$$ Functions $S(Q^2,t,t')\;,g^{00}_{0C}(k\,,Q^2\,,k')$, and the boundary of the integration domain satisfy the conditions of theorem 1$^\circ$, 3$^\circ$, 4$^\circ$. The location of the point, that satisfies the conditions (\[cond1\]), (\[cond2\]), can be obtained by a simple analysis of the function (\[s\]): $(t^0,t'\,^0)=(0,0)$. Let us show that this point satisfies the conditions 2$^\circ$ of the theorem.
It is obvious, that in the point of maximal value $$\left.\frac{\partial S}{\partial n}=\frac{\partial S^*}{\partial
t'}\right|_{(t,t')=(0,0)}= -\frac{r_0^2\;Q}{2 \sqrt{2}}\;\ne\;
0\;.$$ So the condition (\[cond1\]) is satisfied.
Let us calculate now the $B$ matrix from the condition (\[cond2\]). In our case the tangent to the domain of integration boundary vector in the point of maximal value is $\vec{\xi}=(1/\sqrt{2},-1/\sqrt{2})$, i.e. the $B$ matrix is a number: $$\left.\frac{\partial^2 S}{\partial\xi^2}=\frac{\partial^2
S^*}{\partial t^2}\right|_{(t,t')=(0,0)}\;= -r_0^2\;<\;0\;.
\label{B11}$$ We note, that $B$ is negative-definite, i.e. the point $(t^0,t^0\,') = (0,0)$ is really the point of maximal value. So the point (0,0) satisfies the condition 2$^\circ$ of the theorem.
So integral (\[Gcnew\]) satisfies the requirements of the theorem proven in the Sec. 2. Therefore we can apply the asymptotic formula (\[ocenka\]).
Calculating by analogy the other terms of the sum (\[GqGNIP\]) we obtain asymptotic expansions of deuteron form factors in the nonrelativistic impulse approximation: $$G_i^{NR}(Q^2)\sim e^{-\frac{r_0^2
Q^2}{16}}\sum\limits_{m=0}^\infty
\frac{h_{2m}^{NR}}{(2m)!}\;\Gamma(m+\frac{1}{2}), \label{ryadner}$$ $$h_{2m}^{NR}=\left.\sum\limits_{k=0}^\infty\frac{1}{r_0^{2m+2k+3}}\sum\limits_{p=0}^k
b_{kp} \left(\frac{2\sqrt{2}}{Q}\right)^{p+k+1}
\frac{\partial^{2m}}{\partial
t^{2m}}f_i^{NR\;(k-p)}(t,Q^2,0)\right|_{t=0}, \label{koefner}$$ $$b_{k0}=1,\;b_{kp}=b_{k-1\;p}-(k+p-1)b_{k-1\;p-1},\;b_{kk}=(-1)^k(2k-1)!!\;,$$ $$f_i^{NR}(t,Q^2,t')=A_i\sum\limits_{l,l'=0,2}k^{l+2}\,k'\,^{l'+2}\,
\tilde{u}_l\left(k\right)\;\tilde{g}^{ll'}_{0i}(t,Q^2,t')\;\tilde{u}_{l'}
\left(k'\right)\;,$$ with $k=k(t,Q^2,t')\,,k'=k'(t,Q^2,t')$, variables $t,\;t'$ are denoted in (\[tt’\]), $i=C,Q,M$, $\;A_C=1,\;A_Q=2M_d^2/Q^2\,,\;A_M=-M_d$, $$f_i^{NR\;(m)}(t,Q^2,t')=\frac{\partial^m}{\partial
t'\,^{m}}f_i^{NR}(t,Q^2,t')\;.$$ $\tilde{u}_{l,l'}$ is defined by equalities: $$u_0(k)=\tilde{u}_0(k)\;e^{-\frac{r_0^2
k^2}{2}}\;,\;u_2(k)=\tilde{u}_2(k)\;k^2\;e^{-\frac{r_0^2
k^2}{2}}\;. \label{utildeu}$$ Let us perform now the calculation of the relativistic asymptotics of deuteron form factors. To estimate asymptotically integrals (\[GqGRIP\]) we proceed analogously to nonrelativistic case, i.e. we use relativistic analogs of corresponding nonrelativistic formulas (\[Gcnew\])-(\[B11\]). Now the free relativistic charge form factor in (\[GqGRIP\]) at $l = l' = 0$ is given in Ref. [@KrT07]: $$g^{00}_{0C}(s, Q^2, s') = R(s, Q^2, s')\,Q^2 \left[\,(s + s' +
Q^2)\left(G^p_E(Q^2)+G^{n} _E(Q^2)\right)g^{00}_{CE} + \right.$$ $$\left. + \frac{1}{M}\xi(s,Q^2,s')
\left(G^p_M(Q^2)+G^{n}_M(Q^2)\right)g^{00}_{CM}
\right]\;,\label{g00}$$ $G^{p,n}_{E,M}(Q^2)$ are electric and magnetic form factors of proton and neutron respectively, $$g^{00}_{CE} = \left(\frac{1}{2}\cos\omega_1\cos\omega_2 +
\frac{1}{6}\sin\omega_1\sin\omega_2\right)\;,\quad g^{00}_{CM} =
\left(\frac{1}{2}\cos\omega_1\sin\omega_2 -
\frac{1}{6}\sin\omega_1\cos\omega_2\right)\;,$$ $$R(s, Q^2, s') = \frac{(s+s'+Q^2)}{\sqrt{(s-4M^2) (s'-4M^2)}}\,
\frac{\vartheta(s,Q^2,s')}{{[\lambda(s,-Q^2,s')]}^{3/2}}
\frac{1}{\sqrt{1+Q^2/4M^2}}\;,$$ $$\xi(s,Q^2,s')=\sqrt{ss'Q^2-M^2\lambda(s,-Q^2,s')}\;,$$ $\omega_1$ and $\omega_2$ are angles of the Wigner spin rotation, $$\omega_1 =
\arctan\frac{\xi(s,Q^2,s')}{M\left[(\sqrt{s}+\sqrt{s'})^2 +
Q^2\right] + \sqrt{ss'}(\sqrt{s} +\sqrt{s'})}\>,$$ $$\omega_2 = \arctan\frac{ \alpha (s,s') \xi(s,Q^2,s')} {M(s + s' +
Q^2) \alpha (s,s') + \sqrt{ss'}(4M^2 + Q^2)}\>, \label{omega}$$ where $\alpha (s,s') = 2M + \sqrt{s} + \sqrt{s'}$, $\vartheta(s,Q^2,s')= \theta(s'-s_1)-\theta(s'-s_2)$, $\theta$ is a step function, $\lambda(a,b,c)=a^2+b^2+c^2-2(ab+ac+bc)$, $$s_{1,2}=2M^2+\frac{1}{2M^2} (2M^2+Q^2)(s-2M^2) \mp \frac{1}{2M^2}
\sqrt{Q^2(Q^2+4M^2)s(s-4M^2)}\;.$$
To obtain relativistic asymptotic expansion we also perform transition to the new basis (shift and rotation). The function $S$ and the boundary of the integration domain differ from nonrelativistic ones, so it is necessary to perform a special analysis. In other words, instead of change of variables (\[tt’\]) we perform the following replacement: $$s=\frac{1}{\sqrt{2}}\left(t' + \frac{t}{Q}\right) + 2 M^2 +
M\sqrt{Q^2 + 4 M^2}\;,\quad s'=\frac{1}{\sqrt{2}}\left(t' -
\frac{t}{Q}\right) + 2 M^2 + M\sqrt{Q^2 + 4 M^2}\;. \label{tt'rel}$$
Then we obtain asymptotic expansion of the relativistic deuteron form factors by analogy with nonrelativistic case: $$G^R_i(Q^2)\sim
e^{-\frac{r_0^2}{4}(M\sqrt{Q^2+4M^2}-2M^2)}\sum\limits_{m=0}^\infty
\frac{h^R_{2m}}{(2m)!} \Gamma (m+\frac{1}{2}), \label{ryadrel}$$ $$h^R_{2m}=\left.\sum\limits_{k=0}^\infty\sum\limits_{p=0}^{2m}\frac{(-1)^p}{Q^{p+m-\frac{1}{2}}}
\;(2p-1)!!\;
C^{2m}_{2p}\;M^{m-p+\frac{1}{2}}\;\frac{2^{\frac{5}{2}k+4m-2p+\frac{9}{2}}}{r_0^{2k+2m-2p+3}}\;
\frac{\partial^{2m-2p}}{\partial t^{2m-2p}}
f_i^{R\;(k)}(t,Q^2,\varphi(t))\right|_{t=0}, \label{koefrel}$$
$$f_i^R(t,Q^2,t')=A_i\sum\limits_{l,l'=0,2}\tilde{u}_l(k)
\;g^{ll'}_{0i}(t,Q^2,t')\;
\tilde{u}_{l'}(k')\frac{\left(s/4-M^2\right)^{\frac{l+1}{2}}
\left(s'/4-M^2\right)^{\frac{l'+1}{2}}}{\sqrt[4]{s\,s'}},$$ $$f_i^{R\;(m)}(t,Q^2,t')=\frac{\partial^m}{\partial
t'\,^{m}}f_i^{R}(t,Q^2,t')\;.$$ Functions $k=k(s)\,,k'=k'(s')$ are specified in Eq. (\[relnorm\]), $s=s(t,Q^2,t')\,,s'=s'(t,Q^2,t')$, variables $t,\;t'$ are denoted in (\[tt’rel\]), $C^{2m}_{2p}$ are the binomial coefficients.
Asymptotic expansions (\[ryadner\]) and (\[ryadrel\]) are convergent power series in inverse degrees of the parameter $Q$ with known coefficients. The asymptotic expansion of this type is obtained in this work for the first time.
One can see from formulas (\[ryadner\]) and (\[ryadrel\]), that relativistic corrections change the behavior of form factors at high momentum transfer. In particular, exponential multiplier index is $Q^2$ in the nonrelativistic case, but in the relativistic case it is $Q$ at $Q^2\to\infty$. It seems to be a general feature of our relativistic approach to the description of composite systems, in particular, we have obtained the similar result in consideration of asymptotic behavior of the pion form factor in the composite quark model [@KrT98].
Asymptotics of the form factors for the conventional wave functions representation
==================================================================================
In this Section we represent the obtained asymptotical expansions (\[ryadner\]) and (\[ryadrel\]) in terms of initial wave functions in the left side of Eqs. (\[funcu\]), (\[relnorm\]). For this representation it is necessary to replace functions $\tilde{u}_l(k)$ by functions ${u}_l(k)$ in (\[ryadner\]) and (\[ryadrel\]) using (\[relnorm\]), (\[utildeu\]). Keeping the main term on $1/Q$ in asymptotic expansions (\[ryadner\]) and (\[ryadrel\]) one can obtain the next asymptotic formulas in terms of functions $u_l(k)$ and $\varphi_l(s)$ from (\[funcu\]), (\[relnorm\]): $$G_i^{NR}(Q^2)=\left.-A_i\frac{4\sqrt{\pi}}{r_0^3Q}
\sum\limits_{l,l'=0,2}k^2\;k'\,^{2}
u_l(k)\;\tilde{g}^{ll'}_{0i}(t,Q^2,t')\;u_{l'}(k')\right|_{{{t=0}\atop
{t'=0}}}, \label{anywfner}$$ $$G^R_i(Q^2)=\left.-A_i\frac{8\;\sqrt{\sqrt{2}\pi M}}{r_0^3\sqrt{Q}}
\sum\limits_{l,l'=0,2}
\frac{\varphi_l(s)\;g^{ll'}_{0i}(t,Q^2,t')\;\varphi_{l'}
(s')}{\sqrt[4]{s\,s'}}\right|_{{{t=0}\atop {t'=0}}}\;,\quad
i=C,Q,M\;. \label{anywfrel}$$ Let us note, that similar asymptotic representation can be obtained for any finite number of terms in asymptotic expansions (\[ryadner\]), (\[ryadrel\]).
In the modern calculations the deuteron wave functions are usually represented as a discrete superposition of Yukawa-type terms (see, e.g., [@Mac01]): $$u_0(k)=\sqrt{\frac{2}{\pi}}\sum\limits_j
\frac{C_j}{(k^2+m_j^2)},\;\quad
u_2(k)=\sqrt{\frac{2}{\pi}}\sum\limits_j
\frac{D_j}{(k^2+m_j^2)},\; \label{paru}$$ or in the coordinate representation: $$u_0(r) = \sum\limits_j {C_j}{\exp\left(-m_j\,r\right)}\;,$$ $$u_2(r) = \sum\limits_j {D_j}{\exp\left(-m_j\,r\right)} \left[1 +
\frac{3}{m_j\,r} + \frac{3}{(m_j\,r)^2}\right]\;,
\label{parukoord}$$ $$m_j = \alpha + m_0\,(j-1)\;,\quad \alpha =
\sqrt{M\,\left|\varepsilon_d\right|}\;.$$ Coefficients $C_j,\,D_j$, maximal value of the index $j$ and $m_0$ are determined by the best fit of corresponding solution of Schrödinger equation.
The deuteron wave function analytical form (\[parukoord\]) results in the right behavior of the wave functions at large distances: $$u_{0}(r)\;\sim\;\exp(-\,\alpha\,r)\;,\quad
u_{2}(r)\;\sim\;\exp(-\,\alpha\,r)\left(1 + \frac{3}{(\alpha\,r)}
+ \frac{3}{(\alpha\,r)^2}\right)\;. \label{rinfty}$$ The deuteron wave functions behavior at small distances: $$u_0(r)\;\sim\;r\;,\quad u_2(r)\;\sim\;r^3\;, \label{atorigin}$$ is provided by imposing the following conditions on coefficients $C_j$ and $D_j$: $$\sum\limits_j {C_j} = 0\;,\quad
\sum\limits_j {D_j} =
\sum\limits_j {D_j}{m_j^2} = \sum\limits_j \frac{D_j}{m_j^2} =
0\;. \label{cond}$$
Let us substitute the wave functions (\[paru\]) to (\[anywfner\]) and (\[anywfrel\]), and then obtain the first asymptotic terms of the nonrelativistic deuteron form factors:
$$G_C^{NR}\sim
\frac{1}{Q^8}\frac{2^{16}}{\sqrt{\pi}r_0^3}\left[\sum\limits_{j}C_{j}m_j^2\right]^2
\left(G_E^p(Q^2)+G_E^n(Q^2)\right)\;, \label{gnrc}$$
$$G_Q^{NR}\sim
3\,M_d^2\frac{1}{Q^{12}}\frac{2^{\frac{43}{2}}}{\sqrt{\pi}r_0^3}
\left[\sum\limits_{j}C_{j}m_j^2\right]\left[\sum\limits_{j}D_{j}m_j^4\right]
\left(G_E^p(Q^2)+G_E^n(Q^2)\right)\;, \label{gnrq}$$
$$G_M^{NR}\sim
\frac{1}{Q^8}\frac{2^{16}M_d}{\sqrt{\pi}r_0^3M}\left[\sum\limits_{j}C_{j}m_j^2\right]^2
\left(G_M^p(Q^2)+G_M^n(Q^2)\right)\;. \label{gnrm}$$
The dimensional parameter $r_0$ can be found from the expression for the deuteron matter radius in our deuteron model: $$r_m^2=\frac{1}{4}\int_0^\infty(u_0^2(r)+u_2^2(r))r^2dr\;.
\label{rm}$$ One can substitute wave functions of the form (\[funcukoord\]) into this expression. So formula (\[rm\]) specifies an algebraic equation for $r_0$. Solution of this equation can be found numerically.
It should be pointed out that main terms of expansion of charge and magnetic form factors in (\[gnrc\]) - (\[gnrm\]) are determined by $S$-state of deuteron only. The $D$-wave function gives the contribution to the main term of the quadrupole from factor. Its faster decrease at $Q^2\to\infty$ in comparison to the other form factors is a consequence of a faster decrease of a $D$-wave function at small distances in comparison to $S$-wave (\[atorigin\]). From the mathematical point of view the type of leading terms in (\[gnrc\])-(\[gnrm\]) is a consequence of conditions on the coefficients (\[cond\]). The modification of these conditions obviously results in change of the main terms in (\[gnrc\])-(\[gnrm\]). From these formulas it is also noticed that asymptotic expansions for the deuteron form factors contain dependence on the asymptotics of nucleon form factors.
We emphasize, that in the other deuteron asymptotics investigations only the power dependence on the transferred momentum was calculated as a rule. In the present paper we give a rigorous calculation of a multiplicative preasymptotical constant.
One can calculate relativistic asymptotics of form factors by analogy with nonrelativistic case. For this calculation we use the formulas (\[relnorm\]), (\[anywfrel\]), (\[cond\]). As a result we obtain: $$G_{C,M}^R(Q^2)\;\sim\;\frac{Q^3}{2^{\frac{7}{2}}M^3}\,
G_{C,M}^{NR}(Q^2)\;, \label{gcm}$$ $$G_{Q}^R(Q^2)\;\sim\;\frac{Q^4}{2^{\frac{11}{2}}M^4}\,
G_{Q}^{NR}(Q^2)\;. \label{gq}$$ Notice that asymptotic expansions (\[gnrc\])-(\[gnrm\]) and (\[gcm\]),(\[gq\]) are obtained for the first time in our work. It is interesting to compare obtained asymptotic estimations (\[gnrc\]) - (\[gnrm\]), (\[gcm\]),(\[gq\]) with observable behavior of the deuteron characteristics. At present time there exists the experimental information about function $A(Q^2)$ entered the differential cross section of the elastic $ed$-scattering. This function is expressed in terms of the deuteron form factors [@GiG02]. The values of function $A(Q^2)$ are known up to $Q^2\;\simeq\;$ 6 (GeV/c)$^2$. For the comparison with experimental data one needs to specify asymptotics of the nucleon form factors. It is naturally to choose for nucleon form factors the asymptotic which is predicted by the quark model [@GiG02] $G_M^{p,n}\sim {1}/{Q^4}$. Under these conditions the power dependence on $Q^2$ of the function $A(Q^2)$ coincides with experimentally observed one. The physical consequences will be examined in detail in the other paper.
Conclusion
==========
The theorem defining asymptotics of multiple integrals of some special type is proved. With help of the proven theorem the asymptotic expansion of the deuteron electromagnetic form factors at $Q^2\;\to\;\infty$ is calculated for the first time. The expansion is represented as a convergent series on in inverse powers of momentum transfer. The asymptotic of the form factors is found in terms of the conventional representation of the deuteron wave function as a discrete superposition of Yukawa-type terms. The asymptotic behavior of the form factors is calculated in the nonrelativistic impulse approximation and in the relativistic invariant impulse approximation proposed by the authors in the instant form of the Poincare-invariant quantum mechanics previously. It is established that relativistic corrections change the power dependence of the form factors on the momentum transfer at $Q^2\;\to\;\infty$ and slow down its decrease. It is also found that relativistic effects result in the agreement of the theoretical asymptotics and the experimentally observed behavior of the structure function $A(Q^2)$ at highest achieved momentum transfers.
Acknowledgments
===============
This work was supported in part by Russian Foundation of Basic Researches (grant 07-02-00962).
[99]{}
R. Gilman and F. Gross, J.Phys G. [**28**]{}, R37 (2002) \[arXiv:nucl-th/0111015\].
J. Arrington, R. J. Holt, P. E. Reimer [*et al.*]{} Hall A 12 GeV Upgrade (Pre-Conceptual Design Report), Jefferson Lab. 2005.
A. F. Krutov, V. E. Troitsky and N. A. Tsirova, (in preparation).
E. V. Balandina, A. F. Krutov, and V. E. Troitsky, Teor. Mat. Fiz. [**103**]{}, 41 (1995)\[English translation: Theor. Math. Phys. [**103**]{}, 381 (1995)\].
A. F. Krutov and V. E. Troitsky, Phys. Rev. C [**65**]{}, 045501 (2002) \[arXiv:hep-ph/0204053\].
A. F. Krutov and V. E. Troitsky, Phys. Rev. C [**68**]{}, 018501 (2003) \[arXiv:hep-ph/0307217\].
A. F. Krutov and V. E. Troitsky, Teor. Mat. Fiz. [**143**]{}, 258 (2005)\[English translation: Theor. Math. Phys. [**149**]{}, 704 (2005)\] \[arXiv:hep-ph/0412027\].
A. F. Krutov and V. E. Troitsky, Phys. Rev. C [**75**]{}, 014001 (2007) \[arXiv:hep-ph/0607026\].
R. Machleidt, Phys. Rev. C [**63**]{}, 024001 (2001) \[arXiv:nucl-th/0006014\].
M.V. Fedoryuk, [*The saddle point method*]{}. M.:Nauka (1977).
A. D. Jackson and L. C. Maximon, SIAM J. Math. Anal. [**3**]{}, 446 (1972)
A. Stadler and F. Gross, Phys. Rev. Lett. [**78**]{}, 26 (1997).
V.E. Troitsky, in *Proceedings of Quantum Inversion Theory and Applications*, Germany, 1993, edited by H.V. von Geramb, Lecture Notes in Physics [**427**]{} (Springer, Berlin, 1994), p. 50.
A.F. Krutov, D.I. Muravyev, and V.E. Troitsky, J. Math. Phys., [**38**]{}, 2880 (1997).
A. M. Shirokov, A. I. Mazur, S. A. Zaitsev, J. P. Vary and T. A. Weber, Phys. Rev. C [**70**]{}, 044005 (2004) \[arXiv:nucl-th/0312029\].
A. F. Krutov and V. E. Troitsky, Teor. Mat. Fiz. [**116**]{}, 215 (1998)\[English translation: Theor. Math. Phys. [**116**]{}, 907 (1998)\].
[^1]: E-mail: krutov@ssu samara.ru
[^2]: E-mail: [email protected]
[^3]: E-mail: ntsirova@ssu samara.ru
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the version 2.0 of the program package [[<span style="font-variant:small-caps;">GoSam</span>]{}]{} for the automated calculation of one-loop amplitudes. [[<span style="font-variant:small-caps;">GoSam</span>]{}]{} is devised to compute one-loop QCD and/or electroweak corrections to multi-particle processes within and beyond the Standard Model. The new code contains improvements in the generation and in the reduction of the amplitudes, performs better in computing time and numerical accuracy, and has an extended range of applicability. The extended version of the “Binoth-Les-Houches-Accord" interface to Monte Carlo programs is also implemented. We give a detailed description of installation and usage of the code, and illustrate the new features in dedicated examples.'
author:
- Gavin Cullen
- Hans van Deurzen
- Nicolas Greiner
- Gudrun Heinrich
- Gionata Luisoni
- Pierpaolo Mastrolia
- Edoardo Mirabella
- Giovanni Ossola
- Tiziano Peraro
- Johannes Schlenk
- 'Johann Felix von Soden-Fraunhofen'
- Francesco Tramontano
title: '[[<span style="font-variant:small-caps;">GoSam</span>]{}]{}-2.0: a tool for automated one-loop calculations within the Standard Model and beyond'
---
Introduction {#sec:intro}
============
intro
Overview of the program {#sec:overview}
=======================
overview
New features {#sec:newfeatures}
============
Installation and usage {#sec:instantuse}
======================
Installation {#sec:install}
------------
installation
Using [[<span style="font-variant:small-caps;">GoSam</span>]{}]{} {#sec:usage}
-----------------------------------------------------------------
usage
Interfacing to Monte Carlo programs {#sec:blha}
-----------------------------------
blha
Using external model files {#sec:model}
--------------------------
model
Examples {#sec:examples}
========
examples
Conclusions {#sec:conclusion}
===========
conclusion
We would like to thank the Herwig++ members J. Bellm, S. Gieseke, S. Plätzer, D. Rauch and C. Reuschle for fruitful interaction concerning the implementation of the BLHA2 interface. We are grateful to S. Pozzorini and C. Papadopoulos for comparisons and to P.-F. Monni for helpful discussions. We also would like to thank Joscha Reichel for collaboration on the BSM application of GoSam involving spin-2 particles. Finally, we are indebted to Thomas Reiter for setting the groundwork GoSam is based on. The work of G.C. was supported by DFG Sonderforschungsbereich Transregio 9, Computergestützte Theoretische Teilchenphysik. P.M., H.v.D., G.L. and T.P. are supported by the Alexander von Humboldt Foundation, in the framework of the Sofja Kovaleskaja Award Project “Advanced Mathematical Methods for Particle Physics”, endowed by the German Federal Ministry of Education and Research. The work of G.O. was supported in part by the National Science Foundation under Grant PHY-0855489 and PHY-1068550.
Commented example of an input card {#sec:appendix}
==================================
appendixA
Higher rank integrals {#sec:appendixB}
=====================
appendixB
gosam2.bbl
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Two-stage stochastic mixed-integer programming (SMIP) problems with general integer variables in the second-stage are generally difficult to solve. This paper develops the theory of integer set reduction for characterizing the subset of the convex hull of feasible integer points of the second-stage subproblem which can be used for solving the SMIP. The basic idea is to consider a small enough subset of feasible integer points that is necessary for generating a valid inequality for the integer subproblem. An algorithm for obtaining such a subset based on the solution of the subproblem LP-relaxation is then devised and incorporated into the Fenchel decomposition method for SMIP. To demonstrate the performance of the new integer set reduction methodology, a computational study based on randomly generated test instances was performed. The results of the study show that integer set reduction provides significant gains in terms of generating cuts faster leading to better bounds in solving SMIPs than using a direct solver.
[**Keywords:**]{} Stochastic programming, integer programming, integer set reduction, cutting planes, Fenchel decomposition, multidimensional knapsack.
bibliography:
- 'SWFD.bib'
---
[**Integer Set Reduction for Stochastic Mixed-Integer Programming**]{}\
\
Department of Industrial and Systems Engineering, Texas A&M University, 3131 TAMU,\
College Station, TX 77843, USA.\
\
Introduction {#sec-intro}
============
A two-stage stochastic mixed-integer programming (SMIP) problem involves optimizing the here-and-now (first-stage) costs plus expected future (second-stage) costs. Solving SMIP is still challenging and this paper makes strides towards that by introducing the theory of *integer set reduction* for characterizing *subsets* of the convex hull of feasible integer points of the second-stage subproblem that can be used to generate valid inequalities (cutting planes or cuts) for SMIP. The goal of integer set reduction is to speed up the cut generation routines and potentially lead to faster solution times for SMIP than direct solvers. In this paper we consider SMIP problems of the following form:
$$\begin{alignedat}{2}\label{eq-1a}
\text{SIP2: } \operatorname*{\mbox{Max }}\,\, & c^\top x + \mathcal{Q}_{E}(x) && \\
\text{s.t. } & A x \leq b && \\
& x \in X. &&
\end{alignedat}$$
In problem SIP2, $x \in \mathbb{R}_+^{n_1}$ denotes the first-stage decision vector, $c \in \mathbb{R}^{n_1}$ is the first-stage cost vector, $b \in \mathbb{R}^{m_1}$ is the first-stage right hand side, and $A \in \mathbb{R}^{m_1 \times n_1}$ is the first-stage constraint matrix. The set $X$ imposes binary restrictions on all or some components of $x$. The function $\mathcal{Q}_{E}(x)$ denotes the expected second-stage cost based on $x$. The function $\mathcal{Q}_{E}(x)$ is the expected recourse function and is given as follows: $$\label{eqn2}
\mathcal{Q}_{E}(x) = \mathbb{E}_\omega \Phi(q({\omega}),h({\omega})-T({\omega})x,{\omega}),$$ where $q(\omega) \in \mathbb{R}^{n_2}$ is the cost vector, $h(\omega) \in \mathbb{R}^{m_2}$ is the right hand side vector, and $T(\omega) \in \mathbb{R}^{m_2 \times n_1}$ is the technology matrix. The second-stage function $\Phi$ is a value function of a mixed-integer program (MIP) and is given as follows: $$\label{eqn2a}
\Phi(\rho,\tau,{\omega}) = \operatorname*{\mbox{Max }}\{\rho^\top y({\omega}): Wy({\omega}) \leq \tau, 0 \leq y({\omega}) \leq u, y({\omega}) \in Y\}.$$ In the second-stage (scenario) problem , $y({\omega})$ denotes the recourse decision vector and $W \in \mathbb{R}_+^{m_2 \times n_2}$ is the fixed recourse matrix. The vector $u \in \mathbb{Z}_+^{n_2}$ is the upper bound on the second-stage decision variables. It is assumed that $T({\omega}):\Omega \mapsto \mathbb{R}^{m_2 \times n_1} $, $h({\omega}):\Omega \mapsto \mathbb{R}^{m_2}$ and $q({\omega}):\Omega \mapsto \mathbb{R}^{n_2}$ are measurable mappings defined on a probability space $(\Omega,\mathcal{F},\mathbb{P})$. The set $Y$ imposes integer restrictions on all or some components of $y({\omega})$. The function $\mathcal{Q}_{E}(x)$ is the expected recourse function, where $\omega$ is a realization of a multivariate random variable $\tilde{\omega}$, and $\mathbb{E}_\omega$ denotes the mathematical expectation operator.
We consider problem SIP2 under the following assumptions:
- The random variable $\tilde{{\omega}}$ is discrete with finitely many scenarios ${\omega}\in {\Omega}$, each with probability of occurrence $p({\omega})$ such that $\sum_{{\omega}\in {\Omega}}p_{\omega} = 1$.
- The first-stage feasible set $\{A x \leq b, x \in X\}$ is nonempty.
- The right hand side vector $\tau$ and fixed-recourse matrix $W$ are nonnegative, and $W$ is rational.
- The second-stage feasible set $\{W y({\omega}) \leq \tau, 0 \leq y({\omega}) \leq u, y({\omega}) \in Y\}$ and is bounded and nonempty for all $x \in \{A x \leq b, x \in X\}$.
Assumption (A1) is needed for tractability while assumptions (A2) and (A4) are needed to guarantee that the problem has an optimal solution. Assumption (A3) and (A4) are needed for the proposed integer set reduction method to allow for a well-defined problem and finite convergence of the cutting method. Assumption (A4) implies the relatively recourse assumption, i.e., $\mathbb{E}_\omega [\,| \Phi(q({\omega}), h({\omega})-T({\omega})x),\omega |\,] < \infty$ for all $x \in \{A x \leq b, x \in X\}$. Because of assumption (A1), SIP2 can be written in extensive form as a so-called deterministic equivalent problem (DEP) as follows: $$\begin{alignedat}{2}\label{eq-SIP2-dep}
\text{DEP}: \operatorname*{\mbox{Max }}\,\, & c^\top x + \sum_{\omega \in \Omega} p_\omega q(\omega)^\top y(\omega) && \\\text{s.t. } & A x \leq b && \\
& T(\omega) x + W y(\omega) \leq h(\omega) && \\
& x \in X, y(\omega) \in Y. &&
\end{alignedat}$$
In the above formulation, $p_\omega$ denotes the probability of occurrence for the scenario $\omega$, and $\sum{p_{\omega \in \Omega}} = 1$. Even for a reasonable number of scenarios in $\Omega$, DEP is a large-scale MIP. With integer variables in both first and second-stages, a moderate sized DEP may be difficult to solve using a direct solver such as CPLEX [@CPLEX]. This makes a decomposition approach a necessity for most practical sized problems. In SIP2, the type of decision variables (continuous, binary, integer) and in which stage they appear greatly influences algorithm design. The complexity of the solution method depends on the definitions of the sets $X$ and $Y$. When both these sets do not impose integer restrictions on the decision variables, the recourse function $\Phi(\rho,\tau,{\omega})$ is a well-behaved piecewise linear and convex function of $x$. Thus, Benders’ decomposition [@benders1962partitioning] is applicable in this case [@wollmer1980two] and the L-shaped method [@van1969shaped] can be used to solve the problems. Assuming fixed recourse (i.e, the recourse matrix $W$ is independent of $\omega$), the value function of $\Phi(\rho,\tau,{\omega})$ is a piecewise linear function in $x$. Hence, the L-shaped method works by approximating the linear functions from the subproblems by constructing optimality cuts in the first-stage based on the dual values from the subproblems. However, when $Y \in \mathbb{Z}^+$, the linear approximation procedure by L-shaped method is not viable, as the value function is generally discontinuous and is lower semicontinuous [@blair1982value]. Also, the function is non-convex and sub-additive [@schultz1993continuity]. Hence, new algorithms or extensions of the L-shaped method are required to handle integer variables in the second or in both of the stages.
Cutting plane methods that can partially approximate the second-stage problems within the L-shaped method have been proposed for SMIPs with integer variables in the second-stage. In [@caroe1997cutting], a lift-and-project cutting plane approach based on the ideas from [@balas1993lift] is used to solve problems with binary and continuous variables in both the first- and second-stage. In [@sen2005c], for problems with binary variables in first-stage, and binary and continuous variables in second-stage, disjunctive cuts are developed for the second-stage. The work in [@sherali1998reformulation] and [@sherali2002modification] uses the framework of reformulation linearization technique. The algorithm in [@sen2005c] is extended in [@sen2006decomposition] for problems with binary, continuous and discrete variables in the second-stage. Applications of SMIP in supply chain, air traffic control, and auto-carrier vehicle loading problems can be found at [@beier2015nodal], [@corolli2015two], and [@venkatachalam2014algorithms], respectively.
Fenchel cuts are suggested in [@boyd1994fenchel], and a number of characteristics are derived in [@boyd1993solving], [@boyd1994solving] and [@boyd1995convergence]. The most important results from [@boyd1994fenchel], [@boyd1994solving] and [@boyd1995convergence] are that Fenchel cutting planes are facet defining under certain conditions, and the use of Fenchel cuts in a cutting plane approach yields an algorithm with finite convergence. The work also highlights the fact that generating a Fenchel cut for binary programs is computationally expensive in general; therefore, problems with special structure are desirable to achieve faster convergence. Computational experiments demonstrating the effectiveness of Fenchel cuts are presented for knapsack polyhedra in [@boyd1993generating] and for pure binary problems in [@boyd1994solving].
Since the pioneering work in [@boyd1994fenchel], only a few works in the literature have adopted Fenchel cuts. In [@saez2000solving], Fenchel cuts are used to improve the bounds obtained from MIPs using Lagrangian relaxation. Fenchel cuts are used to solve deterministic capacitated facility location problems [@ramos2005solving]. This work compares Fenchel cuts to Lagrangian cuts in finding good relaxation bounds for their problem. In [@boccia2008cut], Fenchel cutting planes are used for finding $p$ median nodes in a graph using a cut and branch approach. Fenchel cuts are first derived for two-stage SMIPs under a stage-wise decomposition setting in [@ntaimo2013fenchel] and are referred to as Fenchel decomposition (FD) cuts. Extensive study of FD cuts for two-stage SMIP with binary decision variables in both first-stage and second-stages are given in [@FDDECOMP] and [@venkatachalam2014algorithms].
This work makes the following contributions to the literature on stochastic programming: (a) deriving *integer set reduction* theory for determining subsets of the second-stage feasible integer set to use for faster cut generation; (b) devising an algorithm for obtaining such subsets based on the solution of the subproblem LP-relaxation; (c) applying the integer set reduction in the context of FD cuts for solving SMIPs with general integer variables in the second-stage; and (d) reporting on a computational study that demonstrates the advantages of integer set reduction. In the literature, Fenchel cuts and FD cuts are derived and used for MIP and SMIP, respectively, with binary variables. This work is the first to derive these cuts for MIP and SMIP with *general* integer variables.
The rest of this paper is organized as follows: Integer set reduction theory is derived in Section 2 and is applied to the FD setting for two-stage SMIP in Section 3. A computational study to illustrate the performance of the new methodology is reported in Section 4. Finally, conclusions and directions for future research are given in Section 5.
Integer Set Reduction for Cut Generation
========================================
In this section we develop the integer set reduction theory to characterize the properties of the convex hull of integer points needed for generating a valid inequality for the second-stage feasible set. We then use the theory to devise an algorithm for obtaining a reduced set of integer points needed for generating a valid inequality based on the second-stage LP-relaxation. To illustrate the concepts, we use simple numerical examples. Next we start with some preliminaries.
Preliminaries
-------------
We will now provide some important definitions needed in the derivation of a valid inequality for SIP2 with general integer variables in the second-stage. Suppressing ${\omega}$ for simplicity in exposition, the feasible set for the second-stage subproblem can be given as follows: $$\label{eqn5}
F^{IP} = \{y:Wy \leq \tau, 0 \leq y \leq u, y \in Y \}.$$ Thus subproblem can now be rewritten as $$\label{eqn2aIP}
\operatorname*{\mbox{Max }}\{\rho^\top y: y \in F^{IP}\}.$$ The LP-relaxation feasible set to $F^{IP}$ can be expressed as $$F^{LP} = \{y: Wy \leq \tau, 0 \leq y \leq u, y \in \mathbb{R}_+^{n_2} \} \notag$$ and the LP-relaxation to given as $$\label{eqn2aLP}
\operatorname*{\mbox{Max }}\{\rho^\top y: y \in F^{LP}\}.$$ We shall denote by $C(F^{IP})$ the convex hull of integer points in $F^{IP}$. Let $F^{IP}_{\upsilon} \subseteq F^{IP}$ and $C(F^{IP}_{\upsilon})$ will denote the convex hull of integer points in $F^{IP}_{\upsilon}$.
Now let $\hat{y} \in F^{LP}$ be the optimal solution to subproblem for a given $x \in \{A x \leq b, x \in X\}$. Our goal is to use the point $\hat{y}$ and restrict the derivation of valid inequalities (cuts) to a relatively small subset of integer points $F^{IP}_{\upsilon}$ instead of $F^{IP}$ so that a generated cut valid for $C(F^{IP}_{\upsilon})$ is valid for $C(F^{IP})$ and cuts off $\hat{y}$. The hope is that doing so would result in reduced cut generation computational time, thus leading to fast cutting plane methods for SIP2. Therefore, it is desirable to have $F^{IP}_{\upsilon} \subset F^{IP}$ such that $|F^{IP}_{\upsilon}| << |F^{IP}|$ so that generating cuts over $F^{IP}_{\upsilon}$ is not expensive.
\[defn1\] An inequality is said to be *valid* for the set $C(F^{IP})$ if it is satisfied by every point in the set. A cut with respect to a point $\hat{y} \notin C(F^{IP})$ is a valid inequality for $C(F^{IP})$ that is violated by $\hat{y}$.
![Separation problem with reduced integer feasible set[]{data-label="sep-fig2"}](sample.pdf "fig:")\
Generating a valid inequality using the subset $F^{IP}_{\upsilon}$ is depicted in Figure \[sep-fig2\]. In the figure, given $\hat{y} \in F^{LP}$, the three points defining the triangle constitute $F^{IP}_{\upsilon}$ and are used to generate the cut (dashed lines) of the form $\pi^\top y \leq \pi_0$. We devise a methodology to obtain $F_{\upsilon}^{IP}$, and subsequently use it to generate a valid inequality. Also, since $F_{\upsilon}^{IP} \subseteq F^{IP}$, we need to form $F_{\upsilon}^{IP}$ such that the generated valid inequality does not cut off any integer points in $F^{IP}$.
In subproblem , the general integer variable vector $y$ is bounded by the vector $u$. Let $I$ be the set of indices of the components of $y$ which are integer variables, and $K$ be the set of indices of the constraints in . Also, let the elements of the matrix $W$ be denoted by $w_{kt}$, where $k \in K$ is the constraint index, and $t \in I$ is the decision variable index. We also make the following assumption regarding $W$:
- The polytope defined by $C(F^{IP})$ is assumed to be full dimensional with dimension $n_2$.
Let $P$ be an index set for integer points in $F^{IP}$ such that for $p \in P$ we have an integer point $y^p \in F^{IP}$. The $i^{th}$ component of $y^p$ will be denoted $y_i^p$. In Figure \[sep-fig2\], for example, the integer point $y^p = (3, 2)$ has components $y_1^p=3$ and $y_2^p=2$. Let $y^{IP}$ be the optimal solution to and $\hat{y}$ be the optimal solution to the LP-relaxation. Also, define $\bar{y} = \lfloor \hat{y}\rfloor$ so that the components $\bar{y}_i$, for all $i = 1, \cdots, n_2$ are initialized as $\bar{y}_i = \lfloor \hat{y}_i\rfloor$. Let $d_{ij}^p$ denote the distance from $y^p$ to the boundary of $F^{LP}$ along the $y^p_j$ axis for all $i, j \in I, i \neq j$. Then $d_{ij}^p$ can be calculated as follows:
$$\label{dval}
\begin{alignedat}{2}
d_{ij}^{p} = \min \left\{ \left( \left[ \tau_k - \sum_{t \in I: t \neq i, t \neq j}w_{kt}\hat{y}_t -w_{ki}\bar{y}_i \right] / w_{kj} \right)_{\forall k \in K} - y_j^p, \, \, u_j - y_j^p \right\} ,
\end{alignedat}$$
where $\tau_k$ is the right hand side for constraint $k \in K$, and $u_j$ is the upper bound for axis $j$. For each $y^p_i$ axis, $i =1,\cdots,n_2$, let $f_i(y^p)$ be the shortest distance from $y^p$ to the boundary of $F^{LP}$ along the $y^p_j$ axis, $i \neq j$. Then $f_i(y^p)$ can be calculated as follows: $$\label{dval1}
\begin{alignedat}{2}
f_i(y^p) = \min_{j \in I, \ i \neq j} \left\lbrace d_{ij}^{p} \right\rbrace.
\end{alignedat}$$ For each $y^p_i$ axis, $i =1,\cdots,n_2$, let $f'_i(y^p)$ be the shortest distance from $y^p$ to the boundary of $F^{LP}$ along the $y^p_i$ axis. Then $f'_i(y^p)$ can be calculated as follows: $$\label{dval2}
\begin{alignedat}{2}
f'_i(y^p) = \min_{j \in I, \ i \neq j} \left\lbrace d_{ji}^{p} \right\rbrace.
\end{alignedat}$$
Next, we establish that there exists a set $F_{\upsilon}^{IP} \subseteq F^{IP}$, which is sufficient for generating a valid inequality for $F^{IP}$.\
\[lemma1\] Given $F^{IP}$, there exists a set $F_{\upsilon}^{IP} \subseteq F^{IP}$ such that $F_{\upsilon}^{IP}$ is sufficient for generating a valid inequality for $F^{IP}$.
Given $F^{IP}$, then either $F_{\upsilon}^{IP} = F^{IP}$ or $F_{\upsilon}^{IP} \subset F^{IP}$. For $F_{\upsilon}^{IP} = F^{IP}$, it is obvious that entire set $F^{IP}$ can be used for generating a valid inequality to cut off a fractional point $\hat{y}$. Now consider the case $F_{\upsilon}^{IP} \subset F^{IP}$. $\vert F^{IP} \vert \geq n_2+1$ as $C(F^{IP})$ is full dimensional by assumption (A5). To generate a valid inequality, $\vert F_{\upsilon}^{IP} \vert \geq n_2$, as $n_2$ affinely independent integer points are needed to construct a facet for $C(F^{IP})$. Then $F_{\upsilon}^{IP}$ can be constructed such that there exists an integer point $\, \, y^{p'} \notin F_{\upsilon}^{IP}$ and $ y^{p'} \in F^{IP}$. Hence, $\vert F^{IP} \vert > \vert F_{\upsilon}^{IP} \vert$ which gives $F_{\upsilon}^{IP} \subset F^{IP}.$
We state the required properties for an integer point $y^p \in F_{\upsilon}^{IP}$ in the following corollary.\
\[corr1\] Let $y^p \in F_{\upsilon}^{IP}$. Then either of the following must be true:
- $d_{ij}^{p} < 1$, $\forall i,j \in I \vert j \neq i$ or
- $d_{ij}^{p} \geq 1$ and any integer point $y^{p'} \in F^{IP}$ such that $y_i^{p'} - y_i^{p} \geq 1, \forall i \in I$ for at least one index $i \in I$ also belongs to $F_{\upsilon}^{IP}$.
When $d_{ij}^{p} < 1$, then there does not exist an integer point$\,\,\,\, y_i^{p} \in F^{IP}$. Alternatively, when $d_{ij}^{p} \geq 1$, then there exists an integer point $ \,\,\, y_i^{p'} \in F^{IP}$ such that $d_{ij}^{p} > d_{ij}^{p'} $. This means that there is an integer point $y_i^{p'}$ between $y_i^{p}$ and the boundary of $F^{LP}$. Since $y_i^{p} < y_i^{p'}$ and $y_i^{p}, y_i^{p'} > 0$, it implies that $d_{ij}^{p} - y_i^{p} > d_{ij}^{p} - y_i^{p'}$ and $y_i^{p'} - y_i^{p} > 0 $. This means that $y_i^{p'} - y_i^{p} \geq 1 $ since $y_i^{p'},y_i^{p} \in F^{IP}$. However by *(ii)*, if $d_{ij}^{p} \geq 1$ and $y_i^{p'} - y_i^{p} \geq 1, \forall i \in I$, then the point $y^{p'} \in F_{\upsilon}^{IP}$.
In Corollary \[corr1\], we simply state the properties for the integer points that define the set $F_{\upsilon}^{IP}$. However, it is desirable to get the smallest possible set $F_{\upsilon}^{IP}$ such that the valid inequality generated based on $F_{\upsilon}^{IP}$ does not cut off any (optimal) integer point in $F^{IP}$. We evaluate each of the components of $y$ and add an integer point $y^p$ to the set $F_{\upsilon}^{IP}$ if it is the closest integer point to $\hat{y}$ for that component, or if all other integer points between $y^p$ and the boundary of $F^{LP}$ are already in the set $F_{\upsilon}^{IP}$.
In the following lemma, we state the requirements for the minimum cardinality for the set $F_{\upsilon}^{IP}$. Ideally, we would like the set $F_{\upsilon}^{IP}$ to have a small number of integer points since generating a cut based on $C(F^{IP})$ may be computationally expensive.\
\[lemma2\] Given $F_{\upsilon}^{IP}$ and the set $D_i = \operatorname*{\mbox{Argmin}}_{y^p} \ \lbrace f_i(y^p) \mid y^p \in F^{IP}, y_i^p > 0 \rbrace$ there exists an integer point $y' \in F_{\upsilon}^{IP}$ for every $i \in I$ such that
$$\label{p2eq}
\begin{alignedat}{2}
y' = \operatorname*{\mbox{argmin}}_{y^k} \,\,\, \lbrace f'_i(y^k) \mid y^k \in D_i \rbrace. \nonumber
\end{alignedat}$$
We prove this result by contradiction. Suppose that the point $y' \in F_{\upsilon}^{IP} $ does not exist. Then this means that a valid inequality would pass through the origin. This implies that $\, \, C(F^{IP})$ is not full dimensional, which is a contradiction due to assumption (A5). Hence, there exists a point $y'$ for each axis $i$.
\[prop1\] The minimum cardinality of the set $F_{\upsilon}^{IP}$ is $n_2$.
By Lemma \[lemma2\], there should be at least one integer point for every component $i \in I$. This implies that $\left\vert I \right\vert=n_2$. Also, since valid inequalities are facets, then we need at least $n_2$ affinely independent points in the set $F_{\upsilon}^{IP}$ for generating the facet.
Integer Set Generation Algorithm
--------------------------------
Based on Lemma \[lemma2\] and Proposition \[prop1\], $\left\vert{F_{\upsilon}^{IP}}\right\vert = n_2$ since subproblem is full dimensional. However, getting the smallest set is not trivial unless we have an oracle providing an ideal interior point $y^p \in F^{IP}$, on which the smallest set $F_{\upsilon}^{IP}$ can be constructed. In the next section, we devise an algorithm to obtain the set $F_{\upsilon}^{IP}$ using Corollary \[corr1\], Lemma \[lemma2\], and Proposition \[prop1\]. In the algorithm, we start with an initial point $y^p \in F_{\upsilon}^{IP}$, where $y^p$ is constructed based on $\hat{y}$. Corollary \[corr1\] and Lemma \[lemma2\] are used to check whether the set of points in $F_{\upsilon}^{IP}$ are sufficient for generating a valid inequality, if not then the set $F_{\upsilon}^{IP}$ is expanded by sequentially adding integer points from the set $F^{IP}$. An algorithm for obtaining the set $F_{\upsilon}^{IP}$ can be stated as follows:
0.5 cm
**Step \[1\] Initialize:** Let $\bar{y}$ be the lower bound of the variables in , and initialized as $\bar{y}_i =\lfloor \hat{y_i}\rfloor, \, \, \forall i =1 \ldots n_2$. Let $K' \subseteq K$ be the subset of indices for the binding constraints at current solution $\hat{y}$. Let $I$ be the set of variable indices, and $K$ be the set of constraint indices for . **Step \[2\] Compute Distance $d_{ij}$:** $d_{ij} = 0, f_k = 0$ (a) *Assign righthand side for constraint index k:* $f_k \leftarrow \tau_k$ (b) *Projections for all other indices except indices i and j:* $f_k \leftarrow f_k - w_{kt} \hat{y}_t$\
(c) *Compute Distance:* $r_k \leftarrow f_k - w_{ki} \bar{y}_i$ $d_{ij} \leftarrow \min \left\lbrace r_k /w_{kj},u_j\right\rbrace $\
\
\
**Step \[3\] Evaluate Bounds:** $z= 0$ $\alpha = 0$ (d) *Evaluate $d_{ij}$:* $\bar{y}_i \leftarrow \bar{y}_i -1$; $\alpha = 1$; $\bar{y}_j \leftarrow \bar{y}_j -1$; $\alpha = 1$; (e) *Check for Integer Points:* $b \leftarrow 1$ $r^{(1)}_k \leftarrow f_k - w_{ki} (\bar{y}_i-b)$; $d^{(1)}_{ij} \leftarrow r^{(1)}_k /w_{kj}$; $\bar{y}_i \leftarrow \bar{y}_i -b$; $\alpha = 1$;\
$ b \leftarrow b + 1$; Re-evaluate $d_{ij}$ using the Step \[2c\]; $z \leftarrow z + 1$;\
\
\
**Step \[4\] Use the computed lower bound** $\bar{y}_i$ for the variable index $i \in I$ in FCG.
In Algorithm \[alg:ISG\], we initialize $\bar{y}_i =\lfloor \hat{y}_i\rfloor$ to obtain the set $F_{\upsilon}^{IP}$. In each iteration, we evaluate two components of $y$ along $i$ and $j$, and the lower bounds for the components are decreased based on the distance of the components from the binding constraints. Each pair of components, $y_i$ and $y_j$ are evaluated in two-dimensional (2D) space. The pair of components $(y_i,y_j)$ are evaluated in the 2D space with the criterion that $\bar{y_i}$ and $\bar{y_j}$ provide at least one integer point for the generation of the valid inequality in $i^{th}$ direction. We also make sure that $\bar{y_i}$ and $\bar{y_j}$ does not remove any integer point in $C(F^{IP})$ so that the generated valid inequality does not cut off any optimal solution.
In Step \[1\], we initialize the parameters using LP-relaxation solution for subproblem given as $\hat{y}_i$ for $i \in I$, where $I$ is the set of decision variable indices. Furthermore, $\bar{y}_i$ is the parameter of the algorithm which we intend to use as lower bound for the $i^{th}$ component in the subproblem . We would like to *increase* the value of $\bar{y}_i$ without cutting off any integer solution for the original problem. Initially $\bar{y}_i$ is set to $\lfloor \hat{y}_i\rfloor$. In Step \[2a\], the right hand side of binding constraint $k \in K'$ is assigned to parameter $f_k$. In Step \[2b\], for any $i,j \in I,$ such that $i \neq j$, we calculate the distance $d_{ij}$ in $(i,j)$ space using equation (\[dval\]). We then evaluate $d_{ij}$ in Step \[2c\]. The parameter $d_{ij}$ is the measure of distance from the other component’s axis to the binding constraint. If $d_{ij} < 1$, then it indicates the absence of an integer point along $i^{th}$ axis, then $\bar{y_i}$ is decreased by one, and $\bar{y}$ will be evaluated again for the binding constraint $k$ using the expanded set $F_{\upsilon}^{IP}$. Hence, we start with a smallest set of integers based on $ \lfloor \hat{y}_i\rfloor$, and as the algorithm progresses, the set $F_{\upsilon}^{IP}$ is expanded by decreasing $\bar{y}$ based on Corollary \[corr1\] and Lemma \[lemma2\].
![Illustration of ISG algorithm[]{data-label="brp-fig1-sp"}](figure2.pdf "fig:")\
In step \[3d\], we check property *(ii)* of Corollary \[corr1\]. If there are any integer points along the component $i$, then $\bar{y}_i$ is reduced to accommodate additional integer points into the set $F_{\upsilon}^{IP}$. We make sure that the reduced set $C(F_{\upsilon}^{IP})$ is sufficient to get the required valid inequality. The complexity of the algorithm is ${\ensuremath{\operatorname{O}\bigl(n^3m\bigr)}}$, where $n$ is the number of variables, and $m$ is the number of constraints. The 2D construction from the entire polytope is similar to variable elimination method described in [@dantzig1973fourier].
Numerical Examples
------------------
In this section, we demonstrate the ISG algorithm using numerical examples. In Example 1, we use an IP subproblem with two decision variables and one constraint to illustrate the generation of a reduced integer set using the ISG algorithm. Example 2 demonstrates the algorithm for a subproblem with two decision variables and two constraints. In both Example 1 and 2, Step\[3e\] of the ISG algorithm is not required, therefore we use Example 3 to demonstrate the significance of this step.\
**Example 1:** Consider the following IP subproblem: $$\begin{alignedat}{2}\label{eq-ip1}
\text{IP1: } \operatorname*{\mbox{Max }}& y_1 + y_2 && \\
\text{s.t. } & 0.4 y_1 + y_ 2 \leq 3.4 && \\
& 0 \leq y_1,y_2 \leq 3 && \\
& y_1,y_2 \in \mathbb{Z}, &&
\end{alignedat}$$
The LP-relaxation to the problem (\[eq-ip1\]) has the optimal solution (3, 2.2). Thus, a valid inequality has to cut off the fractional solution. The steps of Algorithm 1 are as follows:\
- Step \[1\] Initialize: $y^p =(\lfloor 3 \rfloor,\lfloor 2.2 \rfloor) = (3, 2)$. Therefore, ${y}_1^{p} = 3$, ${y}_2^{p} = 2$, $\bar{y}_1 = 3$, $\bar{y}_2 = 2,$ based on $\hat{{y}_1} = 3$ and $\hat{{y}_2} = 2.2$.
- Step \[2\] Compute Distance $d_{12}^p$: $i=1, j=2 , k=1$, $r_1 = 3.4 - 0.4(3) = 2.2$. $d_{12}^p = \min \left\lbrace 2.2/1 - y_2^p, 3 - y_2^p \right\rbrace = 0.2 $.
- Step \[3\] Evaluate Bounds: $z=0, i=1:$ $d_{12}^p < 1$ $\Rightarrow \bar{y}_1 = \bar{y}_1-1 = 3-1=2,$ update $y^p \Leftarrow (2,2)$, and $\alpha= 1$. Since $\alpha= 1$, Re-evaluate $d_{12}^p$: $r_1 = 3.4 - 0.4(2) = 2.6$. $d_{12}^p = \min \left\lbrace 2.6/1 - y_2^p, 3 - y_2^p \right\rbrace = 0.6 $. $z=1,i=1:$ $d_{12}^p < 1$ $\Rightarrow \bar{y}_1 = \bar{y}_1-1 = 2-1=1,$ update $y^p \Leftarrow (1,2)$, and $\alpha= 1$. Since $\alpha= 1$, Re-evaluate $d_{12}^p$: $r_1 = 3.4 - 0.4(1) = 3.0$. $d_{12}^p = \min \left\lbrace 3.0/1 - y_2^p, 3 - y_2^p \right\rbrace = 1.0 $. $z=2,i=1:$ $d_{12}^p \geq 1$, then evaluate the next component $\bar{y}_2$. $\bar{y}_2 = 2$, $r_1 = 3.4 - 2 = 1.4$. $d_{21}^p = \min \left\lbrace 1.4/0.4 - y_1^p, 3 - y_1^p \right\rbrace = \min \left\lbrace 3.5 - y_1^p, 3 - y_1^p \right\rbrace = 2$. $z=0,i=2:$ Since $d_{21}^p \geq 1$, we do not make any changes to $\bar{y}_2 $.
- Step \[4\] New Lower Bounds: $y_1=1$ and $y_2=2$.
The value $d_{12}^p=0.2$ represents the distance between the point $y^p(3, 2)$ and the point $y^p(3, 2.2)$ for a binding constraint in the $y_2$ direction. Since there is no integer point in the direction, the algorithm iterates to reach the point $y^{p'}(2, 2)$, where the distance $d_{12}^{p'} = \min \left\lbrace 2.6/1 - y_2^{p'}, 3 - y_2^{p'} \right\rbrace = 0.6$. The value ‘$0.6$’ is the distance between the point $y^{p'}(2,2)$ and the binding constraint in $y_2^{p'}$ direction. Since $d_{12}^{p'} < 1$, the algorithm is continued to next iteration. The value $d_{12}^{p''} = \min \left\lbrace 3/1 - y_2^{p''}, 3 - y_2^{p''} \right\rbrace = 1$ is the distance between the point $y^{p''}(1,2)$ and the binding constraint at the point $y^{p''}(1,2)$.
The feasible set based on the new point $(1,2)$ as the origin is depicted in Figure \[brp-fig1\]-(a). The reduced feasible set is now used for the generation of a cut, which is depicted in Figure \[brp-fig1\]-(b).
![Example 1 illustration: (a) reduced integer set and (b) cut generated based on the reduced set[]{data-label="brp-fig1"}](3-a.pdf "fig:")\
**Example 2:** Consider the following IP subproblem:
\[eq-ip2\] $$\begin{aligned}
{3}
\text{IP2: } \operatorname*{\mbox{Max }}& y_1 + y_2 && \\ \label{eq-ip2a1}
\text{s.t. } & 0.4 y_1 + y_ 2 \leq 3.4 && \\ \label{eq-ip2a2}
& y_1 + 0.4 y_ 2 \leq 3.4 && \\ \label{eq-ip2a3}
& 0 \leq y_1,y_2 \leq 3 && \\ \label{eq-ip2a4}
& y_1,y_2 \in \mathbb{Z}, &&\end{aligned}$$
The LP-relaxation to the problem (\[eq-ip2\]) has the optimal solution (2.42, 2.42). Thus, a valid inequality has to cut off the fractional solution. The steps of Algorithm 1 are as follows:\
- Step \[1\] Initialize: $y^p =(\lfloor 2.42 \rfloor,\lfloor 2.42 \rfloor) = (2, 2)$. Therefore, ${y}_1^{p} = 2$, ${y}_2^{p} = 2$, $\bar{y}_1 = 2$, $\bar{y}_2 = 2,$ based on $\hat{y}_1 = 2.42$ and $\hat{{y}_2} = 2.42$.
- Step \[2\] Compute Distance : $i=1, j=2 , k=1$, $r_1 = 3.4 - 0.4(2) = 2.6$. $d_{12}^p = \min \left\lbrace 2.6/1 - y_2^p, 3 - y_2^p \right\rbrace = 0.6 $.
- Step \[3\] Evaluate Bounds : $z=0, i=1, k=1:$ $d_{12}^p < 1$ $\Rightarrow \bar{y}_1 = \bar{y}_1-1 = 2-1=1,$ update $y^p \Leftarrow (1,2)$, and $\alpha= 1$. Since $\alpha= 1$, Re-evaluate $d_{12}^p$: $r_1 = 3.4 - 0.4(1) = 3.0$. $d_{12}^p = \min \left\lbrace 3.0/1 - y_2^p, 3 - y_2^p \right\rbrace = 1.0 $. $z=1, i=1, k=1:$ $d_{12}^p \geq 1$, then evaluate the next constraint. $i=1, j=2 , k=2$, $r_2 = 3.4 - 1 = 2.4$. $d_{12}^p = \min \left\lbrace 2.4/0.4 - y_2^p, 3 - y_2^p \right\rbrace = 1.0 $. $z=0, i=1, k=2:$ $d_{12}^p \geq 1$, and all the constraints are evaluated, so we move to the next component $\bar{y}_2$. $z=0,i=2, j=1 , k=1$, Similarly, for $\bar{y}_2 = 2$, $r_1 = 3.4 - 2 = 1.4$. $d_{21}^p = \min \left\lbrace 1.4/0.4 - y_1^p, 3 - y_1^p \right\rbrace = \min \left\lbrace 3.5 - y_1^p, 3 - y_1^p \right\rbrace = 2$. $d_{21}^p \geq 1$, then evaluate the next constraint. $z=0, i=2, j=1 , k=2$, $r_2 = 3.4 - 0.4(2) = 2.6$. $d_{21}^p = \min \left\lbrace 2.6/1.0 - y_1^p, 3 - y_1^p \right\rbrace = \min \left\lbrace 2.6 - y_1^p, 3 - y_1^p \right\rbrace = 1.6$. Since $d_{21}^p \geq 1$, we don’t make any changes to $y_2$.
- Step \[4\] New Lower Bounds: $y_1=1$ and $y_2=2$.
The value $d_{12}^p=0.6$ represents the distance between the point $y^p(2, 2)$ and the point $y^p(2, 2.6)$ for the binding constraint in $y_2$ component’s direction. Since there is no integer point in the direction, the algorithm iterates to reach the point $y^{p'}(1, 2)$, where the distance $d_{12}^{p'} = \min \left\lbrace 3.0/1 - y_2^{p'}, 3 - y_2^{p'} \right\rbrace = 1.0$. The value ‘$1.0$’ is the distance between the point $y^{p'}(1,2)$ and the binding constraint in $y_2^{p'}$ direction. Since $d_{12}^{p'} \geq 1$, the algorithm considers the next constraint. Similarly, the other index $y_2$ is evaluated in $y_1$’s directions for both the constraints and .
The feasible set based on the new origin $(1, 2)$ is depicted in Figure \[brp-fig2\]-(a). The feasible set is used for the generation of valid inequalities, and the generated cut is shown in Figure \[brp-fig2\]-(b). After performing to the ISG algorithm, the new origin is (1, 2).
![Example 2 illustration: (a) reduced integer set and (b) cut generated based on the reduced set[]{data-label="brp-fig2"}](3-b.pdf "fig:")\
**Example 3:** Consider another IP subproblem given as follows:: $$\begin{alignedat}{2}\label{eq-ip3}
\text{IP3: } \operatorname*{\mbox{Max }}& 1.2 y_1 + 3.4 y_2 && \\
\text{s.t. } & 6 y_1 + 5 y_ 2 \leq 37.4 && \\
& 0 \leq y_1,y_2 \leq 5 && \\
& y_1,y_2 \in \mathbb{Z}, &&
\end{alignedat}$$
The LP-relaxation for problem (\[eq-ip3\]) gives the solution $(5,1.48)$. Using Step \[3d\] of the ISG algorithm, the new origin for FCG procedure is shifted to $(4,0)$. However, based on $F_{\upsilon}^{IP}$, the generated valid inequality removes an integer point $(2,5)$ from the solution space. Hence, we use Step \[3e\] to prevent any possibility of removing off integer points from the solution space based on the current reference obtained from step \[3d\]. Thus Step \[3e\] gives the new reference $(2,0)$ based on the possible integer points in $F^{IP}$. The reduced solution space based on the new origin $(4,0)$ without Step (e) is shown in Figure \[brp-fig3\]-(a). The reduced solution space based on Step \[e\] for the generation of a valid inequalities is shown in Figure \[brp-fig3\]-(b).
![Example 3 illustration: (a) reduced integer set and (b) cut generated based on the reduced set[]{data-label="brp-fig3"}](3-c.pdf "fig:")\
The newly computed lower bounds from ISG algorithm will be used for the generation of valid inequalities. The objective of ISG algorithm is to obtain a smaller reduced set $F_{\upsilon}^{IP} \subseteq F^{IP}$ which further gives $C(F_{\upsilon}^{IP}) \subseteq C(F^{IP})$. A reduced set $F_{\upsilon}^{IP}$ is expected to provide better runtime for the generation of a valid inequality.
Fenchel Decomposition Algorithm {#sec-alg}
===============================
To demonstrate how to use the ISG method within an SMIP algorithm, we will apply this method to stage-wise Fenchel decomposition (SFD) for solving SMIPs [@Ntaimo10]. SFD adopts the Benders’ decomposition setting with $x$ as the first-stage decision variable in the master problem, and $y(\omega)$ as the second-stage decision variable in the subproblem. In SIP2, instead of working with the IP subproblem directly, SFD seeks to find the optimal solution via a cutting plane approach on a partial LP-relaxation of SIP2 where only the subproblems are relaxed. Fenchel decomposition (FD) cuts are sequentially generated to recover (at least partially) the convex hull of integer points for each scenario subproblem feasible set. If a subproblem LP has a non-integer solution, an FD is generated and added to cut off the fractional solution. FD cuts are capable of recovering faces of the convex hull of integer programs, which is the special structure for SIP2. The goal is to construct the convex hull of integer points in the neighborhood of the optimal solution so that by solving subproblem LPs with sufficient FD cuts added, we can find the optimal solution without having to resort to a branch-and-bound scheme to guarantee optimality.
Algorithm {#sec-alg}
---------
At a given iteration $k$ of the SFD cutting plane algorithm, the master problem takes the following form:\
\[eq-master-1\] $$\begin{aligned}
z^k = \operatorname*{\mbox{Max }}\ & c^\top x + \theta \notag \\
\text{s.t. } & Ax \leq b \notag\\
& (\eta^{t})^\top x + \theta \geq \gamma^{t} , \ t \in 1,...,k \label{eq-master-1a} \\
& x \in \{0,1\}. \notag
\end{aligned}$$
In the master problem , $\theta$ is the optimality cut decision variable, $\eta \in \Re^{n1}$ is the optimality cut coefficient vector, and $\gamma \in \Re$ is the right hand side. Constraints are the *optimality* cuts, which are computed based on the optimal dual solutions of all the LP-relaxation subproblems. Optimality cuts approximate the value function of the second-stage subproblems. For a first-stage solution $x^k$ from the master problem , the subproblem for each scenario ${\omega}\in \Omega$ is given as follows:
\[eq-suSIP2\] $$\begin{aligned}
\text{SP}({\omega}):\Phi^k_{LP}(\rho,\tau,{\omega}) = \operatorname*{\mbox{Max }}\ & \rho^{\top} y({\omega}) \notag\\
\text{s.t. } & W y({\omega}) \leq \tau \notag \\
& \beta^t({\omega})^{\top} y({\omega}) \leq g({\omega},\beta^t({\omega})), \ t \in \Theta({\omega}) \label{eq-suSIP2a}\\
& 0 \leq y({\omega}) \leq u \notag \\
& y({\omega}) \geq 0. \notag
\end{aligned}$$
Constraints are the Fenchel cuts, and $\Theta({\omega})$ is the index set for algorithm iterations at which a Fenchel cut is generated for each ${\omega}\in \Omega$. Next, we describe how these cuts are generated.
The SFD algorithm starts by initializing data in Step \[1\] and getting an initial solution by solving the LP-relaxation of SIP2 in Step \[2\]. If the initial solution satisfies the integrality restrictions for all subproblems in Step \[3\], i.e., $x \in X$ and $y(\omega) \in Y, \, \forall \, {\omega}\in \Omega$, then the solution is declared $\epsilon$-optimal, and the algorithm terminates. Otherwise, the algorithm continues by calculating and storing the optimality cut coefficients for all subproblems with an integer solution in Step \[4\].
0.5 cm
**Step \[1\] Initialization:** set $k \leftarrow 0, \epsilon > 0, LB \leftarrow -\infty$ and $UB \leftarrow \infty$. **Step \[2\] Get initial solution:** Solve problem (\[eq-master-1\]-\[eq-suSIP2\]) using the L-shaped algorithm to get solution $(\hat{x}^0,\hat{y}^0(\omega))$, objective function value $\varphi^0 = \sum_{\omega \in \Omega}p_{\omega} \Phi^k_{LP}(\rho,\tau,{\omega})$, and dual solutions $\hat{\pi}^k(\omega)$ for each $\omega \in \Omega$.\
**Step \[3\] Check solution integrality:**\
Report $(\hat{x}^k,\hat{y}^k(\omega))$ as optimal.\
**Stop**. **Step \[4\] Calculate and store optimality cuts coefficients for scenarios with integer solution** Calculate and store optimality cut coefficients $\eta(\omega)^k \leftarrow \hat{\pi}(\omega)^{k\top} T(\omega)$ and $\gamma(\omega)^k \leftarrow \hat{\pi}(\omega)^{k\top} h(\omega)$. **Step \[5\] Fenchel cuts and optimality cuts generation:**\
*Compute scenario Fenchel cut coefficients:* Run ISG to get $F_{\upsilon}^{IP}$ and use FCG based on $F_{\upsilon}^{IP}$ to get $\beta(\omega)^k$ and $g(\omega,\beta(\omega)^k)$. Add the cut $\beta(\omega)^{k\top} y(\omega) \leq g(\omega,\beta(\omega)^k)$ to subproblem . Solve the updated subproblem $\text{SP}(\omega)$ and get updated subproblem dual solution $\hat{\pi}(\omega)^k$. Update optimality cut coefficients $\eta^k \leftarrow \eta^k + p_{\omega} \cdot (\hat{\pi}(\omega)^{k})^\top T(\omega)$ and $\gamma^k \leftarrow \gamma^k + p_{\omega} \cdot (\hat{\pi}(\omega)^{k})^\top h(\omega)$.
**Step \[6\] Add optimality cut** $\eta^k x + \theta \geq \gamma^k$ to master problem and update iterator: set $k \leftarrow k+1$. **Step \[7\] Solve master problem** to get a new first-stage solution $\hat{x}^k$ and objective value $z^k$. **Step \[8\] Update lower bound:** Set $LB \leftarrow max\{LB, z^k\}$ **Step \[9\] $\epsilon$-optimality check:** Go to Step \[14\]. **Step \[10\] Solve subproblems:** Solve subproblem to get updated subproblem solution $\hat{y}(\omega)^k$, optimal value $\Phi^k_{LP}(\rho,\tau,{\omega})$ and dual solution $\hat{\pi}(\omega)^k$. Calculate and store optimality cut coefficients $\eta(\omega)^k \leftarrow \hat{\pi}(\omega)^{k\top} T(\omega)$ and $\gamma(\omega)^k \leftarrow \hat{\pi}(\omega)^{k\top} h(\omega)$. **Step \[11\] Subproblem solutions integrality check:** Go to Step \[5\].
**Step \[12\] Update solution and bound information:** $\,\,\,\,$ Update incumbent solution: $x^* \leftarrow x^k$. $\,\,\,\,$ Update upper bound: $UB \leftarrow min \{UB, c^{\top}{x}^k + \sum_{\omega \in \Omega} p_{\omega} \Phi^k_{LP}(\rho,\tau,{\omega}) \}$. **Step \[13\] $\epsilon$-optimality check:** Go to Step \[6\]. **Step \[14\] Declare** $x^*$ $\epsilon$-optimal. **Stop.**
For subproblems with a solution that does not satisfy the integrality requirements, Fenchel cut coefficients $\beta^k(\omega)$, and the right hand side $g(\omega, \beta^k({\omega}))$ are computed for the iteration $k$ in Step \[5\]. A Fenchel cut is added to subproblem $\text{SP}(\omega)$. Next, the dual solution obtained by solving the subproblem is used to generate the optimality cut coefficients. Once all subproblems have been solved at a given iteration, the optimality cut is added to the master problem in Step \[6\]. The iteration counter $k$ is incremented by one, and the master problem is solved again in Step \[7\] to get an updated first-stage solution and objective value.
The lower bound $LB$ is updated in Step \[8\]. The gap between the lower bound $LB$ and the upper bound $UB$ is verified in Step \[9\]. If this gap is small enough, then the incumbent solution is declared $\epsilon$-optimal in Step \[14\], and then the algorithm terminates. Otherwise, all the subproblems are solved again, and optimality cut coefficients are updated for subproblems with an integer solution in Step \[10\]. The integrality of subproblem solutions is verified in Step \[11\]: if a subproblem solution for any given $\omega$ is not integral, the algorithm returns to Step \[5\], to generate and add Fenchel cuts to the subproblems with a non-integer solution, and compute their optimality cut coefficients. Otherwise, the incumbent solution $x^*$ and the upper bound $UB$ are updated in Step \[12\]. The optimality check is done again in Step \[13\]: if it is satisfied, the incumbent solution is $\epsilon$-optimal, and the algorithm is terminated. Otherwise, the algorithm returns to Step \[6\], and the optimality cut is added to the master problem, and its solved again. The algorithm is continued until the termination condition is satisfied.
Fenchel Cut Generation with Reduced Integer Set
-----------------------------------------------
For the sake of completeness, we next present the Fenchel cut generation procedure which is based on [@ntaimo2013fenchel]. The procedure uses a master problem to construct a linear approximation of the subproblem space while the subproblem returns feasible integer points from $F_{\upsilon}^{IP}$. The master problem is given as,
$$\label{eq-cutlpmaster1}
\begin{alignedat}{2}
\delta^{(t)} = && \ \ \operatorname*{\mbox{Max }}_{\beta(\omega) \in \Pi^{\beta}} & \ \ \theta \\
&& \text{s.t. } \ \ \ & -\theta + (\hat{y(\omega)} - y(\omega)^{(\nu)})^\top \beta(\omega)^{(\nu)} \geq 0, \ \nu=1,\cdots,t.
\end{alignedat}$$
where the subproblem at iteration $t$ is solved to get $y(\omega)^{(\nu)}$. $$\label{gee1}
g(\omega,\beta(\omega)) =
\operatorname*{\mbox{Max }}_{y(\omega) \in F_{\upsilon}^{IP}}\left\{ \beta(\omega)^{\top} y(\omega) \right\}.$$
Solving the problems and iteratively will give the optimal $\beta(\omega)$ and $g(\omega,\beta(\omega))$. The algorithm is stated as follows:
0.5 cm
**Step \[1\] Initialization:** Set $t \leftarrow 0, \epsilon > 0, LB \leftarrow -\infty, UB \leftarrow \infty$, and get an initial point $\beta(\omega)^{(0)} \in \Pi^{\beta}$. **Step \[2\] Lower Bound:** $\,\,\,\,$ Use $\beta(\omega)^{(t)}$ to solve problem (\[gee1\]) and get solution $y(\omega)^{(t)}$ and the corresponding objective value $g(\omega,\beta(\omega)^{(t)})$. **Compute lower bound:** $\,\,\,\,$ Let $d^{(t)} \leftarrow (\hat{y(\omega)}-y(\omega)^{(t)})^\top\beta(\omega)^{(t)}$. $\,\,\,\,$ Set $l^{(t+1)} \leftarrow max\{d^{(t)},l^{(t)}\}$. **Update incumbent solution:** $\,\,\,\,$ Set $\mu \leftarrow d^{(t)}$ and $(\beta(\omega)^{*},g(\omega,\beta(\omega)^{*})) \leftarrow (\beta(\omega)^{(t)},g(\omega,\beta(\omega)^{(t)}))$. Use $\hat{y(\omega)}$ and solution ${y}(\omega)^{(t)}$ from subproblem (\[gee1\]) to form and add constraint to the problem (\[eq-cutlpmaster1\]). **Step \[3\] Upper Bound:** $\,\,\,\,$ Solve problem (\[eq-cutlpmaster1\]) to get an optimal solution $(\theta^{(t)},\beta(\omega)^{(t)})$. **Compute upper bound:** $\,\,\,\,$ Set $u^{(t+1)} \leftarrow min\{\theta^{(t)}, u^{(t)}\}$. The incumbent solution is optimal. **Stop.** Set $t \leftarrow t+1$ and go to \[2\].
In Step \[1\] we initialize the parameters for the algorithm. Each component of $\beta(\omega)^{(0)}$ is arbitrarily initialzed to a value with their bounds. Since a subproblem integer program has to be solved many times to generate Fenchel cuts, a linearly constrained domain for $\Pi^{\beta}$ such as the $L^1$ unit sphere or $L^2$ can also be used. Step \[2\] uses $\beta(\omega)^{(0)}$ as coefficients, then subproblem (\[gee1\]) is solved, and the corresponding objective value is stored. Due to ISG, subproblem (\[gee1\]) can be evaluated using $F_{\upsilon}^{IP}$ instead of $F^{IP}$. It should be noted that subproblem (\[gee1\]) is solved as an IP, so the solution $y(\omega)^{(t)}$ is integral. The bounds and incumbent solutions are updated in Step \[2\]. Based on the solution $y(\omega)^{(t)}$ from subproblem (\[gee1\]), the cut is added to master problem (\[eq-cutlpmaster1\]). In Step \[3\], master problem (\[eq-cutlpmaster1\]) is solved and the termination condition is checked. Based on the termination condition, the algorithm either stops or continues.
Computational Study {#sec-lcomptns}
===================
To gain insights into the benefits of imbedding integer set reduction into a cutting plane algorithm, we performed a computational study based on an implementation of the SFD algorithm with the option of turning on ISG. The algorithm was implemented in C++ using the CPLEX 12.5 Callable Library [@CPLEX] in Microsoft Visual Studio 2010. Computations were performed on an ACPI x64 computer with IntelXeonProcessor E5620 (2.4 GHz) and 12GB RAM. CPLEX MIP and LP solvers were used to optimize the master program and subproblems. Two test sets were created based on the two-stage multidimensional knapsack problem. The SFD algorithm was run to solve the test instances to optimality or stopped when a CPU time limit of 3,600 seconds (s) or 7,200s was reached. As a benchmark, the CPLEX MIP solver was applied to the deterministic equivalent problem (DEP) of each test instance. Next we describe how the multidimensional knapsack test instances were generated and then report the computational results in Section \[sec-results\].
Test Instances Generation {#sec-testgen}
-------------------------
We created a test problem with knapsack constraints in the first-stage, and both knapsack and assignment constraints in the second-stage. The first-stage problem is given as follows: $$\begin{alignedat}{2}\label{eq-smip1}
\operatorname*{\mbox{Max }}& \sum_{i=1}^{n_1} c_i^\top x_i + \mathcal{Q}_{E}(x) && \\
\text{s.t. } & \sum_{i=1}^{n_1} x_i \leq b && \\
& x_i \in \{0,1\}, \, \, \forall i=1\ldots n_1. &&
\end{alignedat}$$
In problem , $c \in \Re^{n_1}$ is the first-stage cost vector, and $b \in \Re$ is the first-stage right hand side. The function $\mathcal{Q}_{E}(x)$ is the expected recourse function given as $$\label{eqn2}
\mathcal{Q}_{E}(x) = \mathbb{E}_\omega \Phi(q({\omega}),h({\omega}),T({\omega})x),$$ where ${\omega}$ is a multivariate random variable and $\mathcal{Q}_{E}(.)$ denotes the mathematical expectation operator satisfying $\mathbb{E}_\omega \left[ \mid\Phi(q({\omega}),h({\omega}),T({\omega})x)\mid\right] < \infty$. The underlying probability distribution of ${\omega}$ is discrete with a finite number of realizations (scenarios) with sample space $\Omega$, and corresponding probabilities $p_{\omega}, \omega \in \Omega$. Thus for a given scenario ${\omega}\in \Omega$, the recourse function $\Phi(q({\omega}),h({\omega}),T({\omega})x)$ is given by the following second-stage MIP: $$\begin{alignedat}{2}\label{eq-SIP2}
\Phi(q({\omega}),h({\omega}),T({\omega})x) = \operatorname*{\mbox{Max }}& \sum_{i=1}^{n_2} q({\omega})^{i\top} y({\omega})^{i} && \\ \text{s.t. } & \sum_{i=1}^{n_2} w^{ij} y({\omega})^{i} \leq \sum_{i=1}^{n_1} m \cdot t^i x_i \qquad \forall j=1\ldots m_1 && \\ & \sum_{i=1}^{n_2} v^{ik} y({\omega})^{i} \leq h({\omega})^k \qquad \forall k=1\ldots m_2 && \\ & \sum_{i=1}^{n_2} u^{i\ell} y({\omega})^{i} = r^\ell \qquad \forall \ell=1\ldots m_3 && \\ & 0 \leq y({\omega})^i \leq u^i, y({\omega})^i \in \mathbb{Z^+} \qquad \forall i=1\ldots n_2. &&
\end{alignedat}$$ In formulation , $y({\omega})$ is the recourse decision vector, $q({\omega}) \in \Re^{n_2}$ is the recourse cost vector, $w^{ij} \in \Re^{+}$ is a fixed recourse parameter, and $t^{i} \in \Re^{+}, v^{ik} \in \Re^{+}, u^{i\ell} \in \Re^{+}$ is a parameter taking values 0 or 1, $m$ is a constant, and $h({\omega})^{k} \in \Re^{m_2}, r^{\ell} \in \Re^{m_3}$ are the right hand side parameters. The decision vector $y({\omega})$ is bounded above by vector $u$. Finally, $\mathbb{Z^+}$ is the set of nonnegative integers. Observe that formulation - has knapsack constraints in both the first- and second-stages.
In a supply chain context, the first-stage decision vector $x$ specifies the selection of facilities, mode of transportation, and/or resources. For a realization ${\omega}$, the second-stage decision vector $y({\omega})$ could be the amount of products produced or transported based on the strategic decision $x$ from the first-stage. Additionally, knapsack-type constraints are added to represent capacity limitations in the second-stage.
Test instance data were randomly generated using the uniform distribution ($\mathcal{U}$) with different parameter values. The knapsack weights were generated by sampling from $\mathcal{U}(2, 8)$. Objective function coefficients were generated with the first-stage costs chosen so that they are much larger than second-stage costs. The first- and second-stage objective function coefficients were generated by sampling from $\mathcal{U}(0, 1500)$ and $\mathcal{U}(10, 20)$, respectively. To generate tighter knapsack constraints, the right hand side value of each constraint was generated by finding the maximum knapsack weight ($W_{max}$) for the constraint, and then sampling from $\mathcal{U}(2+(2W_{max}*v_{ub}), 4W_{max}*v_{ub})$, where $v_{ub}$ is the upper bound for the integer variables. We assume that each scenario has equal probability of occurrence.
The problem characteristics are given in Table \[tab:kpdimension\]. The columns of the table are problem name, ‘Scens’ is the number of scenarios, ‘Bvars’ is the number of binary variables, ‘Constr’ is the number of constraints, and ‘Nzeros’ is the number of non-zero elements for each of the problem instances. The problem name has the form $k.m.n.S$, where $k$ stands for ‘knapsack’, $m$ and $n$ is the number of first- and second-stage decision variables, respectively, and $S$ is the number of scenarios. Two test sets, ‘Set 1’ and ’Set 2’ were created, where the first set has relatively smaller size instances compared to the second set. Specifically, Set 1 has test instances with 10 binary and 20 general integer variables in the first- and second-stage, respectively. In this set the test instances were created for 50, 100, 150, and 200 scenarios. Set 2 has test instances with 20 binary and 30 general integer variables in the first- and second-stage, respectively. In the instances, first-stage has 10 constraints and second-stage has 20 and 30 constraints in Set 1 and Set 2, respectively. Test instances for this set were created for 500, 1,000, 1,500, and 2,000 scenarios. Five randomly generated replications were created for each instance size to avoid pathological cases.
[|l|c|c|c|c|c|]{} Problem & Scens & Bvars & Ivars & Constr & Nzeros\
\
k.10.20.50 & 50 & 10 & 1,000 & 1,010 & 12,510\
k.10.20.100 & 100 & 10 & 2,000 & 2,010 & 25,010\
k.10.20.150 & 150 & 10 & 3,000 & 3,010 & 37,510\
k.10.20.200 & 200 & 10 & 4,000 & 4,010 & 50,010\
\
k.10.30.50 & 50 & 10 & 1,500 & 1,510 & 16,510\
k.10.30.100 & 100 & 10 & 3,000 & 3,010 & 33,010\
k.10.30.150 & 150 & 10 & 4,500 & 4,510 & 49,510\
k.10.30.200 & 200 & 10 & 6,000 & 6,010 & 66,010\
Results {#sec-results}
-------
Detailed computational results for Set 1 and Set 2 are reported in Tables \[tab:t1\] and \[tab:t2\], respectively. In the tables, ‘SFD’ and ‘SFD-R’ represent the results using SFD algorithm without ISG and with ISG algorithm, respectively. In the tables the column ‘Instance’ is the name of the test instance. The three columns under SFD and SFD-R, respectively, are as follows: ‘MIPs’ is the number of MIPs solved using the respective procedure; ‘FCuts’ is the number of Fenchel cuts; and ‘%Gap’ is the percentage gap between the lower bound (LB) and the upper bound (UB) value after the stipulated runtime (3600s for Set 1 and 7200s for Set 2). Finally, the last column ‘%Gap’ shows the CPLEX MIP gap after solving the DEP for the designated amount of time.
We can see from Tables \[tab:t1\] and \[tab:t2\] that none of the algorithms is able to solve any single test instance to optimality within the allotted time, an indication of the difficulty of these instances. However, both the SFD and SFD-R algorithms are able to obtain better bounds than using the direct solver applied to the DEP. In Table \[tab:t1\] we see that the SFD-R algorithm provides better performance on average over the SFD algorithms in terms of the percentage gap. The gains are much more significant for Set 1 than for Set 2. The results show that incorporating ISG in the SFD algorithm provides gains in gap reduction for both Set 1 and Set 2. This is an indication that reducing the integer set required for generating cuts in an SMIP algorithm can lead to better bounds for SMIP. To see the performance of each method on a given test instance, we plotted the percentage gap versus the instance in Figure \[fig:gap\]. The graph clearly shows that the SFD-R algorithm provides the best performance overall.
![Instances Gap (%) Runtime (Set 1 - One Hour, Set 2 - Two Hours )[]{data-label="fig:gap"}](gap51.pdf)
We also wanted to look at the number of subproblem MIPs that were solved in the FCG routine under each algorithm. Solving more cut generation MIPs implies fast performance in terms of generating a cut. The results are shown in Figure \[fig:mips\], where ‘\#MIPs’ represents the number of MIPs solved using the SFD (without ISG) and SFD-R (with ISG) algorithms. The results clearly show that the SFD-R algorithm solves relatively more MIPs, and thus generates more FD cuts (Figure \[fig:fc\]), than the SFD algorithm. Finally, we should point out that the larger size test instances (Set 2) generally requires more cuts than the smaller size test instances (Set 1).
![\# MIPs Solved[]{data-label="fig:mips"}](MIPS2.pdf)
![\# Fenchel Cuts[]{data-label="fig:fc"}](Fcuts2.pdf)
Conclusion {#sec-lconcl}
==========
This work introduces a new integer set reduction procedure for cutting plane methods for SMIP with general integer variables in the second-stage. Example illustrations of the new method in the context of generating Fenchel cutting planes are given. The method is then incorporated into the Fenchel decomposition algorithm for SMIP and a computational study is performed to assess the benefits of the new approach. The results from the computational study show that incorporating integer set reduction in the Fenchel decomposition method leads in having better bounds and provides better performance than a direct solver applied to the deterministic equivalent problem. Also, more cuts are generated in a given time period when integer set reduction is used as opposed to when it is not used. Future work along this line of work include extending the integer set reduction procedure to SMIP with arbitrary or general recourse matrices. Another extension is to incorporate and evaluate the new procedure in other cutting plane methods for SMIP such as disjunctive decomposition and dual decomposition.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
T. Garel [^1]\
Service de Physique Théorique, CEA/DSM/SPhT\
Unité de recherche associée au CNRS\
91191 Gif sur Yvette Cedex, France\
title: 'A Newtonian pre-introduction to gravitational lenses'
---
[**Abstract:**]{} Understanding the deflection of light by a massive deflector, as well as the associated gravitational lens phenomena, require the use of the theory of General Relativity. I consider here a classical analogy, based on Newton’s equation of motion for massive particles. These particles are emitted by a distant source and deflected by the gravitational field of a (opaque) star or of a (transparent) galaxy. The dependence of the deviation angle $D$ on the impact parameter $b$, and the - Euclidean - geometry of the (source, deflector, earth) triplet, imply that different particle trajectories may reach an earth based observer. Since $D(b)$ does not depend on the mass of the particles, a (Newtonian) flavor of gravitational lenses phenomena is naively obtained by setting the particles’ velocity equal to the speed of light. Orders of magnitude are obtained through this classical approach, and are compared to the General Relativity results.
\
\
Introduction {#intro}
============
Gravitational lenses phenomena [@Mis_Tho_Whe; @Sch_Ehl; @Zak_Saz; @Bou_Kan; @Adl_Bar; @Wamb] rest on the gravitational deflection of light, and their explanation require the use of General Relativity (the gravitational deflection of light by the sun, of order $10^{-5}$ rd, was indeed a major issue at the very beginning of the theory). This theory uses a rather heavy technical machinery, which is not easily accessible to undergraduate students. One of the motivations of the present paper is to offer, through a classical approach, a feeling (and orders of magnitude) for gravitational lenses phenomena.
Apart from extreme cases (neutron stars, black holes,...), the gravitational defection of light in General Relativity (GR) is small, and can be described by a weak gravity approximation [@Mis_Tho_Whe; @Sch_Ehl; @Zak_Saz]. For quasistationary isolated mass distributions, the evolution of the unit tangent vector $(\vec e(s))$ to a light ray ($\vec r(s)$) is given to lowest order (see below) by \[grav0\] [de(s) ds]{}=-[2 c\^2]{} \_U where $c$ is the speed of light, $U$ the gravitational potential created by static mass distributions, and $\bigl(\vec {\nabla}_{\perp}U=\vec
{\nabla}U -\vec e(\vec e \cdot \vec {\nabla}U)\bigr)$. The weak gravity approximation corresponds to ${U \over c^2}<<1$, and “lowest order” means that the typical velocity $v$ of the mass distribution is small compared to $c$: A more rigorous calculation indeed shows that there appear other terms in the r.h.s of equation (\[grav0\]), among which the lowest order term is of order ${\rm
O}({v \over c}))$ (see eq. 4.17, p. 124 of [@Sch_Ehl]).
Equation (\[grav0\]) bears some resemblance to a classical equation of motion; it is indeed a familiar remark that Newton’s classical equation for a massive particle in the gravitational field of a deflector leads to a mass independent deflection angle (identity of inertial and gravitational masses). In this paper, we study the classical mechanical problem defined by the triplet (particle source (S), deflector ($\Delta$), earth based observer (E)). An important ingredient of this study is the dependence of the gravitational deflection angle $D$ on the impact parameter $b$ of the particles. Since the (Euclidean) distances between (S), ($\Delta$) and (E) are finite, the exact calculation of $D(b)$ is rather involved, even if one takes advantage of the central character of the gravitational force and of some invariant properties (we use here the Runge-Lenz vector and tensor). A rather quantitative approach is used here to derive the main features of the deflection $D(b)$ for the case of spherical deflectors (mass $M$, radius $R$).
To make contact with the (GR) weak gravity approximation, we consider a similar approximation for the classical mechanical study, namely we consider the limit (Gravitational energy $<< $ Kinetic energy). Further, if ($v_{\infty}$) is a measure of the particle velocity as it leaves the source, one has at some point to make the (delicate) correspondence $v_{\infty} \to c$ in the classical mechanical problem. In this paper we follow this view of the gravitational deflection of light to obtain [*orders of magnitude*]{}, using typical undergraduate skills. It is shown that, in some cases, several particle trajectories may reach observer (E). These trajectories translate into a basic gravitational lensing effect: the source (S) may give multiple (photonic) signals for observer (E). The time delay between the reception of these different trajectories (or signals) is also considered.
The plan of the paper is as follows. The gravitational action of the deflector, such as the sun (opaque star) or the Milky Way (transparent galaxy), is studied in Section \[orders\]. For simplicity, we assume there that the source (S) and the observer (E) are both at infinity. When (S) and (E) are at a finite distance from the deflector, a little bit of geometry shows that that (S) may have multiple “images” (Section \[geometry\]). Finally, an estimate of the time lag between the reception of these “images” is obtained, via the analogy of particle mechanics and geometrical optics (Section \[time\]).
Study of the classical mechanical deflection problem {#orders}
====================================================
We consider a source (S) which emits, in an isotropic fashion, isovelocity particles ($\vert \vec v \vert =v_{\infty}$) of mass $m$. Unless otherwise stated, we will consider a point source. We focus our interest on the deflector and take the source and the observer at infinity. Let $\vec v_{\infty}=v_{\infty} \vec u_i$ be the initial velocity of a test particle. The deflector ($\Delta$) may be either opaque or transparent to the particles. The former case will be illustrated by the sun (radius $R_{\odot} \simeq 7 \cdot 10^8$m and mass $M_{\odot} \simeq 2 \cdot 10^{30}$kg), and the latter by a spherical version of the Milky Way [^2] (radius $R_G \simeq 2 \cdot 10^4$ light years (ly) [^3] and mass $M_G \simeq 10^{12} M_{\odot}$).
The opaque detector
-------------------
The impact parameter $b$ and deviation angle $D$ are defined in Figure 1. The study of the classical Kepler problem for $b>R$ [^4] can be found in standard textbooks (see e.g. [@LL]). If $({\vec u_i})$ and $({\vec u_f})$ are the unit vectors along initial and final velocities, we have $\cos D(b) ={\vec u_i} \cdot {\vec u_f}$.
The Runge-Lenz vector of the Kepler problem reads \[RungeLenz\] A=v L- [G]{}Mm [r r]{} where $\vec r$ is the particule position measured from the center of the deflector ($\Delta$), $\vec L=\vec r \times m\vec
v$ its angular momentum, and $\cal G$ the universal gravity constant. As is easily checked, $\vec A$ is an invariant vector of the motion and writing $\vec A_i=\vec A_f$ [@Basa_Bian] leads to \[devia1\] [[tan]{}]{} [D 2]{}= [[G]{}M bv\_\^2]{} The weak gravity condition means that the potential energy $m U(b)$ is a small part of the total energy, that is ${{\cal G}M \over
bv_{\infty}^2}<<1$. The deviation is then given by \[devia2\] D(b)
The (GR) result is obtained from equation (\[grav0\]) as $D_{GR}(b)
\sim {4{\cal G}M \over bc^2}$. Applying these results to the sun at grazing incidence yields $D_{GR}(R_{\odot}) \sim 8 \ 10^{-6}$ rd.
Two remarks are in order (i) The naive identification ($v_{\infty} \to c$), yields an expression which is half the exact (GR) result (ii) In classical mechanics, energy conservation implies that $v^2(r)=v^2_{\infty} +{2{\cal G}M \over r}$. If one identifies $v_{\infty}$ with $c$, the velocity $v(r)$ is greater than $c$, the (weak) correction term being of order ${{\cal G}M \over rc^2}$. For small $D$, the minimum distance $r_0$ approach is of order $b$, leading to a maximum supraluminal correction for the velocity of order ${{\cal G}M \over
bc^2}$, that is small. As a temporary conclusion, I would say that for practical purposes, the weak gravity limit of the classical case, together with the identification $v_{\infty}\to c$, yields a correct order of magnitude for the (GR) value. Some inconsistencies of classical mechanics appear in this identification, but are small in the weak gravity limit.
The transparent deflector
-------------------------
A preliminary remark is that $D(b)$ is not a monotonous function, since it vanishes both for $b=0$ and $b=\infty$. For ($b>R$), the deviation is exactly given by equation (\[devia1\]), the weak gravity approximation corresponding to equation (\[devia2\]).
The case ($b <R$) requires the study of two distinct phases of the motion, since the gravitational field $\vec g(\vec r)$ on the test particle has different expressions, depending on the particle position $\vec r$. We respectively denote these phases by (out) and (in). Their respective contribution to the total deviation $D(b)$ will be denoted by $D_{(out)}$ and $D_{(in)}$. We further assume that the deflecting galaxy is homogeneous, and neglect all collisions with the particles in phase (in).
In phase (in) of the motion, Gauss’ theorem gives \[newton1\] =g\_[(in)]{}(r)=-[[G]{}M R\^3]{}r=-\_0\^2 r This harmonic motion has period $T_0={2\pi \over
\omega_0}=2\pi \sqrt{R^3 \over {\cal G}M}$. The trajectory inside the deflector is elliptic or partially elliptic. One should express the initial and final boundary conditions to get the deflection $D_{(in)}(b)$. These boundary conditions depend on phase (out) of the trajectory ($r>R$), where Newton’s equation reads \[newton2\] =g\_[(out)]{}(r)=-[[G]{}M r\^3]{}r The total deviation $D(b)$ is given by D(b)=D\_[(in)]{}(b)+D\_[(out)]{}(b)
### Phase (out): ($r>R$)
[**[(a) Rigorous solution]{}**]{}
We want to calculate the deviation between initial (source S) and final (entry into deflector ($\Delta$)) points. The final point M has position $\vec r_{in}=\vec {OM}$ and velocity $\vec
v_{in}$ (Figure 2(a)), with $\vert\vec r_{in}\vert=R$, and $\vec
v_{in}=v_0 \vec u_{in}$. Energy conservation yields $v_0^2=v_{\infty}^2+2{{\cal G}M \over R}$. The deviation $D_{out}^{(1)}$ for this part of the (out) trajectory is given by $\cos D_{out}^{(1)}=\vec u_{i} \cdot \vec u_{in}$.
Equating the projections of the Runge-Lenz vector (eq. (\[RungeLenz\])) along $\vec u_i$ for the initial (S) and final (M ) points leads to \[RungeLenz2\] [G]{}Mm= u\_i (v\_[in]{} L-[G]{}Mm [r\_[in]{} R]{})
Defining $\cos (\pi-\Phi)=\vec u_i \cdot {\vec r_{in} \over R}$ (Figure 2(a)), we get \[dev1\] D\_[out]{}\^[(1)]{} =[[G]{}M ]{}(1-)
The exact solution of the Kepler problem (see eq (14,7), p.46 of [@LL]) yields \[phi2\] =-([[b R]{}-d ]{})+([-d ]{}) with $d={{\cal G}M \over bv_{\infty}^2}$.
Taking into account the symmetrical ($\Delta E$) contribution (deviation $D_{out}^{(2)}$) finally gives
\[dout\] D\_[out]{}(b)=2([[G]{}M ]{}(1-)) where $\Phi$ is given by (\[phi2\]).
[**[(b) Weak gravity approximation]{}**]{}
As previously stated, we expect gravitational deviations to be weak. Equation (\[dout\]) shows that the weak gravity (“small ${\cal G}$”) limit can be obtained by setting $d=0$ in eq. (\[phi2\]). The weak gravity approximation therefore reads \[dout2\] D\_[out]{}(b) 2 [[G]{}M ]{}(1-) In particular, the small $b$ limit is given by $D_{out}(b) \sim {b \over
f}$, with $f={R^2v_{\infty}^2 \over {\cal G}M}$.
### Phase (in): ($r<R$)
[**[(a) Rigorous solution]{}**]{}
The geometry of the (in) phase is shown in Figure 2 (b). The angle $\alpha$ is the angle between the radius vector $\vec r_{in}$ and the velocity $\vec v_{in}$ as the particle enters the deflector (Figure 2(a)). From the conservation of the angular momentum, one has \[angular\] mv\_b=mv\_0R where $v_0^2=v_{\infty}^2+2{{\cal G}M \over R}$.
To calculate the deviation $D_{in}(b)$ between the entry and exit points, one may solve the harmonic motion of eq. (\[newton1\]). A more convenient way is to use the Runge-Lenz matrix invariant associated with the harmonic oscillator \[RungeLenz3\] [**A**]{}=[m 2]{}(\_0\^2 **r **r+**v **v) Starting from the invariance of ${\bf A}$, simple calculations [@Siva] show that \[din2\] D\_[(in)]{}(b)=2(-) where ${\rm sin}\alpha ={bv_{\infty} \over Rv_0}$, and ${\rm tan} 2\beta={v_0^2{\rm sin} 2\alpha \over {\omega_0^2 R^2+v_0^2
{\rm cos}2\alpha}}$.********
[**[(b) Weak gravity approximation]{}**]{}
In this approximation, we obtain $\sin \alpha \simeq {b \over R}$, and \[din3\] D\_[in]{}(b) D\_[in]{}(b)=2(-) 2
### Conclusion on the transparent deflector
The total deviation $D(b)$, for $b<R$ and in the weak gravity approximation, is given by eq. (\[dout2\]) and (\[din3\]). Setting $u={b \over R}$, we have \[devia3\] D(b)=D\_[(in)]{}(b)+D\_[(out)]{}(b)=2 [[G]{}M Rv\_\^2]{} ([1-(1-u\^2)\^[3 2]{} u]{})
When $v_{\infty} \to c$, this result can be compared to the (GR) result [@Bou_Kan; @Adl_Bar]. Notwithstanding the (ubiquitous) factor of 2 between classical mechanics and (GR), eq. (\[devia3\]) is in agreement with eq. (7) of reference [@Adl_Bar]. Note that equation (\[devia3\]) can be also be written in a way similar to eq. (\[devia2\]), namely D(b)=2 [[G]{}M(b) bv\_\^2]{} where $M(b)=M (1-\bigl(1-{b^2 \over R^2}\bigr)^{3 \over 2})$ is the partial deflector mass contained in a cylinder of radius $b$.
For small $b$, one gets $D(b) \sim 3 {b \over f}$. The length $f={R^2v_{\infty}^2 \over {\cal G}M}$ can be viewed as a focal length; its order of magnitude for our model galaxy is $2.5 \ 10^9$ ly, much bigger than $R_G \sim 2 \ 10^4$ ly. Note also that $D(b)$ has a maximum for $b \simeq 0.93 R_G$, with $D_{max} \simeq 1.6 \ 10^{-5}$ rd.
Using the results of equations (\[devia2\]),(\[devia3\]), we show in Figure 3 the deviation $D(b)$ as a function of the impact parameter $b$, for opaque and transparent deflectors.
Finite distance geometry and multiple trajectories {#geometry}
==================================================
We now use our results to discuss an experimentally more relevant situation, where both the source (S) and the observer (E) are at a finite distance from the deflector ($\Delta$). The - Euclidean - geometry is shown in Figure 4. We have $r_S=S_0\Delta=SH$, $r_E=\Delta E$. Given the previous orders of magnitude, the angles such as $\beta =\widehat{SE\Delta}$, $\theta=\widehat{\Delta EY}$ and the deviation $D$, are assumed to be small. This implies in particular that $b=\Delta Y=r_E \tan \theta <<r_E, r_S$. We have from (Euclidean) geometry \[sinus\] [[sin]{} SY]{}=[[sin]{} EY]{} From triangle SHY, one has SY\^2=SH\^2+HY\^2=r\_S\^2+(r\_E -(r\_E+r\_S))\^2 yielding for small angles, $SY \simeq r_S (1+O(\theta^2,\beta^2,\theta\beta))$.
From triangle $YE\Delta$ one has EY\^2=r\_E\^2+Y\^2=r\_E\^2 (1+\^2 ) yielding for small angles, $EY \simeq r_E (1+O(\theta^2))$.
Plugging these values in eq. (\[sinus\]), we get to lowest order in the angles $\theta, \beta,..$ \[sinus2\] [ r\_S]{} Since $\widehat{SEY}=\theta-\beta \simeq {b\over r_E} -\beta$, we finally obtain \[deviation3\] D(b) ([b r\_E]{}-) Equation (\[deviation3\]), which expresses the condition that a particle emitted from (S) reaches the earth, is represented by the dotted lines in Figure 3. For an opaque deflector, one may get one or two solutions for $b$. For a transparent deflector, one may get up to three solutions for $b$. Rather than studying the full problem as a function of $\beta,
r_S,r_E,...$, we illustrate some particular situations
A generic case {#generic}
---------------
This case corresponds to a non zero $\beta$ angle (Figure 4). We focus our interest on points (1) and (2) of Figure 3, which are the intersections of the geometrical equation \[deviation4\] D(b) \~[r\_S+r\_E r\_S]{}([b r\_E]{}-) with the $b>R$ gravitational deflection $D(b)$ of Section \[orders\] (see Figure 3). Since we have $D(b) \sim
2 \ {{\cal G}M \over bv_{\infty}^2}$, setting $\theta={b \over r_E}$ leads to \^2--\_E\^2=0 where \[Einstein\] \_E\^2\~2 [[G]{}M v\_\^2]{}[r\_S r\_E(r\_S+r\_E)]{} In this case, there are two trajectories in the $S\Delta E$ plane that reach (E). In photon language, the observer sees two images $(S1)$ and $(S2)$ of the source (S), on opposite sides of the deflector, with \[traj12\] \_[1,2]{}=[ 2]{}=[b\_[1,2]{} r\_E]{} For a transparent deflector, one has another image of the source, corresponding to the point labeled (3) in Figure 3(b).
Einstein rings
--------------
This case corresponds to the alignment of (S), ($\Delta$) and (E) ($\beta=0$). Due to the symmetry of revolution around the $S\Delta E$ axis, all trajectories on the angular cone $\theta=\theta_E ={b \over r_E}$ reach the earth. In photon language, this means that the observer sees a ring image of the point source S. For a transparent deflector, one also has a direct image. For our model galaxy and $r_E \sim r_S \sim
10^9$ ly, a typical value is $\theta_E \sim 10^{-5}$ rd.
The case of a moving deflector
------------------------------
We briefly consider this case (called microlensing), because of its experimental relevance. Since a detailed comparison with the experiments require the use of General Relativity [@Mis_Tho_Whe; @Sch_Ehl; @Zak_Saz], we limit our presentation to orders of magnitude calculations. If the deflector ($\Delta$) moves, with a velocity $v_{\Delta}$, in a direction perpendicular to the (SE) axis (Figure 4), the above calculations suggest the following scenario: for $\beta=0$ (ring image), there is a sudden increase in the signal received by the observer, since two trajectories only survive for $\beta \ne 0$. Physically the transition is gradual, and the observer will receive a gravitationally enhanced signal when the position of deflector $(\Delta)$ is within a distance $b_{E} \sim r_E \theta_E$ from the full alignment position of the previous section [^5] The corresponding time interval is $t_E \sim {b_E \over
v_{\Delta}}$. For distant sources ($r_S >>r_E$), an experimental situation corresponding to a sun-like deflector, with $v_{\Delta}\sim
200$ km s$^{-1}$, $r_E \sim 3.6 \ 10^4$ ly and $v_{\infty}=c$, yields an enhanced signal during an interval $t_E \sim 10^6-10^7$ s, of order one month.
Time lags {#time}
=========
A simple optical analogy
------------------------
We have seen that several trajectories- or light rays- may reach (E) because of the gravitational deflection. Can one further extend the mechanical-optical analogy by finding the time lag between the reception of these trajectories -or light rays- ?
We first consider the transparent deflector for $b<<R$, where we found $D(b)\sim 3{b \over f}$ in section \[orders\]. This result may be compared with the optical deviation of a spherical glass lens, of radius $\rho$ and of optical index $\nu$ which reads \[optik\] D\_[opt]{}=2[-1 ]{} [b ]{}=[b f\_[opt]{}]{} where $b$ is the impact parameter of the light ray and $f_{opt}={\nu
\rho \over 2(\nu -1)}$ is the focal distance of the lens. The comparison of the gravitational and optical deviations suggests that the gravitational deviation may be understood through a gravitational index $n_{grav}$, with $n_{grav} \ne 1$.
Particule trajectories and geometrical optics
---------------------------------------------
The preceding remark can be extended and formalized as follows. Energy conservation for a central potential $U(r)$ reads, in usual polar coordinates E=[m 2]{} [v]{}\^2+ mU(r)=[m 2]{}([r]{}\^2+r\^2 \^2)+ mU(r) Defining ${\rm tan} \Psi(r)={r(\theta) \over r'(\theta)}$, we have E=[[L]{}\^2 2mr\^2[sin]{}\^2(r)]{}+mU(r) which can be rewritten as \[Bouguer\] n\_[U]{}(r) r [sin]{}(r)=([[L]{}\^2 2mE]{})\^[1 2]{} where the “index” $n_U(r)$, associated to the potential $U(r)$ is given by $n_U(r)=\sqrt{1-{mU(r) \over E}}$. Equation (\[Bouguer\]) is analogous to Bouguer’s relation for the propagation of light rays in a spherically symmetric medium of index $n_U(r)$ [@Born].
We will illustrate this analogy with the case $b>R$, where $U(r)=-{{\cal G}M \over r}$. The associated gravitational index reads \[index\] n\_[grav]{}(r)= where we have used the weak gravity approximation. Defining an analog $L_{grav}$ of the optical path, we may express the time lag $\delta t_{12}$ between the reception on earth of trajectories (1) and (2) of section \[generic\] as
\[fermat1\] t\_[12]{}=[L\_2-L\_1 v\_]{}=[\_[(2)]{} n\_[grav]{}(r) ds\_2 -\_[(1)]{} n\_[grav]{}(r) ds\_1 v\_]{} From eqs. (\[index\]) and (\[fermat1\]), one finds that $\delta
t_{12}$ is the sum of a geometrical part $\delta t_{geom}={\int_{(2)}
\ ds_2-\int_{(1)} \ ds_1 \over v_{\infty}}$ and of a gravitational part \[grav2\] t\_[grav]{}=[[G]{}M v\_\^3]{}(\_[(2)]{} [ds\_2 r]{}-\_[(1)]{} [ds\_1 r]{}) The full calculation of the integrals in (\[grav2\]) (see eq. 8.30, p.240 of [@Sch_Ehl]), yields a result that depends only logarithmically on the geometrical parameters ($\beta, r_S,...$). We therefore estimate $
\delta t_{grav} \sim {{\cal G}M \over v_{\infty}^3}$, up to a numerical factor of order one. Setting $v_{\infty}=c$, we find $\delta t_{grav} \sim 10^{-5}$ s for the sun, and $\delta
t_{grav} \sim 10^{7}$ s for our model galaxy. Experiments that confirm the double reception of the same “signal”, with a gravitational time lag of order several months, can be found in references [@Mis_Tho_Whe; @Sch_Ehl; @Zak_Saz; @Adl_Bar; @Wamb].
Finally, it of interest to note that a Fermat approach to the weak gravity approximation of (GR) yields an equation similar to equation (\[index\]), with a (GR) index $n_{GR} \sim 1+2 \ {{\cal
G}M \over c^2r}$ [@Nan_Hel; @Als].
Conclusion
==========
We have studied, at a qualitative level, a classical mechanical introduction to gravitational lens phenomena. This approach rests on the fact that the gravitational deflection of a massive particle by a deflector is independent of the particle mass. It is only an approximation to the theory of General Relativity [@Nan_Hel; @Als], but I believe that this “$\vec F=m \vec a$ optics ” [@Eva_Ros; @Bel_Rod] brings together in a very pedagogical way problems of different origins. In particular, we have derived orders of magnitude for the weak gravity case, that can be compared -up to a factor 2- to the correct (GR) results. As a caveat, we have nevertheless pointed out that the particle velocity may become (weakly) supraluminal, and this (weak) inconsistency with relativity is to be kept in mind.
Beside the study of General Relativity, the interested student can carry further the present approach in several ways. I will only quote here the modeling of gravitational lenses by optical lenses of the appropriate shape [@Adl_Bar], or the link between equations (\[newton1\]) and (\[newton2\]) stemming from conformal transformations [@Mitt_Ste].
It is a pleasure to thank F. Bernardeau for discussions.
[99]{} C. Misner, K. Thorne and J. Wheeler, [*Gravitation*]{} (Freeman, San Francisco 1973). P. Schneider, J. Ehlers and E.E. Falco, [*Gravitational Lenses*]{}, (Springer Verlag, Berlin-New York, 1992). A.F. Zakharov and M.V. Sazhin, “Gravitational microlensing”, Phys. Usp., [**41**]{}, 945-982 (1998). R.R. Bourassa and R. Kantowski, “The theory of transparent garvitational lenses”, Ap. J. [**195**]{}, 13-21 (1975). R.J. Adler, W.C. Barber and M.E. Redar, “Gravitational lenses and plastic simulators”, Am. J. Phys., [**63**]{}, 536-541 (1995) and references therein. J. Wambsganss, “Gravity’s kaleidoscope”, Sci. Am., [**285**]{}, 52-59 (2001). L. Landau and E.M. Lifchitz, [*Mécanique*]{}, Third Edition, Editions Mir, Moscou (1969). L. Basano and A. Bianchi, “Rutherford scattering via the Runge Lenz vector”, Am. J. Phys., [**48**]{}, 400-01 (1980). J. Sivardière, “Comments on the dynamical invariants of the Kepler and harmonic motions”, Eur. J. Phys., [**13**]{}, 64-69 (1992) and references therein. M. Born and E. Wolf, “Principles of optics”(Fourth edition), Pergamon Press, Oxford (1970), eq (7), chapter 3.2.1, page 123. M.J. Nandor and T.M. Helliwell, “Fermat’s principle and multiple imaging by gravitational lenses”, Am. J. Phys., [**64**]{}, 45-49 (1996) and references therein. P.M. Alsing, “The optical-mechanical analogy for stationary metrics in general relativity”, Am. J. Phys., [**66**]{}, 779-790 (1998) and references therein. J. Evans and M. Rosenquist, “$\vec F=m \vec a$ optics”, Am. J. Phys., [**54**]{}, 876-883 (1986) and references therein. C. Bellver-Cebreros and M. Rodriguez-Danta, “Eikonal equation from continuum mechanics and analogy between equilibrium of a string and geometrical light rays”, Am. J. Phys., [**69**]{}, 360-367 (2001) and references therein. L. Mittag and M.J. Stephen, “Conformal transformations and the application of complex variables in mechanics and quantum mechanics”, Am. J. Phys., [**60**]{}, 207-11 (1992).
**Figure Captions**
The deflection geometry for $b>R$, with (S) and (E) at infinity. The vectors $\vec u_i$ and $\vec u_f$ are the unit vectors of the initial (emission) and final (reception) directions. The deviation $D=\arccos (\vec u_i \cdot \vec u_f) $ is a function of the impact parameter $b$. The minimum distance approach is $r_0$, and $r_0 \simeq b$ for small $D$.
[**Figure 2:**]{} The deflection geometry for the transparent deflector
\(a) phase (out) for $b<R$. The vector $\vec u_{in}$ is the unit vector along the velocity as the particule enters the deflector at point M ($\vec v_{in}=v_0 \vec u_{in}$). The corresponding deviation is given by $D_{out}^{(1)}=\arccos (\vec u_i \cdot \vec u_{in})$
\(b) phase (in) for $b<R$. The particle enters deflector ($\Delta$) at point M, and exits at point N. The OX and OY axes are the eigenvectors of the Runge-Lenz tensor ${\bf A}$. Points M and N are symmetric w.r.t. OY. For clarity purposes, the direction of $\vec u_{in}$ has been rotated with respect to Figure 2(a).
The qualitative variation of $D(b)$ (a) opaque deflector (b) transparent deflector (note the maximum for $b \sim R_G$). The dotted lines represent various cases of equation (\[deviation3\]). Solutions (1) and (2) correspond to trajectories (1) and (2) of Figure 4.
[**Figure 4:**]{} A typical geometry for finite distance gravitational deflection. Trajectories (1) and (2), deflected by $(\Delta)$, reach the earth (E): the source (S) has two images (S1) and (S2). For a transparent deflector, one has a third image of the source ( see point (3) in Figure 3(b)), corresponding to a trajectory entering ($\Delta$).
{height="6cm"}
Figure 1
{height="6cm"}
Figure 2
{height="8cm"}
Figure 3
{width="13cm"}
Figure 4
[^1]: Member of CNRS
[^2]: see e.g.
http://hypertextbook.com/facts/2000/AlinaVayntrub.shtml;
http://www.stdimension.de/int/Cartography/mwtour.htm
[^3]: One light year represents a distance of approximately $9.5 \cdot 10^{15}$m.
[^4]: We do not make the distinction between the impact parameter ($b$) and the minimum distance approach ($r_0$). The exact relation $r_0=b\left({1-{\rm
sin}{D \over 2} \over {\rm
cos}{D \over 2}}\right)$ shows that for small $D$, one has $r_0 \simeq
b$.
[^5]: A simple way to calculate the amplification factor is to use the non linear relation $\theta(\beta)$ of equation (\[traj12\]). The flux emitted by an extended source is proportional to $\beta d\beta$, and the flux received by the observer is proportional to $\theta d\theta$. The total amplification factor is given by ${\cal A}=\vert{\theta d\theta \over \beta
d\beta}\vert_{1}+\vert{\theta d\theta \over \beta d\beta}\vert_{2}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recently, the time-reversal violation predicted for the pseudogap phase of the cuprates, which was observed by dichroism experiments using Angle-Resolved Photoemission has also been observed by polarized neutron diffraction. Earlier derivation of dichroism in angle resolved photoemission due to time-reversal violation relied on existence of mirror planes in the crystal. Here the theory of the effect is generalized to the case that mirror plane symmetry is weakly violated due to perturbing potentials such as a superstructure.'
author:
- 'C. M. Varma'
title: Detecting Dichroism in Angle Resolved Photoemission
---
Introduction
============
A theory for the pseudogap phenomena in the Cuprates suggests that it is a phase with broken time-reversal symmetry due to an ordered array of current loops, without change in translational symmetry [@cmv1; @cmv2]. Such an elusive phase can be detected by dichroism in circularly polarized angle resolved photoemission [@cmv3; @simon-varma]. Such experiments were done and found the effect [@kaminski]. Recently the broken symmetry deduced in such experiments has been directly observed by polarized neutron diffraction experiments [@bourges; @mook].
I first summarize the proposal for the dichroic experiment [@simon-varma]: First, one must take note of the [*geometric*]{} dichroic effect which exists with and without any time-reversal breaking. This is the difference in intensity of outgoing photo-current for incident right circular polarized photons and left circular polarized photons if the momentum of the initial electronic state for photoemission is not in the mirror plane of the crystal while the photon is incident in the mirror plane. This effect is null when the momentum of the initial state lies in the mirror plane and it is odd about variations of momentum perpendicular to it. In addition to this, if the crystal is in a phase which breaks time-reversal symmetry accompanied by reflection symmetry breaking about this mirror plane, the difference is non-zero at this mirror plane and is even for variations perpendicular to it. So on cooling from high temperatures where time-reversal is preserved to below a temperature at which time-reversal symmetry is broken, the momentum for which the dichroic effect is zero moves away from the mirror plane. This is the dichroic effect to be looked for to show time-reversal violation as well as to provide the detailed specification of the state through discovery of the [*crystalline*]{} mirror planes about which the reflection symmetry is lost.
It is important to realize (i) that both time-reversal and (some) mirrror plane symmetry must be lost to see the effect and (ii) the dichroic effect is observed because it is linear in the time-reversal violating wavefunction; experiments sensitive only to the $|wavefunction|^2$ (such as ordinary charge diffraction experiments or ARPES with linear polarization) are insensitive to the time-reversal violation as well as the accompanying loss of reflection symmetry of the wave-function. The latter kind of experiment is of-course sensitive to loss of reflection symmetry arising from perturbations in atomic positions or potentials which lower the symmetry of the crystal. Proving time-reversal violation below a certain temperature therefore requires the occurrence of the dichroic effect below that temperature unaccompanied by a loss of symmetry in the latter type of experiment.
A legitimate question can be raised about the dichroism experiments performed: the theory of the effect, as summarized above relies on the existence of mirror planes in the crystal. But due to the superstructure modulation, a buckling of the $BiO$ layer, no mirror plane strictly exists on the surface of the BISCCO crystals on which the ARPES experiments are performed. How then to interpret the observed results? It ought to be mentioned that since the surface reconstruction is observed not to be temperature dependent and the dichroic effect is zero above a certain temperature and changes its magnitude below it, prima facie, the dichroism cannot be due to this loss of symmetry. However, it is worth investigating how the dichroism effect may be modified due to loss of mirror symmetry. This is the aim of this addenda.
The conclusion is that provided the loss of mirror plane symmetry is due to a small perturbation $\epsilon \ll 1$ on the electronic states studied in ARPES, the zero of the dichroic effect shifts linearly in $\epsilon$, while a finite effect due to time-reversal violation occurs at the position where the geometric effect is zero and is even about it. So, the point of zero of the dichroic effect moves on breaking time-reversal symmetry just as it does when there is a [*crystalline*]{} mirror plane. This conclusion is derived below.
Geometric and time-reversal induced dichroism in the absence of a mirror plane
==============================================================================
Suppose a beam of circularly polarized photons of energy $\omega$ shone on a crystal in the direction $\hat{n}$ produces free-electons with momentum $\bf {p}$ and energy $E_ {\bf {p}}$ at the detector. Let $|\bf{k}\rangle$ denote the states of the crystal. Here $\bf{k}$ is the wave-vector in the first Brillouin zone. Also $\bf{k}=\bf{p}$, modulo the reciprocal vectors, because the photon is assumed to have negligible momenta compared to the electrons. Quite generally, |**[p]{}= \_m |[**[p]{}**]{},e+ \_m|[**[p]{}**]{},o;\
|[**[k]{}**]{} = \_m |[**[k]{}**]{},e + \_m |[**[k]{}**]{},o . The division of the wavefunctions into two parts is made such that in reflection on [*any*]{} plane the part labeled $\bf{e}$ does not change sign and the part labeled $\bf{o}$ does: \_m |[**[p]{}**]{}= \_m |[**[p]{}**]{},e- \_m |[**[p]{}**]{},o;\
\_m|[**[k]{}**]{} = \_m |[**[k]{}**]{},e - \_m|[**[k]{}**]{},o .**
If the reflection were on a mirror plane and $\bf{p},\bf{k}$ were in the mirror plane, $\beta_m=\nu_m=0$. This was the case treated in (). Suppose we consider a plane $m$ which looses the symmetry of a mirror plane weakly due to a perturbing potential. Then for $\bf{p},\bf{k}$ in $m$, $\beta_m$ and $\nu_m$ are of $O(\epsilon) \ll 1$. We will also consider further small deviations, also of $O(\epsilon)$ from $\bf{p},\bf{k}$ in $m$.
The matrix element for photemission is = | [**[A\_i.]{}**]{} |[**[k]{}**]{}, where the subscript $i$ refers to circular polarization, either right ($r$) or left ($l$). For simplicity I will consider the situation when the direction $\hat{n}$ of the incident photon is normal to the crystal surface. Then \^[-1]{}\_m(**[A\_.]{})\_m = (**[A\_r.]{}) continues to hold.****
As in Ref.([@simon-varma]), the difference in the outgoing current $J_{\bf p}$ for the two polarizations is calculated from the difference in the (squared) matrix element ${\cal D}_m$. For $\epsilon,\epsilon'\ll1$, so that $\beta_m, \nu_m \ll 1$, we need only keep the first and the third terms of Eq.(8) of Ref.(), so that \_m 4 ( \_m\^\*\_m |\_m|\^2 ,e|[**M**]{}\_r\^\*|[**[k]{}**]{},e ,e|[**M**]{}\_r |[**[p]{}**]{},o + \_m\_m\^\*|\_m|\^2,e|[**M**]{}\_r\^\*|[**[k]{}**]{},e ,o|[**M**]{}\_r|[**[p]{}**]{},e) . Now, unlike in Ref.([@simon-varma]), ${\cal{D}}_m \ne 0$, but of $O(\epsilon)$, when $\bf{p},\bf{k}$ are in $m$. However, suppose $\bf{p},\bf{k}$ are moved by $\delta\bf{p}_{\perp}, \delta\bf{k}_{\perp}$ orthogonal to $m$ so that they lie in a plane $m'$. Then there is an additional contribution to $ {\cal{D}}_m$, which is proportional to $\delta\bf{p}_{\perp}$ and $\delta\bf{k}_{\perp}$, because this part of the wavefunction must change sign for $(\delta\bf{p}_{\perp}, \delta\bf{k}_{\perp}) \to (-\delta\bf{p}_{\perp}, -\delta\bf{k}_{\perp})$. (Due to loss of mirror plane, this statement is true only to $O(\epsilon)$; this makes the conclusions arrived at here violated to $O(\epsilon^2)$). It therefore follows that by adjusting the magnitude of $\delta\bf{p}_{\perp}/\bf{p}$ and $\delta\bf{k}_{\perp}/\bf{k}$ to $O(\epsilon)$ of one or the other sign, the geometric effect can be brought to $0$.
Now, let us consider the effect of time-reversal violation on ${\cal{D}}_m$. The experiment to look for the dichroism effect due to time-reversal violation is done after locating $m'$ as described above. Due to time-reversal violation and attendant [*extra*]{} loss of mirror plane symmetry, the coefficients $\beta_m, \nu_m$ acquire extra amplitudes proportional to the time-reversal breaking order parameter for ${\bf p,k}$ in $m'$. Further, again to $O(\epsilon)$, this effect is even for further small further variations of $\delta{\bf p'}_{\perp},\delta{\bf k'}_{\perp}$ about either side of $m'$.
To summarize: The conditions for the experiment if there are perturbations leading to a weak loss of mirror symmetry are to find incident ${\bf k}$ in $m'$ at which the geometric effect is zero and such that for small perpendicular variations about about such a ${\bf k}$, the effect is odd. This effect should be temperature independent and always present. Then cool below the symmetry breaking temperature and observe a temperature dependent effect which is even for small perpendicular variations of ${\bf k}$ about $m'$.
Experiments
===========
Actually, the experiment which saw the effect [@kaminski] was done precisely in the manner specified above. A geometric effect was found above a certain temperature $T_g$ in each sample (within the uncertainties of the experiments, it coincided with the temperature at which the pseudogap begins to be observed in resistivity and thermodynamic experiments). The zero of the geometric effect, which is independent of T for $T\gtrsim T_g$ was located. Below $T_g$, the dichroic effect was observed in that the zero of dichroism moved as a function of temperature.
Thus in the experiments, the considerations given here to account for any small loss of [*crystalline*]{} mirror plane symmetry were (unwittingly) taken care of. In reality the effect of the super-structure perturbation is so weak that that their effect is invisible in any loss of symmetry in linearly polarized ARPES from states from which the dichroic effect is observed. It was estimated [@kaminski2] that the effect must be less than $0.1\%$, while the dichroic effect observed is 1 to 2 $\%$. Besides, even this small superstructure is not known to have sensitivity to temperature in crystallographic measurements.
Thus there are two different ways that the experiment done is free of the criticizms made on the basis of the superstructure [@armitage; @borisenko; @golden]. First as already noted [@kaminski2], the effect of superstructure is too small and is temperature independent. Second, even if it were significant (but still small), the slightly generalized theory of the effect presented above shows that the procedure followed in the experiment allows deduction of dichroism. Recent direct observation [@bourges; @mook] through polarized neutron diffraction further settle the issue of the broken symmetry in underdoped cuprates.
The calculation presented here may be useful in situations in which crystalline mirror plane symmetry is broken on the same scale as the expected dichroic effect.
[10]{}
C.M.Varma, Phys. Rev. [**B 55**]{}, 14554 (1997); C.M.Varma, Phys. Rev. Letters [**83**]{}, 3538 (1999) C.M.Varma, arXiv.org/cond-mat/ ; submitted to Phys. Rev. B
C.M.Varma, Phys.Rev. [**B 61**]{}, R-3804 (2000) A.Kaminski, et al. Nature, [**416**]{}, 610 (2002). M.E.Simon and C.M. Varma, Phys. Rev. Letters, [**89**]{}, 247003-1 (2002). B.Fauque et al., arXiv.org/cond-mat/0509210 H. Mook (private communication) N.P. Armitage and J. Hu, arXiv.org/cond-mat/0303186. A. Kaminski et al., arXiv.org/cond-mat/0306140 S.V. Borisenko et al., Nature-online doi:10.1038/nature02931 (2004).kamin A. Mans et al., arXiv.org/cond-mat/0508127.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The T2K experiment is a long-baseline neutrino oscillation experiment with the ability to measure precisely $\nu_{\mu}\rightarrow\nu_e$ and $\nu_{\mu}\rightarrow\nu_{\mu}$ oscillations. Control of systematic uncertainties—dominated by flux and cross section uncertainties—is critical for the precision of these measurements. An analysis of charged current $\nu_{\mu}$ interactions at the T2K near detector significantly reduces these uncertainties, from 26.9% to 3.0% for the current $\nu_{\mu}\rightarrow\nu_e$ measurement at T2K.'
address: 'Department of Physics, Imperial College London, London, SW7 2AZ, UK'
author:
- 'A. Kaboth for the T2K Collaboration'
bibliography:
- 'Syst.bib'
title: Systematic Uncertainties at the T2K Experiment for Oscillation Measurements
---
Neutrino Oscillation
====================
Neutrino oscillations are governed by the $3\times3$ Pontecorvo-Maki-Nakagawa-Sakata [@MNS; @Pont3] mixing matrix and parameterized by two mass-squared differences, $\Delta m^2_{21}$ and $\Delta m^2_{32}$; three mixing angles, $\theta_{12}$, $\theta_{23}$, and $\theta_{13}$; and a complex CP-violating phase, $\delta_{CP}$. In recent years, the mass-squared differences and the mixing angles have all been measured to be non-zero [@PhysRevD.86.010001]. However, the phase $\delta_{CP}$ remains unconstrained, and the precision of the measurements of the mixing angles must be improved in order to investigate this phenomenon.
A critical factor in improving oscillation measurements is the control of systematic uncertainties, especially as more powerful neutrino sources are available and statistical uncertainty is thus reduced. Because much remains uncertain in the modeling of neutrino sources and cross sections, modern neutrino experiments use two detectors to constrain their systematic uncertainty by measuring the flavor content of the source close to the source, before neutrinos have had time to oscillate. This work will focus on the method and results of this technique for constraining systematic error in the context of the long-baseline Tokai-to-Kamioka (T2K) experiment.
The T2K Experiment
==================
The T2K experiment [@Abe:2011ks] is a long-baseline experiment starting with a conventional muon neutrino beam of 30 GeV protons incident on a graphite target, produced at the J-PARC facility in Tokai, Japan. The beam is characterized by a near detector complex, 280m downstream of the beam, consisting of an on-axis detector, INGRID [@Abe2012], and an off-axis detector, ND280. Oscillation measurements are made using the Super-Kamiokande (SK) detector [@Fukuda:2002uc], also off-axis from the beam, 295 km from the target. Using an off-axis beam allows for a narrowly focused beam at the maximal oscillation probability, and reduces uncertainty from higher energy interactions.
The ND280 detector, shown in Fig. \[fig:ND280\], is the primary detector for the work discussed here, and is a multi-purpose detector composed of several sub-detectors. There is a central tracker region composed of two fine-grained detectors (FGDs) [@Amaudruz:2012pe] surrounded by three three time projection chambers (TPCs) [@Abgrall:2010hi]. The primary target for neutrino interactions in this analysis are the fine-grained detectors (FGDs), which comprise planes of plastic scintillator bars—providing a carbon interaction target—arranged in alternating directions and read out with wavelength shifting fibers attached to MPPCs. For this analysis, only the upstream FGD was used, with a total fiducial mass of 914 kg. The primary method of determining particle identification and kinematics are the TPCs, which measures the momentum through the track curvature in a 0.2 T magnetic field and the particle identification through the dE/dx energy deposition in the gas.
![ \[fig:ND280\]The T2K off-axis near detector, ND280. Shown in orange are the time projection chambers (TPCs), and in green, the fine grained detectors (FGDs). Also shown is the $\pi^0$ detector (P0D) [@Assylbekov201248], the electromagnetic calorimeters [@allan2013electromagnetic]; and the UA1 magnet yoke and solenoid coils with muon range detector [@smrd], surrounding the inner detectors.](nd280.pdf){width="40.00000%"}
Beam and Related Systematics
============================
A full account of the T2K beam and beam simulation can be found in [@PhysRevD.87.012001]. The neutrino flux is predicted by modeling the interactions of the 30 GeV proton beam on the graphite target using FLUKA2008 [@Ferrari:2005zk]. The fluxes of the outgoing pions and kaons are tuned using data from the NA61/SHINE experiment [@Abgrall:2011ae; @Abgrall:2011ts] in momentum and angle bins. The decay and propagation of the hadrons through the magnetic horns, decay volume, and beam dump of the hadrons are modeled with GEANT3 [@GEANT3] and GCALOR [@GCALOR].
Systematic uncertainty is calculated on the beam from five sources: the proton beam, monitored in the proton beam line; the horn current uncertainty; alignment uncertainties on the target and horn; the beam direction, which is measured by the INGRID detector and beam line muon monitors; and the hadron production uncertainties from the NA61/SHINE tuning. These uncertainties are evaluated in bins of energy at ND280 and SK for the four neutrino species of the beam: $\nu_{\mu}$, $\bar{\nu}_{\mu}$, $\nu_e$, and $\bar{\nu}_e$. Of these uncertainties, the largest contributor is the hadron production uncertainty, as shown in Fig. \[fig:fluxuncertaintycontrib\]. The uncertainties are provided to analyses as a covariance matrix, which allows the analysis at ND280 to constrain the flux uncertainties at SK.
![ \[fig:fluxuncertaintycontrib\]Contributions to the flux uncertainty from various sources for the $\nu_{\mu}$ beam component at ND280. The dominant uncertainty comes from the tuning applied from NA61/SHINE data.](hadronuncertainty.pdf){width="45.00000%"}
Cross Section Model and Related Systematics
===========================================
Neutrino interactions at T2K are simulated using the NEUT neutrino interaction generator [@Hayato:2009]. At the typical energies of the T2K beam, the dominant charged current (CC) interaction mode is charged current quasi-elastic (CCQE), with significant contribution from charged current single resonant pion production (CC1$\pi$). Because neutrino interactions at ND280 and SK are on different nuclei—carbon and oxygen, respectively—there are parameters that can be shared between the two detectors, relating to the underlying physics of neutrino interactions, as well as parameters that pertain only to one detector or the other, relating to the nuclear model.
The most prominent of the first category of parameters are the two axial masses, $M_A^{QE}$ and $M_A^{RES}$, which govern the shape of the CCQE and CC1$\pi$ interactions. In addition to these, there are three CCQE normalization parameters in bins of energy of 0–1.5, 1.5–3.5, and $>3.5$ GeV, and two CC1$\pi$ normalizations for interactions of less than and greater than 2.5 GeV. There is also a normalization of neutral current $\pi^0$ (NC$\pi^0$) production. This cross section is of particular importance, as NC$\pi^0$ events form a significant background for $\nu_{\mu}\rightarrow\nu_e$ oscillation analyses. These eight cross section parameters are the cross section parameters that are propagated in the covariance matrix between ND280 and SK.
Parameters that are evaluated separately at the two detectors include uncertainties on the binding energy and fermi momentum of nucleons from the fermi gas model of the nucleus. Additionally, a systematic is evaluated on the difference between the fermi gas model and a spectral function model of the nucleus. Also included are uncertainties on the ratios of $\sigma_{\nu_{e}}/\sigma_{\nu_{\mu}}$ and $\sigma_{\bar{\nu}_{\mu}}/\sigma_{\nu_{\mu}}$ in oscillation analyses.
The initial values and prior uncertainties for the cross section parameters are driven by data as much as feasible. In particular, the value of $M_A^{QE}$ and the values of the single resonant pion production parameters ($M_A^{RES}$, CC1$\pi$ normalization, and NC1$\pi^0$ normalization) are determined by constraints from the MiniBooNE data [@mb-ccqe; @mb-cc1pi0].
ND280 Detector Systematics
==========================
There are many sources of systematic uncertainty in the detector modeling at ND280: track kinematics determination—e.g., from uncertainties in the magnetic field; the efficiencies of each of the sub-detectors, as well as the matching efficiency between the sub-detectors; external background uncertainties—e.g. interactions from outside the FGD fiducial volume, event pileup, and interactions in the rock upstream of the detector; and monte carlo modeling—e.g., the modeling of the reinteraction rates of pions exiting primary neutrino interactions. These uncertainties are measured as far as is possible with external datasets (for example, cosmic muons used for efficiency and matching uncertainties) and the dominant uncertainties come from events outside the fiducial volume for muon momenta below 400 MeV/c and from pion reinteraction rates above 400 MeV/c. Typical detector uncertainties are smaller than flux and cross section uncertainties, with values of a few percent.
ND280 Event Selection and Parameter Constraints
===============================================
The event selection aims to constrain the flux and cross section systematics described above by selecting a CC $\nu_{\mu}$ interaction sample, and subdividing it according to the topology of the interaction. These sub-samples better constrain the cross section systematics on specific interaction modes. The selection begins by identifying muons as the highest momentum negative-curvature track emerging from the fiducial volume of the upstream FGD, with an energy deposit consistent with that of a muon in the next most downstream TPC. This sample is then divided into CC0$\pi$ (enhanced in CCQE interactions), CC1$\pi^{+}$ (enhanced in single resonant pion interactions), and CC other samples. The presence of a $\pi^{+}$ can be determined in three ways: a track emerging from the FGD to the TPC that has positive curvature and energy deposit consistent with that of a pion, a track fully contained in the FGD with energy deposit consistent with a pion, or a time-delayed energy deposit that is consistent with a $\pi^+\rightarrow\mu^+\rightarrow e^+$ decay chain. Only the TPC method can conclusively determine the charge of a charged pion. Neutral pions are further identified with energy deposits in the TPC consistent with an electron. The CC0$\pi$ sample is required to have no identified pions of any charge in the event, the CC1$\pi^+$ sample must have one and only one identified positive pion, and the CC other sample has all remaining events. The purities are 72.6%, 49.4%, and 73.8%, respectively. These samples are fit, binned in momentum and angle with respect to the beam axis, using the flux, cross section, and detector systematic uncertainties described previously. Fig. \[fig:CCmomdist\] shows the momentum distributions before and after the fit for data and MC for all three subsamples. Table \[tab:ND280rates\] shows the number of events for data and pre- and post-constraint MC for the three subsamples.
![The momentum distributions for the (a) CC0$\pi$, (b) CC1$\pi^+$, and (c) CC other ND280 samples. Data is shown in black points, blue line shows the MC prediction before fitting the data, and red line shows the MC prediction after fitting the data.[]{data-label="fig:CCmomdist"}](ND280prepostmom.pdf){width="\textwidth"}
CC0$\pi$ CC1$\pi$ CC Other
----------------------- ---------- ---------- ----------
Data 16912 3936 4062
Prediction (pre-fit) 20016.2 5059.4 4602.1
Prediction (post-fit) 16802.5 3970.3 4006.0
: Event rates for the ND280 detector, showing the number of data events, the number of MC predicted events before and after the fit to data.[]{data-label="tab:ND280rates"}
Fig. \[fig:SKfluxconstraint\] shows the impact of the constraint on the flux and cross section parameters propagated to SK. There is both significant reduction in the uncertainty on each parameter, and also significant shifts in the parameters. Most shifts are within the systematic error assigned to them before fitting the data, but the value of $M_{A}^{RES}$ moves significantly away from its prior value. This shift is primarily driven by the CC1$\pi^+$ sample; this sample is dominated by single resonant pion events, and, as shown in Table \[tab:ND280rates\], the initial predicted value of the number of events is higher than the number of data events.
![The effect of the ND280 analysis on the parameter uncertainties propagated to oscillation analyses. Left: the impact on the flux normalizations for the muon neutrino flux at SK. Right: the impact on the cross section parameters. The value of the axial masses is in GeV; all other parameters are normalizations.[]{data-label="fig:SKfluxconstraint"}](FluxXsecFit.pdf){width="95.00000%"}
The constraint from the ND280 data significantly reduces the error on the number of predicted events at SK for oscillation analyses, especially because there are negative correlations between the constraints on the propagated flux and cross section parameters. Table \[tab:NSK\] shows the magnitude of this effect, compared with other sources of error in the analysis; using the ND280 data reduces the impact of the propagated parameters from the largest source of error by far to the smallest source of error.
Uncertainty Source Without ND280 Constraint With ND280 Constraint
----------------------------------- -------------------------- -----------------------
Flux and Cross section from ND280 26.9% 3.0%
Other Cross section 7.5% 7.5%
SK Detector 3.5% 3.5%
Total 28.2% 8.8%
: Percentage error on the number of events expected in a $\nu_{\mu}\rightarrow\nu_e$ oscillation analysis at SK, for a value of $\sin^2 2\theta_{13}=0.1$.[]{data-label="tab:NSK"}
Conclusions and Future Thoughts
===============================
The sources of systematic uncertainty at the T2K experiment are many, and, before data constraints, can be quite large. The T2K near detector, ND280, has been effectively employed to select a sample of charged current muon neutrino interactions that tightly constrain these uncertainties. The division of the sample into subsamples that address different neutrino interaction modes is particularly useful. Further work to reducing these uncertainties will come from studying other interaction modes, $\nu_e$ interactions, and interactions on materials other than carbon at ND280.
References {#references .unnumbered}
==========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Due to an exponential rise in the social media user base, incidents like hate speech, trolling, cyberbullying are also increasing and that lead hate speech detection problem reshaped into different tasks like Aggression detection, Fact detection. This paper attempt to study the effectiveness of text representation schemes on two tasks namely: User Aggression and Fact Detection from the social media contents. In User Aggression detection, The aim is to identify the level of aggression from the contents generated in the Social media and written in the English, Devanagari Hindi and Romanized Hindi. Aggression levels are categorized into three predefined classes namely: ‘Non-aggressive‘, ‘Overtly Aggressive‘, and ‘Covertly Aggressive‘. During the disaster-related incident, Social media like, Twitter is flooded with millions of posts. In such emergency situations, identification of factual posts is important for organizations involved in the relief operation. We anticipated this problem as a combination of classification and Ranking problem. This paper presents a comparison of various text representation scheme based on BoW techniques, distributed word/sentence representation, transfer learning on classifiers. Weighted $F_1$ score is used as a primary evaluation metric. Results show that text representation using BoW performs better than word embedding on machine learning classifiers. While pre-trained Word embedding techniques perform better on classifiers based on deep neural net. Recent transfer learning model like ELMO, ULMFiT are fine-tuned for the Aggression classification task. However, results are not at par with pre-trained word embedding model. Overall, word embedding using fastText produce best weighted $F_1$-score than Word2Vec and Glove. Results are further improved using pre-trained vector model. Statistical significance tests are employed to ensure the significance of the classification results. In the case of lexically different test Dataset, other than training Dataset, deep neural models are more robust and perform substantially better than machine learning classifiers.'
author:
-
title: An Empirical Evaluation of Text Representation Schemes on Multilingual Social Web to Filter the Textual Aggression
---
NLP: Natural Language Processing; BoW :Bag-of-Word; CNN:Convolution Neural Network;LSTM :Long Short-Term Memory.
Aggression, Trolling, Bag-of-Words, Word Embedding, Transfer learning, Word2Vec, Glove, fastText, Significance Test, Wilcoxon signed-rank, Student t-test
Introduction {#intro}
============
The Social Web is a great source for studying human interaction and behavior. In the last few years, there is an exponential growth in Social Media user base. Sensing content of Social Media like Facebook, Twitter, by the smart autonomous application empower its user community with real-time information which is unfolded across the different part of the world. Social media provide the easiest and anonymous platform for common people to voice their opinion or view on a various entity like celebrity, politician, product, stock market etc or any social movement. Sometime such opinions might be aggressive in nature and propagate hate in the social media community.
With the unprecedented increase in the user base of the social media and its availability on the Smartphones, incidents like Hate speech, trolling, Cyberbullying, and Aggressive posts are increasing exponentially. A smart autonomous system is required which enable surveillance on the social media platform and detect such incidents. Some of the researchers look posts from the aspect like aggression @Kumar18 to filter the contents. some of the posts contain words which might be qualified as either highly or overly aggressive or have hidden aggression. Sometimes posts do not have any aggression. Based on these, posts or comments are categorized into three classes namely: ‘Overtly Aggressive‘, ‘Covertly Aggressive‘ and ‘Non-aggressive‘ @Kumar18. Henceforth, in the rest of the paper, we will denote these classes by these abbreviations namely: OAG, CAG, NAG respectively.Table \[tab:Sample-post\] shows the sample posts belonging to these classes.
\[tab:Sample-post\]
Social Media, specifically Microblog has proved its importance during the disaster-related incidents like an earthquake, Hurricane and floods [^1]. Organizations involved in relief operation actively track posts related to situational information posted on Facebook and Twitter during the disaster. However, At the same time, social media is flooded with lots of prayer and condolence messages. Posts which contain factual information are extremely important for the organization involved in post-disaster relief operations for coordination. Filtering and Ranking of the posts containing factual information will be very useful to them. We believe that this is the special problem of the Sentiment Analysis task. We consider this problem as a combination of two-class classification problem: factual posts and nob-factual posts plus Ranking. Table \[tab:Sample-post-so\] shows the example of the posts of belong to these class.
\[tab:Sample-post-so\]
The Text representation of social web content plays a pivotal role in any NLP task. Bag-of-word is the oldest and simple technique to represent the document or post into a fixed length vector. The BoW techniques generate very sparse and high dimensional space vector. Text representation using distributed word/sentence representation or word embedding is gain rapid momentum recently. In this paper, one of the objectives is to find the best text representation scheme to model social web content for the machine learning classifier and deep neural net. Various Text representation scheme based on BoW, word embedding and are studied empirically. We have reported result on popular word embedding technique like Word2vec, Glove and fastText on standard machine learning classifier like Multinomial Naive Bayes (MNB), Logistic Regression(LR), K-Nearest neighbors KNN, Support Vector Classifier (SVC), Decision Tree (DT),Stochastic Gradient Descent(SGD), Random forest (RF), Ridge, AdaBoost, Perceptron, Deep neural net based on LSTM, CNN and Bidirectional LSTM. Results are also reported on Doc2vec embedding, a popular sentence or paragraph embedding technique for the above classifiers.
Transfer Learning is well practiced in the area of computer vision. However, in the NLP, transfer learning has limited application in the form of pre-trained word vector which is used to initialize the weights of the embedding layer of the deep neural network. With the advent of transfer learning method like ELMO [@peters2018deep], ULMFiT [@howard2018universal] claimed substantial improvement in the performance of various NLP tasks like Sentiment Analysis, Question/Answering, Textual Entailment empirically. The main idea behind these methods is to train language model on the large corpus and fine tune on the task-specific corpus. In this paper, We have evaluated the performance of these methods in the Aggression classification tasks.
Research Questions
------------------
In this study, experiments are performed on the benchmark dataset with to answer the following questions
- Which is the best Text Representation scheme to model text from the Social Web?
- Does pre-trained language model based on transfer learning better than pre-trained word embedding based on shallow transfer learning on Social media data?
- Does Making too Deep Neural net make sense?
To answer all research question listed above, experiments are performed on two tasks namely: Aggression detection (Trolling Aggression and Cyberbullying (TRAC) dataset) [@trac2018dataset] and Fact detection (FIRE iRMDI Dataset)[@fire2018-irdimis]. In this paper, we present exhaustive benchmarking of text representation schemes on these datasets. Our results reveal that fastText with pre-trained vector along with CNN outperform standard machine learning classifiers based on BoW Model and marginally perform better than Word2vec and Glove. Paragraph vector or Doc2vec [@le2014distributed] perform very poor on our dataset and turn out to be the worst text representation scheme among all. We also found that model based on the deep neural net is more robust than machine learning classifier when tested on lexically different dataset than training Dataset. i.e. deep neural model substantially outperforms machine learning classifier on Twitter test Dataset while trained on Facebook Dataset in this evaluation.
To validate our claims, statistical significance tests are performed on weighted $F_1$-score of the classifier for each text representing scheme. Statistical inference is used to check evidence to support or reject these claims. Significance tests like Wilcoxon signed-rank and Student t-test were carried out by comparing weighted $F_1$ score all the text representation scheme with the fastText pre-trained vector. In most of the cases, p-values are less than 0.05.
The rest of the paper is organized as follows: In section \[Related-Work\], we review the relevant works in the area of Sentiment analysis and hate speech detection. Section \[Dataset\] contains the detail information about the various benchmark Datasets used in the experiments. Various Text Representation schemes are described in section \[Text-rep-scheme\]. We formally describe the evaluation task and models in section \[task\]. We report results in section \[sec:results\] and present detail result analysis in section \[lab:ra\]. We conclude the discussion and provide insight for the future work in section \[lab:con\].
Related Work {#Related-Work}
============
Bag-of-Words (BoW) [@harris1954distributional] is the oldest technique to represent the text of the documents in fixed-length vectors with high dimensionality. @mikolov2013efficient proposed two architecture namely: skip-gram(SG) and continuous-bag-of word (CBOW) to learn high quality low dimensional word embedding. However, to generate sentence vector often, average or mean of word vector are considered. Doc2vec or paragraph vector @le2014distributed proposed Paragraph2vec (Doc2vec) which is the extension of the Word2vec to learning document level embedding. It is an unsupervised method which learns document vector from paragraph, sentence or document. @pennington2014glove proposed word embedding based on the co-occurrence matrix. @lau2016empirical have performed a comprehensive evaluation of Doc2Vec on two tasks namely: Forum Question Duplication and Semantic Textual Similarity (STS) task. Authors claimed that Doc2Vec performs better than Word2vec provided that models trained on large external corpora, and can be further improved by using pre-trained word embedding. They have published the hyper-parameter for the Doc2Vec embedding. Our work is similar to this but we have reported the evaluation of all the text representation scheme including doc2vec on TRAC dataset[@trac2018dataset] on each classifier.
Hate speech is a type of language which is used to incite or spread violence towards the group of people based on the gender, community, race, religion. Sentiment analysis and hate-speech are closely related in fact sentiment analysis techniques are used in hate speech detection. Initially, Sentiment Analysis problem is formulated as a binary classification problem for predicting the election results or detecting political opinion [@maynard2011automatic; @conover2011political; @conover2011predicting; @tumasjan2010predicting] on Twitter. Then after, It turned into the multi-class classification problem with the introduction of the neutral label. Soon, Researchers come with different notion like aggression [@Kumar18], cyberbullying[@xu2012learning], sarcasm, trolling. Semeval (International workshop on semantic evaluation)[@rosenthal2017semeval] is one of the popular competition on sentiment analysis which is started since 2013. TRAC [^2](Trolling, aggression, cyberbullying) workshop [@Kumar18] co-located with the International Conference of Computational Linguistics (COLING 2018) redefine hate speech detection task in terms of three type of aggression namely: Non-Aggression (NAG), Overly-Aggression(OAG) and Covertly Aggression (CAG).
Sentiment Analysis
------------------
During the initial year, there is a lack of standard dataset for comparative performance analysis. International Workshop on Semantic Evaluation 2013 (SemEval-2013) [@hltcoe2013semeval] was the first forum who developed standard tweet dataset for the benchmarking of the various sentiment analysis system. Most of the team who had participated in the competition used supervised approaches based on SVM, Naive Bayes, and Maximum Entropy. some of the team had used ensemble classifier and rule-based classifier. @mohammad2013nrc was the top team of the Semeval-2013 challenge. They have incorporated various semantic and lexicon based sentiment features for the experiment and SVM was used for the classification. Deep learning and word embedding had shown its footprints in SemEval-2015 [@rosenthal2015semeval]. Team UNITN [@severyn2015unitn] was the second team in the message polarity task. They have build convolution neural network for the sentiment classification. They have used an unsupervised neural language model to initialize word embeddings that are further tuned by deep learning model on a distant supervised corpus [@severyn2015unitn]. In fourth edition SemEval-2016 [@nakov2016semeval],Team SwissCheese [@deriu2016swisscheese] was the first ranked team with $F_1$ score around 63.3 %. Their approach was based on 2-layer convolution neural networks whose predictions are combined using a random forest classifier. SemEval-2017 [@rosenthal2017semeval] was the fifth edition, Team DataStories [@baziotis2017datastories] was the top-ranked team with AvgRec= 68.1 and $F_1$ around=67.7 %. They use Long Short-Term Memory (LSTM) networks augmented with two kinds of attention mechanisms, on top of word embedding pre-trained on a big collection of Twitter messages without using any hand-crafted features.
Hate Speech/Cyberbullying/Aggression Detection
----------------------------------------------
Hate Speech Detection research attracts researchers from the diverse background like Computational linguistic, computer science, social science. The actual term hate speech was coined by @warner2012detecting. Various Authors used different notion like offensive language [@razavi2010offensive], Cyberbullying [@xu2012learning], Aggression [@trac2018dataset]. @davidson2017automated studied tweet classification of hate speech and offensive language and defined hate speech as following: language that is used to expresses hatred towards a targeted group or is intended to be derogatory, to humiliate, or to insult the members of the group. Authors observed that offensive language often miss-classified as hate speech. They have trained a multi-class classifier on N-gram features weighted by its TF-IDF weights and PoS tags. In addition to these, features like sentiment score of each tweet, no of hashtags, URLS, mentions are considered. Authors concluded that Logistic regression and Linear SVM perform better than NB, Decision Tree, Random Forests. @schmidt2017survey perform comprehensive survey on hate speech. They have identified features like Surface features, sentiment, word generalization,lexical, linguistics etc. can be used by classifier.
Cyberbullying is the type bullying that occurs on social media platform or app via cellphone or any internet enabled device. @xu2012learning introduces Cyberbullying to the NLP community. They have performed various binary classification on tweets text with bullying perspective to determine whether the user is cyberbully or not. They reported binary classification accuracy around 81%. @kwok2013locate, authors performed classification using NB classifier on tweets based on two classes :racist and non-racists and achieved accuracy around 76 %. @burnap2015cyber, authors studied cyber hate on Twitter. They have used various classifier like SVM, BLR, RFDT, Voting base ensemble for the binary classification achieved best F1-score of 0.77 in the voted ensemble.@malmasi2017detecting, authors have used NLP based lexical approach to address the multi-class classification problem. They have used character N-gram, word N-gram and word skip-gram feature for the classification.
@schmidt2017survey, have described the key areas that have been explored to detect hate speech. They have surveyed different types of features used for hate speech classification. They have categorized features in Simple surface features, word generalization features, sentiment features, linguistic features, lexical resources features, Knowledge-based features, and Meta-Information features Simple surface features include features like character level unigram/n-gram, word generalization features include features like the bag-of-words, clustering, word embedding, paragraph embedding. Linguistic features include PoS tag of tokens. list of bad words or hate words can be considered as a lexical resource. @malmasi2018challenges, tried to address the problem of discriminating profanity from the hate speech in the social media posts. n-grams, skip-gram and clustering based word representation features are considered for the 3-class classification.The Author use SVM and advance ensemble based classifier for this task and achieved 80 % accuracy.
@aroyehun2018aggression performed translation as data augmentation strategy. TRAC Dataset [@trac2018dataset] was also augmented using translation and pseudo labeled using an external dataset on hate speech. they have reported best performance with LSTM and $F1$ score around 0.6415 on TRAC English dataset [@trac2018dataset]. @arroyo2018cyberbullying implement ensemble of the Passive-Aggressive (PA) and SVM classifiers with character n-grams. TF-IDF weighting used for feature representation. FIRE initiative also gave importance text representation in Indian language since its inception. [@majumder2008text][@majumdar2007initiative].
Dataset {#Dataset}
=======
Experiments are performed on standard benchmarked Datasets to evaluate the performance of various text representation scheme. For User Aggression detection problem, Trolling, Aggression and Cyberbullying TRAC [@trac2018dataset] is considered for the experiments which contain post in English and code-mixed Hindi. For the Factual Detection task, experiments are performed on FIRE IRMiDis Dataset.
TRAC Dataset
------------
TRAC (Trolling, Aggresion and Cyberbullying) consist of 15,001 aggression-annotated Facebook Posts and Comments each in Hindi (Romanized and Devanagari script) and English for training and validation [@trac2018dataset].
\[tab:class-stats\]
\[tab:test-data-stats\]
FIRE IRMiDis Dataset
--------------------
Forum for Information Retrieval Evaluation, have introduced Microblog track since 2016 as Information Retrieval from Microblogs during Disasters (IRMiDis). IRMiDis track [@fire2018-irdimis] of FIRE is organized with the objective to extract factual or fact-checkable tweets during the disaster which might be helpful to the victims or the people who are involved in the relief operation. Dataset contain tweets which are downloaded from the Twitter during Nepal earthquake 2015. Following are the example of factual or fact-checkable and non-fact-checkable tweet.Table \[tab:class-stats-Fire\] shows a detail statistics of FIRE IRMiDis Dataset. As we look at the table, There are only 83 tweets is annotated with objective class. not a single tweet is annotated from the subjective class.
\[tab:class-stats-Fire\]
Text Representation Schemes {#Text-rep-scheme}
===========================
The main objective of this paper is an identification of the best text representation scheme for the Social media text which is very sparse and noisy in nature. Text representation is about representing documents in a numerical way so that they can be feed as an input to the classifier. This numerical representation is in the form of the vectors which together form matrices. Essentially, There are two types of text representation scheme :(i) Bag-of-words(BoW) (ii) Distributed Word/sentence representation. BoW with count vector and TF/IDF weighting , various word embedding techniques(Word2Vec, Glove, fastText), and sentence or paragraph embedding (Doc2Vec) are studied.
Bag-of-Word Model for Text Representation
-----------------------------------------
The Bag-of-words is the simple technique to represent the document or social media posts in the vector form and also a very common feature extraction method from the text. Word count or TF/IDF weight of each n-gram word can be used as a features. The dimension of the vector is equal to the size of vocabulary of the text corpus or dataset which results in very high dimensional sparse document vector. It is the common method used for the text representation in order to perform various NLP task like text classification, clustering. However,the BoW methods ignore the word order which may lead to loss of the context.
Word Embedding for Text representation
--------------------------------------
Word Embedding is the text representation technique to represent the word in the low dimensional space so that semantically similar word have similar representation. Major word embedding techniques like Word2vec learn word embedding using shallow neural network. The fastText, extension of Word2vec, consider the morphological structure of the word.
### Word2Vec
Word2vec [@mikolov2013efficient] is the unsupervised and predictive neural word embedding technique to learn the word representation in the low dimensional space. Word2vec is a two-layer neural net that take text corpus as an input and output is a set of vectors. two novel model architectures: Skipgram and CBOW(Continuous bag of words) are proposed for computing continuous vector representations of words from very large data sets.
### Glove
GloVe stands for Global vector for \[Word Representation\] [@pennington2014glove]is an unsupervised method for learning word embedding. A Co-occurrence word matrix is created from the text corpus for the training and is reduced in low dimensional space which explain the variance of high dimensional data and provide word vector for each word.
### fastText
fastText [@bojanowski2017enriching] is the neural word embedding technique which learn distributed low dimensional word embedding. Word2vec, Glove consider each word as single unit and ignore the morphological structure of the word. They are not able generate word embedding for the unseen or out of vocabulary word during the training. fastText overcome this limitation of Word2vec and GLOVE by considering each word as N-gram of characters. A word vector for a word is computed from the sum of the n-gram characters. The range of N is typically 3 to 6. Since user on social media often make spelling error, typos, fastText will be more effective then rest of two.
Paragraph vector/Doc2vec
------------------------
Paragraph Vector is an unsupervised algorithm that learns fixed-length feature representations from variable-length pieces of texts, such as sentences, paragraphs, and documents [@le2014distributed]. Paragraph vector represents each document by a dense vector which is trained to predict words in the document. Authors believe that Paragraph vector have the potential to overcome the weaknesses of bag-of-words models and claimed that Paragraph Vectors outperform bag-of-words models as well as other techniques for text representations. Paragraph vector model is also referred as doc2vec model. Henceforth, we will refer paragraph and Doc2vec interchangeably. Doc2vec model have two architecture namely : (i) DM: This is the Doc2Vec model analogous to CBOW model in Word2vec. The paragraph vectors are obtained by training a neural network on the task of inferring a center word based on context words and a context paragraph. (ii) DBOW: This is the Doc2Vec model analogous to Skip-gram model in Word2Vec. The paragraph vectors are obtained by training a neural network on the task of predicting a probability distribution of words in a paragraph given a randomly-sampled word from the paragraph.
Transfer Learning
------------------
Transfer Learning in NLP is not as matured as compare to in Computer Vision. Transfer learning is a method in which model is trained on large corpus for a particular task and use this pre-trained model for the similar task. There are two way to use transfer learning in NLP (i) Use of Pre-trained word embedding to initialize first layer of neural network model which can be called as shallow representation. (ii) Use the full model and fine tune for the task specific in supervise learning way.
Word2vec, Glove and fastText provide pre-trained word vector trained on the large corpus. Google Word2vec pre-trained model have word vector for 3 million words with size 300 and trained on Google news. Glove pre-trained model available with different embed size and trained on common crawl, Twitter. We have use Glove pre-trained model with vocabulary size 2.2 million and trained on common crawl. fastText pre-trained models are available in 157 language. We have use fasttext pre-trained vector for Englsih and Hindi language trained on commnon crawl and wikipedia.
Recently, transfer learning in NLP done in new way; First language model is trained on large text corpus in unsupervise way and fine tune on specific task like text classification on labeled data. @peters2018deep author argued that word representation is depend upon the context. So each word has different word vector depending upon the position of the word in the sentence. Essentially Each word has dynamic word vector with respect to the context as opposed to the traditional word embedding techniques which always give same word vector ignoring the context. Embedding from Language Models (ELMos) use languge model for the word embedding. @howard2018universal author propose Universal Language Model Fine-Tuning for Text Classification (ULMFiT) which is bi-LSTM model that is trained on a general language modeling (LM) task and then fine tuned on text classification. Results are reported on both transfer learning model on TRAC dataset [@trac2018dataset].
Evaluation Tasks {#task}
================
We have benchmarked various text representation scheme on two specialized NLP task namely: aggression detection and fact detection. Text Representation scheme are evaluated on machine learning and deep neural model.
Aggression Detection task
-------------------------
The objective of this task is to identify type of aggression present in the text in both Englsih and code-mixed Hindi language. Aggression are classified into three level namely: ‘Overtly Aggressive‘ (OAG), ‘Covertly Aggressive‘ (CAG) and ‘Non-aggressive‘ (NAG). We have implemented all standard machine learning classifiers like Multinomial Naive Bayes (MNB), Logistic Regression(LR), K-Nearest neighbors (KNN), Support Vector Classifier (SVC), Decision Tree (DT),Stochastic Gradient Descent(SGD), Random forest (RF), Ridge, AdaBoost, Perceptron, and various voting based ensemble with different text representation schemes like count based, TF/IDF and word embedding to prepare baseline results. Various word embedding techniques like Word2Vec, Glove, fastText, Paragraph2Vec are studied.
### Problem statement
Basically Aggression detection is a Text classification problem. Formally, the task of Text Classification is stated as follows. Given a set of social media feed and a set of classes, We need to compute a function of the form: $$C=f(T,\Omega)$$
where f is the multi-class classifier that is computed using training data, T is the numeric representation of the text of the dataset, $\Omega$ is the set of parameters of the classifier and C is the pre-define class-labels.
### Model Architectures and Hyperparameters
In this subsection, we will discuss the architecture and hyperparameters of our deep neural model used for the classification. Model learns feature from the input texts Ther is no need to design hand-crafted features which used to encode text into feature vector.
#### Bidirectional LSTM
The first model is based on the Bidirectional LSTM include embedding layer with embed size 300, convert each word from the post into a fixed length vector. short posts are padded with zero values. Subsequent layers includes Bidirectional LSTM layer with 50 memory units followed by one-dimensional global max pooling layer, a hidden layer with size 50 and output layer with softmax activations. ReLU activation function is used for the hidden layer activation. A drop out layer is added between the last two layers to counter the overfitting with parameter 0.1. Hyperparameters are as follows: Sequence length is fixed at 1073 word; maximum length of posts in the dataset. No of features is equal to half of total vocabulary size. Models are trained for 10 epoch with batch size 128. Adam optimization algorithm is used to update network weights.
#### Single LSTM with higher dropout
This model is based on the Long Short Term Memory, a type of recurrent neural network with higher dropout. This model is having one embedding layer, one LSMT layer with a size 64 memory unit, and one fully connected hidden layer with Relu activation and size 256 and an output layer with softmax activation. Hyperparameters are same as discussed in the previous model. A dropout layer is added between the hidden layer and an output layer with drop out rate 0.2 to address the overfitting issue.
#### CNN Model
This model includes one embedding layer whose weights are initialized with fastText pre-trained vector with embed size is 300, followed by one-dimensional convolution layer with 100 filters of height 2 and stride 1 to target biagrams. In addition to this, Global Max Pooling layer added to fetch the maximum value from the filters which are feed to the fully connected hidden layer with size 256, followed by output layer. ReLU and softmax activation function are used for the hidden layer and output layer respectively.
#### CNN model with Multiple Convolution layer
This model includes embedding layer with embed size 300. Three one dimensional convolution layers with size 100 and different filters with height 2,3,4 to target bigrams, trigrams, and four-grams features, followed by max pooling layer which concatenate max pooled result from each of one-dimensional convolution layer. The final two layers include a fully connected hidden layer with size 250 and output layer with ReLu and softmax activation. A Drop out layer is added between the last two layer with rate 0.2. Hyperparameters are same as discussed in the first model. This model is similar to proposed by [@zhang2015sensitivity].
Factual Post/Tweet Detection from Social Media
----------------------------------------------
During the emergency situation like earthquake or floods, Microblog plays a very important role as an anonymous communication medium. The various entity like, Volunteers, NGOs involved in relief operation always look for real-time information which contains facts instead of prayer and condolence messages. In more technical term, these agencies are looking for factual information from Microblog instead of the subjective information. In addition to this, the system should generate rank-list of the tweets based upon the worthiness of facts. we considered this problem as a binary classification problem plus pure IR Ranking problem. two classes can be labeled as factual and non-factual.
the IRMiDis dataset [^3],which was prepared from the tweet posted during Nepal earthquake 2015 [@fire2018-irdimis] is considered for the experiment. There are only 83 fact checkable tweets in the dataset. Non-factual tweets are not available. Total no of tweets in the dataset is more than 50000.
### Preparation of Training Data
Due to the unavailability of adequate training data, The first task is to prepare training data to train the deep neural model. We randomly choose 100 tweets from the dataset and labeled as a non-fact-checkable tweet and 83 fact-checkable tweets present in the dataset labeled as fact-checkable. We have trained our Convolution neural network on these training data and tested the model on the remaining 50000 tweets. At this stage we are not interested in the class but, we have sorted all the tweets based upon the predicted probability of the fact-checkable class and selected top 2000 tweets. We have randomly selected tweets and gave relevance judgment based upon availability of factual information in first 1000 tweets and manually extracted 300 tweets as non-fact-checkable tweets to minimize the false positives. Remaining 1700 tweets labeled as fact-checkable tweets. We selected the last 1700 tweets with the least probability of the class fact-checkable and labeled them as non-fact-checkable tweets. So our Training corpus has 1783 fact-checkable and 2000 non-fact-checkable tweets.\
### Proposed Approach
We have used word embedding to represent the text instead of bags-of-words. fastText [@mikolov2018advances] pre-trained vector with 300 dimensions is used to initialize the weight matrix of the embedding layer of the network. We trained our CNN model on this training corpus with 10-fold cross-validation.The Model gives validation accuracy around 94%. Finally, we run the model on the entire corpus and sorted the tweet based upon the predicted probability of the Fact-checkable class. Essentially this approach termed as weakly-supervise classification.
Results {#sec:results}
=======
In this section, we first present results of classifiers TRAC dataset [@trac2018dataset] with different text representation scheme. Latter we present result on FIRE IRMiDis 2018 Dataset. Tweets are very noisy in nature contains user mentions, Hashtags, Emojis, and URLs. We do not perform any kind of text pre-processing on tweets in experiments with deep neural models. In experiments with machine learning classifier, before classification, Hashtag symbol \# and User mentions are dropped from the tweets. Non-ASCII characters and stop-words are removed from tweet text [@modha2016daiict].
Results On TRAC Dataset
-----------------------
Precision, Recall, and F1-score are the standard metrics which are used to evaluate the classifier performance. We have evaluated 16 classifiers performance on 4 Datasets (2 English+2 Hindi) with 10 Text Representation scheme(8 in the case of Hindi Dataset). Looking at such massive experiment, it is difficult to report results in all the above metrics. Therefore, Results are reported in terms of weighted F1-score only which is the function of Precision and Recall. Classifiers results based on LSTM and CNN on BoW text representation schemes are not possible due to the high dimensionality. Bernoulli classifier is used instead of Naive Bayes Classifier in case of text representation schemes other than BoW. Since word vectors might have negative weights, it is impossible to calculate probabilities with negative weights. Skip-gram variant of Word2Vec and fastText is used in this experiment instead of continuous bag-of-word. Table \[tab:res-en-fb-mc\] and \[tab:res-en-tw-mc\] shows results on Facebook and Twitter English Dataset with BoW and word embedding while Table \[tab:res-en-fbtw-dnn\] present result with pre-trained word embedding with same dataset. Table \[tab:res-hi-fb-mc\] and \[tab:res-hi-tw-mc\] shows results on Facebook and Twitter code-mixed Hindi Dataset. Only fastText provide pre-trained word vector [@mikolov2018advances] for the Hindi language. Exhaustive evaluation is performed with all classifiers with respect to each text representation schemes. Experiments are also performed with the new transfer learning model like ELMO and ULMFIT. Table \[tab:res:transfer\] presents results on both Facebook and Twitter English Datasets. Figure \[fig:hf\] and figure \[fig:ht\] display the heatmap of the results achieved by classifiers on each text representation scheme.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[**Classifier**]{} [**Count-vector**]{} [**TF/IDF**]{} [**W2Vec**]{} [**Glove**]{} [**Fasttext**]{} [**doc2vec-dmc**]{} [**doc2vec-dbow**]{}
-------------------- ---------------------- ---------------- --------------------------------------------- --------------- ------------------ --------------------- ----------------------
[NB]{} 0.5571 [**0.5596**]{} 0.4870 0.3873 0.5035 0.4585 0.4634
[LR]{} 0.5953 [**0.6046**]{} 0.5675 0.5358 0.5400 0.5266 0.5139
[KNN]{} 0.5466 [**0.5428**]{} 0.5061 0.5130 0.5113 0.5114 0.5095
[SVC]{} 0.5801 [**0.5902**]{} 0.5369 0.5037 0.5137 0.5388 0.5033
[DT]{} [**0.5269**]{} 0.5055 0.4468 0.4067 0.5002 0.4198 0.4198
[SGD]{} 0.5706 [**0.5938**]{} 0.4647 0.3571 0.5167 0.5060 0.3521
[ RF]{} [**0.5621**]{} 0.5582 0.5199 0.4752 0.5513 0.4230 0.4210
[Ridge]{} [**0.6009**]{} 0.5999 0.5347 0.5336 0.5225 0.5385 0.5083
[AdaB]{} [**0.6210**]{} 0.6141 0.5491 0.4932 0.5644 0.4689 0.4852
[Perce.]{} 0.5387 [**0.5491**]{} 0.5230 0.4020 0.4848 0.3800 0.3253
[ANN]{} [**0.5703**]{} 0.5350 0.5350 0.5037 0.5380 0.4980 0.4401
[Ensemble]{} 0.58 [**0.5900**]{} 0.5558 0.4067 0.5617 0.4980 0.4401
LSTM **0.5649 & 0.5454 & 0.5062 & &\
BLSTM & & & **0.5759 & 0.4760 & 0.5641 & &\
CNN & & & 0.5515 & 0.5365 & **0.5638 & &\
NCNN & & & **0.5919 & 0.5488 & 0.4849 & &\
********
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[tab:res-en-fb-mc\]
\[tab:res-en-tw-mc\]
[lcccccc]{}
& &\
[**Classifier**]{} & [**p-Word2vec**]{} & [**p-Glove**]{} & [**p-Fasttext**]{} & [**p-Word2vec**]{} & [**p-Glove**]{} & [**p-Fasttext**]{}\
[NB]{} & 0.5342 & 0.5373 & [**0.5519**]{} & 0.4152 & [**0.4527**]{} & 0.4276\
[LR]{} & 0.5799 & 0.6050 & [**0.6045**]{} & 0.4197 & [**0.4527**]{} & 0.4441\
[KNN]{} & 0.4981 & [**0.5103**]{} & 0.4819 & 0.3405 & [**0.3959**]{} & 0.3912\
[SVC]{} & 0.5832 & 0.5678 & [**0.6120**]{} & 0.4446 & [**0.4581**]{} & 0.4350\
[DT]{} & 0.4700 & 0.4515 & [**0.4900**]{} & 0.3640 & [**0.3949**]{} & 0.3632\
[SGD]{} & 0.5019 & 0.5521 & [**0.5360**]{} & 0.3692 & 0.3793 & [**0.3852**]{}\
[RF]{} & 0.5402 & 0.5338 & [**0.5505**]{} & 0.3394 & [**0.3716**]{} & 0.3687\
[Ridge]{} & 0.5829 & 0.5952 & [**0.6140**]{} & 0.4092 & [**0.4530**]{} & 0.4461\
[AdaB]{} & 0.5713 & 0.5781 & [**0.5907**]{} & 0.4241 & [**0.4261**]{} & 0.4033\
[Perce.]{} & 0.5114 & 0.5201 & [**0.5660**]{} & [**0.4224**]{} & 0.4118 & 0.4049\
[ANN]{} & 0.5025 & 0.5498 & [**0.5722**]{} & 0.3728 & 0.3722 & [**0.4842**]{}\
[ Ensemble]{} & 0.5300 & 0.5500 & [**0.5558**]{} & 0.3728 & 0.3722 & [**0.4500**]{}\
[LSTM]{} & 0.4979 & 0.4979 & [**0.6178**]{} & 0.5537 & 0.5518 & [ 0.5541]{}\
[ BLSTM]{} & 0.5501 & [**0.6062**]{} & 0.6000 & 0.5359 & [**0.5466**]{} & 0.5423\
[ CNN]{} & 0.4749 & 0.5405 & [**0.6407**]{} & 0.5226 & [**0.5667**]{} & 0.5520\
[ NCNN]{} & 0.5169 & [**0.5883**]{} & 0.5600 & 0.5384 & 0.5067 & [**0.5407**]{}\
\[tab:res-en-fbtw-dnn\]
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[**Classifier**]{} [**Count-vector**]{} [**TF/IDF**]{} [**W2Vec**]{} [**Glove**]{} [**Fasttext**]{} [**p-fastText**]{} [**doc2vec-dmc**]{} [**doc2vec-dbow**]{}
-------------------- ---------------------- ---------------- --------------- --------------- ----------------------------------------------------- -------------------- --------------------- ----------------------
[ NB]{} 0.5535 [**0.6031**]{} 0.3001 0.372 0.2959 0.3176 0.3459 0.4736
[ LR]{} 0.5855 [**0.6134**]{} 0.5779 0.464 0.5457 0.5518 0.3894 0.4380
[ KNN]{} 0.3340 0.1721 0.4998 0.425 [**0.5106**]{} 0.4909 0.3768 0.4038
[ SVC]{} 0.5556 [**0.5862**]{} 0.4806 0.373 0.5186 0.5442 0.3879 0.4344
[ DT]{} [**0.5307**]{} 0.5025 0.4629 0.388 0.4392 0.4288 0.3485 0.3485
[ SGD]{} 0.5533 [**0.5922**]{} 0.3912 0.393 0.3670 0.4746 0.3331 0.4134
[RF]{} 0.5473 [**0.5473**]{} 0.5374 0.440 0.5047 0.4788 0.3512 0.3477
[Ridge]{} 0.5780 [**0.5850**]{} 0.5293 0.381 0.5092 0.5544 0.3866 0.4292
[AdaB]{} [**0.5373**]{} 0.5233 0.5342 0.479 0.5336 0.4913 0.3751 0.4214
[Perce.]{} 0.5213 [**0.5598**]{} 0.4232 0.364 0.3763 0.4873 0.2661 0.3282
[ANN]{} [**0.5703**]{} 0.5350 0.5455 0.5037 [**0.5842**]{} 0.5190 0.4091 0.4440
[Ensemble]{} 0.5700 [**0.6087**]{} 0.5558 0.4067 0.534 0.5612 0.4980 0.4401
LSTM 0.5649 0.590 **0.6021 & 0.5916 & &\
BLSTM & & & 0.5759 & 0.527 & 0.5770 & **0.5900 & &\
CNN & & & 0.5515 & 0.566 & 0.5950 & **0.6081 & &\
NCNN & & & 0.5919 & 0.573 & 0.5912 & **0.5965 & &\
********
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[tab:res-hi-fb-mc\]
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[**Classifier**]{} [**Count-vector**]{} [**TF/IDF**]{} [**W2Vec**]{} [**Glove**]{} [**Fasttext**]{} [**p-fastText**]{} [**doc2vec-dmc**]{} [**doc2vec-dbow**]{}
-------------------- ---------------------- ------------------------------------------------------------------------------------------------ ---------------- --------------- ------------------ -------------------- --------------------- ----------------------
[ NB]{} 0.2970 0.2902 0.3215 0.273 [**0.3359**]{} 0.2897 0.3270 0.3205
[LR]{} [**0.3787**]{} 0.3724 0.2819 0.279 0.3184 0.3524 0.2438 0.2833
[KNN]{} 0.2527 0.2553 [**0.3704**]{} 0.334 0.3381 0.2917 0.3051 0.3299
[ SVC]{} 0.3781 **0.3886 & 0.2821 & 0.261 & 0.3087 & 0.3472 & 0.2580 & 0.2905\
[ DT]{} & 0.3685 & [**0.3936**]{} & 0.3572 & 0.326 & 0.3475 & 0.3473 & 0.2988 & 0.2988\
[ SGD]{} & [**0.3996**]{} & 0.3993 & 0.2822 & 0.287 & 0.2739 & 0.3163 & 0.2605 & 0.2588\
[ RF]{} & 0.3585 & [**0.3737**]{} & 0.3286 & 0.344 & 0.3449 & 0.3288 & 0.2981 & 0.2988\
[ Ridge]{} & 0.3616 & [**0.3872**]{} & 0.2811 & 0.242 & 0.3346 & 0.3361 & 0.2549 & 0.2875\
[ AdaB]{} & 0.1886 & 0.1903 & 0.3256 & [**0.362**]{} & 0.3261 & 0.3441 & 0.2614 & 0.2933\
[ Perce.]{} & [**0.3931**]{} & 0.3868 & 0.2802 & 0.329 & 0.2787 & 0.3835 & 0.2616 & 0.2752\
[ ANN]{} & 0.430 & [**0.44**]{} & 0.3163 & 0.3552 & 0.2399 & 0.3593 & 0.2419 & 0.3132\
[ Ensemble]{} & 0.4400 & [**0.4600**]{} & 0.4500 & 0.3938 & 0.3426 & 0.3555 & 0.3230 & 0.3281\
LSTM & & & 0.3840 & 0.376 & 0.3667 & **0.4600 & &\
BLSTM & & & 0.2846 & 0.318 & 0.3005 & **0.4600 & &\
CNN & & & 0.3323 & 0.317 & 0.2669 & **0.4992 & &\
NCNN & & & 0.3338 & 0.380 & 0.3494 & **0.4600 & &\
**********
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[tab:res-hi-tw-mc\]
\[tab:res:transfer\]
\[tab:res:comp-peer\]
![Heatmap on English Facebook Test Dataset Results.[]{data-label="fig:hf"}](f1-heat-new-fb-en.jpeg)
![Heatmap on English Twitter Test Dataset Results.[]{data-label="fig:ht"}](f1-heat-new-twitter-en.jpeg)
Information Retrieval from Microblogs during Disasters (IRMiDis) Dataset
------------------------------------------------------------------------
As discussed in the previous section \[intro\], This task is classification plus Ranking task. Table \[tab:res:IRMiDis\] shows our system results on IRMiDis dataset [@fire2018-irdimis] along with the rest of teams. nDCG overall is the primary metric for the evaluation. Our system substantially outperforms rest of team in the most of the metrics which justifies our claim established on TRAC dataset [@trac2018dataset]
\[tab:res:IRMiDis\]
Result Analysis {#lab:ra}
===============
In this section, we will present the comprehensive result analysis and try to answer the research questions which framed before the experiments were performed. As we look at the table \[tab:res-en-fb-mc\] \[tab:res-en-tw-mc\], and \[tab:res-en-fbtw-dnn\], Overall, LSTM and CNN with pre-trained fastText word embedding marginally outperform (around 2 % to 4%) standard machine learning classifiers and ensemble of classifier with respect to weighted $F_1$- score on Facebook English corpus and substantially outperforms on Twitter English corpus. By and large similar results observed on code-mixed Hindi corpus as shown in table \[tab:res-hi-fb-mc\] and \[tab:res-hi-tw-mc\].
\[tab:res-ana-trac-fb-en\]
Table \[tab:res-ana-trac-fb-en\] present the detailed comparative results of two classifiers: CNN model with fastText pre-trained vector and the logistic regression with TF/IDF weighting on TRAC Facebook English dataset. The CNN Model classify Facebook posts better than logistic regression at the individual class level and overall. It has been quite evident that posts belong to CAG class are hard to classify and @malmasi2017detecting reported that the same observation. Table \[tab:res-ana-cnn-cag\] show posts which are miss-classified by logistic regression however, CNN model correctly classified them into the CAG class.
\[tab:res-ana-cnn-cag\]
Significance Test
-----------------
To support our claim drawn in the previous section, significance tests, like Wilcoxon signed-rank test and Student t-test were carried out by comparing Weighted $F_1$ score of each classifier for each text representation scheme with fastText pre-trained vector scheme. Table \[tab:res:stats-en\] and \[tab:res:stats-Hi\] summarizes the p-values of statistical significance tests on English and Hindi Dataset respectively. In Wilcoxon signed-rank test, p-values of the results is less than 0.05 for Facebook English dataset and Twitter Hindi Dataset. However, On Twitter English dataset and Facebook Hindi Dataset, some of the p-values are higher than 0.05. In student t-test, we get mixed bag results. By and large, our results are statistically significant.
\[tab:res:stats-en\]
\[tab:res:stats-Hi\]
In the following subsection, we will try to answer all the research questions framed during the experiments were planned.
Best Text Representation scheme to model the text from Social web
------------------------------------------------------------------
Text Representation is the primary task for to address any NLP task like Question/answering, classification etc. As dicusses in section \[Text-rep-scheme\], There are basically two text representing scheme:Bag-of-Word(BoW) with countvector, TF/IDF weighting and word embedding. Word2Vec [@mikolov2013efficient], Glove [@pennington2014glove], and fastText [@mikolov2018advances], an extension of Word2vec are popular word embedding techniques.
Results clearly show that models with fastText pre-trained vector outperform Glove pre-trained vector on Facebook test dataset as well as the Twitter test dataset. the main reason behind the outperformance of fastText over Glove and Word2vec is that The fastText consider each word as N-gram characters. A word vector for a word is computed from the sum of the n-gram characters. Glove and Word2vec consider each word as a single unit and provide a word vector for each word. Since Facebook users make a lot of mistakes in spelling, typos, fastText is more convenient than Glove [@majumder2018filtering]. from Figure \[fig:hf\] and \[fig:ht\] shows that BoW is still effective text representation scheme for the standard machine learning classier which takes hand-crafted feature and n-grams as inputs. Logistic Regression and Support Vector perform better than other classifiers in English as well as Hindi Dataset. Adaboost performs better than LR and SVC on Facebook English Dataset but substantially underperform them on rest of three Datsets. Our participation [@majumder2018filtering] in TRAC competition [@Kumar18] FIRE Information Retrieval from Microblogs during Disasters [@fire2018-irdimis] track where our team performed well and secured top position.
Transfer Learning Model vs Pre-trained Word Embedding Model
-----------------------------------------------------------
Transfer learning is focused on storing knowledge gained while solving one problem and applying it to a different but related problem. On many occasion, NLP researchers face the problem of unavailability of sufficient labeled data to train the model. With the advent of new transfer learning method like ELMO [@peters2018deep] and Universal language model fine-tuning for Text Classification (ULMFiT) [@howard2018universal] attract interest among NLP Researchers. These models are trained or large text corpus. @howard2018universal claimed that these model can be fine-tuned on the task-specific corpus. We have used these transfer learning model on TRAC English dataset [@trac2018dataset] and results are presented in table \[tab:res:transfer\]. one can observe that results are substantially lower than the results reported in Table \[tab:res-en-fb-mc\], \[tab:res-en-tw-mc\], and \[tab:res-en-fbtw-dnn\] where pre-trained word vectors are used to initialize the first layer of deep neural model and rest of the network is trained from scratch achieve better results than transfer learning model. @howard2018universal termed use of pre-trained vector as shallow representation.
In these experiments, we trained different classifier models on Facebook posts. Table \[tab:res-en-tw-mc\] \[tab:res-hi-tw-mc\] shows the results on Twitter dataset [@trac2018dataset]. There is lexical difference between Facebook and Twitter posts. From the results shown in Table \[tab:res-en-tw-mc\] \[tab:res-en-fbtw-dnn\] and table \[tab:res-hi-tw-mc\], one can conclude that weighted $F_1$ score of standard machine learning classifiers are substantially lower in Twitter Dataset as compare to Facebook Dataset. While deep learning models perform better than machine learning classifiers for the Twitter Dataset. Thus, Deep learning models are more robust than machine learning classifier across diverse datasets.
Does Deeper Neural Net make Sense
---------------------------------
To answer this question, we designed first CNN model with one convolution layer and other CNN model with 3 convolution layer with different filters height. As we look at results shown in Table \[tab:res-en-fb-mc\],\[tab:res-en-tw-mc\],\[tab:res-en-fbtw-dnn\], and \[tab:res-hi-fb-mc\], one can conclude that by and large weighted $F_1$ score lower for CNN model with multiple convolution layer than CNN model with single convolution layer.
Conclusion {#lab:con}
==========
In this Paper, Multilingual Social media stream is studied with special kind of text features: Aggression and fact perspective. Exhaustive experiments are performed to benchmark the text representation scheme on machine learning classifiers and deep neural nets. From the results, we conclude that deep Neural model with pre-trained word embedding is the better choice than machine earning classifier and transfer learning model. Word embedding is the better text representative scheme than Bag-of-words for the deep neural models. In fact, performance can be improved with the help of fastText pre-trained vector. However, machine learning classifiers perform better in BoW with TF/IDF weighting than word embedding. We also concluded that higher drop out will help to counter model overfitting and improvise a standard evaluation metrics. CNN and LSTM are the better models for these datasets. On the English test corpus, we obtained a better weighted $F_1$ score for NAG class and poor weighted $F_1$ score for CAG class which supports the previous [@malmasi2017detecting] findings. For the Facebook Hindi test corpus, the same seems not to be true. We obtained a better F1 score for CAG class than NAG class. It is also to be noted that the model leads to poor result on Twitter test data since the training corpus was created from Facebook. In such cases, deep neural models substantially outperform machine learning classifiers. Significance test confirms these claims with 95 % confidence interval in most the cases. Our work shows what kind of problems are moving into the center of attention for research in machine learning. Using deep learning models, there is great potential to solve some of these problems, yet still, the performance is far from perfect. Model transfer between problems and the application of derived knowledge in user interfaces are areas directions for future work.
Aroyehun, Segun Taofeek and Gelbukh, Alexander (2018). *Aggression detection in social media: Using deep neural networks, data augmentation, and pseudolabeling*.*Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying (TRAC-2018)*, pp.90–97.
Arroyo-Fern[á]{}ndez, Ignacio and Forest, Dominic and Torres-Moreno, Juan-Manuel and Carrasco-Ruiz, Mauricio and Legeleux, Thomas and Joannette,Karen (2018). *Cyberbullying Detection Task: the EBSI-LIA-UNAM System (ELU) at COLING’18 TRAC-1*.*Proceedings of the First Workshop on Trolling,Aggression and Cyberbullying (TRAC-2018)*, pp 140–149.
Basu, Moumitaand Ghosh, Saptarshi and Ghosh, Kripabandhu (2018). *Overview of the FIRE 2018 track: Information Retrieval from Microblogs during Disasters (IRMiDis)*. *Proceedings of FIRE 2018 - Forum for Information Retrieval Evaluation*, Gujrat, India, December .
Baziotis, Christos and Pelekis, Nikos and Doulkeridis, Christos (2017). *Datastories at semeval-2017 task 4: Deep lstm with attention for message-level and topic-based sentiment analysis*. *Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017)*, pp.747–754
Bojanowski, Piotr and Grave, Edouard and Joulin, Armand and Mikolov, Tomas (2017) *Enriching word vectors with subword information*, *Transactions of the Association for Computational Linguistics*, vol-5,pp.[135–146]{}, MIT Press.
Burnap, Pete and Williams, Matthew Ltitle (2015). *Cyber hate speech on twitter: An application of machine classification and statistical modeling for policy and decision makingPolicy & Internet*, vol-7 number 2, pp.223–242, Wiley Online Library.
Conover, Michael and Ratkiewicz, Jacob and Francisco, Matthew R and Gon[ç]{}alves , Bruno and Menczer, Filippo and Flammini, Alessandro (2011). *Political polarization on twitter.*,*Icwsm*, vol-133, pp.89–96.
Conover, Michael D and Gon[ç]{}alves, Bruno and Ratkiewicz, Jacob and Flammini, Alessandro and Menczer, Filippo (2011). *Predicting the political alignment of twitter users*, *Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE Third Inernational Conference on Social Computing (SocialCom), 2011 IEEE Third International Conference on*, pp.192–199.
Davidson, Thomas and Warmsley, Dana and Macy, Michael and Weber, Ingmar (2017). *Automated Hate Speech Detection and the Problem of Offensive Langemuage*. *Proceedings of ICWSM*.
Deriu, Jan and Gonzenbach, Maurice and Uzdilli, Fatih and Lucchi, Aurelien and Luca, Valeria De and Jaggi, Martin (2016). *Swisscheese at semeval-2016 task 4: Sentiment classification using an ensemble of convolutional neural networks with distant supervision*. *Proceedings of the 10th international workshop on semantic evaluation*,pp.1124–1128.
Hltcoe, J (2013).*Semeval-2013 task 2: Sentiment analysis in Twitter*,vol-312 Atlanta, Georgia, USA.
Howard, Jeremy & Ruder, Sebastian [2018]{}. *Universal language model fine-tuning for text classification"*,*arXiv preprint arXiv:1801.06146*,
Harris, Zellig S (1954),*Distributional structure* Word,[10]{}, number=[2-3]{},pp.146–162, *1954*, Taylor & Francis.
Kumar, Ritesh and Reganti, Aishwarya N. and Bhatia, Akshit and Maheshwari,Tushar (2018), *Aggression-annotated Corpus of Hindi-English Code-mixed Data*, *Proceedings of the 11th Language Resources and Evaluation Conference (LREC)*, Miyazaki, Japan.
Kumar, Ritesh and Ojha, Atul Kr. and Malmasi, Shervin and Zampieri Marcos (2018). *[Benchmarking Aggression Identification in Social Media]{}, *Proceedings of the First Workshop on Trolling, Aggression and Cyberbulling (TRAC)**, Santa Fe, USA
Kwok, Irene and Wang, Yuzhou (2013),,*Twenty-Seventh AAAI Conference on Artificial Intelligence*.
Lau, Jey Han and Baldwin, Timothy (2016). *An empirical evaluation of doc2vec with practical insights into document embedding generation*. *arXiv preprint arXiv:1607.05368*.
Le, Quoc and Mikolov, Tomas (2014) *Distributed representations of sentences and documents*.*International Conference on Machine Learning*, pp.1188–1196
Majumder, Prasenjit and Mandl, Thomas and Modha Sandip (2018)*Filtering Aggression from the Multilingual Social Media Feed* *Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying (TRAC-2018)*, pp. [199–207]{}
Malmasi, Shervin and Zampieri, Marcos (2017)*Detecting Hate Speech in Social Media* *Proceedings of the International Conference Recent Advances in Natural Language Processing (RANLP)*, pp.[467–472]{}.
Malmasi, Shervin and Zampieri, Marcos (2018).*Challenges in Discriminating Profanity from Hate Speech* *Journal of Experimental & Theoretical Artificial Intelligence* pp.[1–16]{}, vol-30, issue-2,Taylor & Francis.
Maynard, Diana and Funk, Adam (2011) *Automatic detection of political opinionsin tweets*, *Extended Semantic Web Conference*,pp. [88–99]{}, *Springer*.
Mikolov, Tomas and Chen, Kai and Corrado, Greg and Dean, Jeffrey (2013). *Efficient estimation of word representations in vector space*,*arXiv preprint arXiv:1301.3781*.
Mikolov, Tomas and Grave, Edouard and Bojanowski, Piotr and Puhrsch, Christian and Joulin, Armand (2018). *Advances in Pre-Training Distributed Word Representations*,*Proceedings of the International Conference on Language Resources and Evaluation (LREC 2018)*.
Modha, Sandip and Agrawal, Krati and Verma, Deepali and Majumder, Prasenjit and Mandalia, Chintak 2016. *DAIICT at TREC RTS 2016: Live Push Notification and Email Digest.*,TREC.
Mohammad, Saif M and Kiritchenko, Svetlana and Zhu, Xiaodan (2013). *NRC-Canada: Building the state-of-the-art in sentiment analysis of tweets.arXiv preprint arXiv:1308.6242*.
Nakov, Preslav and Ritter, Alan and Rosenthal, Sara and Sebastiani, Fabrizio and Stoyanov, Veselin (2016). *SemEval-2016 task 4: Sentiment analysis in Twitter*,*Proceedings of the 10th international workshop on semantic evaluation (semeval-2016)*, pp. [1–18]{}.
Pennington, Jeffrey and Socher, Richard and Manning, Christopher *Glove: Global vectors for word representation* *Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP)*, pp.[1532–1543]{}.
Peters, Matthew E and Neumann, Mark and Iyyer, Mohit and Gardner, Matt and Clark, Christopher and Lee, Kenton and Zettlemoyer, Luke (2018). *Deep contextualized word representations* *arXiv preprint arXiv:1802.05365*.
Razavi, Amir H and Inkpen, Diana and Uritsky, Sasha and Matwin, Stan (2010) *Offensive language detection using multi-level classification*, *Canadian Conference on Artificial Intelligence* pp.[16–27]{}, Springer
Rosenthal, Sara and Nakov, Preslav and Kiritchenko, Svetlana and Mohammad, Saif and Ritter, Alan and Stoyanov, Veselin (2015). *Semeval-2015 task 10: Sentiment analysis in twitter* *Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015)*, pp.[451–463]{}.
Rosenthal, Sara and Farra, Noura and Nakov, Preslav (2017). *SemEval-2017 task 4: Sentiment analysis in Twitter* *Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017)*, pp.[502–518]{}.
Schmidt, Anna and Wiegand, Michael (2017).*A Survey on Hate Speech Detection Using Natural Language Processing*,*Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media. Association for Computational Linguistics* [Valencia, Spain]{}, pp.[1–10]{},
Severyn, Aliaksei and Moschitti, Alessandro (2015) *Unitn: Training deep convolutional neural network for twitter sentiment classification*, *Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015)*, pp. [464–469]{}.
Tumasjan, Andranik and Sprenger, Timm Oliver and Sandner, Philipp G and Welpe, Isabell M (2010). *Predicting elections with twitter: What 140 characters reveal about political sentiment.*,*Icwsm*, vol-10, number-[1]{}, pp. [178–185]{}.
Warner, William and Hirschberg, Julia (2012) *Detecting hate speech on the world wide web*, *Proceedings of the Second Workshop on Language in Social Media*,pp [19–26]{},[Association for Computational Linguistics]{}.
Xu, Jun-Ming and Jun, Kwang-Sung and Zhu, Xiaojin and Bellmore, Amy (2012). *Learning from bullying traces in social media*, *Proceedings of the 2012 conference of the North American chapter of the association for computational linguistics: Human language technologies*, pp.[656–666]{}, [Association for Computational Linguistics]{}
Zhang, Ye and Wallace, Byron (2015). *A Sensitivity Analysis of (and Practitioners’ Guide to) Convolutional Neural Networks for Sentence Classification*,*arXiv preprint arXiv:1510.03820*. Majumder, Prasenjit and Mitra, Mandar and Pal, Dipasree and Bandyopadhyay, Ayan and Maiti, Samaresh and Mitra, Sukanya and Sen, Aparajita and Pal, Sukomal.*Text collections for FIRE*, *Proceedings of the 31st annual international ACM SIGIR conference on research and development in information retrieval*,pp.[699–700]{},[ACM]{}
Majumdar, P and Mitra, Mandar and Parui, Swapan K and Bhattacharya, *Initiative for indian language ir evaluation*, *The First International Workshop on Evaluating Information Access (EVIA)*
[^1]: https://phys.org/news/2018-08-social-media-bad-disaster-zones.html
[^2]: https://sites.google.com/view/trac1
[^3]: https://sites.google.com/site/irmidisfire2018/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Strongly disordered spin chains invariant under the SO(N) group are shown to display random-singlet phases with emergent SU(N) symmetry without fine tuning. The phases with emergent SU(N) symmetry are of two kinds: one has a ground state formed of randomly distributed singlets of strongly bound pairs of SO(N) spins (a ‘mesonic’ phase), while the other has a ground state composed of singlets made out of strongly bound integer multiples of N SO(N) spins (a ‘baryonic’ phase). The established mechanism is general and we put forward the cases of $\mathrm{N}=2,3,4$ and $6$ as prime candidates for experimental realizations in material compounds and cold-atoms systems. We display universal temperature scaling and critical exponents for susceptibilities distinguishing these phases and characterizing the enlarging of the microscopic symmetries at low energies.'
author:
- 'V. L. Quito'
- 'P. L. S. Lopes'
- 'José A. Hoyos'
- 'E. Miranda'
bibliography:
- 'bibliog\_son.bib'
title: 'Emergent SU(N) symmetry in disordered SO(N) spin chains'
---
*Introduction.* The process of symmetry breaking, as the energy of a given system is lowered, plays a central role in our current understanding of both high-energy physics (as in the standard model) and condensed matter physics (with universality and classification of phases) [@AndersonBook; @Laughlin04012000]. A less noticed (and explored) scenario is that of symmetry emergence, in which the lowering of the system’s energy allows for ground states and excitations which are symmetric under a larger group of transformations than their corresponding microscopic Hamiltonian. A basic mechanism by which this can happen can be understood in the renormalization group framework by means of fixed points characterized by a symmetry which is broken only by irrelevant perturbations. There remains, nevertheless, a widespread lack of recognizable generic processes or patterns, so systems which realize this type of physics are found by trial and error (see [@zamolodchikov89; @PhysRevLett.115.166401; @schmalianbatista08; @Coldeaetal2010; @Damle2002; @Senthietal2004; @Groveretal2014; @batistaortiz04; @fidkowski-etal-prb09; @Yipetal2015; @zohar2016; @PhysRevB.55.8295; @linetal98] for examples). In scenarios dominated by disorder, the situation is even more clouded. It was in this context that, in Ref. [@Quito_PhysRevLett.115.167201], it was shown that generic disordered SU(2)-symmetric spin-1 chains exhibit emergent SU(3)-symmetric random-singlet phases (RSPs) [@madasguptahu]. There, it was also noted that in the pioneering work by Fisher on disordered XXZ spin-1/2 chains [@fisher94-xxz], there was also the emergence of SU(2) symmetric RSPs; SU(2) is explicitly broken down to U(1) in the microscopic XXZ Hamiltonian. What was *not* noted, however, is that in *both* cases the emergent SU(N) symmetry materialized out of systems with manifest SO(N) invariance, with $N=3$ and $2$, respectively.
This situation, which at first might be naively thought of as just a coincidence, uncovers, on the contrary, a consistent pattern. It is the aim of this Letter to show that generic disordered magnetic chains invariant under the SO(N) group, in its defining vector representation, display emergent SU(N)-symmetric phases via a unified route for any $\mathrm{N}\geq2$; we denote this process by $\text{SO\ensuremath{\left(\mathrm{N}\right)}}\overset{\textrm{emerg}}{\longrightarrow}\text{SU\ensuremath{\left(\mathrm{N}\right)}}$. Our pattern of symmetry emergence contains two phases: (i) an obvious SU(N) generalization of the SU(2)-symmetric random singlet phase of the Heisenberg chain of Ref. [@fisher94-xxz], (ii) a phase whose ground state also consists of random SU(N)-symmetric singlets, but which are composed of $k$N original SO(N) ‘spins’, with $k$ an arbitrary integer. Separating the two phases there is a critical point with manifest SU(N) symmetry. In the particular case of $\text{SO\ensuremath{\left(3\right)}}\overset{\textrm{emerg}}{\longrightarrow}\text{SU\ensuremath{\left(3\right)}}$ of Ref. [@Quito_PhysRevLett.115.167201] [\[]{}previously interpreted as $\text{SU}\text{\ensuremath{\left(2\right)}}_{\textrm{spin-1}}\overset{\textrm{emerg}}{\longrightarrow}\text{SU\ensuremath{\left(3\right)}}$[\]]{}, particular versions of these phases were dubbed “mesonic” and “baryonic” random singlet phases, respectively. Furthermore, every one-dimensional RSP encountered so far [@PhysRevB.70.180401; @Bonesteel2007; @fidkowski-etal-prb08; @fidkowski-etal-prb09] seems to find a counterpart in one of the permutation-symmetric multicritical points described by Damle and Huse [@PhysRevB.66.104425; @PhysRevLett.89.277203], each one indexed by an integer $n$. The SO(N) baryonic RSPs we found realize all of these Damle-Huse points with $n=\mathrm{N}$ in an extended phase (see also the discussion in [@QuitoLopes2016PRB])
While SO(N) magnetism may sound exotic at first, such systems can be realized in several ways, either by exploiting explicit breaking of a larger SU(N) isotropy or, more interestingly, by taking advantage of the isomorphisms between orthogonal ($\mathrm{so}\left(\mathrm{N}\right)$) and unitary ($\mathrm{su}\left(\mathrm{N}\right)$) algebras at low N values. Some examples, summarized in Table \[tab:Summary-known-models\], follow:
\(i) The first two mentioned cases, that of the XXZ spin-1/2 Heisenberg chain [@fisher94-xxz] and of spin-1 bilinear and biquadratic Hamiltonians [@Quito_PhysRevLett.115.167201] can be realized in solid state [@Toskovic2016] and, in principle, in cold atom systems [@PhysRevLett.93.250405; @PhysRevA.68.063602], respectively. The former has a Hamiltonian with broken SU(2)-symmetry which, in fact, corresponds to an SO(2) symmetric Hamiltonian. The latter is realized explicitly as the most general SU(2)-symmetric Hamiltonian with spin-1 representations, but due to the algebra isomorphism $\mathrm{so}\left(3\right)\sim\mathrm{su}\left(2\right)$, it corresponds also to the most general SO(3)-symmetric Hamiltonian in the defining vector representation.
\(ii) Through the isomorphism $\mathrm{so}\left(4\right)\sim\mathrm{su}\left(2\right)\otimes\mathrm{su}\left(2\right)$, SO(4)-symmetric magnetism is realized by the well-known KugelKhomskii Hamiltonian [@0038-5670-25-4-R03], commonly used in the description of $e_{g}$ orbitals in transition metal oxides [@Tokura_2000], with $\mathrm{su}\left(2\right)$-spin ($\mathbf{S}$) and $\mathrm{su}\left(2\right)$-orbital ($\mathbf{T}$) degrees of freedom $$\begin{aligned}
H_{KK} & = & \sum_{i}\left[J_{i}\left(\mathbf{S}_{i}\cdot\mathbf{S}_{i+1}+\mathbf{T}_{i}\cdot\mathbf{T}_{i+1}\right)\right.\nonumber \\
& & +\left.8D_{i}\left(\mathbf{S}_{i}\cdot\mathbf{S}_{i+1}\right)\left(\mathbf{T}_{i}\cdot\mathbf{T}_{i+1}\right)\right].\label{eq:KK}\end{aligned}$$
\(iii) There are proposals to realize SU(N) magnetism with arbitrary N in fermionic alkaline-earth cold atomic systems in representations other than the fundamental one [@Gorshkov2010]. Exploiting the isomorphism $\mathrm{so}\left(6\right)\sim\mathrm{su}\left(4\right)$, disordered SU(4) magnetic chains in the self-conjugate representation realize an SO(6)-symmetric chain in its defining representation. In this case, according to our mechanism, disordered SU(4) symmetric chains would realize $\ensuremath{\text{SU\ensuremath{\left(4\right)}}\overset{\mbox{emerg}}{\longrightarrow}\text{SU\ensuremath{\left(6\right)}}}$[^1].
\(iv) Random SO(2S+1) chains can, in fact, be designed by fine-tuning in *any* disordered rotation invariant spin-S system. Such generic spin-S chains have been previously studied by some of us [@quitoprb2016], but the SO(N) phases of these systems were not, at that point characterized.
--------------------------------------------------------------------------------------------------------------
$\begin{array}{c} Possible realizations $\begin{array}{c}
\text{\ensuremath{\mathrm{Hamiltonian}}}\\ \text{Emergent}\\
\mathrm{symmetry} \text{symmetry}
\end{array}$ \end{array}$
-------------------------------------------- --------------------------------------------- -------------------
SO(2) anisotropic spin-1/2 systems SU(2)
SO(3) generic spin-1 systems SU(3)
SO(4) $e_{g}$ orbitals in transition metal oxides SU(4)
SO(6) cold fermionic alkaline-earth atoms SU(6)
--------------------------------------------------------------------------------------------------------------
: List of the most relevant SO(N)-symmetric one-dimensional models described by Eq. with their possible physical realizations and the corresponding emergent symmetry in the limit of strong disorder. \[tab:Summary-known-models\]
We will first describe the general model and our results for the disordered $\text{SO\ensuremath{\left(N\right)}}\overset{\textrm{emerg}}{\longrightarrow}\text{SU\ensuremath{\left(N\right)}}$ mechanism. After that, we will give the finer technical details of our work.
![(a) Phase diagram of the strongly-disordered one-dimensional SO(N)-symmetric Hamiltonian of Eq. . Points in the circle refer to the angle $\tan\theta\equiv K_{i}^{\left(2\right)}/K_{i}^{\left(1\right)}$, which is taken to be constant despite the randomness in $K_{i}^{\left(1,2\right)}$. The blue and the green regions realize two distinct random-singlet phases, *both* with emergent SU(N) symmetry. In the blue region, SU(N) singlets are built out of SO(N) ‘spin’ pairs [\[]{}‘mesons’, shown in panel (b)[\]]{}. The green region has SU(N) singlets made of $k$N ‘spins’ (with $k=1,2,\ldots$) [\[]{}‘baryons’, shown in panel (c)[\]]{}. The arrows indicate the renormalization group flow. Red and white stars represent stable and unstable fixed points, respectively. The black (for any N) and the yellow (for even N) regions are not addressed in this work. \[fig:Full-phase\]](Phase_diag8){width="1\columnwidth"}
*Model and results.* The $\mathrm{N}\left(\mathrm{N}-1\right)/2$ SO(N) generators [\[]{}SO(N) ‘spins’[\]]{} will be denoted by $L^{ab}$, with $a,b$ in the range $a=1,\ldots,\mathrm{N}$ and $a<b$ [^2]. We will take them in the defining representation, which is spanned by a basis $\left|c\right\rangle $, $c=1,\ldots,N$. Each $L^{ab}$ generates rotations in the $ab$ plane. For $N=4$, for example, $L^{23}$ rotates a four-dimensional vector in the $\left(2,3\right)$ Cartesian plane, while components 1 and 4 are kept fixed. In general, $$iL^{ab}\left|c\right\rangle =\delta^{ac}\left|b\right\rangle -\delta^{bc}\left|a\right\rangle .\label{eq:generatoraction}$$ The $L^{ab}$ operators obey the so(N) Lie algebra
$$\left[L^{ab},L^{cd}\right]=i\left(\delta^{bc}L^{ad}+\delta^{ad}L^{bc}-\delta^{ac}L^{bd}-\delta^{bd}L^{ac}\right),\label{eq:SON-algebra}$$
with $\mbox{Tr}\left(L^{ab}L^{cd}\right)=2\delta^{ac}\delta^{bd}$.
An SO(N)-symmetric Hamiltonian is built as a sum over pairs of SO(N) spins. In the defining representation, the most general pair term contains only bilinear and biquadratic terms [@Tu_PhysRevB.78.094404; @QuitoLopes2016PRB]. In one dimension and considering only nearest-neighbor interactions we have $H=\sum_{i}H_{i}$ where $$H_{i}=J_{i}\mathbf{L}_{i}\cdot\mathbf{L}_{i+1}+D_{i}\left(\mathbf{L}_{i}\cdot\mathbf{L}_{i+1}\right)^{2},\label{eq:Hamilt0}$$ where $\mathbf{L}_{i}\cdot\mathbf{L}_{i+1}=\sum_{a<b}L_{i}^{ab}L_{i+1}^{ab}$ and $J_{i},\,D_{i}$ are random couplings of $i$-th link. For later convenience, we will recast $H$ in terms of the linear combinations $K_{i}^{\left(1\right)}=J_{i}-\frac{N-2}{2}D_{i}$ and $K_{i}^{\left(2\right)}=\frac{N-2}{2}D_{i}$,
$$H_{i}=K_{i}^{\left(1\right)}\hat{O}_{i,i+1}^{\left(1\right)}+K_{i}^{\left(2\right)}\hat{O}_{i,i+1}^{\left(2\right)},\label{eq:Hamilt}$$
where $\hat{O}_{i,i+1}^{\left(1\right)}=\mathbf{L}_{i}\cdot\mathbf{L}_{i+1}$ and $\hat{O}_{i,i+1}^{\left(2\right)}=\mathbf{L}_{i}\cdot\mathbf{L}_{i+1}+\frac{2}{N-2}\left(\mathbf{L}_{i}\cdot\mathbf{L}_{i+1}\right)^{2}$.
We choose a parametrization of Eq. in terms of the polar coordinates $\left(r_{i},\theta_{i}\right)$ in the $\left(K_{i}^{\left(1\right)},K_{i}^{\left(2\right)}\right)$ plane, so that $\tan\theta_{i}\equiv K_{i}^{\left(2\right)}/K_{i}^{\left(1\right)}$. For simplicity, we focus on random couplings $K_{i}^{\left(1\right)}$ and $K_{i}^{\left(2\right)}$ *with a fixed ratio throughout the chain*, i.e., $\theta_{i}=\theta\ \forall i$ (the general case is discussed in the accompanying paper [@QuitoLopes2016PRB]). In the regime of strong disorder, RSPs are found at low energies. The phase is determined by $\theta$, as displayed in a circle, see Fig. \[fig:Full-phase\]. The basins of attraction, delineated by the colors and arrows in Fig. \[fig:Full-phase\], are found via a strong-disorder renormalization group (SDRG) treatment [@madasgupta; @madasguptahu; @fisher94-xxz; @fishertransising2]. The green and blue regions are both characterized by *infinite effective disorder* at long length scales [@fisher94-xxz]. More interestingly, both the blue and the green regions of Fig.\[fig:Full-phase\] correspond to phases with emergent SU(N) symmetry.
RSPs are characterized by a ground state formed by a collection of singlets. In the blue region, these random singlets are formed by spin pairs [\[]{}SO(N) ‘mesons’[\]]{}, as in the random Heisenberg chain studied by Fisher [@fisher94-xxz] [\[]{}see Fig. \[fig:Full-phase\](b)[\]]{}.
In the green region of Fig. \[fig:Full-phase\], on the other hand, the ground state consists of a collection of singlets formed out of $k$N ($k=1,2,\dots$) original SO(N) spins [\[]{}SO(N) ‘baryons’[\]]{} as depicted in Fig. \[fig:Full-phase\](c). The same relation between energy and length scales . Note that the structure of these RSPs is the same as the Damle-Huse multicritical points [@PhysRevB.66.104425; @PhysRevLett.89.277203].
The emergent SU(N) symmetry in each of these phases arises because, as it turns out, the strongly entangled SO(N) singlets, be they pairs or N-tuples, are also SU(N) singlets. Likewise, the original spins into which these singlets are broken at energies above zero also transform as SU(N) spins. As these two types of objects ultimately determine the low-energy properties, the latter will reflect this enhanced symmetry group. For example, the susceptibilities of the SU(N) operators (which can be constructed from linear *or bilinear* combinations of the SO(N) operators, as we will show) will also have the quoted behavior with the same exponent in each phase. The same is true of the correlation function distributions. These two types of phases and their properties had been described before by two of us in disordered spin chains
These are our main results. Their derivation relies on the application of an elegant Lie algebra machinery to the SDRG. In what follows we outline and motivate the results, relegating the full details to a longer and more pedagogic exposition [@QuitoLopes2016PRB].
*SDRG details.* The SDRG method is based on an iterative removal of degrees of freedom in real space following an energy hierarchy dictated by the largest local 2-site gap. Each iteration step consists of (i) the decimation of the pair with largest gap $\Omega$ by a projection of its Hilbert space onto its ground multiplet and (ii) the renormalization of the remaining couplings between this sub-space to the adjacent spins using perturbation theory. When applied sequentially, this process translates into a flow of the distribution of coupling constants. While the form of the Hamiltonian and the connectivity of the chain is preserved, new multiplets belonging to any one of the anti-symmetric SO(N) representations appear throughout the flow. As a consequence, the full characterization of the phases involves a flow of representation distributions.
Using Eq. , the decimation rules can be written in closed form [@QuitoLopes2016PRB]. Crucially, the decimations of the angles $\theta_{i}$ do not involve the radial variables $r_{i}$. Suppose the largest gap occurs between spins 2 and 3. If the ground multiplet of $H_{2,3}$ is not a singlet, it belongs to one of the $\mathrm{int}\left(\mathrm{N}/2\right)$ anti-symmetric representation of SO(N), and spins $2$ and $3$ are replaced by a new spin in that representation. The couplings in links 1 and 3 are renormalized according to
$$\tan\tilde{\theta}_{1,3}=\pm\tan\theta_{1,3}.\label{eq:deci1ord}$$
The choice of sign is determined by the representations being decimated as well as their ground state multiplet [@QuitoLopes2016PRB]. If the ground multiplet of $H_{2,3}$ is a singlet, spins $2$ and $3$ are effectively removed. In this case, a new coupling between spins 1 and 4 is created with [@QuitoLopes2016PRB] $$\tan\tilde{\theta}=-\left(\frac{N+2}{N}\right)\frac{\frac{N-2}{N+2}-\tan\theta_{2}}{1-\tan\theta_{2}}\tan\theta_{1}\tan\theta_{3}.\label{eq:deci2ord}$$
In the blue mesonic region of Fig. \[fig:Full-phase\](a), the ground multiplets are always singlets and it follows trivially from Eq. that $\theta=0$ and $-\pi/2$ (points 3 and 4 of the Figure) and $\theta=-\pi/4$ are fixed points of the flow. The same equation can be used to show that points 3 and 4 are stable whereas $\theta=-\pi/4$ is unstable.
In the green baryonic region of Fig. \[fig:Full-phase\](a) both types of decimations occur and the analysis is more involved. The pair of angles $\theta=\pm\pi/2$ taken together are fixed points and singlets are formed out of $k\mathrm{N}\text{ }\left(k=1,2,\ldots\right)$ of SO(N) spins. There are several paths by which this can happen and an illustrative example is shown in Fig. \[fig:baryonic\] for SO(4). In this case, the RG flow involves two anti-symmetric representations depicted by Young tableaux with 1 or 2 stacked boxes. Note how the angle can switch back and forth from $\pi/2$ to $-\pi/2$ depending on the representations involved. This is the fixed point 1 in Fig. \[fig:Full-phase\](a). A stability analysis shows that point 1 is a stable fixed point. Similarly, the extremities of the green region $\theta=\pm\pi/4$ and $\theta=\pm3\pi/4$ are unstable fixed points since, crucially, they lead to Hamiltonians with exact SU(N) symmetry and this symmetry is preserved by the SDRG flow.
![\[fig:baryonic\]An example of singlet formation for SO(4) at the fixed point 1 of Fig. \[fig:Full-phase\](a), with $\theta_{i}$ indicated on each bond.](4sites_baryon3){width="1\columnwidth"}
We now show that the SO(N)-symmetric Hamiltonian of Eq. can be viewed as an SU(N)-anisotropic problem. The $\mathrm{N}^{2}-1$ generators $\left\{ \Lambda_{i}\right\} $ of the fundamental representation of the SU(N) group are traceless Hermitian matrices, normalized as $\mbox{Tr}\left[\Lambda_{i}^{\left(a\right)}\Lambda_{j}^{\left(b\right)}\right]=2\delta^{ab}\delta_{ij}$. We can break this set in a subset of $\mathrm{N}\left(\mathrm{N}-1\right)/2$ **purely imaginary** ** **anti-symmetric** matrices, the generators of SO(N), and another subset of $\mathrm{N}\left(\mathrm{N}+1\right)/2-1$ **real** ** **traceless symmetric** ones, which are SO(N) second-rank tensors [\[]{}see the form of $\hat{O}_{i,i+1}^{\left(2\right)}$ after Eq. . The Hamiltonian is then equivalent to an SU(N)-anisotropic Hamiltonian,
$$\begin{aligned}
H_{i} & = & K_{i}^{\left(1\right)}\sum_{a=1}^{d_{\text{SO(N)}}}\Lambda_{i}^{\left(a\right)}\Lambda_{i+1}^{\left(a\right)}+K_{i}^{\left(2\right)}\sum_{a=d_{\text{SO(N)}}+1}^{\mathrm{N}^{2}-1}\Lambda_{i}^{\left(a\right)}\Lambda_{i+1}^{\left(a\right)},\,\,\label{eq:SUN_aniso}\end{aligned}$$
with $d_{SO(N)}=\frac{N\left(N-1\right)}{2}.$ We can immediately find the expected SU(N)-symmetric points: $K_{i}^{\left(1\right)}=\pm K_{i}^{\left(2\right)}$. That the choice with a minus sign is also SU(N)-symmetric can be seen from the transformation $\Lambda_{i}^{\left(a\right)}\to-\Lambda_{i}^{\left(a\right)*}\equiv\tilde{\Lambda}_{i}^{\left(a\right)}$ on every other chain site, which changes an SU(N) representation into its conjugate and absorbs the minus sign. This case corresponds to having SU(N) (anti-)fundamental representations on odd (even) sites.
The location of these angular fixed points sets the topology of the flow, as shown by the arrows in Fig. \[fig:Full-phase\](a). Although the $\theta$-distribution starts as a delta function, it broadens under the SDRG flow. The existence of the stable fixed points, however, forces the distribution to narrow back down to a delta function at one of the points 1, 3 or 4. Point 2 and its associated black region are outside the scope of this paper as symmetric representations of SO(N) are generated by the flow. The yellow region between the generalized AKLT point $\theta_{VBS}=\arctan\left[\left(\mathrm{N}-2\right)/\left(\mathrm{N}+2\right)\right]$ (blue pentagon) [@PhysRevLett.59.799] and $\pi/4$ flows to the fixed point 4 for odd N. For even N, the procedure becomes ill-defined in this region, and our method cannot be applied [@QuitoLopes2016PRB].
The renormalization of radial variables depends explicitly on the representations being decimated as well as the effective ones being introduced. A systematic derivation of such rules will be given elsewhere [@QuitoLopes2016PRB], but up to pre-factors, the rules are the ones derived in Ref. [@PhysRevLett.89.277203]. The distribution of $r_{i}$ broadens without limit and flows to an infinite disorder form given by $P\left(r\right)\sim r^{\alpha_{i}\left(\Omega\right)-1}$. Here $\alpha_{i}\left(\Omega\right)=\left(\psi_{i}^{-1}-1\right)/\left|\ln\Omega\right|$, $i=B$ or $M$ in the green or blue region, respectively, and $\Omega$ is the decreasing cutoff of the distribution [@fisher94-xxz; @fishertransising2; @igloi-review; @Quito_PhysRevLett.115.167201; @QuitoLopes2016PRB].
In the blue region, adjacent spins always form a singlet and no other representation appears in the flow. The ground state structure is shown in Fig. \[fig:Full-phase\](b). In contrast, in the green region decimations with ground multiplets belonging to any one of the $\mathrm{int}\left(\mathrm{N}/2\right)$ antisymmetric representations of SO(N) are generated.After an initial transient, each one of them is equally populated in the renormalized system [^3]. A singlet only forms out of $k\mathrm{N}\text{ }\left(k=1,2,\ldots\right)$ SO(N) spins, leading to the ground state structure in Fig. \[fig:Full-phase\](c). The different singlet structures lead to different physical properties at finite energies, as discussed above. The apparently intricate combinations leading to singlet formation out of $k\mathrm{N}$ SO(N) spins can be easily understood at the exact SU(N) point $\theta=\pm\pi/4$: only with $k\mathrm{N}$ SU(N) fundamentals can one form an SU(N) singlet [@PhysRevB.70.180401]. The stable fixed points that attract the flow are adiabatically connected to these SU(N) points and have the same ground state structure.
The emergent SU(N) symmetry, as mentioned, relies on the fact that free spins and frozen singlets, the building blocks of the renormalized system, transform as SU(N) fundamentals and singlets, respectively. If we now recall that some of the SU(N) generators $\Lambda_{i}^{\left(a\right)}$ with $a\in\left[\mathrm{N}\left(\mathrm{N}-1\right)/2+1,\mathrm{N}^{2}-1\right]$ are actually 2nd-rank SO(N) tensors (see Eqs. and ), it follows that susceptibilities and correlation functions built with these quadratic SO(N) operators are governed by the same power laws as those of the SO(N) generators. Measuring SO(N) susceptibilities may sound as a challenging task. Yet, we point that this can be envisaged at least for the case of $N=3$. In this case these susceptibilities are just regular magnetic susceptibilities for spin-1 operators. [@Quito_PhysRevLett.115.167201] The susceptibilities for 2nd rank operators in this case are nothing but quadrupolar susceptibilities; protocols for their measurements have recently been proposed at least in two dimensions by considering cross responses between magnetic probes and strain. [@Patri_Multipolar]
*Conclusions.* Our study of random SO(N)-symmetric chains unveils a unified mechanism of symmetry emergence in a large and diverse set of realizable physical situations. Some possible realizations had been previously studied ($\mathrm{N}=2,3$) but new ones ($\mathrm{N}=4,6$) are here introduced. Crucial to the mechanism is the existence of explicit SU(N)-symmetric points in the parameter space whose ground states are adiabatically connected (no local-gap closing) to those of a finite region: symmetry emergence requires no fine tuning. Disorder is the ingredient responsible for filtering, from the set of SO(N) representations, those which find correspondence in the SU(N) group.
*Acknowledgment -* We thank Gabe Aeppli for discussions. VLQ and PLSL contributed equally to this work. VLQ acknowledges financial support from the NSF Grant No. DMR-1555163 and the National High Magnetic Field Laboratory through DMR-1157490, the State of Florida and the Aspen Center for Physics, supported by NSF grant PHY-1607611, for hospitality. PLSL is supported by the Canada First Research Excellence Fund. JAH and EM acknowledge financial support from FAPESP, CNPq and Capes.
[^1]: The proposal from ref. [@Gorshkov2010] generates SU(N)-symmetric spin Hamiltonians, with arbitrary N, in perturbation theory in $1/U$ (where $U$ is the usual Hubbard on-site interaction) in the Mott insulating limit. To lowest order, only the Heisenberg term appears. By symmetry, however, other SU(N)-invariant terms (biquadratic, etc.) are also allowed and appear in higher orders of perturbation theory. Systems which are closer to the Mott transition and at weaker interaction strengths should therefore be described by these more general SU(N) Hamiltonians.
[^2]: We can adhere to this convention if we define $L^{ab}=-L^{ba}$ whenever $a>b$.
[^3]: With one exception: for even N, the self-conjugate representation is half as likely as any of the others.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show how to adapt a $0-f$ self-referencing technique [@Fuji-RGAUYTHK-2005njp; @Rauschenberger-FHYUGHK-2006lpl] to provide a single shot absolute Carrier Envelope Phase (CEP) measurement by using the CEP reference provided by difference frequency generation (DFG) between the spectral wings of the fundamental pulse. Usually, the beat between the input pulse and the DFG signal then provides feedback with which to stabilize the CEP slip in a pulse train. However, with a simple extension we can get a single shot absolute CEP measurement. Success relies on having well characterized input pulses, and the use of accurate propagation models through the nonlinear crystal – these enable us to construct a mapping between the experimental measurement and the CEP of the optical pulse.'
author:
- 'S. B. P. Radnor'
- 'P. Kinsler'
- 'G. H. C. New'
bibliography:
- '/home/physics/\_work/bibtex.bib'
title: 'Proposal for absolute CEP measurement using $0$-to-$f$ self-referencing'
---
=10000
Note: The research in this paper comprised part of S. B. P. Radnor’s PhD dissertation [@Radnor-2007phd], and was completed in conjunction with the other authors. It was initially reported in 2007 at ECLEO in Munich. This text was eventually finalised by P. Kinsler.
Introduction {#S-Introduction}
============
The management of Carrier Envelope Phase (CEP) is of paramount importance in attosecond physics and frequency metrology. In the absence of intervention, the CEP shifts from one pulse to the next in a pulse train, a feature that is unacceptable in experminents involving few-cycle pulses. The first step is to stablize the CEP, the second to measure the CEP offset once stabilized, and the third to be able to create any desired CEP to order.
Recently, a $0-f$ self-referencing technique for CEP stabilization was developed [@Rauschenberger-FHYUGHK-2006lpl; @Fuji-RGAUYTHK-2005njp] as an alternative to the $f-2f$ schemes [@Fortier-JC-2003ol] previously used for CEP stabilization. The scheme was based on difference frequency generation (DFG) [@Boyd-NLO] between the spectral wings of the input pulse, a process in which the overall CEP is cancelled out. The beat between the DFG signal and the extreme low-frequency wing of the pulse then provides feedback to the source laser to enable elimination of the CEP drift.
Measurements of the absolute CEP have been based on a number of techniques including photo-ionization [@Apolonski-DPKHULTBHK-2004prl] high harmonic generation [@Haworth-CRKMT-2007np], and plasma generation [@Kress-LTDGZEMMUR-2006np]. Some innovative methods for single-shot CEP measurement based on spectral interference have also been suggested [@Mehendale-MLVC-2000ol; @Kakehata-TKTFHT-2001ol]. These methods rely on CEP dependent interference occurring between various harmonics. In the case of Mehendale et al. [@Mehendale-MLVC-2000ol] this involves interference between the second and third harmonics, whereas the work by Kakehata et al. [@Kakehata-TKTFHT-2001ol] relies on interference between a delayed fundamental and its second harmonic. Both mechanisms work on the basis of a relative CEP dependent relationship being enforced by the nonlinearity. Though the schemes are interesting, it is not clear how sensitive their interference assumptions are to intensity fluctuations, propagation distance, and so on.
In the present paper, we show how the $0-f$ technique [@Rauschenberger-FHYUGHK-2006lpl; @Fuji-RGAUYTHK-2005njp] can be extended to enable a measurement of absolute CEP to be made [@Radnor-2007phd]. Provided the evolution of the pulse within the nonlinear crystal used for the DFG can be accurately mapped, we show that the CEP can be recovered from the detailed characteristics of the beat signal used as the feedback source in the original experiment. We test the robustness of this technique to phase and intensity variations.
In section \[S-CEP\], we establish a rigorous definition of CEP that provides a sound basis for section \[S-Scheme\], where we discover how the absolute CEP can be recovered from an interferometric measurement. In section \[S-modelling\], we describe the numerical techniques needed to extract the CEP value from the interference record, and in section \[S-Testing\] we test their reliability. In section \[S-Measuring\], we show how this enables one to measure the absolute CEP of an input pulse, followed in section \[S-Conclude\] by our conclusions.
Carrier envelope phase {#S-CEP}
======================
Developments in ultrafast optical pulses have led to the production of sub-cycle pulses. In these limits, robust definitions are needed to fully characterise the pulse, as common descriptors can become ambiguous or fail. Perhaps the best example of this ambiguity is the representation of a pulse with a carrier and envelope, where even as early as 1946 it was known that carrier envelope decompositions (of radar pulses) were not unique [@Gabor-1946jiee]. Brabec and Krausz went some way towards dealing with these issues by suggesting a definition for the central frequency, and stating that an envelope definition is only valid if it remains invariant under a change of phase [@Brabec-K-1997prl].
The most natural way of defining CEP would appear to be based on a time-domain picture in which the time interval between the peak of the pulse and the closest maximum or minimum is measured as a fraction of the optical period. However, for more complex pulse shapes there is no unambiguous way to determine peak of the pulse[^1], or to determine the period when the spectral bandwidth is broad. Further, a fixed pulse envelope in concert with a varying carrier phase generates a pulse that does not guarantee a constant energy or satisfaction of the the zero-force condition (see e.g. [@Milosevic-PBB-2006jpb]). We therefore adopt a purely spectral approach in which we define the absolute spectral CEP of a pulse using the equation $$\begin{aligned}
\Psi_0(\omega)
&=&
\phi_0
+
\psi(\omega)
.
\label{eqn-phase-single-pulse}\end{aligned}$$ The CEP we wish to measure is $\phi_0$, while the relative spectral phase $\psi(\omega)$ is set to zero at a chosen centre frequency $\omega_0$. Since $\psi(\omega_0)=0$, it follows that $\phi_0=\Psi(\omega_0)$. We note that $\psi(\omega)$ is a measurable quantity even for ultrashort pulses – It is possible to determine the relative spectral phases of few-cycle pulses [@Cheng-FSLSK-1999ol; @Kobayashi-SF-2001ieeqe][^2]. This can be achieved to an accuracy of 0.04 rads using the SPIDER technique [@Anderson-AKW-2000apb]. This $\psi(\omega)$ tells us the relative CEP of all frequency components of the pulse, something which we use in the scheme presented in this paper.
Having considered a single pulse, we now need to consider a pulse train, the spectrum of which is a comb of equally spaced frequency components with the lowest frequency tooth at $\omega_{cep}$. The phase slip between pulses is $\Delta = \omega_{cep} \tau$, where $\tau$ is the time interval between sucessive pulses in the train. Hence the *absolute* spectral phase of the $n$-th pulse in a train is $$\begin{aligned}
\Psi_n(\omega)
&=&
\phi_n
+
\psi(\omega)
\label{eqn-phase-nth-pulse}
,\end{aligned}$$ where $\phi_n = \phi_0 + n \Delta$.
Scheme {#S-Scheme}
======
If we know the relative spectral phase $\psi(\omega)$, *and* we can determine the absolute CEP of any one frequency, then we can calculate the absolute CEP of any other. The advantage of a DFG process [@Boyd-NLO] is that it provides us with just such a reference CEP.
To get efficient DFG, a nonlinear crystal (such as MgO:LN) is periodically poled to phase match a particular frequency mixing process. The poling period is chosen so that selected frequencies on the upper ($\omega_{+}$) and lower ($\omega_{-}$) wings of the pulse spectrum are phase matched for DFG at $\omega_d=\omega_{+}-\omega_{-}$, as illustrated on fig. \[fig-diagram\]. The nonlinear interaction generates a polarization term whose phase $\Psi_{p}$ is $$\begin{aligned}
\Psi_{p}
&=&
\Psi_n(\omega_{+})
-
\Psi_n(\omega_{-})
~~~~
=
\psi(\omega_{+})
-
\psi(\omega_{-})
.
\label{eqn-Psi-polariz}\end{aligned}$$ Since the unknown offsets $\phi_0$ and $\Delta$ have canceled out, $\Psi_{p}$ can be calculated from $\psi(\omega)$, and provides a CEP reference. Unfortunately, this cannot be measured directly, so we have to analyse how both the incident pulse and the DFG propagates through the crystal, and how they interfere and are measured.
*First,* let us consider the DFG component. The polarization term (with absolute phase $\Psi_{p}$) will then generate a DFG signal that *exits* the crystal with a phase shifted by an amount $\delta_p$, so that $$\begin{aligned}
\Psi_{d}
&=&
\Psi_{p} + \delta_p
.
\label{eqn-Psi-DFG}\end{aligned}$$
![Diagram of the DFG signal being generated, comparing the phase matched DFG component to the wing of the input pulse. []{data-label="fig-diagram"}](fig01-diagram){width="0.70\columnwidth"}
We can see how this shift $\delta_p$ arises in the idealised case where a detectable DFG signal could be generated from a thin layer of dispersionless $\chi^{(2)}$ medium. Here $\delta_p=\pi/2$, since the DFG signal is just the integral of the driving polarization. In general, however, $\delta_p$ will be a complicated function of both $\Psi_{p}$ and the pulse intensity and profile, since it results from propagation through a crystal which is both nonlinear and dispersive.
*Second,* we need to consider the wing of the input pulse spectrum at frequency $\omega_d$, which is co-propagating with the DFG component discussed above. This part of the pulse had an initial phase $\Psi_n(\omega_{d})$, but this changes as it propagates through the crystal, and on exit it has a phase $\Psi'_n(\omega_{d})$ that has been shifted by $\delta_d$, i.e. $$\begin{aligned}
\Psi'_n(\omega_{d})
&=&
\Psi_n(\omega_{d}) + \delta_d
.
\label{eqn-Psi-propagate}\end{aligned}$$ Although $\delta_d$ will predominantly arise from the dispersion, but there will also be additional contributions, due to other DFG processes (i.e. those not involving $\omega_{+}, \omega_{-}$), from SPM, four-wave mixing or other processes. However, with careful design these can be minimized, and so will only add an unimportant (but nevertheless calculable) offset to $\delta_d$.
*Lastly,* we apply our understanding of both the DFG and pulse propagation to determine the interference between them, and what would be measured on a photodetector. As the pulse exits the crystal, the DFG component (with phase $\Psi_{d}$) will interfere with the wing of the input pulse at that same frequency, which now has phase $\Psi'_n(\omega_{d})$. Thus the photodetector sees an interference between the DFG and the spectral wing at frequency $\omega_d$. From this interference it is possible to infer the relative phase $\delta_i$ between the contributions, which is $$\begin{aligned}
\delta_i
&=&
\Psi'_n(\omega_{d})
-
\Psi_{d}
\\
&=&
\Psi_n(\omega_{d})
+
\delta_d
-
\Psi_{d}
.
\label{eqn-Psi-interference}\end{aligned}$$ In order to optimize the visibility of this interference, we need to ensure that the DFG signal and the amplitude of the wing of the input pulse at that frequency are comparable; if either is too dominant, the CEP sensitive modulation of the interference will be less detectable.
In existing $0-f$ CEP stabilization experiments[@Rauschenberger-FHYUGHK-2006lpl; @Fuji-RGAUYTHK-2005njp], the evolving interference signal resulting from the train of CEP-slipping pulses produces a beat signal dependent on the CEP slip. This beat is then used in feedback designed to reduce the CEP slip to zero, thus *stabilizing* the CEP of the pulses in the train to a fixed (but unknown) value $\Psi_n(\omega)=
\phi_0 + \psi(\omega)$.
*The scheme presented here works because we incorporate additional information based on knowledge of how the pulse propagates through the crystal.* This means that we can predict the phase shifts $\delta_p$ and $\delta_d$, at which point the interference measurement (i.e. of $\delta_i$) turns a simple CEP stabilization into a CEP measurement.
To make it clear how the CEP measurement is constructed, we now show how the various phases in the scheme are related. We are specifically interested in the CEP $\phi_0$ of the central frequency $\omega_0$ of the pulse; so it is useful to write down how the relative phases between the low frequency ($\omega_d$) wing of the input pulse and its centre are $$\begin{aligned}
\Psi_n(\omega_0)
&=&
\Psi_n(\omega_{d})
% +
% \psi(\omega_0)
-
\psi(\omega_{d})
.\end{aligned}$$ Now, by substituting in the preceeding collection of phase relationships (eqns. (\[eqn-Psi-polariz\], \[eqn-Psi-DFG\], \[eqn-Psi-propagate\], \[eqn-Psi-interference\])), we can get $$\begin{aligned}
\Psi_n(\omega_0)
=
\phi_n
&=&
% \Psi_n(\omega_{d})
% +
% \psi(\omega_0)
% -
% \psi(\omega_{d})
%\\
%&=&
% \Psi'_n(\omega_{d}) - \delta_{d}
% +
% \psi(\omega_0)
% -
% \psi(\omega_{d})
%\\
%&=&
% \Psi_{d} + \delta_{i} - \delta_{d}
% +
% \psi(\omega_0)
% -
% \psi(\omega_{d})
%\\
%&=&
% \Psi_{p}
% + \delta_{p}
% + \delta_{i}
% - \delta_{d}
% +
% \psi(\omega_0)
% -
% \psi(\omega_{d})
%\\
%&=&
\psi(\omega_{+})
-
\psi(\omega_{-})
+
\delta
% +
% \psi(\omega_0)
-
\psi(\omega_{d})
,
\label{eq-absphi}\end{aligned}$$ where the sum of all the nonlinearity and propagation phase shifts is $\delta = \delta_{p} + \delta_{i} - \delta_{d}$.
Thus if we \[a\] know the relative phase spectrum $\psi(\omega)$ of the input pulse(s), \[b\] understand the phase shifts introduced by propagation through the crystal ($\delta_d$) and of the DFG ($\delta_p$), and then can \[c\] *measure* the interference phase difference ($\delta_{i}$), we will know every part of the RHS of eqn. (\[eq-absphi\]) – i.e. we know the absolute CEP $\phi_n$ of the incoming pulse.
The most challenging part of the CEP measurement is determining the $\delta_{p}$ and $\delta_{d}$ contributions. Fortunately, we can avoid having to calculate them individually by calculating them all at the same time in a numerical simulation, as discussed in the following section.
Modeling {#S-modelling}
========
The modeling is a crucial part of the scheme, since it allows us to determine how differing input CEPs map onto the detected interference measurements. After choosing our crystal and evaluating its parameters, and characterising the pulses in our pulse train (particularly $\psi(\omega)$, we run a set of simulations over the range of CEPs. The results can then be used to build a map between the input CEP and the interference signal. To do this we need to take the spectrum of each output pulse from a simulation, and integrate over the detector response. In our results, we assume the spectral response of an InGaAs-Hamamatsu PD as used by Fuji and others [@Fuji-RGAUYTHK-2005njp; @Rauschenberger-FHYUGHK-2006lpl]. We then need to check the mapping, and ensure that we will get the required level of discrimination between interference measurements from pulses with different CEP’s.
To do the propagation part of the modeling, we solve Maxwell’s equations using the PSSD technique [@Tyrrell-KN-2005jmo] for a chosen crystal thickness. In existing $0-f$ experiments, the crystal thickness is typically of the order of millimetres, so diffraction is negligible. We consider MgO:LN, with parameters taken from Further, with a suitable choice of nonlinear crystal (i.e. MgO:LN, and parameters from [@Handbook-NLOCX; @Zheng-WPCF-2002josab]), the nonlinear $\chi^{(2)}$ interaction occurs only in the extraordinary polarization (i.e. is $e+e\rightarrow e$), allowing us to further reduce Maxwell’s equations to $$\begin{aligned}
\label{eq:Max}
\frac{\partial E_{x}}{\partial z}
&=&
-\mu_{0}\frac{\partial H_{y}}{\partial t}
\\
\frac{\partial H_{y}}{\partial z}
&=&
-
\epsilon_{0}\frac{\partial}{\partial t}\left[E+\chi^{(1)}*E
+
\chi^{(2)}E^{2}
+
\chi^{(3)}E^{3}\right]
,
\label{eqn-maxwellseqns}\end{aligned}$$ where $\chi^{(1)}$ contains linear dispersion, and any nonlinear response is assumed to be instantaneous.
In existing experiments [@Rauschenberger-FHYUGHK-2006lpl; @Fuji-RGAUYTHK-2005njp], a MgO:LN crystal is periodically poled at $11.21\mu$m, and is optimized for DFG between the wings of the fundamental: $\omega_{+}(3.04\times 10^{15})-\omega_{-}(1.885\times 10^{15})
=\omega_{d}(1.155\times 10^{15})$ \[rad s$^{-1}$\]. The peak pulse power was calculated to be $\simeq5\times 10^{11} \textrm{W/cm}^{2}$, with a duration of $\sim$ 6fs (830 nm carrier). At these pulse powers in MgO:LN, the relative nonlinear strengths in the crystal at the pulse peaks are $\chi^{(2)} E = 0.082$ and $\chi^{(3)} E^{2} = 0.0027$. This means that the self-phase modulation (SPM) distance is $L_{SPM}=0.26$mm, implying significant SPM over a 2mm crystal, along with other $\chi^{(3)}$ effects such as 4-wave mixing, further complicating propagation and DFG. Modelling of extreme SPM (only) and sensitivity to CEP has been considered by Kinsler [@Kinsler-2007-phspm] and also Genty et al.[@Genty-KKD-2008oe].
With these parameter values, the pulses inside the crystal decohere rapidly because of their high intensity and short duration. In combination with the relatively small energy content within the spectral wings, the DFG does not continue to grow throughout the whole crystal, but does so only over several coherence lengths. Nevertheless, the DFG signal can still be made large enough for the photodiode (PD) to detect a beat against the wing of the input spectrum.
In our modeling, we retained the bulk of these parameter values to match experiment, but adjusted the crystal thicknesses to $\sim 100\mu$m. This is because thinner crystals still create sufficient DFG, but also have the advantage of producing a cleaner CEP to interference signal mapping.
Testing the CEP response at $\omega_{d}$ {#S-Testing}
========================================
In order to make our scheme work, we need to guarantee that the response of the measured interference signal depends on the CEP in a reliable way, *and* that it is sufficiently insensitive to other pulse characteristics, such as intensity. Obviously these will depend on the particular design of the experiment, e.g. the size of the chosen nonlinear crystal, the periodic-poling length, pulse frequency, and so on. In this section, we use the parameters described in the previous section to test the stability of the pulse propagation and interference signal against CEP variation and intensity fluctuations.
Response to CEP slip $\Delta$ {#Ss-Testing-ce}
-----------------------------
To test whether the interference signal will behave as expected, we did a set of simulations. Each simulations started with the same parameters, except for a cumulative “shot-to-shot” CEP slip of $\Delta = \pi/10$. Fig. \[fig-CEP\_10\] shows example input pulses.
Let us consider an idealized case, with only DFG and a linearly propagated input pulse present. Here we should see two regimes when comparing the phase spectrum of output pulses generated by input pulses differing by a CEP slip $\Delta$.
1. The spectrum of the input pulse dominates the contribution from the DFG signal, typically this occurs closer to the centre of the pulse spectrum, i.e. $\omega>\omega_{1}$ In this case, the CEP of the pulse dominates, so that the phase difference between subsequent pulses will just be the inter-pulse CEP slip $\Delta$.
2. The DFG signal dominates the contribution from the input pulse, typically this occurs for low frequencies, i.e. $\omega<\omega_{1}$. Since the DFG is insensitive to the input CEP, the phase difference between subsequent pulses will be zero.
In between these two regimes will be a transition region where the two are comparable; this is just the regime in which we look for the interference between the incoming pulse (with its phase slip $\Delta$, as described in point 1 above); and the phase stabilised DFG (with no phase slip, as described in point 2 above). Both regions, and the transition region of interference between them can be seen on fig. \[fig:cep\]. It is important to note that, the change between different pairs of simulations is small – irrespective of the absolute CEP values chosen for the two pulses. The only departures are at the narrow spikes caused by the (expected) strong CEP sensitivity near the nodes of the pulse. As an aside, we could (if desired) also estimate the linearity of the response to the CEP slip as done by Kinsler [@Kinsler-2007-phspm] in the extreme nonlinear regime; that work also suggests more systematic tests of simulation pairs to investigate the CEP dependence.
Here, however, we are satisfied by the fact that fig. \[fig:cep\] not only shows the predicted DFG phase stabilized region at low frequencies, but also the existance of a transition region where strongly modulated phase sensitive intereference takes place.
Response to intensity fluctuations {#Ss-Testing-I}
----------------------------------
Our scheme relies on nonlinear interactions, but these are strongly intensity dependent. This means that intensity changes between pulses in the train might change the interference signal in a way that masks the CEP sensitivity of the interference signal. We now test the effect of intensity fluctuations by fixing the CEP for a set of simulations, whilst making shot-to-shot changes in intensity spanning a range of $\sim \pm 1$%.
The results of this simulation set are displayed on fig. \[fig:cep\], which demonstrates that intensity variation has a relatively weak effect in the interference region. This means that the mapping between photodetector signal and absolute CEP will be *insensitive* to the intensity fluctuations in the pulse train. As a result, for our parameters, we can disregard the effect of intensity fluctuations when it comes to reconstructing the CEP of an individual pulse. In a more extreme nonlinear case this is not always true, see e.g. [@Kinsler-2007-phspm]; but here the nonlinear phase shifts (e.g. those due to SPM) are not strongly intensity dependent.
Measuring the Absolute CEP {#S-Measuring}
==========================
In the previous section we demonstrated that (or our chosen parameters) not only was the interference in the $\omega_{d}$ DFG region sensitive to CEP in a controllable way, the effect would not be masked by intensity fluctuations. Since we see this clear dependence on CEP, it is possible to determine the absolute CEP from the interference signal – as long as the pulse intensity and crystal parameters are appropriately matched. For example, the CEP dependence is better behaved at some distances than at others – for our chosen parameters, it happens that distances of $\sim 50, 100\mu$m give good results.
Fig. \[fig:wow1\] shows the CEP dependent structure at a propagation distance of 50$\mu$m, where each curve is the intensity for a different input CEP. Using this, we can then integrate that spectral behaviour over the response of the photodetector to generate our mapping. The mapping corresponding to fig. \[fig:wow1\] and our chosen photodiode is shown on fig. \[fig:wow2\].
However, because each interference signal value is not unique, we have only determined the CEP to within $\pi$. To complete the determination of the CEP to within a $2\pi$ range we need to take two such measurements under slightly different conditions, e.g. using different propagation lengths.
Summarizing, for this CEP measurement scheme to succeed, we must have accurate knowledge of the following three things:
1. **The phase spectrum $\psi(\omega)$ of the pulse.** In eqn.(\[eq-absphi\]) we see that it depends on the sums or difference of four values of $\psi(\omega)$, thus compounding any uncertainty in its determination.
2. **The intensity fluctuations of the pulse.** The intensity must be controlled sufficiently well so that the SPM or XPM effects are minimized, and do not significantly alter the propagation of the pulse through the crystal.
3. **Pulse propagation.** This can be done easily with PSSD simulation code (or similar), but we need accurate information on the initial conditions of the pulses in the pulse train.
Remarkably, this can be done with a small extension to current $0-f$ self-referencing methods, which are currently only used to stabilize the CEP. The extension is to numerically model the propagation of the pulse through the nonlinear crystal in order to determine the mapping between the photdetector signal and the input CEP.
Conclusion {#S-Conclude}
==========
We have demonstrated how an ordinary $0-f$ self-referencing scheme can be easily extended to measure absolute CEP, rather than just being used as a CEP stabilization tool. The scheme relies on a phase stable signal being passively produced through DFG, and does not require strong field physics to operate. Instead, we propose to numerically model the propagation of the pulse through the nonlinear crystal; and to use the information gained to determine the mapping between the detected interference signal and absolute CEP of the input pulse.
[^1]: It is of course possible to invent schemes for generating a suitable centre position, e.g. by calculating a weighted average over its intensity profile (see e.g. [@Brabec-K-1997prl]) – however this can perform poorly for pulses with satellite peaks
[^2]: In Kobayashi et al.’s Fig. 4(b), $\psi$ extends from 550nm to 800nm. In our scheme, $\psi$ will most likely need to be characterised further down into the low frequency wing – our $\omega_{+}/\omega_d$ ratio is $2.6$; for Kobayashi et al. the ratio was only $1.5$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Low-lying $\Lambda$ baryons with spin 1/2 are analyzed in full (unquenched) lattice QCD. We construct $2 \times 2$ cross correlators from flavor SU(3) “octet” and “singlet” baryon operators, and diagonalize them so as to extract information of two low-lying states for each parity. The two-flavor CP-PACS gauge configurations are used, which are generated in the renormalization-group improved gauge action and the ${\mathcal O}(a)$-improved quark action. Three different $\beta$’s, $\beta = 1.80$, 1.95 and 2.10, are employed, whose corresponding lattice spacings are $a = 0.2150$, 0.1555 and 0.1076 fm. For each cutoff, we use four hopping parameters, ($\kappa_{\rm val}, \kappa_{\rm sea}$), which correspond to the pion masses ranging about from 500 MeV to 1.1 GeV. Results indicate that there are two negative-parity $\Lambda$ states nearly degenerate at around 1.6 GeV, while no state as low as $\Lambda (1405)$ is observed. By decomposing the flavor components of each state, we find that the lowest (1st-excited) negative-parity state is dominated by flavor-singlet (flavor-octet) component. We also discuss meson-baryon components of each state, which has drawn considerable attention in the context of multi-quark pictures of $\Lambda (1405)$.'
author:
- 'Toru T. Takahashi'
- Makoto Oka
title: 'Low-lying $\Lambda$ Baryons with spin 1/2 in Two-flavor Lattice QCD'
---
Introduction
============
$\Lambda (1405)$ is one of the most interesting hadrons and attracting much interest from several view points. $\Lambda (1405)$ is the lightest negative-parity baryon, even though it has one valence strange quark, which is heavier than the up and down quarks. Among the $J^P = 1/2^-$ baryons, $\Lambda (1405)$ is isolated from the others, much lighter than the non-strange counterpart $N(1535)$. It has no spin-orbit partner in the vicinity, as the lowest spin $3/2^-$ state is $\Lambda (1520)$. Furthermore, the structure of $\Lambda (1405)$ remains mysterious. On one hand, $\Lambda (1405)$ is interpreted as a flavor-SU(3)-singlet three-quark state in conventional quark models. On the other hand, $\Lambda (1405)$ could be interpreted as an antikaon-nucleon $\bar KN$ molecular bound state (B.E. $\sim$ 30 MeV). The binding energy of $\bar KN$ implies a strong attractive force between $\bar K$ and $N$ [@Sakurai:1960ju; @Dalitz:1967fp], which may cause a new type of dense hadronic matter, kaonic nuclei or kaonic nuclear matter [@Akaishi:2002bg; @Yamazaki:2002uh; @Akaishi:2005sn]. We also expect that such $\bar KN$ bound states with large binding energies can be regarded as compact 5-quark states. The 5-quark picture of $\Lambda(1405)$ has advantages that all five quarks can be placed in the lowest-lying $L=0$ state to form a negative-parity baryon, and also that it requires no spin-orbit partner of $\Lambda (1405)$.
The property of $\Lambda (1405)$ can therefore be an important clue to new paradigm in hadron physics. We here study properties of $\Lambda(1405)$ using the lattice QCD formulation. Lattice QCD is nowadays employed as a very powerful tool for nonperturbative analysis directly based on QCD, and expected to cast light on the nature of $\Lambda (1405)$ in a model-independent way. Though several lattice QCD studies on $\Lambda (1405)$ have been performed so far [@Melnitchouk:2002eg; @Nemoto:2003ft; @Burch:2006cc; @Ishii:2007ym], most of them are based on quenched QCD and few lattice QCD studies succeeded in reproducing the mass of $\Lambda (1405)$. Moreover, little has been discussed on the lattice about the possible mixing of flavor-SU(3)-octet and -singlet components induced by the symmetry breaking.
Several possible reasons for the failure of reproducing $\Lambda (1405)$ in lattice QCD were suggested through these studies, such as missing meson-baryon components due to quenching, exotic (non-3-quark type) structure of $\Lambda (1405)$, or insufficiency of the lattice volume in the simulations. Resolving such difficulties requires unquenched lattice QCD calculation on a larger lattice volume with varieties of interpolating operators.
In this paper, we aim at clarifying the properties of $\Lambda (1405)$ with two-flavor full lattice QCD, adopting the “octet” and “singlet” baryon operators to construct correlation matrices, which enables us to extract the low-lying spectrum as well as the mixing between octet and singles components in $\Lambda (1405)$.
Lattice QCD setups
==================
simulation conditions
---------------------
We adopt the renormalization-group improved gauge action and the ${\mathcal O}(a)$-improved quark action. Simulation parameters are listed in Table \[simparam\]. We adopt three different $\beta$’s, $\beta = 1.80$, 1.95 and 2.10, and corresponding lattice spacings are $a = 0.2150$, 0.1555 and 0.1076 fm [@AliKhan:2001tx], which are determined so that the empirical $\rho$-meson mass is reproduced. The hopping parameters for strange quark $\kappa_s$ are set to be 0.1431, 0.1393, and 0.1373 at $\beta$=1.80, 1.95, and 2.10 so that the Kaon mass is reproduced. We note here that, if we adopt $\phi$-meson mass as an input, $\kappa_s$ would be slightly different from those determined by the Kaon mass. The dynamics of $\Lambda$ resonances would be closely connected to Kaon dynamics, and hence we adopt Kaon mass as an input in this study. We employ four different hopping parameters ($\kappa_{\rm val}, \kappa_{\rm sea}$) for each cutoff. Corresponding pion masses range approximately from 500MeV to 1.1 GeV at each $\beta$.
$\beta$ $c_{\rm SW}$ $N_s^3\times N_t$ $a$ \[fm\] $N_sa$ \[fm\] $\kappa_s$
--------- -------------- ------------------- ------------ --------------- ------------
1.80 1.60 $12^3\times 24$ 0.2150 2.580 0.1431
1.95 1.53 $16^3\times 32$ 0.1555 2.488 0.1393
2.10 1.47 $24^3\times 48$ 0.1076 2.582 0.1373
: \[simparam\] List of the present simulation parameters. $\beta$ and $c_{\rm SW}$ are the parameters for the renormalization-group improved gauge action and the ${\mathcal O}(a)$-improved quark action. $N_s$ and $N_t$ denote the spatial and temporal lattice extents, and the lattice spacings $a$ and the strange-quark hopping parameters $\kappa_s$ are determined so that the empirical $\rho$-meson mass and Kaon mass are reproduced [@AliKhan:2001tx].
baryonic operators
------------------
In order to extract the low-lying states in $S=-1$ and isosinglet channel, we construct $2\times 2$ cross correlators from the following “singlet” and “octet” operators, $$\begin{aligned}
{\eta}_{\bf 1}(x)
\equiv
\Lambda_1(x)
=
\frac{1}{\sqrt{3}}
\epsilon^{abc}
\left\{
u^a(x)[d^{T b}(x) C \gamma_5 s^c(x)]
\right. \nonumber \\ \left.
+
d^a(x)[s^{T b}(x) C \gamma_5 u^c(x)]
+
s^a(x)[u^{T b}(x) C \gamma_5 d^c(x)]
\right\}
\label{defeta1}\end{aligned}$$ $$\begin{aligned}
{\eta}_{\bf 8}(x)
\equiv
\Lambda_8(x)
=
\frac{1}{\sqrt{6}}
\epsilon^{abc}
\left\{
u^a(x)[d^{T b}(x) C \gamma_5 s^c(x)]
\right. \nonumber \\ \left.
+
d^a(x)[s^{T b}(x) C \gamma_5 u^c(x)]
-2
s^a(x)[u^{T b}(x) C \gamma_5 d^c(x)]
\right\}
\label{defeta2}\end{aligned}$$ It is easy to check that ${\eta}_{\bf 1}(x)$ (${\eta}_{\bf 8}(x)$) belongs to the singlet (octet) irreducible representation of the flavor SU(3). The cross correlator of ${\eta}_{\bf 1}$ and ${\eta}_{\bf 8}$ vanishes in the flavor-SU(3) symmetric limit ($m_u=m_d=m_s$). We define cross correlators as $${\mathcal M}(x,y)_{IJ}
\equiv
{\rm Tr}
\frac{1\pm\gamma_4}{2}
\langle
{\eta}_I(x)
\bar{\eta}_J(y)
\rangle,$$ where the trace is taken over spinor indices. The parity-projection operator, $P(\pm) \equiv \frac{1\pm\gamma_4}{2}$, is inserted to separate the positive and negative parity states. In the actual simulation, we adopt point-type operators for the sink, $\eta(x)$, and extended operators, which are smeared in a gauge-invariant manner, for the source, $\bar\eta(y=0)$. Smearing parameters are chosen so that root-mean-square radius is approximately 0.5 fm.
diagonalization of cross correlators
------------------------------------
We here consider a general situation where we have a set of $N$ independent operators, $\eta^{\rm snk}_I$ for sinks and $\eta^{\rm src \dagger}_I$ for sources in order to construct correlation matrices ${\mathcal M}_{IJ}(t)\equiv\langle \eta^{\rm snk}_I(t)
\eta^{\rm src \dagger}_J(0)\rangle$, which can be decomposed into the sum over the energy eigenstates $|i \rangle$ as $$\begin{aligned}
{\mathcal M}_{IJ}(t)
&\equiv&
{\mathcal M}_{IJ}(t,0)
=
\langle \eta^{\rm snk}_I(t) \eta^{\rm src \dagger}_J(0)\rangle
\nonumber \\
&=&
\sum_{i,j}
(C^\dagger_{{\rm snk}})_{Ii}
\Lambda(t)_{ij}
(C_{{\rm src}})_{jJ} \nonumber \\
&=&
(C^\dagger_{\rm snk} \Lambda(t) C_{\rm src})_{IJ},\end{aligned}$$ where the small letters ($ij$) are the indices for the intermediate mass eigenstates and $$\begin{aligned}
(C^\dagger_{\rm snk})_{Ii}\equiv \langle {\rm vac} | \eta^{\rm snk}_I | i \rangle,
\quad
(C_{\rm src})_{jI}\equiv \langle j | \eta^{\rm src \dagger}_J | {\rm vac} \rangle, \end{aligned}$$ are the coupling matrices. $\Lambda(t)$ is a diagonal matrix given by the energy eigenvalue, $E_i$, of the state $i$. $$\Lambda(t)_{ij}\equiv \delta_{ij} e^{- E_it}.
\label{dmatrix}$$ From the product $${\mathcal M}^{-1}(t+1){\mathcal M}(t)
=C_{\rm src}^{-1}\Lambda(-1)C_{\rm src},$$ we extract the effective eigen-energies (or the effective masses after the zero-momentum projection) {$E_i$} at the time slice $t$ as the logarithm of eigenvalues {$e^{E_i}$} of the matrix ${\mathcal M}^{-1}(t+1){\mathcal M}(t)$.
Besides overall constants, $(C_{\rm src})^{-1}$ and $(C^\dagger_{\rm snk})^{-1}$ can be obtained as right and left eigenvectors of ${\mathcal M}^{-1}(t+1){\mathcal M}(t)$ and ${\mathcal M}(t){\mathcal M}(t+1)^{-1}$ respectively, since $${\mathcal M}^{-1}(t+1){\mathcal M}(t)(C_{\rm src})^{-1}
=(C_{\rm src})^{-1}\Lambda(-1)$$ and $$(C^\dagger_{\rm snk})^{-1}{\mathcal M}(t){\mathcal M}(t+1)^{-1}
=\Lambda(-1)(C^\dagger_{\rm snk})^{-1}$$ hold. One can also directly construct optimal source and sink operators, ${\cal O}^{\rm src \dagger}_i$ and ${\cal O}^{\rm snk}_i$, which couple dominantly (solely in the ideal case) to $i$-th lowest state, as $${\cal O}^{\rm src \dagger}_i=\sum_J
\eta_J^{\rm src \dagger} (C_{\rm src})^{-1}_{Ji}$$ and $${\cal O}^{\rm snk}_i=\sum_J
(C^\dagger_{\rm snk})^{-1}_{iJ} \eta_J^{\rm snk}.$$ When an optimal correlator which corresponds to a single-state propagation is needed, one can extract it by sandwiching ${\mathcal M}(t)$ between $(C^\dagger_{\rm snk})^{-1}$ and $(C_{\rm src})^{-1}$ since $$(C^\dagger_{\rm snk})^{-1} {\mathcal M}(t)(C_{\rm src})^{-1}
=\Lambda(t)
\label{diag01}$$ is diagonal. We note here that, if the correlation matrix ${\mathcal M}(t)$ is hermitian, one can determine $(C_{\rm src})^{-1}$ and $(C_{\rm snk})^{-1}$ up to overall phase factors so that Eq.(\[diag01\]) is satisfied.
These prescriptions are valid only in the $t$-range where the number of the relevant physical states is no more than $N$, since we prepare a set of just $N$ independent operators. In practice, higher excited states may come in, and such contamination manifests itself as a $t$-dependence in several “constants” mentioned above. In order to avoid the contamination of the excited states, we identify a $t$-window in which the “constants” are (almost) $t$-independent. All the physical quantities are extracted within such a $t$-window.
different choices in formulation
--------------------------------
We here discuss the difference in formulation. In general, we solve generalized eigenvalue problems [@Luscher:1990ck; @Perantonis:1990dy], $$\begin{aligned}
&&{\mathcal M}^{-1}(t'){\mathcal M}(t)
=(C_{\rm src}^{-1})\Lambda(t-t')(C_{\rm src}), \\
&&{\mathcal M}(t){\mathcal M}^{-1}(t')
=(C^\dagger_{\rm snk})\Lambda(t-t')(C^\dagger_{\rm snk})^{-1},\end{aligned}$$ where we have ambiguity in the choice of $t'$. As long as a system is dominated by the states as many as the rank of correlation matrices, choice dependence vanishes. In practice, we may have excited-state contaminations, and choice dependence can emerge. In our case, $|t-t'|$ is fixed to 1, and we take as large $t$ as possible: We seek plateau regions and fit data in that window. The authors in Ref. [@Blossier:2009kd] investigated possible subleading contaminations by higher excited states and the dependence on the choice of $t$ and $t'$.
To estimate the systematic errors coming from such ambiguity, we try four different choices, $|t-t'|$=1, 2, 3, and 4, and see the dependence. Namely, we solve the eigenvalue problems $${\mathcal M}(t){\mathcal M}^{-1}(t+m)
=(C^\dagger_{\rm snk})\Lambda(-m)(C^\dagger_{\rm snk})^{-1},$$ with $m\leq 4$. We plot in Fig. \[tcheckexample\] $(C_{\rm snk})_{0{\bf 1}}$/$(C_{\rm snk})_{0{\bf 8}}$, which can be a measure of flavor contents in an eigenstate and will be discussed later. They are obtained at $\kappa_{\rm sea}, \kappa_{\rm val} = 0.1367$ at $\beta = 2.10$. Though they are all t-independent in an ideal case, there remain considerable contaminations especially with small $m$. On the other hand, it reaches plateau region much faster with larger $m$. However larger $m$ results in larger statistical fluctuations and diagonalization sometimes fails at relatively small $t$. In our case, as long as we identify plateaus properly, resulting values are all consistent with each other.
![\[tcheckexample\] $(C_{\rm snk})_{0{\bf 1}}$/$(C_{\rm snk})_{0{\bf 8}}$ obtained by solving eigenvalue problem, ${\mathcal M}(t){\mathcal M}^{-1}(t+m)
=(C^\dagger_{\rm snk})\Lambda(-m)(C^\dagger_{\rm snk})^{-1}$ with $m\leq 4$ are plotted as a function of $t$. They are obtained at $\kappa_{\rm sea}, \kappa_{\rm val} = 0.1367$ at $\beta = 2.10$. The horizontal lines denote the fitted value and the corresponding error evaluated by fitting plateaus.](tcheck.eps "fig:") ![\[tcheckexample\] $(C_{\rm snk})_{0{\bf 1}}$/$(C_{\rm snk})_{0{\bf 8}}$ obtained by solving eigenvalue problem, ${\mathcal M}(t){\mathcal M}^{-1}(t+m)
=(C^\dagger_{\rm snk})\Lambda(-m)(C^\dagger_{\rm snk})^{-1}$ with $m\leq 4$ are plotted as a function of $t$. They are obtained at $\kappa_{\rm sea}, \kappa_{\rm val} = 0.1367$ at $\beta = 2.10$. The horizontal lines denote the fitted value and the corresponding error evaluated by fitting plateaus.](tcheck2.eps "fig:")
Lattice QCD Results
===================
hadronic masses
---------------
![\[effmassexample\]The effective mass plots for ground and 1st-excited states in the positive- (upper panel) and negative-parity (lower panel) channels, which are obtained with $\kappa_{\rm sea}, \kappa_{\rm val} = 0.1367$ at $\beta = 2.10$. The horizontal lines are fits of the eigen-energies in the plateau region for each state.](pLm_effmass.eps "fig:") ![\[effmassexample\]The effective mass plots for ground and 1st-excited states in the positive- (upper panel) and negative-parity (lower panel) channels, which are obtained with $\kappa_{\rm sea}, \kappa_{\rm val} = 0.1367$ at $\beta = 2.10$. The horizontal lines are fits of the eigen-energies in the plateau region for each state.](nLm_effmass.eps "fig:")
In Fig. \[effmassexample\], as typical results, effective masses are plotted for the ground and the 1st-excited states obtained from positive- and negative-parity projected data with $\kappa_{\rm sea}, \kappa_{\rm val} = 0.1367$ at $\beta = 2.10$. We can find $t$-windows in which the effective masses exhibit plateaus. We extract eigen-energies by fitting lattice QCD data in the $t$-range where both of the ground- and 1st-excited states exhibit plateaus. In Table \[hadronicmass\], we list all the hadronic masses extracted in this study. Figs. \[posmass\] and \[negmass\] show the eigen-energies in the positive- and negative-parity channels, respectively, plotted as functions of the squared pion mass $m_\pi^2$. Filled circles denote the energies of the ground states, and open squares those of 1st-excited states. The solid curves represent quadratic fits as a function of squared pion mass $m_\pi^2$. Two solid lines at the vertical axes indicate the masses of $\Lambda(1115)$ and $\Lambda(1600)$ in Fig. \[posmass\] (positive-parity states), and $\Lambda(1405)$, $\Lambda(1670)$ and $\Lambda(1800)$ in Fig. \[negmass\] (negative-parity states).
The obtained masses of the positive-parity ground state agree very well with the mass of the ground-state $\Lambda(1115)$ at all three $\beta$’s. On the other hand, the 1st-excited state in this channel lies much higher than $\Lambda(1600)$, which is the 1st excited state experimentally observed so far. The same tendency was reported in Ref. [@Burch:2006cc], and the situation is similar to the case of the Roper resonance, which is the non-strange SU(3) partner of $\Lambda (1600)$ [@Sasaki:2001nf].
In the negative-parity channel (Fig.\[negmass\]), the ground- and the 1st-excited states always have close energies at all the $\beta$’s and $\kappa$’s. The eigen-energies have similar quark-mass dependences, and the mass splittings are almost quark mass independent. The chirally extrapolated values both lie around the mass of $\Lambda(1670)$ rather than $\Lambda(1405)$. Similarly to the previous studies, we do not reproduce the mass of $\Lambda(1405)$ in our calculation. While in quenched simulations in Refs.[@Melnitchouk:2002eg; @Nemoto:2003ft] such failure was regarded as an evidence of possible meson-baryon molecule components in $\Lambda(1405)$, our present [*unquenched*]{} simulation contains effects of dynamical quarks and thus should incorporate meson-baryon molecular states. We will come back to this point later in section V.
![\[posmass\] Masses of the lowest two positive-parity $\Lambda$ states as functions of the squared pion mass. The filled circles (open squares) denote the masses of lowest (1st-excited) state. Two solid curves represent quadratic functions used in the chiral extrapolation. Two dashed lines indicate the $\pi\Sigma$ and the $\bar K N$ thresholds in the presence of the relative momentum $p=\frac{2\pi}{L}$. Two solid lines on the vertical axes show the experimentally observed masses of $\Lambda(1115)$ and $\Lambda(1600)$. ](1224_pos.eps "fig:") ![\[posmass\] Masses of the lowest two positive-parity $\Lambda$ states as functions of the squared pion mass. The filled circles (open squares) denote the masses of lowest (1st-excited) state. Two solid curves represent quadratic functions used in the chiral extrapolation. Two dashed lines indicate the $\pi\Sigma$ and the $\bar K N$ thresholds in the presence of the relative momentum $p=\frac{2\pi}{L}$. Two solid lines on the vertical axes show the experimentally observed masses of $\Lambda(1115)$ and $\Lambda(1600)$. ](1632_pos.eps "fig:") ![\[posmass\] Masses of the lowest two positive-parity $\Lambda$ states as functions of the squared pion mass. The filled circles (open squares) denote the masses of lowest (1st-excited) state. Two solid curves represent quadratic functions used in the chiral extrapolation. Two dashed lines indicate the $\pi\Sigma$ and the $\bar K N$ thresholds in the presence of the relative momentum $p=\frac{2\pi}{L}$. Two solid lines on the vertical axes show the experimentally observed masses of $\Lambda(1115)$ and $\Lambda(1600)$. ](2448_pos.eps "fig:")
![\[negmass\] Masses of the lowest two negative-parity $\Lambda$ states as functions of the squared pion mass. The filled circles (open squares) denote the masses of the lowest (1st-excited) state. Two solid curves represent quadratic functions used in chiral extrapolation. Two dashed lines indicate the $\pi\Sigma$ and $\bar K N$ thresholds. Two solid lines on the vertical axes show the experimentally observed masses of $\Lambda(1405)$, $\Lambda(1670)$ and $\Lambda(1800)$. ](1224_neg.eps "fig:") ![\[negmass\] Masses of the lowest two negative-parity $\Lambda$ states as functions of the squared pion mass. The filled circles (open squares) denote the masses of the lowest (1st-excited) state. Two solid curves represent quadratic functions used in chiral extrapolation. Two dashed lines indicate the $\pi\Sigma$ and $\bar K N$ thresholds. Two solid lines on the vertical axes show the experimentally observed masses of $\Lambda(1405)$, $\Lambda(1670)$ and $\Lambda(1800)$. ](1632_neg.eps "fig:") ![\[negmass\] Masses of the lowest two negative-parity $\Lambda$ states as functions of the squared pion mass. The filled circles (open squares) denote the masses of the lowest (1st-excited) state. Two solid curves represent quadratic functions used in chiral extrapolation. Two dashed lines indicate the $\pi\Sigma$ and $\bar K N$ thresholds. Two solid lines on the vertical axes show the experimentally observed masses of $\Lambda(1405)$, $\Lambda(1670)$ and $\Lambda(1800)$. ](2448_neg.eps "fig:")
$\beta$ $\kappa$ $m_\pi$ $m_K$ $m_N$ $m_\Sigma$
--------- ---------- ---------- ----------- ----------- ------------ -- -- --
1.80 0.1409 1.156(1) 1.087(01) 2.260( 5) 2.201(05)
0.1430 0.983(1) 0.976(01) 2.024(07) 2.019(06)
0.1445 0.822(1) 0.872(01) 1.813(09) 1.844(07)
0.1464 0.531(2) 0.672(02) 1.405(11) 1.530(09)
[C.L.]{} — 0.291(30) 1.078(41) 1.269(17)
1.95 0.1375 0.894(1) 0.834(01) 1.707(06) 1.468(06)
0.1390 0.729(1) 0.725(01) 1.475(05) 1.464(06)
0.1400 0.596(1) 0.635(01) 1.289(05) 1.318(05)
0.1410 0.427(1) 0.533(01) 1.051(08) 1.134(07)
[C.L.]{} — 0.455(16) 0.777(27) 0.919(18)
2.10 0.1357 0.630(1) 0.565(1) 1.194(05) 1.131(06)
0.1367 0.517(1) 0.491(1) 1.016(05) 0.990(05)
0.1374 0.424(1) 0.435(1) 0.902(05) 0.907(05)
0.1382 0.295(1) 0.363(1) 0.734(06) 0.784(05)
[C.L.]{} — 0.312(9) 0.579(32) 0.674(26)
: \[hadronicmass\] Hadronic masses (in lattice units) obtained for four different $\kappa$’s at each $\beta$. The $m_\pi$, $m_K$, $m_N$, and $m_\Sigma$ denote the masses of pion, kaon, nucleon, and $\Sigma$ baryon, respectively. $E_i^\pm$ represents the eigen-energy of the $i$-th state in positive ($+$) or negative ($-$) parity channel with $I=0$ and $S=-1$. “C.L.” in each block shows the results extrapolated to the chiral limit with quadratic functions.
$\beta$ $\kappa$ $E_0^+$ $E_1^+$ $E_0^-$ $E_1^-$
--------- ---------- ----------- ----------- ----------- -----------
1.80 0.1409 2.201(03) 3.53(34) 2.794(27) 2.897(27)
0.1430 2.013(04) 3.53(40) 2.583(31) 2.662(26)
0.1445 1.836(04) 2.99(18) 2.407(20) 2.477(21)
0.1464 1.533(07) 2.22(19) 2.074(27) 2.131(26)
[C.L.]{} 1.284(14) 1.40(19) 1.801(31) 1.858(41)
1.95 0.1375 1.640(3) 2.465(28) 2.051(13) 2.088(12)
0.1390 1.457(3) 2.292(72) 1.847(13) 1.889(13)
0.1400 1.309(2) 2.094(18) 1.692(09) 1.737(09)
0.1410 1.134(3) 1.899(19) 1.483(12) 1.538(10)
[C.L.]{} 0.925(6) 1.656(22) 1.239(30) 1.303(23)
2.10 0.1357 1.133(09) 1.677(28) 1.384(13) 1.437(12)
0.1367 0.994(05) 1.526(15) 1.246(09) 1.293(07)
0.1374 0.919(05) 1.469(18) 1.158(10) 1.203(10)
0.1382 0.778(06) 1.325(15) 1.019(14) 1.078(18)
[C.L.]{} 0.653(52) 1.214(69) 0.896(38) 0.977(31)
: \[hadronicmass\] Hadronic masses (in lattice units) obtained for four different $\kappa$’s at each $\beta$. The $m_\pi$, $m_K$, $m_N$, and $m_\Sigma$ denote the masses of pion, kaon, nucleon, and $\Sigma$ baryon, respectively. $E_i^\pm$ represents the eigen-energy of the $i$-th state in positive ($+$) or negative ($-$) parity channel with $I=0$ and $S=-1$. “C.L.” in each block shows the results extrapolated to the chiral limit with quadratic functions.
$\beta$ $\kappa$ $g_0^+\times 10^2$ $g_1^+\times 10^2$ $g_0^-\times 10^2$ $g_1^-\times 10^2$
--------- ---------- -------------------- -------------------- -------------------- --------------------
1.80 0.1409 4.91(303) 0.30(01) $-$10.7(036) 6.46(291)
0.1430 0.12(013) 0.02(00) $-$0.64(035) 0.40(027)
0.1445 $-$3.82(195) $-$0.33(01) 15.2(023) $-$10.5(018)
0.1464 0.10(126) $-$1.39(06) 35.7(077) $-$21.9(063)
[C.L.]{} $-$8.36(800) $-$1.66(27) 65.3(116) $-$43.1(113)
1.95 0.1375 7.22(066) 0.46(01) $-$33.4(145) 22.2(109)
0.1390 0.58(029) 0.11(00) $-$2.99(141) 4.25(167)
0.1400 $-$2.46(022) $-$0.32(00) 11.4(013) $-$8.83(120)
0.1410 $-$7.02(081) $-$1.14(01) 26.3(031) $-$15.2(025)
[C.L.]{} $-$9.12(323) $-$1.81(24) 39.9(049) $-$19.8(072)
2.10 0.1357 10.07(128) 0.78(03) $-$22.3(038) 12.4(036)
0.1367 5.17(031) 0.37(01) $-$17.9(061) 8.21(367)
0.1374 $-$0.77(010) $-$0.09(00) 2.19(051) $-$1.26(044)
0.1382 $-$5.44(112) $-$1.11(04) 19.8(036) $-$15.2(040)
[C.L.]{} $-$14.4(052) $-$1.81(47) 42.5(103) $-$33.3(017)
: \[hadronicmass2\] The ratios $g_i^\pm$ (multiplied by $10^2$) defined in Eqs.(\[gminus0\])-(\[gplus1\]). “C.L.” shows the values extrapolated to the chiral limit using quadratic functions.
lattice-spacing dependences
---------------------------
![\[massvsa\] Masses of the ground and the 1st-excited states in the chiral limit for the positive (upper) and negative (lower) parity $\Lambda$ states plotted as functions of the lattice spacing $a$. The horizontal lines indicate the masses of the experimentally observed $\Lambda$ states. ](mass_vs_a_positive.eps "fig:") ![\[massvsa\] Masses of the ground and the 1st-excited states in the chiral limit for the positive (upper) and negative (lower) parity $\Lambda$ states plotted as functions of the lattice spacing $a$. The horizontal lines indicate the masses of the experimentally observed $\Lambda$ states. ](mass_vs_a_negative.eps "fig:")
Fig. \[massvsa\] shows the chirally extrapolated masses of the lowest two states in each parity v.s. the lattice spacing $a$. Except for the 1st-excited positive-parity state at the largest $a$, all the data show little lattice-cutoff dependence. The horizontal lines indicate the masses of the experimentally observed $\Lambda$ states. The positive-parity ground state agrees well with the mass of $\Lambda(1115)$, while the lowest two negative-parity states lie around $\Lambda(1670)$.
possible contaminations by scattering states
--------------------------------------------
Since our calculations contain dynamical u and d quarks, scattering states could come into spectra. $N\bar K$ and $\Sigma\pi$ thresholds are drawn in Fig. \[posmass\] and Fig.\[negmass\] as dashed lines. The energy of the 1st-excited state at the lightest u- and d-quark masses could be contaminated by such scattering states. Several strategies have been adopted to discriminate resonances from scattering states: study of volume dependence of the masses, and the spectral weights or boundary condition dependence of the eigen-energies. They were tested and applied to the pentaquark baryon [@Sasaki:2006jn; @Mathur:2003zf; @Csikor:2003ng; @Mathur:2004jr; @Ishii:2004qe; @Takahashi:2005uk].
There are two meson-baryon channels relevant in the present calculation, $\pi\Sigma$ and $\bar K N$. In terms of the valence quark contents, we have 5 different thresholds (2 for $\bar K N$, 3 for $\pi\Sigma$). In this paper, in order to distinguish resonance states from all these scattering states, we impose the following boundary condition on the quark fields, $$\psi(x+L)
=
e^
{\frac23\pi i}
\psi(x).$$ Under such boundary condition for quark fields, a hadronic state $\phi_{3k+n}(x)$ which contains $3k+n$ valence quarks obeys $$\phi_{3k+n}(x+L)
=
e^
{\frac23n\pi i}
\phi_{3k+n}(x),$$ and can have spatial momenta, $$\begin{aligned}
p_{\rm lat}
=
\frac{2\pi}{L}m+\frac{2n\pi}{3L}
\ \
(m \in {\rm Z})\end{aligned}$$
As a result, only states which consist of 3k valence quarks can be zero-momentum states. Since other quark combinations inevitably have non-vanishing spatial momenta, their energies are raised up. As long as we employ three-quark operators for baryon creation/annihilation, hadronic states appearing in scattering states should contain one or two valence quark(s), since sea-quark pairs themselves cannot carry flavors.
![\[compare\] Lowest two eigen-energies in the $(J^P,S) = (1/2^-,-1)$ channel at $\beta =2.10$ and $\kappa =0.1382$ under the normal (periodic) and the twisted boundary conditions. The circles (squares) are for the ground (1st-excited) states. The lower and upper solid lines respectively represent the threshold energies of $\Sigma\pi$ and $N\bar K$ states evaluated with normal/twisted boundary condition.](2448comp_negative.eps)
Now we turn back to our full-QCD calculations. We impose the following boundary condition to all the valence quark fields ($q=u,\ d,\ s$); $$q(x+L)
=
e^
{\frac23\pi i}
q(x).$$
We plot in Fig. \[compare\] the eigen-energies under the periodic and the twisted boundary conditions, which are obtained at $\beta =2.10$ and $\kappa =0.1382$. The open and filled circles (squares) are for the ground (1st-excited) states. The lower and upper solid lines respectively represent the threshold energies of $\Sigma\pi$ and $N\bar K$ states with normal/twisted boundary condition. The threshold energies are raised up in the case of twisted boundary condition, whereas the lowest two eigen-energies remain unchanged. Thus we conclude that the observed states are insensitive to boundary conditions, and contaminations from meson-baryon scattering states are small. Namely the results indicate that these states are regarded as compact resonance states.
Similar results are obtained also for the positive-parity states. Their energies are insensitive to the boundary conditions. (The energies of the positive-parity 1st-excited states at the smallest $\kappa$ are higher than p-wave thresholds (with relative momenta $p = \frac{2\pi}{L}$) of $\Sigma\pi$ or $N\bar K$ states.)
Flavor structures
=================
The chiral unitary approach [@Jido:2003cb] has suggested that $\Lambda (1405)$ is not a single pole but a superposition of two independent resonance poles. The structure of $\Lambda (1405)$ is now attracting much interest, and desired to be clarified in a model independent manner. We investigate the flavor structure of the ground and the 1st-excited states obtained from the cross correlators of two operators.
In order to clarify the flavor structure of the low-lying states, we evaluate the overlaps of the obtained states with the source and sink operators. We define the magnitudes of the singlet ($I={\bf 1}$) and the octet ($I={\bf 8}$) components in the $i$-th state ($i=0$ for the ground state and $i=1$ for the 1st excited state) by $$\langle i | \eta_I^\dagger | {\rm vac}\rangle = (C_{\rm})_{iI}.$$ Actually, we can only obtain the ratio between $$\langle i | \eta_{\bf 1}^\dagger | {\rm vac}\rangle = (C_{\rm})_{i{\bf 1}}$$ and $$\langle i | \eta_{\bf 8}^\dagger | {\rm vac}\rangle = (C_{\rm})_{i{\bf 8}},$$ since the overall factors in $C_{\rm src}$ cannot be determined in our setups. We evaluate $g_0$ and $g_1$ defined as $$\begin{aligned}
g_0^-
&\equiv&
C_{0{\bf 8}}/C_{0{\bf 1}}
\label{gminus0}
\\
g_1^-
&\equiv&
C_{1{\bf 1}}/C_{1{\bf 8}},
\label{gminus1}\end{aligned}$$ Both $g_0^-$ and $g_1^-$ vanish when the SU(3)$_F$ symmetry is exact, showing that the ground (1st excited) state is purely flavor singlet (octet) in the limit. For the positive-parity states, we similarly define $$\begin{aligned}
g_0^+
&\equiv&
C_{0{\bf 1}}/C_{0{\bf 8}}
\label{gplus0}
\\
g_1^+
&\equiv&
C_{1{\bf 8}}/C_{1{\bf 1}}.
\label{gplus1}\end{aligned}$$ In this case, we exchange the denominator and the numerator since the ground state is flavor octet and the 1st excited state is flavor singlet in the SU(3)$_F$ limit.
![\[gpos\] $g_i^+$ are plotted as a function of $m_\pi^2$. Solid lines denote quadratic functions used in chiral extrapolation. ](1224r2_pos.eps "fig:") ![\[gpos\] $g_i^+$ are plotted as a function of $m_\pi^2$. Solid lines denote quadratic functions used in chiral extrapolation. ](1632r2_pos.eps "fig:") ![\[gpos\] $g_i^+$ are plotted as a function of $m_\pi^2$. Solid lines denote quadratic functions used in chiral extrapolation. ](2448r2_pos.eps "fig:")
![\[gneg\] $g_i^-$ are plotted as a function of $m_\pi^2$. Solid lines denote quadratic functions used in chiral extrapolation. ](1224r2_neg.eps "fig:") ![\[gneg\] $g_i^-$ are plotted as a function of $m_\pi^2$. Solid lines denote quadratic functions used in chiral extrapolation. ](1632r2_neg.eps "fig:") ![\[gneg\] $g_i^-$ are plotted as a function of $m_\pi^2$. Solid lines denote quadratic functions used in chiral extrapolation. ](2448r2_neg.eps "fig:")
We show $g_0^\pm$ and $g_1^\pm$ as a function of pion-mass squared for each lattice cutoff in Fig. \[gpos\] and Fig. \[gneg\] . As is expected from the symmetry, such mixing coefficients cross zero at the flavor symmetric limit ($\kappa_u=\kappa_d=\kappa_s$). The data show smooth quark-mass dependence toward the chiral limit in all the lattice-cut-off cases and the dependences are almost lattice-cut-off independent. We find that the magnitude of operator mixing gets larger for larger flavor-symmetry breaking. It is interesting that the 1st-excited state in positive-parity channel is dominated by the flavor-singlet component showing almost no contaminations of octet components.
We perform quadratic fits of these mixing coefficients as $$\begin{aligned}
g_0(m_\pi^2)
=
a_{2,0} m_\pi^2
+
a_{1,0} m_\pi
+
a_{0,0} \\
g_1(m_\pi^2)
=
a_{2,1} m_\pi^2
+
a_{1,1} m_\pi
+
a_{0,1}\end{aligned}$$ and list the extrapolated values in Table \[hadronicmass2\]. Even in the chiral limit, $g_i\ (i=1,2)$ in negative-parity channel remain less than 0.6, which indicates that the ground (1st-excited) state is dominated by the singlet (octet) component in the present quark-mass range.
On the other hand, the mixings of the singlet and octet components are generally weaker in the positive parity $\Lambda$’s. It is natural because the splitting of the two states are large in the SU(3)$_F$. Nevertheless, it is interesting to see that the ground state, corresponding to $\Lambda(1115)$, contains significant (about 10%) flavor singlet component in the chiral limit, which was not expected from the simple quark model picture.
In practice, we need to interpret such overlap coefficients into physical amplitudes. In the Euclidean-time range we consider $$\begin{aligned}
\eta_{\bf 1} | {\rm vac} \rangle
=
c_0|0\rangle + g_1c_1|1\rangle
\\
\eta_{\bf 8} | {\rm vac} \rangle
=
g_0c_0|0\rangle + c_1|1\rangle\end{aligned}$$ are satisfied. We here assume that each state can be expanded in terms of the singlet and octet basis states as $$\begin{aligned}
&&|0\rangle
=
M_{01} |\Lambda_1\rangle
+
M_{08} |\Lambda_8\rangle \\
&&|1\rangle
=
M_{11} |\Lambda_1\rangle
+
M_{18} |\Lambda_8\rangle.\end{aligned}$$ We note that $|\Lambda\rangle$’s in this expression are no longer energy eigenstates in QCD, and they are restricted to elements that interpolating operators have overlaps with; $|\Lambda\rangle\propto\eta^\dagger|{\rm vac}\rangle$. (Actually we can introduce other singlet and octet bases, for example, whose spatial distributions are different. If we adopt more interpolating fields such as those with spatial derivatives or with exotic spinor structures, we can introduce much more basis states in this argument.)
We here extract $M_{ij}$’s, using the conditions, $\langle \Lambda_8 | \eta_{\bf 1} | {\rm vac} \rangle = 0$ and $\langle \Lambda_1 | \eta_{\bf 8} | {\rm vac} \rangle = 0$, giving $$\begin{aligned}
&&\frac{M_{08}}{M_{18}}
=
-g_1R,
\\
&&\frac{M_{11}}{M_{01}}
=
-g_0R^{-1}\end{aligned}$$ with $R = c_1/c_0$. $R^2$ can be estimated from the 2-point correlators, since optimized operators satisfy $$\begin{aligned}
&&(\eta_{\bf 1}-g_1\eta_{\bf 8})| {\rm vac}\rangle
=
c_0(1-g_0g_1)|0\rangle
\\
&&(\eta_{\bf 8}-g_0\eta_{\bf 1})| {\rm vac}\rangle
=
c_1(1-g_0g_1)|1\rangle.\end{aligned}$$ Using the normalization relations, $$\begin{aligned}
&&
M_{01}^2+M_{08}^2 = 1
\\
&&
M_{11}^2+M_{18}^2 = 1,\end{aligned}$$ we determine $M_{ij}^2$, the amplitude of $j$-plet component in $i$-th state as follows: $$\begin{aligned}
M_{01}^2
&=&
\frac{1-g_1^2R^2}{1-(g_0g_1)^2}
\\
M_{08}^2
&=&
\frac{g_1^2(R^2-g_0^2)}{1-(g_0g_1)^2}
\\
M_{11}^2
&=&
\frac{g_0^2(1-g_1^2R^2)}{R^2\{1-(g_0g_1)^2\}}
\\
M_{18}^2
&=&
\frac{R^2-g_0^2}{R^2\{1-(g_0g_1)^2\}}\end{aligned}$$ The amplitudes $M_{ji}^2$ obtained on the finest lattice are plotted in Fig. \[FcompoN\]. In the present quark-mass range, ground (1st-excited) state is dominated by flavor-singlet (octet) components.
![\[FcompoN\] $M_{ij}^2$ obtained at $\beta = 2.1$ are plotted as a function of squared pion mass $m_\pi^2$. $M_{ij}$ are normalized so that $M_{i1}^2+M_{i8}^2=1$. ](f0m-state.eps "fig:") ![\[FcompoN\] $M_{ij}^2$ obtained at $\beta = 2.1$ are plotted as a function of squared pion mass $m_\pi^2$. $M_{ij}$ are normalized so that $M_{i1}^2+M_{i8}^2=1$. ](f1m-state.eps "fig:")
So far, we have argued the internal structures from the view point of flavor contents. It would be interesting and useful to interpret these flavor-mixing amplitudes in terms of a hadronic basis, since hadronic-basis expansion of unstable particles could be directly compared with experiments and sometimes beneficial for model calculations. Each flavor element can be decomposed into hadronic bases as $$\begin{aligned}
|\Lambda_1\rangle
&=&
W_{N\bar K,1}|N\bar K\rangle
+
W_{\Sigma\pi,1}|\Sigma\pi\rangle
+ ....
\\
|\Lambda_8\rangle
&=&
W_{N\bar K,8}|N\bar K\rangle
+
W_{\Sigma\pi,8}|\Sigma\pi\rangle
+ ....\end{aligned}$$ In this paper we restrict ourselves only to the $N\bar K$ and $\Sigma\pi$ bases. The SU(3) Clebsh-Gordan coefficients tell us [@Amsler:2008zzb] $$\begin{aligned}
&&
| \Lambda_1 \rangle
\propto
\nonumber \\
&&
\frac{1}{\sqrt{8}}
\left(
\sqrt{2}|N\bar K\rangle
+
\sqrt{3}|\Sigma\pi\rangle
-
|\Lambda\eta\rangle
-
\sqrt{2}|\Xi K\rangle
\right) \\
&&
| \Lambda_8 \rangle
\propto
\nonumber \\
&&
\frac{g_D}{\sqrt{20}}
\left(
\sqrt{ 2}|N\bar K\rangle
-
\sqrt{12}|\Sigma\pi\rangle
-
\sqrt{ 4}|\Lambda\eta\rangle
-
\sqrt{ 2}|\Xi K\rangle
\right)
\nonumber \\
&+&
\frac{g_F}{\sqrt{12}}
\left(
\sqrt{6}|N\bar K\rangle
+
\sqrt{6}|\Xi K\rangle
\right),\end{aligned}$$ and the following expressions follows; $$\begin{aligned}
W_{N\bar K,1}
&=&
\sqrt{\frac14} \\
W_{\Sigma\pi,1}
&=&
\sqrt{\frac38} \\
W_{N\bar K,8}
&=&
\frac{1}{\sqrt{g_D^2+g_F^2}}
\left(
\sqrt{\frac{1}{10}}g_D+\sqrt{\frac12}g_F
\right) \\
W_{\Sigma\pi,8}
&=&
\frac{1}{\sqrt{g_D^2+g_F^2}}
\left(
-\sqrt{\frac{3}{5}}g_D+\sqrt{\frac12}g_F
\right).\end{aligned}$$ Here, $g_D$ and $g_F$ are unknown coupling constants corresponding to two possible SU(3) constructions of the meson-baryon octet states, ${\rm Tr}\left(\left\{\bar B, B \right\} M\right)$ and ${\rm Tr}\left(\left[\bar B, B \right] M\right)$, respectively. We finally reach the probabilities of $N\bar K$ and $\Sigma\pi$ states in $i$-th state, $P_{N\bar K}^i$ and $P_{\Sigma\pi}^i$ , $$\begin{aligned}
P_{N\bar K}^i &=& \left[\sum_{r=1,8}M_{ir}W_{N\bar K,r} \right]^2
\\
P_{\Sigma\pi}^i &=& \left[\sum_{r=1,8}M_{ir}W_{\Sigma\pi,r}\right]^2.\end{aligned}$$ Defining $\alpha$ by $$\left(
\cos\alpha,\sin\alpha
\right)
\equiv
\left(
\frac{g_D}{\sqrt{g_D^2+g_F^2}},\frac{g_F}{\sqrt{g_D^2+g_F^2}}
\right)$$ we obtain $P_{N\bar K}^i(\alpha)$ and $P_{\Sigma\pi}^i(\alpha)$ as functions of the angle $\alpha$, where we substitute $g_0$ and $g_1$ extrapolated to the chiral limit at $\beta =2.10$ (the finest lattice). We simply adopt $R^2=1$, since $R^2$ does not deviate from unity, and is almost quark-mass independent; $R^2=$1.002(29), 1.015(21), 1.030(29), and 1.024(29) at $\kappa_{\rm
u,d}=$0.1357, 0.1367, 0.1374, and 0.1382 at this $\beta$. The relative signs in $M_{ij}$ are determined so that $M_{i1}M_{i8}(C)_{i1}(C)_{i2}>0$.
We focus on the meson-baryon components of the low-lying states in negative-parity channel, since they are of interest here. We show in Fig. \[MBcompoN\], $P_{N\bar K}^i(\alpha)$ and $P_{\Sigma\pi}^i(\alpha)$ normalized as $P_{N\bar K}^i(\alpha) + P_{\Sigma\pi}^i(\alpha) = 1$ as a function of $\alpha$. Though $g_i$’s are determined from those obtained at $\beta=2.10$, the behaviors do not change significantly even if we adopt extrapolated values on coarser lattices. Now that the only unknown parameter is $\alpha$ representing $g_D$ and $g_F$ in the $J^P=1/2^-$ channel. Unfortunately, the couplings, $g_D$ and $g_F$, should be determined up to including their signs in order to estimate the meson-baryon components in each state in this strategy. They can in principle be determined from lattice QCD computations and their determination is left for forthcoming studies.
![\[MBcompoN\] $P_{N\bar K}^i(\alpha)$ and $P_{\Sigma\pi}^i(\alpha)$ of the ground state (upper panel) and of the 1st excited state (lower panel) as functions of $\alpha$. They are determined from the $g_i^-$’s obtained at $\beta=2.10$ (the finest lattice). ](0m-state.eps "fig:") ![\[MBcompoN\] $P_{N\bar K}^i(\alpha)$ and $P_{\Sigma\pi}^i(\alpha)$ of the ground state (upper panel) and of the 1st excited state (lower panel) as functions of $\alpha$. They are determined from the $g_i^-$’s obtained at $\beta=2.10$ (the finest lattice). ](1m-state.eps "fig:")
Conclusion and Discussions
==========================
An unquenched lattice QCD study for the low-lying $\Lambda$ baryon states in the $S=-1$, $I=0$ and $J=1/2$ channel has been performed, focusing especially on the masses and structures of the two lowest negative-parity $\Lambda$’s. We have constructed $2 \times 2$ cross correlators of the “octet” and “singlet” baryonic operators, and diagonalized them so that we can properly extract the information of the low-lying two states. In our measurements, we have adopted the 2-flavor gauge configurations generated by CP-PACS collaboration with the renormalization-group improved gauge action and the ${\mathcal O}(a)$-improved quark action. We have performed simulations at three different $\beta$’s, $\beta = 1.80$, 1.95 and 2.10, whose corresponding lattice spacings are $a = 0.2150$, 0.1555 and 0.1076 fm, employing four different hopping parameters ($\kappa_{\rm val}, \kappa_{\rm sea}$) for each cutoff, corresponding to the pion masses ranging from 500 MeV to 1.1 GeV.
It is shown that the mass of the ground-state $\Lambda(1115)$ is reproduced well at all three $\beta$’s, while the 1st-excited positive-parity $\Lambda$ state lies much higher than the experimentally observed $\Lambda(1600)$. The same tendency was reported in Ref. [@Burch:2006cc], and also in the case of the Roper resonance, $N^*(1440)$, which is the non-strange SU(3) partner of $\Lambda (1600)$ [@Sasaki:2001nf]. It should also be noted that no $\pi\Sigma$ or $\bar K N$ scattering states are seen in the calculation.
Our results for negative-parity $\Lambda$ states indicate that there are two $1/2^-$ states nearly degenerate at around $1.6 -1.7$GeV, while no state as low as $\Lambda (1405)$ is observed. We have revealed the flavor structures of these states from the lattice data for the first time. It is found that the lowest negative-parity $\Lambda$ state is dominated by flavor-singlet component. The second state, which is less than 100 MeV above the ground state, is predominantly flavor octet. Thus we find that the first two negative-parity $\Lambda$’s have different flavor structures.
They, however, come significantly heavier than the experimentally observed lowest-mass state, $\Lambda(1405)$. One naive possibility is that the obtained states do not correspond to the physical $\Lambda(1405)$, but describe excited $\Lambda$ states. In fact, the 2nd and 3rd negative-parity states lie at 1670 MeV and 1800 MeV, both of which are three-star states in the Particle Data Group classification. The present lattice data are consistent with these excited states. In the non-relativistic quark model approach, each of these states is classified as a flavor octet P-wave baryon. We, however, have shown that the lower state is dominated by a flavor-singlet component. So, the results here predict one flavor-singlet state and one flavor-octet state in the vicinity. The reason for missing $\Lambda(1405)$ state will be its poor overlap with three-quark operators. In fact, the inclusion of dynamical quarks does not strongly enhance the signals of possible meson-baryon scattering states. (See the later discussions.) If $\Lambda (1405)$ is predominantly a meson-baryon molecular state, such overlaps would be naturally small.
The other possibility is, of course, that the lowest and the 2nd-lowest states describe physical $\Lambda(1405)$ and $\Lambda(1670)$ but the masses have been overestimated. Considering that our simulation contains two-flavor dynamical quarks, the failure of obtaining a light $\Lambda$ state could be attributed either to (1) strange-quark quenching, (2) insufficient lattice volume or (3) lack of chiral symmetry. In fact, (1) seems to make the masses of octet baryons in positive parity channel slightly ($\alt 10\%$) overestimated in the present setups [@AliKhan:2001tx]. On the other hand, the deficiencies,(2) and (3), may cause the lowest state not properly reproduced, supposing that the main component of $\Lambda (1405)$ is a meson-baryon molecular state. In order to check whether this conjecture is correct, simulations with light dynamical quarks ($m_\pi \ll 500$ MeV) and larger volume ($L \gg 2.5$ fm) will be required.
Upon the above conjecture, we can further consider one interesting scenario that these states both correspond to the physical $\Lambda(1405)$. Then, the results may support its double pole structure proposed by the chiral unitary approach [@Jido:2003cb]. In our results, the lowest two states are almost degenerate at all the $\beta$’s (lattice spacing) and $\kappa$’s (quark masses). Namely, the obtained two states are the signature of the double-pole resonance, but the mass has not yet been reproduced because of the deficiency stated above.
It is also important to note the missing meson-baryon scattering states. Despite that dynamical up and down quarks are included and the meson-baryon thresholds appear around/below the obtained eigen-energies, we have found no clear signal of the meson-baryon scattering states. This fact is supported by the observation that the effective mass plots under the normal and twisted boundary conditions show no prominent differences. We show in Fig.\[bccompare2\] the effective mass plots of the ground state negative-parity $\Lambda$ under the normal and twisted b.c.’s. They are obtained on the $24^3\times 48$ lattice with the largest hopping parameter. Thus we conclude that no scattering states appear in the present spectrum. Also in Ref. [@Bulava:2009jb], in which excited-state nucleon spectrum was systematically and extensively investigated with two-flavors of dynamical quarks, no clear signal of scattering states was found and the importance of multi-quark operators was raised.
A similar situation can be found in the computation of the Wilson loops, whose expectation values give us the potential between a (heavy fundamental) quark and an antiquark. In the presence of dynamical quarks, such an interquark potential should saturate and flatten at some interquark distance, where the confining string is broken and a quark-antiquark pair is created. However, no one has ever observed successfully such a “string breaking” effect in the Wilson loop computations [@Heller:1994rz; @Bolder:2000un]. One possible reason for this phenomenon is that the Wilson loop itself has poor overlaps with such broken strings.
In the $\Lambda$ spectrum, we may similarly conjecture that the scattering states or broad resonances, which are sensitive to boundary conditions, have little overlap with the 3-quark interpolating field operators. More detailed investigation would be needed for clarification of scattering states purely induced by dynamical quarks.
![\[bccompare2\] Effective masses of the negative-parity $\Lambda$ ground state under the normal (periodic) and the twisted boundary conditions. ](bccompare2.eps)
The flavor structures of the $\Lambda$ states have been very well clarified using the variational method. The octet and the singlet components are mixed when the flavor-SU(3) symmetry is broken. Actually, the ground (1st-excited) state is dominated by singlet (octet) component, and the contamination by another representation is at most 20% (5% when squared) in our present analysis. From these findings, we expect that the flavor-SU(3) symmetry is not largely broken. A similar conclusion was also derived for the study of the meson-baryon coupling constants in lattice QCD [@Erkol:2008yj].
Because the SU(3) breaking effect seems small, the analyses without SU(3) mixings adopted so far [@Melnitchouk:2002eg; @Nemoto:2003ft; @Burch:2006cc; @Ishii:2007ym] make sense to some extent. The mixings, however, get larger towards the chiral limit, variational analyses could be essentially needed when we adopt much lighter quarks. We also find the meson-baryon contents in each state strongly depend on the meson-baryon couplings and their signs. Precise determination of the meson-baryon contents will require reliable determination of the couplings up to signs, for which further lattice QCD calculations may be helpful.
All the numerical calculations were performed on NEC SX-8R at CMC, Osaka university and BlueGene/L at KEK. The unquenched gauge configurations employed in our analysis were all generated by CP-PACS collaboration [@AliKhan:2001tx]. This work was supported in part by the Yukawa International Program for Quark-Hadron Sciences (YIPQS), by the Japanese Society for the Promotion of Science under contract number P-06327 and by KAKENHI (17070002, 19540275, 20028006 and 21740181).
[8]{}
J. J. Sakurai, Annals Phys. [**11**]{}, 1 (1960). R. H. Dalitz, T. C. Wong and G. Rajasekaran, Phys. Rev. [**153**]{}, 1617 (1967). Y. Akaishi and T. Yamazaki, Phys. Rev. C [**65**]{}, 044005 (2002). T. Yamazaki and Y. Akaishi, Phys. Lett. B [**535**]{} (2002) 70. Y. Akaishi, A. Dote and T. Yamazaki, Phys. Lett. B [**613**]{}, 140 (2005) \[arXiv:nucl-th/0501040\]. W. Melnitchouk [*et al.*]{}, Phys. Rev. D [**67**]{}, 114506 (2003) \[arXiv:hep-lat/0202022\]. Y. Nemoto, N. Nakajima, H. Matsufuru and H. Suganuma, Phys. Rev. D [**68**]{}, 094505 (2003) \[arXiv:hep-lat/0302013\]. T. Burch, C. Gattringer, L. Y. Glozman, C. Hagen, D. Hierl, C. B. Lang and A. Schafer, Phys. Rev. D [**74**]{}, 014504 (2006) \[arXiv:hep-lat/0604019\]. N. Ishii, T. Doi, M. Oka and H. Suganuma, Prog. Theor. Phys. Suppl. [**168**]{}, 598 (2007) \[arXiv:0707.0079 \[hep-lat\]\]. A. Ali Khan [*et al.*]{} \[CP-PACS Collaboration\], Phys. Rev. D [**65**]{}, 054505 (2002) \[Erratum-ibid. D [**67**]{}, 059901 (2003)\] M. Luscher and U. Wolff, Nucl. Phys. B [**339**]{}, 222 (1990). S. Perantonis and C. Michael, Nucl. Phys. B [**347**]{}, 854 (1990). B. Blossier, M. Della Morte, G. von Hippel, T. Mendes and R. Sommer, JHEP [**0904**]{}, 094 (2009) \[arXiv:0902.1265 \[hep-lat\]\]. S. Sasaki, T. Blum and S. Ohta, Phys. Rev. D [**65**]{}, 074503 (2002) \[arXiv:hep-lat/0102010\]. S. Sasaki and T. Yamazaki, Phys. Rev. D [**74**]{}, 114507 (2006) \[arXiv:hep-lat/0610081\]. N. Mathur [*et al.*]{}, Phys. Lett. B [**605**]{}, 137 (2005) \[arXiv:hep-ph/0306199\]. F. Csikor, Z. Fodor, S. D. Katz and T. G. Kovacs, JHEP [**0311**]{}, 070 (2003) \[arXiv:hep-lat/0309090\]. N. Mathur [*et al.*]{}, Phys. Rev. D [**70**]{}, 074508 (2004) \[arXiv:hep-ph/0406196\]. N. Ishii, T. Doi, H. Iida, M. Oka, F. Okiharu and H. Suganuma, Phys. Rev. D [**71**]{}, 034001 (2005) \[arXiv:hep-lat/0408030\]. T. T. Takahashi, T. Umeda, T. Onogi and T. Kunihiro, Phys. Rev. D [**71**]{}, 114509 (2005) \[arXiv:hep-lat/0503019\]. D. Jido, J. A. Oller, E. Oset, A. Ramos and U. G. Meissner, Nucl. Phys. A [**725**]{} (2003) 181 \[arXiv:nucl-th/0303062\].
C. Amsler [*et al.*]{} \[Particle Data Group\], Phys. Lett. B [**667**]{}, 1 (2008). J. M. Bulava [*et al.*]{}, Phys. Rev. D [**79**]{}, 034505 (2009) \[arXiv:0901.0027 \[hep-lat\]\]. U. M. Heller, K. M. Bitar, R. G. Edwards and A. D. Kennedy, Phys. Lett. B [**335**]{}, 71 (1994) \[arXiv:hep-lat/9401025\]. B. Bolder [*et al.*]{}, Phys. Rev. D [**63**]{} (2001) 074504 \[arXiv:hep-lat/0005018\]. G. Erkol, M. Oka and T. T. Takahashi, Phys. Rev. D [**79**]{}, 074509 (2009) \[arXiv:0805.3068 \[hep-lat\]\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we show that Dirichlet heat kernel estimates for a class of (not necessarily symmetric) Markov processes are stable under non-local Feynman-Kac perturbations. This class of processes includes, among others, (reflected) symmetric stable-like processes on closed $d$-sets in $\bR^d$, killed symmetric stable processes, censored stable processes in $C^{1, 1}$ open sets as well as stable processes with drifts in bounded $C^{1, 1}$ open sets.'
author:
- '[**Zhen-Qing Chen**]{}[^1], [^2] and [^3]'
date: '(December 11, 2011)'
title: '**Stability of Dirichlet heat kernel estimates for non-local operators under Feynman-Kac perturbation**'
---
[**AMS 2010 Mathematics Subject Classification**]{}: Primary 60J35, 47G20, 60J75; Secondary 47D07, 47D08
[**Keywords and phrases**]{}: fractional Laplacian, symmetric $\alpha$-stable process, symmetric stable-like process, censored stable process, relativistic symmetric stable process, heat kernel, transition density, Dirichlet heat kernel, Feynman-Kac perturbation, Feynman-Kac transform
Introduction {#s:1}
============
Recently, sharp two-sided Dirichlet heat kernel estimates have been obtained for several classes of discontinuous processes (or non-local operators), including symmetric stable processes [@CKS], censored stable processes [@CKS1], relativistic stable processes [@CKS2], and stable processes with drifts [@CKS4]. Although the proofs in these papers share a general road map, there are many distinct difficulties and the actual arguments are specific to the underlying processes. The main purpose of this paper is to establish a stability result for the sharp Dirichlet heat kernel estimates of a family of discontinuous processes under non-local Feynman-Kac perturbations. Here for a discontinuous Hunt process $X$, a non-local Feynman-Kac transform is given by $$T_t f (x) =\E_x \Big[
\exp \Big(A_t +\sum_{s\leq t} F(X_{s-}, X_s) \Big)f(X_t)\Big],$$ where $A$ is a continuous additive functional of $X$ having finite variations each compact time interval and $F(x, y)$ is a measurable function that vanishes along the diagonal. The approach of this paper is quite robust that it applies to a class of not necessarily symmetric Markov processes which includes all the four families of processes mentioned above in bounded $C^{1,1}$ open sets.
Transformation by multiplicative functionals is one of the most important transforms for Markov processes (see, for example, [@Chu; @sharpe]). Non-local Feynman-Kac transforms are particular cases. They play an important role in the probabilistic as well as analytic aspect of potential theory, and also in mathematical physics. For example, it is shown in [@CS03b] that relativistic stable processes can be obtained from the symmetric $\alpha$-stable processes through Feynman-Kac transformations. We refer the reader to [@ChZ; @Sim] for nice accounts on Feynman-Kac semigroups of Brownian motion. In particular, it is shown in [@BM; @Sim] that under a certain Kato class condition, the integral kernel (called the heat kernel) of the Feynman-Kac semigroup of Brownian motion admits two-sided Gaussian bound estimates. In [@R], sharp two-sided estimates on the densities of (local) Feynman-Kac semigroups of killed Brownain motions in $C^{1, 1}$ domains were established. Non-local Feynman-Kac semigroups for symmetric stable processes and their associated quadratic forms were studied in [@S1; @S2]. By combining some ideas from [@Zq2] with results from [@CK], it was proved in [@S] that, under a certain Kato class condition, the heat kernel of the non-local Feynman-Kac semigroup of a symmetric stable-like process $X$ on $\bR^d$ is comparable to that of $X$. The symmetry condition on $F(x, y)$ plays an essential role in the argument of [@S]. The nonsymmetric pure jump case for stable-like processes is dealt with in [@W]. For recent development in the study of non-local Feynman-Kac transforms for general symmetric Markov processes, we refer the reader to [@CFKZ1; @CFKZ2] and the references therein. We also mention that the stability of Martin boundary under non-local Feynman-Kac perturbation is addressed in [@CKi]. To the best of the authors knowledge, Dirichlet heat kernel estimates for (either local or non-local) Feynman-Kac semigroups of discontinuous processes is studied here for the first time. The main challenge in studying Dirichlet heat kernel estimates of Feynman-Kac semigroups is to get exact boundary decay behavior of the heat kernels. While our main interest is in the Dirichlet heat kernel estimates for Feynman-Kac semigroups, our theorem also covers the whole space case as well as “reflected" stable-like processes on subsets of $\bR^d$. In particular, our result recovers and extends the main results of [@S; @W] where $D=\bR^d$. Even in the whole space case, our approach is different from those in [@S; @W].
Setup and main result
---------------------
In this paper we always assume that $\alpha\in (0, 2)$, $d \ge 1$, $D$ is a Borel set in $\bR^d$. For any $x\in D$, $\delta_D(x)$ denotes the Euclidean distance between $x$ and $D^c$. We use “$:=$" to denote a definition, which is read as “is defined to be". For $a, b\in \bR$, $a\wedge b:=\min \{a, b\}$ and $a\vee b:=\max \{a, b\}$. The Euclidean distance between $x$ and $y$ is denoted as $|x-y|$.
For $\gamma \geq 0$, let $$\psi_\gamma(t, x, y):= \left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma \left(1\wedge
\frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma,
\qquad t>0, \, x, y \in D.$$ Throughout this paper, $X$ is a Hunt process on $D$ with transition semigroup $\{ P_t: t\ge 0\}$ that admits a jointly continuous transition density $p_D(t, x, y)$ with respect to the Lebesgue measure and that there exist $C_0>1$ and $\gamma\in [0, \alpha\wedge d)$ such that $$\label{e:hke}
C_0^{-1} \psi_\gamma(t, x, y)q(t,x,y) \le p_D(t, x, y) \le C_0
\psi_\gamma(t, x, y)q(t,x,y)$$ for all $(t, x, y)\in (0, 1]\times D\times D$, where $$\label{e:defq}
q(t,x, y)
:=t^{-d/\alpha}\wedge\frac{t}{|x-y|^{d+\alpha}}.$$ It is easy to see that under this assumption, $X$ is a Feller process satisfying the strong Feller property. It is easy to see that, by increasing the value of $C_0$ if necessary, $$\label{e:hkein1}
C_0^{-1} \le \int_{\bR^d} q(t,x,y)dy \le C_0
\quad \text{ for all }(t, x, y)\in (0, \infty)\times \bR^d.$$ Thus $$\label{e:hkein2}
\int_{D} p_D(t, x, y) dy \le C_0^2 \left(1\wedge
\frac{\delta_D
(x)}{t^{1/\alpha}}\right)^\gamma
\quad \text{ for all }(t, x)\in (0, 1]\times D.$$
Note that $X$ is not necessarily symmetric. We further assume that $X$ has a Lévy system $(N, t)$ where $N=N(x, dy)$ is a kernel given by $$N(x, dy)=\frac{c(x, y)}{|x-y|^{d+\alpha}}dy,$$ with $c(x, y)$ a measurable function that is bounded between two positive constants on $D\times D$. That is, for any $x\in D$, any stopping time $T$ (with respect to the filtration of $X$) and any non-negative measurable function $f$ on $D\times D$ with $f(y, y)=0$ for all $y\in D$ that is extended to be zero off $D\times D$, $$\label{e:levy}
\E_x
\left[\sum_{s\le T} f(X_{s-}, X_s) \right]= \E_x \left[
\int_0^T \left( \int_{D} f(X_s, y) \frac{c(X_s, y)}{|X_s-y|^{d+\alpha}} dy \right) ds \right].$$ By increasing the value of $C_0$ if necessary, we may and do assume that $$\label{e:1.5a}
1/C_0 \leq c(x, y) \leq C_0 \qquad \hbox{for } x, y\in D.$$
Recall that an open set $D$ in $\bR^d$ (when $d\ge 2$) is said to be a $C^{1,1}$ open set if there exist a localization radius $ r_0>0 $ and a constant $\Lambda_0>0$ such that for every $z\in\partial D$, there exist a $C^{1,1}$-function $\phi=\phi_z: \bR^{d-1}\to \bR$ satisfying $\phi (0)=0$, $\nabla\phi (0)=(0, \dots, 0)$, $\| \nabla
\phi \|_\infty \leq \Lambda_0$, $| \nabla \phi (x)-\nabla \phi (w)|
\leq \Lambda_0 |x- w|$, and an orthonormal coordinate system $y=(y_1,
\cdots, y_{d-1}, y_d):=(\wt y, \, y_d)$ such that $ B(z, r_0 )\cap
D=B(z, r_0 )\cap \{ y: y_d > \phi (\wt y) \}$. We call the pair $(r_0, \Lambda_0)$ the characteristics of the $C^{1,1}$ open set $D$. By a $C^{1,1}$ open set in $\bR$ we mean an open set which can be expressed as the union of disjoint intervals so that the minimum of the lengths of all these intervals is positive and the minimum of the distances between these intervals is positive.
It follows from [@CKS; @CKS1; @CKS4; @CK] that the following are true:
[(i)]{} the (reflected) symmetric stable-like process on any closed $d$-subset $D$ in $\bR^d$ (see Subsection \[S:4.1\] for the definition of $d$-set) satisfies the conditions and with $\gamma=0$ and $c(x, y)$ a symmetric measurable function that is bounded between two positive constants;
[(ii)]{} the killed symmetric $\alpha$-stable process on a $C^{1, 1}$ open set $D$ satisfies the conditions and with $\gamma=\alpha/2$ and $c(x, y)=c$;
[ (iii)]{} when $d \ge 2$ and $\alpha \in (1,2)$, the killed symmetric $\alpha$-stable process with drift in a bounded $C^{1, 1}$ open set $D$ satisfies the conditions and with $\gamma=\alpha/2$ and $c(x, y)=c$; and
[(iv)]{} when $\alpha \in (1,2)$, the censored $\alpha$-stable process in a $C^{1, 1}$ open set $D$ satisfies the conditions and with $\gamma=\alpha-1$ and $c(x, y)=c$.
By a signed measure $\mu$ we mean in this paper the difference of two nonnegative $\sigma$-finite measures $\mu_1$ and $\mu_2$ in $D$. We point out that $\mu=\mu_1-\mu_2$ may not be a signed measure in $D$ in the usual sense as both $\mu_1 (D)$ and $\mu_2(D)$ may be infinite. However, there is an increasing sequence of subsets $\{F_k, k\geq 1\}$ whose union is $D$ so that $\mu_1(F_k)+\mu_2(F_k)<\infty$ for every $k\geq 1$. So when restricted to each $F_k$, $\mu$ is a finite signed measure. Consequently, the positive and negative parts of $\mu$ are well defined on each $F_k$ and hence on $D$, which will be denoted as $\mu^+$ and $\mu^-$, respectively. We use $|\mu|=\mu^++\mu^-$ to denote the total variation measure of $\mu$. Taking such an extended view of signed measures is desirable when one studies the correspondence between signed measures and continuous functions of finite variations or the correspondence between signed smooth measures and continuous additive functionals of finite variations for a Hunt process. For a signed measure $\mu$ on $D$ and $t>0$, we define $$N^{\alpha, \gamma}_{\mu}(t)=\sup_{x\in
D}\int^t_0\int_{D}\left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma q(s, x,
y)|\mu|(dy)ds.$$
\[d:kc\] A signed measure $\mu$ on $D$ is said to be in the Kato class ${\bf K}_{\alpha, \gamma}$ if $\lim_{t\downarrow 0}N^{\alpha,
\gamma}_{\mu}(t)=0$.
Note that if $N^{\alpha, \gamma}_\mu(t)<\infty$ for some $t>0$, then $|\mu|$ is a Radon measure on $D$. We say that a measurable function $
g$ belongs to the Kato class ${\bf K}_{\alpha, \gamma}$ if $
g(x)dx
\in {\bf K}_{\alpha, \gamma}$ and we denote $N^{\alpha, \gamma}_{
g (x) dx}$ by $N^{\alpha, \gamma}_
g$. It is well known that any $\mu\in {\bf K}_{\alpha, \gamma}$ is a smooth measure in the sense of [@FG]. Moreover, using the fact that $X$ has a transition density function under each $\P_x$, one can show that the continuous additive functional $A^{\mu}_t$ of $X$ with Revuz measure $\mu \in {\bf K}_{\alpha, \gamma}$ can be defined without exceptional set, see [@FOT pp. 236–237] for details. Concrete conditions for $\mu\in {\bf K}_{\alpha, \gamma}$ are given in Proposition \[P:4.1\].
For any measurable function $F$ on $D\times D$ vanishing on the diagonal, we define $$N^{\alpha, \gamma}_F(t) := \sup_{y\in D}
\int_0^t \int_{
D\times D}
\Big(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\Big)^\gamma \,
q(s, y, z ) \,
\Big(1 + \frac{|z-w| \wedge t^{1/\alpha}}{|y-z|} \Big)^\gamma
\, \frac{|F|(z,w) + |F|(w,z)}{|z-w|^{d+\alpha}} dw dz ds.$$
\[d:nlkc\] Suppose that $F$ is a measurable function on $D\times D$ vanishing on the diagonal. We say that $F$ belongs to the Kato class ${\bf
J}_{\alpha, \gamma}$ if $F$ is bounded and $\lim_{t\downarrow 0}N^{\alpha, \gamma}_{F}(t)=0$.
It follows immediately from the two definitions above that if $F\in
{\bf J}_{\alpha, \gamma}$, then the function $$z\mapsto \int_D\frac{|F|(z,w) + |F|(w,z)}{|z-w|^{d+\alpha}} dw$$ belongs to ${\bf K}_{\alpha, \gamma}$. See Proposition \[p:newqw1\] for a sufficient condition for $F\in
{\bf J}_{\alpha, \gamma}$.
It is easy to check that if $F$ and $G$ belong to ${\bf J}_{\alpha,
\gamma}$ and $c$ is a constant, then the functions $cF, e^{F}-1,
F+G$ and $FG$ all belong to ${\bf J}_{\alpha, \gamma}$. Throughout this paper, we will use the following notation: For any given measurable function $F$ on $D\times D$, $F_1(x, y)$ denotes the function $e^{F(x, y)}-1$.
For any signed measure $\mu$ on $D$ and any measurable function $F$ on $D\times D$ vanishing on the diagonal, we define $$N^{\alpha, \gamma}_{\mu, F}(t) :=N^{\alpha, \gamma}_{\mu}(t)+
N^{\alpha, \gamma}_F(t).$$ When $\mu\in{\bf K}_{\alpha, \gamma}$ and $F$ is a measurable function with $F_1\in {\bf J}_{\alpha, \gamma}$, we put $$A^{\mu, F}_t
=A^\mu_t+\sum_{0<s\le t}F(X_{s-}, X_s).$$ For any nonnegative Borel function $f$ on $D$, we define $$T^{\mu, F}_t f(x)
=\E_x\left[\exp(A^{\mu, F}_t) f(X_t)\right], \quad t\ge 0,
x\in D.$$ Then $(T^{\mu, F}_t: t\ge 0)$ is called the Feynman-Kac semigroup of $X$ corresponding to $\mu$ and $F$. The main purpose of this paper is to establish the following result. Recall that $\gamma\ge 0$ and $C_0\geq 1$ are the constants in and . For any bounded function $F$ on $D\times D$, we use $\|F\|_\infty$ to denote $\|F\|_{L^\infty (D \times D)}$.
\[t:main\] Let $d\geq 1$, $\alpha \in (0, 2)$ and $\gamma \in [0, \alpha \wedge d)$. Suppose $X$ is a Hunt process in a Borel set $D\subset \bR^d$ with a jointly continuous transition density $p_D(t, x, y)$ satisfying , and . If $\mu$ is a signed measure in ${\bf K}_{\alpha, \gamma}$ and $F$ is a measurable function so that $F_1:=e^F-1\in {\bf J}_{\alpha, \gamma}$, then the non-local Feynman-Kac semigroup $(T^{\mu, F}_t: t\ge 0)$ has a continuous density $q_D(t, x, y)$, and for any $T>0$, there exists a constant $C=C(d, \alpha, \gamma, C_0,
N^{\alpha, \gamma}_{\mu, F_1}, \|F_1\|_\infty,
T) >0$ such that for all $(t, x, y)\in (0, T]\times D\times D$, $$q_D(t, x, y) \le C \psi_\gamma(t, x, y)q(t,x,y).$$ If $\mu \in {\bf K}_{\alpha, \gamma}$ and $F\in {\bf J}_{\alpha, \gamma}$, then there exists a constant $\wt C= \wt C(d, \alpha, \gamma, C_0,
N^{\alpha, \gamma}_{\mu, F},
\|F\|_\infty, T) >1$ such that for all $(t, x, y)\in (0, T]\times D\times D$, $$\wt C^{-1} \psi_\gamma(t, x, y)q(t,x,y) \leq
q_D(t, x, y) \le \wt C \psi_\gamma(t, x, y)q(t,x,y).$$
Here and in the sequel, the dependence of the constant $C$ on $N^{\alpha, \gamma}_{\mu, F_1}$ and $\|F_1\|_\infty$ means that the value of the constant $C$ depends only on a specific upper bound for the rate of the function $N^{\alpha, \gamma}_{\mu, F_1}(t)$ going to zero as $t\to 0$ and on a specific upper bound for $\|F_1\|_\infty$. When $D=\bR^d$ and $\gamma=0$, Theorem \[t:main\] in particular recovers and extends the main results of [@S; @W].
Approach
--------
To explain our approach, we first recall the definition of the Stieltjes exponential. If $K_t$ is a right continuous function with left limits on $\bR_+$ with $K_0=1$ and $\Delta K_t:=K_t-K_{t-}>-1$ for every $t>0$, and if $K_t$ is of finite variation on each compact time interval, then the Stieltjes exponential $\Exp (K)_t$ of $K_t$ is the unique solution $Z_t$ of $$Z_t=1+ \int_{(0, t]} Z_{s-} dK_s, \quad t> 0.$$ By [@sharpe (A4.17)], $$\label{e:stie0}
\Exp (K)_t = e^{K^c_t} \prod_{0<s\leq t} (1+\Delta K_s),$$ where $K_t^c$ denotes the continuous part of $K_t$. Clearly $\exp (K_t)\geq \Exp (K)_t$ with the equality holds if and only if $K_t$ is continuous. The reason of $\Exp (K)_t$ being called the [*Stieltjes*]{} exponential of $K_t$ is that by [@DD] we have $$\label{e:stie1}
\Exp (K)_t =1+ \sum_{n=1}^\infty
\int_{(0, t]} dK_{t_n} \int_{(0, t_n]} dK_{t_{n-1}}
\cdots \int_{(0, t_2]} dK_{t_1}.$$ The advantage of using the Stieltjes exponential $\Exp (K)_t$ over the usual exponential $\exp (K_t)$ is the identity , which allows one to apply the Markov property of $X$.
Recall that $F_1(x, y)=e^{F(x, y)}-1$. In view of , we can express $\exp (A^{\mu, F}_t)$ in terms of the Stieltjes exponential: $$\exp (A^{\mu, F}_t ) = \Exp \Big( A^\mu+ \sum_{s\leq \cdot }
F_1 (X_{s-}, X_s) \Big)_t \qquad \hbox{for } t\geq 0.$$ Applying with $K_t:=A^\mu_t+\sum_{s\leq t}
F_1 (X_{s-}, X_s)$ and using the Markov property of $X$, we have for any bounded $f\geq 0$ on $D$, $$\begin{aligned}
T^{\mu, F}_t f (x)
&=& \E_x\left[\exp (A^{\mu, F}_t )f(X_t)\right]
=\E_x \left[f(X_t) \,\Exp \Big( A^\mu + \sum_{s\leq \cdot }
F_1 (X_{s-}, X_s) \Big)_t \right] \nonumber \\
&=& P_t f(x) + \E_x \left[ f(X_t)\sum_{n=1}^\infty
\int_{(0, t]} dK_{t_n} \int_{(0, t_n]} dK_{t_{n-1}}
\cdots \int_{(0, t_2]} dK_{t_1}\right] . \label{e:newq3}\end{aligned}$$
It can be shown that, for $\mu\in{\bf
K}_{\alpha, \gamma}$ and measurable function $F$ with $F_1
\in {\bf J}_{\alpha, \gamma}$, there is some constant $T_0>0$ so that we can change the order of the expectation and the infinite sum when $t\leq T_0$. Hence we have for every $t\leq T_0$, $$\begin{aligned}
\label{e:1.11}
T^{\mu, F}_t f(x) &=& P_t f(x) + \sum_{n=1}^\infty \E_x \left[ f(X_t)
\int_{(0, t]} dK_{t_n} \int_{(0, t_n]} dK_{t_{n-1}}
\cdots \int_{(0, t_2]} dK_{t_1}\right] \nonumber \\
&=&P_t f(x) + \sum_{n=1}^\infty \E_x \left[
\int_{(0, t]} P_{t-t_n} f(X_{t_n})dK_{t_n} \int_{(0, t_n]} dK_{t_{n-1}}
\cdots \int_{(0, t_2]} dK_{t_1} \right].\end{aligned}$$ Note that by , for any bounded function $g$, $$\begin{aligned}
\label{e:1.12}
& \E_x \left[ \int_{(0, s]} g(X_{r}) dK_r \right]
=\E_x \left[ \int_{(0, s]} g(X_{r}) dA^\mu_r
+ \sum_{r\leq s}
g(X_{r}) F_1(X_{r-}, X_r)\right] \nonumber \\
=& \int_0^s \int_D p_D(r, x, y) g(y) \mu (dy) dr
+ \E_x \left[
\int_0^s \left( \int_{D} F_1(X_r, y)g(y) \frac{c(X_r, y)}{|X_r-y|^{d+\alpha}} dy \right) dr \right] \nonumber \\
=&\int_0^s \int_D p_D(r, x, y) g(y) \mu (dy) dr
+
\int_0^s \int_D p_D(r, x, z) \left( \int_{D} F_1(z, y) g(y)
\frac{c(z, y)}{|y-z|^{d+\alpha}} dy \right) dz dr .\end{aligned}$$ This together with motives us to define $p^0 (t, x, y):=p_D(t, x, y)$ and, for $k \ge 1$ $$\begin{aligned}
p^k (t, x, y)&=&\int^t_0 \left(\int_{D}p_D(s, x,
z)p^{k-1}(t-s, z, y)\mu(dz) \right)ds\nonumber \\
&&+ \int^t_0 \left(\int_{D\times D} p_D( s, x,
z)\frac{c(z, w) F_1(z, w)}{|z-w|^{d+\alpha}}p^{k-1}(t-s, w, y) dzdw
\right)ds.\label{e:i0}\end{aligned}$$ One then concludes from that $$T^{\mu, F}_t f(x) = \int_D q_D(t, x, y) f(y) dy,$$ where $$\label{e:1.14}
q_D(t, x, y) := \sum_{k=0}^\infty p^k(t, x, y).$$ We then proceed to establish the following key estimates: there exist constants $T_1\in (0, T_0]$, $c>0$ and $0<\lambda <1$ such that $$\label{e:1.15}
|p^k(t, x, y)| \leq (\lambda^k+ck \lambda^{k-1} )p_D(t, x, y)
\qquad \hbox{on } (0, T_1] \times D \times D
\hbox{ for every } k\geq 1.$$ From this we can deduce that for every $t\in (0, T_1]$, $$\label{e:1.15a}
q_D(t, x, y)=
\sum_{k=0}^\infty p^k(t, x, y)
\leq \left(\frac1{1-\lambda}+ \frac{c}{(1-\lambda)^2}\right)\, p_D(t, x, y),$$ and, under the assumption $F\in {\bf J}_{\alpha, \gamma}$, that $$q_D(t, x, y) \ge 2^{-2(\lambda+c)}p_D(t, x, y),$$ which establish Theorem \[t:main\] for $t\leq T_1$. The general case of $t\leq T$ follows from an application of the Chapman-Kolmogorov equation.
The key to establish the estimate are two integral forms of the 3P inequality given in Lemma \[l:3p\] and Theorem \[t:G3p\] below. For a killed Brownian motion in a smooth domain, the following form of 3P inequality is known (see [@KS1; @R1]): for any $0< c < a \wedge (b-a)$, there exists $M=M(a,b,c)>0$ such that for every $0<s<t$ and $x,y,z \in D$ $$\label{e:ndf1}
\frac{p^W_a(t-s, x,
z) p^W_b(s, z,y)}{p^W_a(t, x, y)} \le M \frac{\delta_D(z)}{\delta_D(x)}
p^W_c(t-s, x, z)+ M\frac{\delta_D(z)}{\delta_D
(y)} p^W_c(s, y, z)$$ where $ p^W_c(t,x,y) := \psi_1(t, x, y) t^{-d/2} e^{-c|x-y|^2/t}. $ For symmetric $\alpha$-stable processes in $\bR^d$, one has the following form of 3P inequality (see [@BJ] and below): $$\label{e:ndf2}
\frac{ q(s,x,z) q(t-s, z,y)}{ q(t,x,y)}\le
c\left(q(s,x,z)+ q(t-s, z,y) \right) \quad \text{for every }0<s<t \text{ and } x,y,z \in \bR^d .$$ The above 3P type inequalities and played essential roles in establishing the heat kernel estimates in [@BJ; @KS1; @R1]. It seems that, for the processes we are dealing with in this paper, the above two types of 3P inequalities are not true in general. Moreover, we need a 3P type estimate on $p_D(t-s, x,
z) p_D(s, w,y)/p_D(t, x, y)$, where $z \not= w$.
The rest of the paper is organized as follows. In Section \[s:2\], we prove some key inequalities, including two forms of the 3P inequality. The main estimates and Theorem \[t:main\] will be established in Section \[s:3\]. In the last section, we give some applications of our main results.
In this paper, we will use capital letters $
\wt C, C, C_0,
C_1, C_2, \dots$ to denote constants in the statements of results, and their values will be fixed. The lower case letters $c_1, c_2,
\dots$ will denote generic constants used in proofs, whose exact values are not important and can change from one appearance to another. The labeling of the lower case constants starts anew in each proof. For two positive functions $f$ and $g$, we use the notation $f\asymp g$, which means that there are two positive constants $c_1$ and $c_2$ whose values depend only on $d, \alpha$ and $\gamma$ so that $c_1 g\leq f \leq c_2 g$.
3P inequalities {#s:2}
===============
In this section we will establish some key inequalities which will be essential in proving Theorem \[t:main\]. The main results of this section are Lemma \[l:4-1\], Theorem \[t:key\], Lemma \[l:4-2\] and Theorem \[t:G3p\]. Throughout this section, $D$ is a Borel set in $\bR^d$.
The following elementary facts will be used several times in this section.
\[l:ineq2-1\] For any $s, t>0$ and $(y,z) \in D \times D$, we have $$\label{e:elineq1}
1\wedge \frac{\delta_D(z)}{t^{1/\alpha}} =
\frac{\delta_D(y)}{t^{1/\alpha}}\,
\left(\frac{\delta_D(z)\wedge
t^{1/\alpha}}{\delta_D(y)}\right)$$ and $$\label{e:elineq2}
\left(1\wedge \frac{\delta_D(y)}{s^{1/\alpha}}\right) \left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right) \le
2 \left(1+\frac{|y-z|}{s^{1/\alpha} +\delta_D (y)}\right)\left(1 \wedge
\frac{\delta_D(y)}{t^{1/\alpha}}\right) .$$
The identity is clear, so we only need to prove . Since $\delta_D(z) \le |y-z|+\delta_D(y)$, we see that $$\begin{aligned}
1\wedge \frac{\delta_D(z)}{t^{1/\alpha}}
\le 1\wedge
\left(\left(\frac{|y-z|+\delta_D(y)}{\delta_D(y)}\right)\frac{\delta_D(y)}{t^{1/\alpha}}\right)\le
\left(1+\frac{|y-z|}{\delta_D(y)}\right)
\left(1\wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right).\end{aligned}$$ Thus, applying the elementary inequality $$\label{e:ineq*}
\frac{a}{a +b} \le 1\wedge\frac{a}{b}
\le \frac{2a }{a+ b}, \quad a, b>0$$ we get $$\begin{aligned}
\left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right) \left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right) &\le& \left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right)\left(1+
\frac{|y-z|}{\delta_D(y)}\right)
\left(1\wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right)\\
&\le&2\left(1+\frac{|y-z|}{s^{1/\alpha}+\delta_D(y)}\right)
\left(1 \wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right).\end{aligned}$$
Using and , we get that $$\label{e:ineq1}
\frac{t}{(t^{1/\alpha}+|x-y|)^{d+\alpha}}
\le q(t,x,y) \le 2^{d+\alpha} \frac{t}{(t^{1/\alpha}+|x-y|)^{d+\alpha}}.$$
\[l:4-1\] For any $\gamma\in [0, 2\alpha)$, there exists a constant $C_1:=C_1(d, \alpha, \gamma)>1$ such that for all $(t, y,z )\in (0,
\infty)\times D \times D$, $$\label{e:ineqmain1}
\left(1\wedge \frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma
\int^{t/2}_0\psi_\gamma(s, z, y) q(s, z, y) ds
\le C_1\left(1\wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma
\int^{t/2}_0 \left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma
q(s, z, y) ds.$$
The inequality holds trivially when $\gamma =0$ with $C_1=1$ so for the rest of the proof, we assume $\gamma \in (0, 2\alpha)$. The inequality is obvious if $\delta_D(y) \ge t^{1/\alpha} $ or $\delta_D(z) \le 2\delta_D(y)$. So we will assume $\delta_D(y) < t^{1/\alpha} \wedge (\delta_D(z)/2)$ throughout this proof. Note that in this case, $$\label{e:tyu1}
|z-y|\geq \delta_D(z)-\delta_D (y) \geq \frac{\delta_D(z)}2\geq
\delta_D(y).$$ By , we have $$\begin{aligned}
&&\int_{(t/2)\wedge |z-y|^\alpha}^{t/2}\left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma\psi_\gamma(s, z, y)
q(s, z, y) ds\nonumber\\ &\le& 2^{2\gamma}\left(1\wedge
\frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma\int_{(t/2)\wedge
|z-y|^\alpha}^{t/2}\left(1\wedge
\frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma q(s, z, y)
ds, \label{e:e2}\end{aligned}$$ while by $$\begin{aligned}
& \left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma
\int^{(t/2)\wedge |z-y|^\alpha}_0\psi_\gamma(s, z, y)
q(s, z, y)ds \nonumber \\
\le&\left( \frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma
\left(\frac{\delta_D(z)\wedge
t^{1/\alpha}}{\delta_D(y)}\right)^\gamma \int^{(t/2)\wedge
|z-y|^\alpha}_0 \left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma q(s, z, y)ds .
\label{e:new}\end{aligned}$$ In view of , and , $$\begin{aligned}
&\int^{(t/2)\wedge |z-y|^\alpha}_0 \left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma
q(s, z, y)ds \nonumber \\
\asymp & \int_0^{(t/2)\wedge \delta_D(y)^\alpha}
\frac{s}{|z-y|^{d+\alpha}} ds
+ \int_{(t/2)\wedge \delta_D(y)^\alpha}^{(t/2)\wedge |z-y|^\alpha}
\left(\frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma \frac{s}{|z-y|^{d+\alpha}} ds \nonumber \\
\asymp& \frac1{|z-y|^{d+\alpha}}
\left( \left((t/2)\wedge \delta_D(y)^\alpha \right)^2 +
\delta_D(y)^\gamma \left( \left((t/2)\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha} -
\left((t/2)\wedge \delta_D(y)^\alpha \right)^{2-\gamma/\alpha} \right) \right) \nonumber \\
\asymp& \frac1{|z-y|^{d+\alpha}}
\Big( \left((t/2)\wedge \delta_D(y)^\alpha \right)^2 \nonumber \\
\quad&+
(\delta_D (y)\wedge (t/2)^{1/\alpha})^\gamma
\left( \left((t/2)\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha} -
\left((t/2)\wedge \delta_D(y)^\alpha \right)^{2-\gamma/\alpha} \right) \Big) \nonumber \\
\asymp& \frac{ (\delta_D (y)\wedge (t/2)^{1/\alpha})^\gamma \,
\left((t/2)\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha}}
{|z-y|^{d+\alpha}}
\,
\asymp\, \frac{ \delta_D (y)^\gamma \,
\left((t/2)\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha}}
{|z-y|^{d+\alpha}} . \label{e:2.8}\end{aligned}$$ On the other hand, using we have $$\begin{aligned}
&\int^{(t/2)\wedge |z-y|^\alpha}_0 \left(1\wedge
\frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma
q(s, z, y)ds \nonumber \\
\asymp & \int_0^{(t/2)\wedge (\delta_D(z)/2)^\alpha}
\frac{s}{|z-y|^{d+\alpha}} ds
+ \int_{(t/2)\wedge (\delta_D(z)/2)^\alpha}^{(t/2)\wedge |z-y|^\alpha} \left(\frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma \frac{s}{|z-y|^{d+\alpha}} ds \nonumber \\
\asymp& \frac1{|z-y|^{d+\alpha}}
\left( \left(\frac{t}2\wedge \left(\frac{\delta_D(z)}2\right)^\alpha \right)^2 +
\delta_D(z)^\gamma \left( \left(\frac{t}2\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha} -
\left(\frac{t}2\wedge \left(\frac{\delta_D(z)}2\right)^\alpha \right)^{2-\gamma/\alpha} \right) \right) \nonumber \\
\ge& \frac1{|z-y|^{d+\alpha}}
\left(\left(\frac{t}2\wedge \left(\frac{\delta_D(z)}2\right)^\alpha \right)^2 \right.\nonumber \\ \quad&+
\left.\left( \frac{\delta_D(z)}2 \wedge \left(\frac{t}2\right)^{1/\alpha}\right) ^\gamma
\left( \left(\frac{t}2\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha} -
\left( \frac{t}2 \wedge \left(\frac{\delta_D(z)}2\right)^\alpha \right)^{2-\gamma/\alpha} \right)\right) \nonumber \\
\asymp& \frac{ (\delta_D(z) \wedge t^{1/\alpha})^\gamma \,
\left((t/2)\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha}}
{|z-y|^{d+\alpha}} . \label{e:2.9}\end{aligned}$$ One then deduces from – and the assumption $\delta_D(y) \leq t^{1/\alpha}$ that $$\begin{aligned}
&& \left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma
\int^{(t/2)\wedge |z-y|^\alpha}_0\psi_\gamma(s, z, y)
q(s, z, y)ds \nonumber \\
&\leq & c_1 \left(1\wedge
\frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma
(\delta_D(z) \wedge t^{1/\alpha})^\gamma
\frac{\left((t/2)\wedge |z-y|^\alpha\right)^{2-\gamma/\alpha}}
{|z-y|^{d+\alpha}}\\
&\leq & c_2 \int^{(t/2)\wedge |z-y|^\alpha}_0 \left(1\wedge
\frac{\delta_D(z)}{
s^{1/\alpha}}\right)^\gamma
q(s, z, y)ds.\end{aligned}$$ This combining with establishes the inequality .
It follows from and that for every $0<s<t$, and $x, y, z\in \bR^d$, $$\begin{aligned}
&&\frac{q(s,x,z) q(t-s, z,y)}{q(t,x,y)}\nonumber\\
& \le&
4^{d+\alpha} \frac{s(t-s)}{t}
\left(\frac{ t^{1/\alpha}+ |x-y|}{ (s^{1/\alpha}+
|x-z|)((t-s)^{1/\alpha} + |y-z|)} \right)^{d+\alpha} \nonumber\\
&\le&
4^{d+\alpha} (s \wedge (t-s))
\left(\frac{ (s+(t-s))^{1/\alpha}+ |x-z|+|y-z| }{
(s^{1/\alpha}+ |x-z|)((t-s)^{1/\alpha} + |y-z|)} \right)^{d+\alpha}\nonumber\\
&\le&
2^{(d+\alpha)(3+1/\alpha)} (s \wedge (t-s))
\left(\frac{1}{(s^{1/\alpha}+|x-z|)^{d+\alpha}}+
\frac{1}{((t-s)^{1/\alpha}+|y-z|)^{d+\alpha}} \right)\nonumber\\
&\le& 2^{(d+\alpha)(3+1/\alpha)}
\left(q(s,x,z)+ q(t-s, z,y) \right).\label{e:ppp}\end{aligned}$$ (See also [@BJ].)
Now we are ready to prove one form of the 3P inequality. Note that the right hand side of the 3P inequality below has the term $q(s, x, z)+q(s, z,y)$ rather than $q(t-s, x, z)+q(s, z,y)$.
\[l:3p\] For every $\gamma\in [0, \alpha)$, there exists a constant $C_2:=
C_2(d, \alpha, \gamma)>0$ such that for all $(t, x,y, z) \in (0, \infty) \times D\times D \times D$, $$\begin{aligned}
\int^t_0\frac{\psi_\gamma(t-s, x, z)q(t-s, x, z)
\psi_\gamma(s, z,y) q(s, z,y)}{ \psi_\gamma(t, x, y)
q(t, x, y)}ds
\le C_2\, \int^{t}_0 \left(1\wedge \frac{\delta_D(z)}
{s^{1/\alpha}}\right)^\gamma (q(s, x, z)+q(s, z,y))ds.\end{aligned}$$
When $\gamma =0$, the desired inequality follows from with $C_2=2^{(d+\alpha)(3+1/\alpha)}$. So for the rest of the proof, we assume $\gamma \in (0, \alpha)$. Let $$J(t, x, y, z):= \int^t_0\psi_\gamma(t-s, x, z)
q(t-s, x, z) \psi_\gamma(s, z,y) q(s, z,y) ds.$$ Since $$\begin{aligned}
J(t, x, y, z)
&\le & c_1\left(1\wedge \frac{\delta_D(x)}
{t^{1/\alpha}}\right)^\gamma \left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma q(t, x, z) \int^{t/2}_0
\psi_\gamma(s, z, y) q(s, z, y)
ds \\
&& + c_1\left(1\wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma
\left(1\wedge \frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma
q(t, z, y) \int^{t}_{t/2}\psi_\gamma(t-s,
x, z) q(t-s, x,z) ds,\end{aligned}$$ we have by Lemma \[l:4-1\] that $$\begin{aligned}
J(t, x, y, z)
&\le& c_2 \psi_\gamma(t, x, y)
\int^{t/2}_0\left(1\wedge
\frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma q(t-s, x, z) q(s, z, y)
ds \\
&&+ c_2\psi_\gamma(t, x, y) \int^{t}_{t/2}\left(1\wedge
\frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma q(t-s, x, z) q(s, z, y)
ds .\end{aligned}$$ It then follows from that $$\begin{aligned}
J(t, x, y, z)
&\le& c_3 \psi_\gamma(t, x, y) q(t, x, y) \int^{t/2}_0
\left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma
(q(t-s, x, z)+q(s, z, y)) ds \\
&& + c_3 \psi_\gamma(t, x, y) q(t, x, y) \int_{t/2}^t\left(1\wedge
\frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma ( q(t-s, x,z)
+q(s, z, y)) ds \\
&\le& c_4 \psi_\gamma(t, x, y) q(t, x, y) \int^{t}_0 \left(1\wedge
\frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma (q(s, z, y)+q(s, x,z)) ds.\end{aligned}$$ Here in the last inequality, we used the fact that $$\int_0^{t/2} \left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma
q(t-s, x, z) ds \leq c_5 \int_{t/2}^t \left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma
q(s, x, z) ds$$ and $$\int_{t/2}^t\left(1\wedge
\frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma q(s, z, y) ds
\leq c_5 \int_{t/2}^t\left(1\wedge
\frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma q(s, z, y) ds.$$ The above two inequalities can be easily verified by using the facts that $q(s, x, y)\asymp q(t, x, y)$ for $s\in [t/2, t]$ and that $$\int_0^{t/2} \left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma ds
\leq \frac{\alpha}{\alpha-\gamma} 2^{\gamma/\alpha-1} t \left(1\wedge \frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma
\leq c_6 \int_{t/2}^t \left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma ds,$$ which follows easily from the assumption $\gamma \in (0, \alpha)$ by a direct calculation. This completes the proof of the lemma.
The above 3P inequality immediately implies the following theorem, which will be used later.
\[t:key\] For every $\gamma\in [0, \alpha)$, there exists a constant $C_3=
C_3 (d, \alpha, \gamma)>0$ such that for any measure $\mu$ on $D$ and any $(t, x,y)\in (0, \infty)\times D \times D$, $$\begin{aligned}
&&\int^t_0\int_D\psi_\gamma(t-s, x, z) \psi_\gamma(s, z,y) q(t-s, x,
z) q(s, z,y) \mu(dz)ds\nonumber\\
&\le&C_3\, \psi_\gamma(t, x, y) q(t,x,y) \sup_{u\in
D}\int^t_0\int_{D}\left(1\wedge
\frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma q(s,u,z) \mu(dz)ds.\end{aligned}$$
The results of the remainder of this section are geared towards dealing with the discontinuous part of $A^{\mu, F}$.
\[l:4-2\] For every $\gamma \in [0, 2\alpha)$, there exists a constant $C_4:=C_4(d, \alpha, \gamma)>1$ such that for all $(t, y,z,w)\in (0,
\infty)\times D \times D \times D$, $$\begin{aligned}
& \left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma\int^{t/2}_0
\psi_\gamma(s, w, y)
q(s, w, y) ds\nonumber \\
\le& C_4\left(1\wedge
\frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma \left(1
+ \frac{|y-z| \wedge |z-w| \wedge t^{1/\alpha}}{|y-w|}
\right)^\gamma\int^{t/2}_0 \left(1\wedge
\frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma
q(s, w, y) ds.\label{e:ineqmain2}\end{aligned}$$
The desired inequality holds trivially for $\gamma =0$ with $C_4=1$ so for the rest of the proof we assume $\gamma \in (0, 2\alpha)$. The inequality is obvious if $\delta_D(y) \ge t^{1/\alpha}$ or $\delta_D(z) \le 2\delta_D(y)$, so we will assume $\delta_D(y) < t^{1/\alpha} \wedge (\delta_D(z)/2)$ in the remainder of this proof. Note that in this case $$\label{e:2.14}
|y-z|\geq \delta_D(z)-\delta_D(y) \geq \frac{\delta_D(z)}2 \geq \delta_D (y),$$ By , and our assumption $\delta_D(y) < t^{1/\alpha}$, we have that $$\label{e:2.13}
\left( 1 \wedge \frac{\delta_D(z)}{t^{1/\alpha}}\right)
\left( 1 \wedge \frac{\delta_D(y)}{s^{1/\alpha}}\right)
\leq 2 \, \frac{\delta_D(y)}{t^{1/\alpha}}
\left( \frac{\delta_D (z)\wedge t^{1/\alpha}}
{s^{1/\alpha}+\delta_D (y)} \right) =2 \left( 1 \wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right)
\left( \frac{\delta_D (z)\wedge t^{1/\alpha}}
{s^{1/\alpha}+\delta_D (y)}\right).$$ When $s\geq |y-w|^\alpha$, by , $$\frac{\delta_D (z)\wedge t^{1/\alpha}}{s^{1/\alpha}+\delta_D (y)}
\leq 2\, \frac{|y-z|\wedge t^{1/\alpha}}{|y-w|}
\leq 2 \left( 1+ \frac{|y-z|\wedge |z-w|
\wedge t^{1/\alpha}}{|y-w|}
\right), \nonumber$$ where the last inequality is due to the fact $|y-z| \le |y-w| + ( |y-z| \wedge |z-w|)$. This together with implies that $$\begin{aligned}
& \left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma
\int^{ t/2 }_{(t/2)\wedge |y-w|^\alpha}
\left(1\wedge \frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma
\left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma
q(s, w, y) ds\nonumber \\
\le& 4^\gamma \left(1 \wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma
\left(1 + \frac{|y-z| \wedge |z-w| \wedge t^{1/\alpha}}{|y-w|}
\right)^\gamma
\int^{ t/2 }_{(t/2)\wedge |y-w|^\alpha}\left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s, w, y) ds.\label{e:2.16}\end{aligned}$$ On the other hand, by , $$\begin{aligned}
&& \int_0^{(t/2)\wedge |y-w|^\alpha}
\left( \frac{\delta_D (z)\wedge t^{1/\alpha}}
{s^{1/\alpha}+\delta_D (y)}\right)^\gamma \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s, w, y) ds \nn \\
&\leq & 2^\gamma \int_0^{(t/2)\wedge |y-w|^\alpha}
\left( \frac{|y-z|\wedge t^{1/\alpha}}
{s^{1/\alpha} }\right)^\gamma \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma\, \frac{s}{|y-w|^{d+\alpha}} ds \nn \\
&= & c_1 \, \frac{\left( |y-z|\wedge t^{1/\alpha}\right)^\gamma}{|y-w|^{d+\alpha}}
\int_0^{(t/2)\wedge |y-w|^\alpha}
{s^{1-\gamma/\alpha} } \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma ds.
\label{e:2.15}\end{aligned}$$ We claim that $$\begin{aligned}
\int_0^{(t/2)\wedge |y-w|^\alpha}
{s^{1-\gamma/\alpha} } \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma ds
\asymp \left(\frac{t}2\wedge |y-w|^\alpha\right)^{-\gamma/\alpha}
\int_0^{(t/2)\wedge |y-w|^\alpha}
{s } \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma ds . \label{e:2.811}
\end{aligned}$$ The case $\delta_D(w) > (t/2)^{1/\alpha}$ is clear. If $\delta_D(w) \le |y-w|\wedge (t/2)^{1/\alpha} $, $$\begin{aligned}
&\int_0^{(t/2)\wedge |y-w|^\alpha}
{s^{1-\gamma/\alpha} } \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma ds \nonumber \\
= & \int_0^{\delta_D(w)^\alpha}
s^{1-\gamma/\alpha} ds
+ \delta_D(w)^\gamma\int_{\delta_D(w)^\alpha}^{(t/2)\wedge |y-w|^\alpha}
s^{1-2\gamma/\alpha} ds \nonumber \\
\asymp& \
\delta_D(w)^{2\alpha-\gamma} +
\delta_D(w)^\gamma
\left( \left(\frac{t}2\wedge |y-w|^\alpha\right)^{2-2\gamma/\alpha} - \delta_D(w)^{2(\alpha-\gamma)} \right) \nonumber \\
\asymp&\delta_D(w)^\gamma
\left(\frac{t}2\wedge |y-w|^\alpha\right)^{2-2\gamma/\alpha}\nonumber \\
=&\left(\frac{t}2\wedge |y-w|^\alpha\right)
^{-\gamma/\alpha} \left(\delta_D(w)\right)^\gamma
\left(\frac{t}2\wedge |y-w|^\alpha\right)^{2-\gamma/\alpha} \nonumber \\
\asymp&\left(\frac{t}2\wedge |y-w|^\alpha\right)
^{-\gamma/\alpha} \int_0^{(t/2)\wedge |y-w|^\alpha}
{s } \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma ds
. \nonumber\end{aligned}$$ The remaining case $ |y-w| < \delta_D(w)\le (t/2)^{1/\alpha}$ is simpler. Thus we have proved the claim . Now by and , $$\begin{aligned}
&& \int_0^{(t/2)\wedge |y-w|^\alpha}
\left( \frac{\delta_D (z)\wedge t^{1/\alpha}}
{s^{1/\alpha}+\delta_D (y)}\right)^\gamma \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s, w, y) ds \\
&\leq & c_2 \left( \frac{ |y-z|\wedge t^{1/\alpha}}{|y-w|\wedge t^{1/\alpha}}\right)^\gamma \int_0^{(t/2)\wedge |y-w|^\alpha}
\left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma \frac{s}{|y-w|^{d+\alpha}} ds \\
&\leq & c_2 \left( 1+ \frac{ |y-z|\wedge t^{1/\alpha}}{|y-w| }\right)^\gamma \int_0^{(t/2)\wedge |y-w|^\alpha}
\left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s, w, y) ds \\
&\leq & 2c_2 \left( 1+ \frac{ |y-z|\wedge |z-w|\wedge t^{1/\alpha}}
{|y-w| }\right)^\gamma \int_0^{(t/2)\wedge |y-w|^\alpha}
\left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s, w, y) ds .\end{aligned}$$ Here again the last inequality is due to the fact that $|y-z| \le |y-w| + ( |y-z| \wedge |z-w|)$. This together with and establishes the inequality . In the remainder of this section. we use the following notation: For any $(x, y)\in D\times D$, $$\begin{aligned}
V_{x,y}&:=&{\{(z,w)\in D\times D:
|x-y| \ge 4 (|y-w| \wedge |x-z|)
\}},\\
U_{x,y}&:=&(D \times D) \setminus
V_{x,y}.\end{aligned}$$ Recall that, for any bounded function $F$ on $D\times D$ we use $\|F\|_\infty$ to denote $\|F\|_{L^\infty (D \times D)}$.
Now we are ready to prove the following generalized 3P inequality.
\[t:G3p\] For every $\gamma \in [0, \alpha \wedge d)$, there exists a constant $C_5:=C_5(\alpha, \gamma, d)>0$ such that for any nonnegative bounded function $F(x,y)$ on $D\times D$, the following are true for $(t, x, y)\in (0, \infty) \times D\times D$.
\(a) If $|x-y| \le t^{1/\alpha}$, then $$\begin{aligned}
&&\int^{t}_{0}\int_{D\times D} \frac{\psi_\gamma(t-s, x, z)q(t-s,
x, z)
\psi_\gamma(s, w,y) q(s, w,y)}
{ \psi_\gamma(t, x, y) q(t, x,
y)} \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber\\
&\le& C_5\int^{t}_{0}\int_{D\times D} \left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma
{q(s, x,z)}\left(1 + \frac{ |z-w|
\wedge t^{1/\alpha}}{|x-z|} \right)^\gamma
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber\\
&&+ C_5\int^{t}_{0}\int_{D\times D} \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma
{q(s, y,w)}\left(1 + \frac{ |z-w|
\wedge t^{1/\alpha}}{|y-w|} \right)^\gamma
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds.\end{aligned}$$
\(b) If $|x-y| > t^{1/\alpha}$, then $$\begin{aligned}
&&\int^{t}_{0}\int_{U_{x,y}} \frac{\psi_\gamma(t-s, x, z)q(t-s,
x, z)\psi_\gamma(s, w,y) q(s, w,y)}
{ \psi_\gamma(t, x, y) q(t, x,
y)} \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber\\
&\le& C_5\int^{t}_{0}\int_{U_{x,y}} \left(1\wedge \frac{\delta_D(z)}{s^{1/\alpha}}\right)^\gamma
{q(s, x,z)}\left(1 + \frac{ |z-w|
\wedge t^{1/\alpha}}{|x-z|} \right)^\gamma
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber\\
&&+ C_5\int^{t}_{0}\int_{U_{x,y}} \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma
{q(s, y,w)}\left(1 + \frac{ |z-w|
\wedge t^{1/\alpha}}{|y-w|} \right)^\gamma
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds .\end{aligned}$$
\(c) If $|x-y| > t^{1/\alpha}$, then $$\int^{t}_{0}\int_{V_{x,y}} \frac{\psi_\gamma(t-s, x, z)q(t-s,
x, z) \psi_\gamma(s, w,y) q(s, w,y)}
{ \psi_\gamma(t, x, y) q(t, x,
y)} \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds \, \le \, C_5 \|F\|_{\infty }.$$
By Lemma \[l:4-2\], we get that $$\begin{aligned}
& \int^t_0\int_{D\times D}
\frac{
\psi_\gamma(t-s, x, z)
q(t-s, x, z) \psi_\gamma(s, w,y) q(s, w,y)}
{\psi_\gamma(t, x, y)}
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber\\
\le& c_1 \int_{D\times D} \int^{t/2}_{0} \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s,
w,y)q(t-s, x,z)\left(1 + \frac{ |z-w| \wedge t^{1/\alpha}}{|y-w|}
\right)^\gamma
ds \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw\nonumber\\
&+ c_1\int_{D\times D} \int^{t}_{t/2} \left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma q(s, w,y)
q(t-s, x,z)\left(1 + \frac{|z-w|\wedge t^{1/\alpha}}{|x-z|}
\right)^\gamma
ds \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw. \label{e:newe1}\end{aligned}$$ If $|x-y| \le t^{1/\alpha}$ and $s\in (0, t/2]$, we have $ q(t-s, x,z)
\le 2^{d/\alpha}q(t,x,y)$, and if $|x-y| \le t^{1/\alpha}$ and $s\in (t/2, t]$, we have $ q(s, w,y)
\le 2^{d/\alpha}q(t,x,y)$. Thus (a) follows immediately from .
In the remainder of this proof, we fix $(t,x,y)\in (0, \infty) \times D \times D$ with $|x-y|> t^{1/\alpha}$. Let $$\begin{aligned}
U_1&:=&{\{(z,w)\in D\times D: |y-w| >
4^{-1}|x-y|,
|y-w| \ge |x-z| \}}, \\
U_2&:=&{\{(z,w)\in D\times D: |x-z| >
4^{-1}|x-y| \}}.\end{aligned}$$ Since $ q(t-s, x,z) \le 4^{d+\alpha} q(t,x,y)$ for $(s, z, w)\in (0, t)\times U_2$, by Lemma \[l:4-2\], we have $$\begin{aligned}
& \int^{t/2}_0\int_{U_2}
\frac{\psi_\gamma(t-s, x, z)q(t-s, x, z) \psi_\gamma(s, w,y) q(s, w,y)}{\psi_\gamma(t, x, y)}
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber\\
\le& c_2 \int_{U_2} \int^{t/2}_{0} \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s,
w,y)q(t-s, x,z)\left(1 + \frac{ |z-w| \wedge t^{1/\alpha}}{|y-w|}
\right)^\gamma
ds \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw\nonumber\\
\le& c_3 q(t, x,y)\int_{U_2} \int^{t/2}_{0} \left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s,
w,y)\left(1 + \frac{ |z-w| \wedge t^{1/\alpha}}{|y-w|}
\right)^\gamma
ds \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw\label{e:new_1}\end{aligned}$$ and, similarly $$\begin{aligned}
& \int^t_{t/2}\int_{U_1}
\frac{\psi_\gamma(t-s, x, z)
q(t-s, x, z) \psi_\gamma(s, w,y) q(s, w,y)}{\psi_\gamma(t, x, y)}
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds\nonumber\\
\le& c_4\int_{U_1} \int^{t}_{t/2} \left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma q(s, w,y)
q(t-s, x,z)\left(1 + \frac{|z-w|\wedge t^{1/\alpha}}{|x-z|}
\right)^\gamma
ds \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw\nonumber\\
\le& c_5 q(t, x,y) \int_{U_1} \int^{t}_{t/2} \left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
q(t-s, x,z)\left(1 + \frac{|z-w|\wedge t^{1/\alpha}}{|x-z|}
\right)^\gamma
ds \frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw. \label{e:new_2}\end{aligned}$$
On the other hand, we observe that, since $ q(s, w,y) \le 4^{d+\alpha} q(t,x,y)$ for $(s, z, w)\in (0, t/2]\times U_1$, $$\begin{aligned}
& \int^{t/2}_0\int_{U_1}\psi_\gamma(t-s, x, z)
q(t-s, x, z) \psi_\gamma(s, w,y) q(s, w,y)
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds\\
\le&
c_6 \psi_\gamma(t, x, z)q(t, x, y)
\int_{U_1}q(t, x, z)
\int^{t/2}_0 \psi_\gamma(s, w,y) ds
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw.\end{aligned}$$ Now, applying the inequality $$\int^{t/2}_0 \psi_\gamma(s, w,y) ds \le
\int^{t/2}_0 \left(1\wedge \frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma ds
\le
\frac{\alpha}{\alpha-\gamma} 2^{\gamma/\alpha-1}
t\left(1\wedge \frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma,$$ we get $$\begin{aligned}
& \int^{t/2}_0\int_{U_1}\psi_\gamma(t-s, x, z)
q(t-s, x, z) \psi_\gamma(s, w,y) q(s, w,y)
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds\nonumber\\
\le&
c_7 \psi_\gamma(t, x, y)q(t, x, y)
\int_{U_1}q(t, x, z)
t\left(1\wedge \frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdw\nonumber\\
\le&
c_8 \psi_\gamma(t, x, y)q(t, x, y)
\int_{U_1}
\int_0^{t/2} q(t-s, x, z)\left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
\frac{F(z,w)}{|z-w|^{d+\alpha}}ds dzdw.\label{e:new_3}\end{aligned}$$ Similarly $$\begin{aligned}
& \int^t_{t/2}\int_{U_2}\psi_\gamma(t-s, x, z)
q(t-s, x, z) \psi_\gamma(s, w,y) q(s, w,y)
\frac{F(z,w)}{|z-w|^{d+\alpha}} dzdwds\nonumber\\
\le& c_{9}\psi_\gamma(t, x, y)q(t, x, y)\int_{U_2} \int^{t}_{t/2}
\left(1\wedge \frac{\delta_D(w)}{s^{1/\alpha}}\right)^\gamma q(s, w,y)
\frac{F(z,w)}{|z-w|^{d+\alpha}}ds dzdw. \label{e:new_4}\end{aligned}$$ Since $U_{x, y}=U_1\cup U_2$, from –, we know that (b) is true.
Note that for $(z,w)\in V_{x,y}$, we have $|z-w| \ge |x-y| -(|x-z|+|y-w|) \ge 2^{-1} |x-y|.$ Thus, by Lemma \[l:4-2\] and , it is easy to see that $$\begin{aligned}
& \int^t_0\int_{V_{x,y}}\frac{\psi_\gamma(t-s, x, z)
q(t-s, x, z) \psi_\gamma(s, w,y) q(s, w,y)}{\psi_\gamma(t, x, y)}
\frac{F(z,w)}{q(t,x,y)|z-w|^{d+\alpha}} dzdwds\\
\le& c_{10} \|F\|_\infty t^{-1}\int_{V_{x,y}} \int^{t/2}_{0} {q(s,
w,y)q(t-s, x,z)}\left(1 +\frac{ t^{1/\alpha}}{|y-w|}
\right)^\gamma dsdzdw\\
&+ c_{10} \|F\|_\infty t^{-1} \int_{V_{x,y}} \int^{t}_{t/2} {q(s,
w,y)q(t-s, x,z)}\left(1 +\frac{ t^{1/\alpha}}{|x-z|}
\right)^\gamma dsdzdw\\
\le &c_{11} \|F\|_\infty t^{-1} \int^{t}_{0}\left(\int_{D} {q(s,
w,y)}\left(1 +\frac{ t^{1/\alpha}}{|y-w|}
\right)^\gamma dw + \int_{D} {q(s,
x,z)}\left(1 +\frac{ t^{1/\alpha}}{|x-z|}
\right)^\gamma dz \right) ds.\end{aligned}$$ Since, using $\gamma \in (0, \alpha \wedge d)$, $$\begin{aligned}
&&\int^{t}_{0}\left(\int_{D} {q(s,
w,y)}\left(1 +\frac{ t^{1/\alpha}}{|y-w|}
\right)^\gamma dw + \int_{D} {q(s,
x,z)}\left(1 +\frac{ t^{1/\alpha}}{|x-z|}
\right)^\gamma dz \right) ds\\
&\le & 2^{d+\alpha+1}\int_0^t
\int_{\bR^d} \frac{s}{(s^{1/\alpha}+|w|)^{d+\alpha}}
\left(\frac{ t^{1/\alpha}}{|w|} \right)^\gamma dwds\\
&= & c_{12}\left(\int^\infty_{0} \frac{u^{d-1-\gamma}
du}{(1+u)^{d+\alpha}} \right) t^{\gamma /\alpha} \int_0^t
s^{-\gamma /\alpha} ds \le c_{13} t,\end{aligned}$$ (c) follows immediately.
Heat kernel estimates {#s:3}
=====================
In this section we give the proof of our main result, Theorem \[t:main\]. Throughout this section, we fix $\gamma\in [0, \alpha\wedge d)$. Recall the definition of $p^k(t, x, y)$ given by .
Using , , Theorems \[t:key\] and \[t:G3p\], we can choose a constant $$\label{e:M1}
M= M (\alpha, \gamma, d, C_0)>
\frac{\alpha}{\alpha-\gamma} 2^{2\gamma/\alpha+d+\alpha+1} C_0^4
(C_1 \vee C_4)$$ such that for any $\mu$ in ${\bf K}_{\alpha, \gamma}$, any measurable function $F$ with $F_1=e^F-1 \in {\bf J}_{\alpha, \gamma}$ and any $(t, x,y)\in
(0, 1]\times D
\times D$, $$\begin{aligned}
&&\int^t_0\int_D\psi_\gamma(t-s, x, z) \psi_\gamma(s, z,y) q(t-s, x,
z) q(s, z,y) |\mu|(dz)ds\nonumber\\
&\le&M\, p_D(t, x, y) N_\mu^{\alpha, \gamma}(t),\label{e:M2}\\
&& \int^{t}_{0}\int_{D\times D} {\psi_\gamma(t-s, x,
z)q(t-s, x, z)
\psi_\gamma(s,w,y)
q(s,w,y)}\frac{c(z, w) |F_1|(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber \\
&\le& M\, p_D(t, x, y)(N_{F_1}^{\alpha, \gamma}(t)+
\|F_1\|_{\infty}\text{{\bf 1}}_{\{ |x-y| > t^{1/\alpha}
\}})\label{e:M3}\end{aligned}$$ and $$\begin{aligned}
&& \int^{t}_{0}\int_{
U_{x,y}} {\psi_\gamma(t-s, x, z)q(t-s, x, z)
\psi_\gamma(s,w,y)
q(s,w,y)}
\frac{c(z, w)|F_1|(z,w)}{|z-w|^{d+\alpha}} dzdwds \nonumber\\
&\le& M\, p_D(t, x, y)N_{F_1}^{\alpha, \gamma}(t). \label{e:M4}\end{aligned}$$
In the remainder of this section, we fix a signed measure $\mu\in{\bf K}_{\alpha, \gamma}$, a measurable function $F$ with $F_1=e^F-1 \in {\bf J}_{\alpha, \gamma}$ and the constant $M>0$ in .
\[l:i1\] For every $k \ge 0$ and $(t, x)\in (0, 1]\times D$, $$\begin{aligned}
\label{e:i1}
\int_D|p^k(t, x, y)|dy \le C_0^2M^k
\,\left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma \left(N^{\alpha,
\gamma}_{\mu, F_1}(t)\right)^k.\end{aligned}$$
We use induction on $k \ge 0$. By , is clear when $k=0$. Suppose is true for $k-1\ge 0$. Then by we have $$\begin{aligned}
\int_D p^k(t, x, y) dy
=&\int^{t/2}_0 \left(\int_{D}p^0(t-s, x, z)\Big(\int_D
p^{k-1}(s, z, y)dy\Big)\mu(dz) \right)ds \\
&+\int^{t/2}_0 \left(\int_{D}\int_{D} p^0(t-s, x,
z)\frac{c(z, w)F_1(z,w)}{|z-w|^{d+\alpha}} \int_D
p^{k-1}(s, w, y)dy) dzdw \right)ds\\
&+\int^t_{t/2} \left(\int_{D}p^0(t-s, x,
z)\Big(\int_Dp^{k-1}(s, z, y)dy\Big)\mu(dz) \right)ds \\
&+ \int^t_{t/2} \left(\int_{D}\int_{D} p^0(t-s, x,
z)\frac{c(z, w)F_1( z,w)}{|z-w|^{d+\alpha}} \int_D p^{k-1}(s, w, y)dy)
dzdw \right)ds.\end{aligned}$$ Thus using and our induction hypothesis, we have $$\begin{aligned}
&\int_D |p^k(t, x, y)| dy\\
\le& 2^{\gamma/\alpha} C_0^3 M^{k-1}\big(N^{\alpha, \gamma}_{\mu,
F_1}(t)\big)^{k-1} \left( \left(1\wedge \frac{
\delta_D(x)}{t^{1/\alpha}}\right)^\gamma\int^{t/2}_0 \left(\int_{D}
\left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
q(t-s, x, z)|\mu|(dz) \right)ds \right. \\
& + C_0 \left(1\wedge \frac{
\delta_D(x)}{t^{1/\alpha}}\right)^\gamma\int^{t/2}_0
\left(\int_{D}\int_{D}
\left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
q(t-s, x, z)
\frac{|F_1|(z,w)}
{|z-w|^{d+\alpha}} dzdw \right)ds\\
& + \int_{D}\left(1\wedge
\frac{\delta_D(z)}{t^{1/\alpha}}\right)^\gamma \int^t_{t/2}
\psi_\gamma(t-s, x,
z) q(t-s, x, z)ds |\mu|(dz) \\
& \left. + C_0 \int_{D}\int_{D} \left(1\wedge \frac{\delta_D(w)}{t^{1/\alpha}}
\right)^\gamma
\int^t_{t/2} \psi_\gamma(t-s, x, z)
q(t-s, x,z)
ds \frac{|F_1|(z,w)}{|z-w|^{d+\alpha}} dzdw \right) .\end{aligned}$$ Applying , Lemmas \[l:4-1\] and \[l:4-2\], the above is less than $$\begin{aligned}
& 4^{-1}C_0 M^{k}\big(N^{\alpha, \gamma}_{\mu,
F_1}(t)\big)^{k-1} \left( \left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma\int^{t/2}_0
\left(\int_{D} \left(1\wedge
\frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
q(t-s, x, z)|\mu|(dz) \right)ds \right. \\
&+ C_0 \left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma \int^{t/2}_0
\left(\int_{D}\int_{D}
\left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
q(t-s, x, z)
\frac{|F_1|(z, w)}{|z-w|^{d+\alpha}} dzdw \right)ds\\
&+ \int_{D}
\left(1\wedge \frac{\delta_D(x)}{t^{1/\alpha}} \right)^\gamma
\int^t_{t/2} \left(1\wedge
\frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
q(t-s, x, z)ds |\mu|(dz) \\
& + C_0 \int_{D}\int_{D} \left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma\left(1 + \frac{|x-w|
\wedge |z-w|\wedge t^{1/\alpha}}{|x-z|}
\right)^\gamma
\\
&\left.\quad \times \int^{t}_{t/2}
\left(1\wedge \frac{\delta_D(z)}{(t-s)^{1/\alpha}}\right)^\gamma
q(t-s, x, z)
\frac{|F_1|(z,w)}{|z-w|^{d+\alpha}} dzdw \right) \\
\le&C_0^2 M^{k}\,\left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma (N^{\alpha,
\gamma}_{\mu, F_1}(t))^k.\end{aligned}$$
\[l:i2\] For every $k \ge 0$ and $(t, x, y)\in (0, 1]\times D\times D$, $$\begin{aligned}
\int_0^t \int_D p_D(t-s, x, z) dz \int_D
|p^k(s, w, y)| dw ds
\,\le\,t\,
\frac{\alpha}{\alpha-\gamma}2^{2\gamma/\alpha}
\, C_0^4\, M^k\,
\psi_\gamma (t,x,y)\big(N^{\alpha,
\gamma}_{\mu, F_1}(t)\big)^k.\end{aligned}$$
By and Lemma \[l:i1\], $$\begin{aligned}
&&\int_0^t \int_D p_D(t-s, x, z) dz \int_D
|p^k(s, w, y)|
dw
ds \\
&=& \int_0^{t/2} \int_D p_D(t-s, x, z) dz \int_D
|p^k(s, w, y)| dw
ds+\int_{t/2}^t \int_D p_D(t-s, x, z) dz \int_D
|p^k(s, w, y)| dw
ds\\
&\le&
C_0 \int_0^{t/2} \int_D \left(1\wedge
\frac{\delta_D(x)}{(t-s)^{1/\alpha}}\right)^\gamma q(t-s, x, z) dz
\int_D
|p^k(s, w, y)| dw ds\\
&&+
C_0 \int_{t/2}^t \int_D\left(1\wedge \frac{\delta_D(x)}
{(t-s)^{1/\alpha}}\right)^\gamma q(t-s, x, z) dz \int_D
|p^k(s, w, y)| dw ds\\
&\le&
C_0^3 M^{k} \int_0^{t/2} \int_D \left(1\wedge
\frac{\delta_D(x)}{(t-s)^{1/\alpha}}\right)^\gamma q(t-s, x, z) dz
\,\left(1\wedge \frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma
\big(N_{\mu, F_1}(s)\big)^k ds\\
&&+
C_0^3 M^{k} \int_{t/2}^t \int_D\left(1\wedge
\frac{\delta_D(x)} {(t-s)^{1/\alpha}}\right)^\gamma q(t-s, x, z)
dz \,\left(1\wedge \frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma
\big(N_{\mu, F_1}(s)\big)^k ds\\
&\le&2^{\gamma/\alpha}
C_0^3M^{k} \big(N^{\alpha, \gamma}_{\mu,
F_1}(t)\big)^k \left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma
\left(\int_0^{t/2}\left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma ds \right) \int_D
q(t-s, x, z) dz \\
&&+2^{\gamma/\alpha}
C_0^3M^{k} \big(N^{\alpha, \gamma}_{\mu,
F_1}(t)\big)^k \left(1\wedge
\frac{\delta_D(y)}{t^{1/\alpha}}\right)^\gamma \int_D
\left(\int_{t/2}^t\left(1\wedge
\frac{\delta_D(x)}{(t-s)^{1/\alpha}}\right)^\gamma ds \right)q(t-s,
x, z) dz.\end{aligned}$$ Using $$\int_0^{t/2} \left(1\wedge
\frac{\delta_D(x)}{s^{1/\alpha}}\right)^\gamma ds
\le\frac{\alpha}{\alpha-\gamma}2^{\gamma/\alpha-1}
t \left(1\wedge
\frac{\delta_D(x)}{t^{1/\alpha}}\right)^\gamma,$$ we get that $$\begin{aligned}
&&\int_0^t \int_D p_D(t-s, x, z) dz \int_D p^k(s, w, y) dw
ds \\
&&\,\le\,
\frac{\alpha}{\alpha-\gamma}2^{2\gamma/\alpha}
C_0^3 M^{k} \big(N^{\alpha, \gamma}_{\mu,
F_1}(t)\big)^k \left(1\wedge \frac{\delta_D(y)}
{t^{1/\alpha}}\right)^\gamma \int_D q(t-s, x, z) dz.\end{aligned}$$ Applying , we have proved the lemma.
\[l:i3\] For $k \ge 0$ and $(t, x, y)\in (0, 1]\times D\times D$ we have $$\label{e:i3-1}
|p^k(t, x, y)| \le p^0(t,x,y) \left(\big(C_0^2MN^{\alpha, \gamma}_{\mu,
F_1}(t)\big)^k +k \|F_1\|_\infty C_0^2M \big(C_0^2MN^{\alpha, \gamma}_{\mu,
F_1}(t)\big)^{k-1} \right).$$
We use induction on $k \ge 0$. The $k=0$ case is obvious. Suppose that is true for $k-1\ge 0$. Recall that $$\begin{aligned}
V_{x,y}={\{(z,w)\in D\times D:
|x-y|\ge 4(|y-w|
\wedge |x-z|) \}}, \quad
U_{x,y}=(D \times D) \setminus
V_{x,y}.\end{aligned}$$ Applying , , and , we have by our induction hypothesis $$\begin{aligned}
&&
|p^{k}(t, x, y)|
\,\le\,\int^t_0 \left(\int_{D}p^0(t-s, x,
z)
|p^{k-1}(s, z, y)|
|\mu|(dz) \right)ds \\
&&+ \int^t_0 \left(\int_{
U_{x,y}} p^0(t-s, x, z)\frac{c(z, w) |F_1(z,
w)|}{|z-w|^{d+\alpha}}
| p^{k-1}(s, w, y) |dzdw \right)ds\\
&&+\int^t_0 \left(\int_{
V_{x,y}} p^0(t-s, x, z)\frac{c(z, w)|F_1(z,
w)|}{|z-w|^{d+\alpha}}
| p^{k-1}(s, w, y) | dzdw \right)ds\\
&\le& \left(\big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}
+(k-1) \|F_1\|_\infty C_0^2M \big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-2} \right) \\
&&\qquad\times \int^t_0 \left(\int_{D}p^0(t-s, x, z)p^{0}(s, z,
y)|\mu|(dz) \right) ds \\
&&+\left(\big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}
+(k-1)\|F_1\|_\infty C_0^2M \big(C_0^2MN^{\alpha,
\gamma}_{\mu, F_1}(t)\big)^{k-2} \right) \\
&&\qquad\times \int^t_0 \left(\int_{
U_{x,y}} p^0(t-s, x,
z)\frac{c(z, w)|F_1|(z,
w)}{|z-w|^{d+\alpha}}p^{0}(s, w, y) dzdw \right)ds \\
&&+ \int^t_0 \left(\int_{
V_{x,y}} p^0(t-s, x, z)\frac{c(z, w)|F_1|(z,
w)}{|z-w|^{d+\alpha}}|p^{k-1}(s, w, y)| dzdw \right)ds\\
&\le& p^0(t,x,y) \left(\big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}
+(k-1) \|F_1\|_\infty C_0^2M \big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-2}
\right) C_0^2 M N^{\alpha, \gamma}_{\mu, F_1}(t) \\
&&+C_0\frac{2^{d+\alpha}\|F_1\|_\infty}{|x-y|^{d+\alpha}} \int^t_0 \left(\int_{D\times D} p^0(t-s, x, z)
|p^{k-1} (s, w, y)|
dzdw \right)ds.\end{aligned}$$ Applying Lemma \[l:i2\] and using , we get that if $|x-y|^\alpha \ge t$, $$\begin{aligned}
&&C_0\frac{2^{d+\alpha}\|F_1\|_\infty}{|x-y|^{d+\alpha}}\int^t_0
\left(\int_{D\times D} p^0(t-s, x, z)
|p^{k-1} (s, w, y)|
dzdw \right)ds\\
&\le& \psi_\gamma (t,x,y) \frac{ t }{|x-y|^{d+\alpha}}
\|F_1\|_\infty
C_0^5\frac{\alpha}{\alpha-\gamma}
2^{d+\alpha+\gamma/\alpha} M^{k-1} \big(N^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}\\
&\le& p^0 (t,x,y)
\|F_1\|_\infty
C_0^6\frac{\alpha}{\alpha-\gamma}
2^{d+\alpha+\gamma/\alpha} M^{k-1} \big(N^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}\\
&\le& p^0 (t,x,y) \|F_1\|_\infty C_0^2 M^{k} \big(N^{\alpha,
\gamma}_{\mu, F_1}(t)\big)^{k-1} \\
&\le& p^0 (t,x,y) \|F_1\|_\infty C_0^2 M \big(C_0^2 M N^{\alpha,
\gamma}_{\mu, F_1}(t)\big)^{k-1}.\end{aligned}$$ Thus is true for $k$ when $|x-y|^\alpha \ge t$.
If $|x-y|^\alpha \le t$, using , , and , we have by our induction hypothesis $$\begin{aligned}
&&
|p^{k}(t, x, y)|
\,\le\,\int^t_0 \left(\int_{D}p^0(t-s, x,
z)
|p^{k-1}(s, z, y)|
|\mu|(dz) \right)ds \\
&&+ \int^t_0 \left(\int_{
D \times D} p^0(t-s, x, z)\frac{c(z, w) |F_1(z,
w)|}{|z-w|^{d+\alpha}}
| p^{k-1}(s, w, y) |dzdw \right)ds\\
&\le& \left(\big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}
+(k-1) \|F_1\|_\infty C_0^2M \big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-2} \right) \\
&&\qquad\times \int^t_0 \left(\int_{D}p^0(t-s, x, z)p^{0}(s, z,
y)|\mu|(dz) \right) ds \\
&&+\left(\big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}
+(k-1)\|F_1\|_\infty C_0^2M \big(C_0^2MN^{\alpha,
\gamma}_{\mu, F_1}(t)\big)^{k-2} \right) \\
&&\qquad\times \int^t_0 \left(\int_{
D \times D} p^0(t-s, x,
z)\frac{c(z, w)|F_1|(z,
w)}{|z-w|^{d+\alpha}}p^{0}(s, w, y) dzdw \right)ds \\
&\le& p^0(t,x,y) \left(\big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-1}
+(k-1) \|F_1\|_\infty C_0^2M \big(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t)\big)^{k-2}
\right) C_0^2 M N^{\alpha, \gamma}_{\mu, F_1}(t).\end{aligned}$$ The proof is now complete.
Since $F_1 \in {\bf J}_{\alpha, \gamma}$, there is $t_1 :=t_1(d, \alpha,
\gamma, C_0, M, N^{\alpha, \gamma}_{\mu, F_1}, \|F_1\|_\infty)\in (0, 1)$ so that $$N^{\alpha, \gamma}_{\mu, F_1}(t_1) \le
\big(3C_0^2M\big)^{-1} \wedge \big(9(C_0^2M)^2
\|F_1\|_\infty\big)^{-1}.$$ It follows from Lemma \[l:i3\] that for every $(t, x,y) \in (0, t_1] \times D \times D$, $$\begin{aligned}
&&\sum_{k=0}^\infty |p^k(t, x, y)| \,=\, p^0(t, x, y) +
\sum_{k=1}^\infty |p^k(t, x, y)|\nonumber\\
&\le& p^0(t, x, y) +p^0(t,x,y) \left( \sum_{k=1}^\infty
(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t))^k +\|F_1\|_\infty C_0^2M
\sum_{k=1}^\infty k
(C_0^2MN^{\alpha, \gamma}_{\mu, F_1}(t))^{k-1} \right) \nonumber\\
&\le& p^0(t, x, y) +p^0(t,x,y) \left( \frac12
+\frac94\|F_1\|_\infty C_0^2M \right) \nonumber \\
&\le& p^0(t,x,y) \left( \frac32
+\frac94\|F_1\|_\infty C_0^2M \right). \label{e:newq1}\end{aligned}$$ Hence, Fubini’s Theorem, and yield and . Thus we conclude from and that
\[t:ub\] There exist $t_1:=t_1
(d, \alpha, \gamma, C_0, M, N^{\alpha, \gamma}_{\mu, F_1}, \|F_1\|_\infty)\in (0, 1)$ and a positive constant $C_6:=C_6
(d, \alpha,
\gamma, C_0, M, \|F_1\|_\infty)$ such that the Feynman-Kac semigroup $T^{\mu, F}_t$ corresponding to $\mu$ and $F$ has a continuous density $q_D(t, x, y)$ for $t \le t_1$ and $$\label{e:hke1}
q_D(t, x, y) \le C_6 \psi_\gamma(t, x, y)q(t,x,y)$$ for every $(t,x,y) \in (0, t_1] \times D \times D$.
For the lower bound estimate, we need to assume that $F$ is a function in ${\bf J}_{\alpha, \gamma}$.
\[t:lb-1\] Suppose that $\mu \in {\bf K}_{\alpha, \gamma}$ and $F$ is a function in ${\bf J}_{\alpha, \gamma}$. Then there exist constants $t_2:=t_2(\alpha, \gamma, C_0, M, N^{\alpha, \gamma}_{\mu, F}, \|F\|_\infty)\in (0, 1)$ and $C_7:=C_7(\alpha, \gamma, C_0, M, N^{\alpha, \gamma}_{\mu, F}, \|F\|_\infty)>1$ such that $$\label{e:hke2}
C_7^{-1} \psi_\gamma(t, x, y)q(t,x,y) \leq q_D(t, x, y)
\leq
C_7 \psi_\gamma(t, x, y)q(t,x,y)$$ for every $(t,x,y) \in (0, t_2] \times D \times D$.
Since $F$ is a bounded function in ${\bf J}_{\alpha, \gamma}$, so is $F_1:=e^F-1$ with $|F_1 (x, y) | \leq e^{\| F\|_\infty} |F| (x, y)$ and $N^{\alpha, \gamma}_{F_1}\leq e^{\| F\|_\infty} N^{\alpha, \gamma}_{F}$. Thus the upper bound estimate in follows directly from Theorem \[t:ub\]. To establish the lower bound, we define for $(t, x, y)\in (0, \infty)\times D\times D$, $$\begin{aligned}
{\tilde p}^1(t, x, y)&=&\int^t_0\left(\int_Dp^0(t-s, x,
z)p^0(s, z, y)|\mu|(dz) \right)ds\\
&&+ \int^t_0\left(\int_D\int_Dp^0(t-s, x, z)\frac{c(z, w)|F|(z,
w)}{|z-w|^{d+\alpha}}p^0(s, w, y)dzdw \right)ds.\end{aligned}$$ Then for any bounded Borel function $f$ on $D$ and any $(t, x)\in
(0, \infty)\times D$, we have $$\E_x \left[ A^{|\mu|, |F|}_tf(X_t) \right]
=\int_D
{\tilde p}^1(t, x, y)f(y)dy.$$ Applying Lemma \[l:i3\] with $|\mu|$ and $|F|$ in place of $\mu$ and $F_1$, we have $${\tilde p}^1(t, x, y)\,\le\,(C^2_0MN^{\alpha, \gamma}_{\mu, F}(1)+C^2_0M\|F\|_\infty) p^0(t, x, y)
\,=:\,(k/2)p^0(t, x, y)$$ for all $(t, x , y)\in
(0,1]\times D\times D$. Hence we have for all $(t, x , y)\in
(0,1]\times D\times D$, $$\label{e:lbd1}
p^0(t, x, y)- \frac1k
{\tilde p}^1(t, x, y)\ge \frac12p^0(t, x,
y).$$ Using the elementary fact that $$1-A^{|\mu|, |F|}_t/k\le \exp\left(-A^{|\mu|, |F|}_t/k\right)\le \exp\left( A^{\mu, F}_t/k \right),$$ we get that for any $B(x, r)\subset D$ and any $(t, y)\in
(0, 1]\times D$, $$\frac1{|B(x, r)|}\E_y\left[\big(1-A^{|\mu|, |F|}_t/k \big)
{\bf 1}_{B(x,
r)}(X_t) \right] \le \frac1{|B(x, r)|}
\E_y\left[\exp (A^{\mu, F}_t/k)
{\bf 1}_{B(x, r)}(X_t) \right].$$ Thus, by and Hölder’s inequality, we have $$\begin{aligned}
&&\frac12\frac1{|B(x, r)|}\E_y\left[
{\bf 1}_{B(x, r)}(X_t) \right]\,\le \, \frac1{|B(x, r)|}\E_y\left[\exp (A^{\mu, F}_t/k)
{\bf 1}_{B(x,
r)}(X_t) \right]\\
&&\le
\left(\frac1{|B(x, r)|}\E_y\left[\exp (A^{\mu, F}_t)
{\bf 1}_{B(x, r)}(X_t) \right] \right)^{1/k}
\left(\frac1{|B(x, r)|} \E_y\left[
{\bf 1}_{B(x, r)}(X_t) \right]\right)^{1-1/k}.\end{aligned}$$ Therefore $$\frac1{2^k}\frac1{|B(x, r)|} \E_y\left[
{\bf 1}_{B(x, r)}(X_t) \right]
\le \frac1{|B(x, r)|}
\E_y\left[\exp (A^{\mu, F}_t )
{\bf 1}_{B(x, r)}(X_t)
\right].$$ We conclude by sending $r\downarrow 0$ that for every $(t, x , y)\in
(0,1]\times D\times D$, $2^{-k}p^0(t, x, y)\le q_D(t, x, y)$.
Combining the two theorems above with the semigroup property, we immediately get the main result of this paper, Theorem \[t:main\].
Applications {#s:4}
============
In this section, we will apply our main result to (reflected) symmetric stable-like processes, killed symmetric $\alpha$-stable processes, censored $\alpha$-stable processes and stable processes with drifts. We first record the following two facts.
Suppose that $d\geq 2$ and $\alpha \in (0, 2)$. A signed measure $\mu$ on $\bR^d$ is said to be in Kato class $\bK_{d, \alpha}$ $$\lim_{r\to 0} \sup_{x\in \bR^d} \int_{B(x, r)} \frac1{|x-y|^{d-\alpha}} |\mu| (dy) =0.$$ A function $
g$ on $\bR^d$ is said to be in $\bK_{d, \alpha}$ if $
g(x) dx\in \bK_{d, \alpha}$.
\[P:4.1\] Suppose that $d\geq 2$ and $\alpha \in (0, 2)$.
[(i)]{} Let $D$ be an arbitrary Borel subset of $\bR^d$. $\mu \in {\bf K}_{\alpha, 0}$ if and only if ${\bf 1}_D \mu \in
\bK_{d, \alpha}$. Hence for every $\mu \in \bK_{d, \alpha}$, $\mu |_D
\in {\bf K}_{\alpha, \gamma}$ for every $\gamma \geq 0$. In particular, $L^\infty (D; dx) \subset {\bf K}_{\alpha, \gamma}$ and $L^p(D; dx) \subset {\bf K}_{\alpha, \gamma}$ for every $p>d/\alpha$ and $\gamma \geq 0$.
[(ii)]{} Suppose that $D$ is a bounded Lipschitz open set in $\bR^d$ and $\gamma \in (0, \alpha)$. Let $
g$ be a function defined on $D$. If there exist constants $c>0$, $\beta\in (0, \, \gamma + (\alpha -\gamma)/d)$ and a compact subset $K$ of $D$ such that $
{\bf 1}_K(x)
g(x) \in \bK_{d, \alpha}$ and $$|
g(x)|\le
c \delta_D (x)^{-\beta} \quad \hbox{for } x\in D\setminus K,$$ then $
g \in {\bf K}_{\alpha, \gamma}$.
\(i) By the proof of [@Zh Theorem 2], we have that $\mu\in \bK_{d, \alpha}$ if and only if $$\lim_{t\to 0} \sup_{x\in \bR^d} \int_0^t \int_{\bR^d}
q(s, x, y) \mu (dy) ds =0.$$ This implies that $\mu \in {\bf K}_{\alpha, 0}$ if and only if ${\bf 1}_D \mu \in \bK_{d, \alpha}$. In particular we have for every $\mu \in \bK_{d, \alpha}$, $\mu |_D \in {\bf K}_{\alpha, \gamma}$ for every $\gamma \ge 0$. Clearly $L^\infty (D; dx)\subset \bK_{d, \alpha}$. Using Hölder’s inequality, one concludes that $L^p(\bR^d; dx) \subset \bK_{d, \alpha}$ for every $p>d/\alpha$.
\(ii) Let $
g$ be a function defined on $D$ such that there exist constants $
c_1>0$, $\beta\in (0, \, \gamma + (\alpha -\gamma)/d)$ and a compact subset $K$ of $D$ so that $
{\bf 1}_K(x)
g(x) \in \bK_{d, \alpha}$ and $|
g(x)|\le
c_1\delta_D (x)^{-\beta}$ for $x\in D\setminus K$. In view of (i), it suffices to show that $
{\bf 1}_{D\setminus K}
g\in
{\bf K}_{\alpha, \gamma}$. Note that $$\begin{aligned}
&& \sup_{x\in D}\int^t_0\int_{D\setminus K}\left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma
q(s, x, y)|
g(y)| dy ds \nn \\
&\leq & c_1 \sup_{x\in D}\int^t_0\int_{D\setminus K}\left(1\wedge
\frac{\delta_D(y)}{s^{1/\alpha}}\right)^\gamma \delta_D (y)^{-\beta}
q(s, x, y) dy ds \nn \\
&\le
& c_1 \sup_{x\in D}
\int_{D\setminus K} \left( \int_0^{\delta_D(y)^\alpha \wedge t}
\left( s^{-d/\alpha} \wedge \frac{s}{|x-y|^{d+\alpha}}\right) ds \right) \delta_D(y)^{-\beta} dy \nn \\
&&+
c_1 \sup_{x\in D}
\int_{D\setminus K} \left( \int_{\delta_D(y)^\alpha \wedge t}^t
s^{-\gamma /\alpha}
\left( s^{-d/\alpha} \wedge \frac{s}{|x-y|^{d+\alpha}}\right) ds \right) \delta_D(y)^{\gamma -\beta} dy
\,
=:\, I+II. \label
{e:4.2-1}\end{aligned}$$ Here $$\begin{aligned}
I &\leq & c_1 \sup_{x\in D}
\left(\int_D \int_0^{\delta_D (y)^\alpha \wedge |x-y|^\alpha
\wedge t} \frac{s}{|x-y|^{d+\alpha}} ds \,\delta_D(y)^{-\beta} dy \right.
\nn \\
&& +
\left. \int_D \int_{\delta_D (y)^\alpha \wedge |x-y|^\alpha
\wedge t}^{\delta_D(y)^\alpha \wedge t} s^{-d/\alpha} ds \, \delta_D(y)^{-\beta} dy \right)\nn \\
&\leq& c_2 \sup_{x\in D} \int_D \left(
\frac{(\delta_D (y)^\alpha \wedge |x-y|^\alpha
\wedge t)^2 \delta_D (y)^{-\beta}}{|x-y|^{d+\alpha}}
+
{\bf 1}_{\{ |x-y|< \delta_D (y) \wedge t^{1/\alpha}\}} \, \frac{ \delta_D (y)^{-\beta}}{ |x-y|^{d-\alpha}} \right) dy \nn \\
&\leq & c_2 \sup_{x\in D} \int_D
\left( \frac{( |x-y|
\wedge t^{1/\alpha})^{2\alpha-\beta}} {|x-y|^d}
+{\bf 1}_{\{ |x-y|< \delta_D (y) \wedge t^{1/\alpha}\}}
\frac{1}{ |x-y|^{d-\alpha+\beta}} \right) dy \nn \\
&\leq &
2c_2t^{(\alpha-\beta)/(2\alpha)} \sup_{x\in D} \int_D \frac{1}{|x-y|^{d-(\alpha-\beta)/2}} dy
= c_3\, t^{(\alpha-\beta)/(2\alpha)},
\label{e:4.3}\end{aligned}$$ while $$\begin{aligned}
II
&\leq & c_1 \sup_{x\in D}
\int_{D} \left( \int_{\delta_D(y)^\alpha \wedge t}^t
{\bf 1}_{\{s<|x-y|^\alpha\}} \frac{s^{1-\gamma /\alpha}}{|x-y|^{d+\alpha}}
ds \right) \delta_D(y)^{\gamma -\beta} dy \nn \\
&& + c_1 \sup_{x\in D}
\int_{D} \left( \int_{\delta_D(y)^\alpha \wedge t}^t
{\bf 1}_{\{s\geq |x-y|^\alpha\}} \, s^{-(d+\gamma )/\alpha}
ds \right) \delta_D(y)^{\gamma -\beta} dy \nn \\
&\leq & c_4 \sup_{x\in D} \int_{D}
\frac{\left(|x-y|\wedge t^{1/\alpha}\right)^{2\alpha-\gamma}}{|x-y|^{d+\alpha}} \,
{\bf 1}_{\{ \delta_D(y)< |x-y| \wedge t^{1/\alpha}\}} \,
\delta_D(y)^{\gamma -\beta} dy \nn\\
&& + c_4 \sup_{x\in D} \int_D |x-y|^{\alpha -d -\gamma}
{\bf 1}_{\{ |x-y| \le t^{1/\alpha}\}}
\delta_D(y)^{\gamma -\beta} dy \nn\\
&\leq & c_4 t^{\delta/\alpha}
\sup_{x\in D} \int_D \frac{1}{|x-y|^{d-\alpha +\eps} \delta_D (y)^{\beta -\gamma} } \, dy, \label{e:4.4}\end{aligned}$$ where $\delta := (\alpha -\gamma -d(\beta -\gamma))/2>0$ and $\eps := (\alpha +\gamma -d(\beta -\gamma))/2>0$. Note that $\eps + \delta= \alpha-d(\beta-\gamma)$ and $\eps-\delta =\gamma$. Let $p=d/(d-\alpha+\eps+\delta/2)$ and $q= d/(\alpha-(\eps + \delta/2))$ so that $1/p+1/q=1$. Since $D$ is a bounded Lipschitz open set, $p(d-\alpha+\eps)<d$ and $q(\beta-\gamma)<1$, we have by Young’s inequality, $$\sup_{x\in D} \int_D \frac{1}{|x-y|^{d-\alpha +\eps}
\delta_D (y)^{\beta -\gamma} } \, dy
\leq \sup_{x\in D} \int_D \left(\frac1{p} \frac1{|x-y|^{p(d-\alpha+\eps)}}
+ \frac1{q} \frac{1}{\delta_D(y)^{q(\beta-\gamma)}} \right) dy
<\infty.$$ This together with – implies that $\lim_{t\to 0} N^{\alpha, \gamma}_{
{\bf 1}_{D\setminus K} g(x)}(t)=0$; that is, ${\bf 1}_{D\setminus K}
g \in {\bf K}_{\alpha, \gamma}$. This completes the proof of the proposition.
\[p:newqw1\] Suppose $\gamma \in [0, \alpha \wedge d)$ and $|F|(z,w) \le A(|z-w|^\beta \wedge 1)$ for some $A>0$ and $\beta>\alpha$. Then there exists $
C_8=
C_8 (\beta, d, \alpha, \gamma)>0$ such that for every arbitrary Borel subset $D$ of $\bR^d$, $$\label{e:works for F}
N^{\alpha, \gamma}_F(t) \,\le\,
C_8\, A t.$$ This in particular implies that $F\in {\bf J}_{\alpha, \gamma}$.
By , we have that $$\begin{aligned}
&\int_0^t\int_{\bR^d \times \bR^d}q(s, y, w )\left(1
+
\frac{ |z-w| \wedge t^{1/\alpha}}{|y-w|} \right)^\gamma
\frac{|F|(z,w) + |F|(w,z)}{|z-w|^{d+\alpha}} dzdwds\\
\le &2 A\left(\int_{\bR^d} (|z|^\beta \wedge 1) |z|^{-d-\alpha} dz \right)
\int_0^t\int_{\bR^d}q(s, y, w )\left(1
+
\frac{ |z-w|
\wedge t^{1/\alpha}}{|y-w|} \right)^\gamma dwds\\
\le & c_1A \left(\int_{B(0, 1)} \frac{dz}{|z|^{d+\alpha-\beta}} +
\int_{B(0, 1)^c} \frac{dz}{|z|^{d+\alpha}} \right) \int_0^t
\left(1+ \int_{D} 2^{d+\alpha} \frac{st^{\gamma/\alpha }}
{(s^{1/\alpha}+|y-w|)^{d+\alpha} |y-w|^\gamma} \right) dwds\\
\le & c_2A \int_0^t \left(1+ t^{\gamma /\alpha} s
\int^{\infty}_{0} \frac{r^{d-1} }{r^{\gamma} (s^{1/\alpha}+r)^{d+\alpha}} dr \right) ds\\
\le & c_2A t +c_3A
\left(\int^\infty_{0} \frac{u^{d-1-\gamma} }{
(1+u)^{d+\alpha}} du \right) t^{\gamma /\alpha} \int_0^t s^{-\gamma
/\alpha} ds \le c_4A t\end{aligned}$$ where the assumption $\gamma \in [0, \alpha \wedge d)$ is used the last inequality. This establishes .
Stable-like processes on closed $d$-sets {#S:4.1}
----------------------------------------
A Borel subset $D$ in $\bR^d$ with $d\ge 1$ is said to be a $d$-set if there exist constants $r_0>0$, $
\sC_2>
\sC_1>0$ so that $$\label{eqn:dset}
\sC_1\, r^d\le | B(x,r) \cap
D
|\le
\sC_2 \, r^d \qquad \hbox{for all }~x\in
D \hbox{ and } 0<r\leq r_0,$$ where for a Borel set $A\subset \bR^d$, we use $|A|$ to denote its Lebesgue measure. The notion of a $d$-set arises both in the theory of function spaces and in fractal geometry. It is known that if $
D$ is a $d$-set, then so is its Euclidean closure $\overline
D$. Every uniformly Lipschitz open set in $\bR^d$ is a $d$-set, so is its Euclidean closure. It is easy to check that the classical von Koch snowflake domain in $\bR^2$ is an open $2$-set. A $d$-set can have very rough boundary since every $d$-set with a subset of zero Lebesgue measure removed is still a $d$-set.
Suppose that $D$ is a closed $d$-set $D\subset \bR^d$ and $c(x, y)$ is a symmetric function on $D \times D$ that is bounded between two strictly positive constants $
\sC_4>
\sC_3>0$, that is, $$\label{eqn:1.4}
\sC_3 \leq c(x, y) \leq
\sC_4 \qquad \hbox{for a.e. } x, y \in D.$$ For $\alpha\in (0, 2)$, we define $$\begin{aligned}
\sF &=& \left\{ u \in L^2(D;
dx): \, \int_{D\times D}
\frac{(u(x)-u(y))^2}
{|x-y|^{d+\alpha}} \, dx \, dy < \infty \right\}
\label{e:DF1} \\
\sE(u,v)&=& \frac 12 \int_{D\times D} (u(x)-u(y))(v(x)-v(y)) \frac{
c(x, y) } {|x-y|^{d+\alpha}}\, dx \, dy, \quad u, v \in \sF.
\label{e:DF2}\end{aligned}$$ It is easy to check that $(\sE, \sF)$ is a regular Dirichlet form on $L^2(D, dx)$ and therefore there is an associated symmetric Hunt process $X$ on $D$ starting from every point in $D$ except for an exceptional set that has zero capacity. The process $X$ is called a symmetric $\alpha$-stable-like process on $D$ in [@CK]. When $c(x, y)$ is a constant function, $X$ is the reflected $\alpha$-stable process appeared in [@BBC]. Note that when $D=
\bR^d$ and $c(x, y)$ is a constant function, then $X$ is nothing but a symmetric $\alpha$-stable process on $\bR^d$.
It follows as a special case from [@CK Theorem 1.1] that the symmetric stable-like process $X$ on a closed $d$-set in $\bR^d$ has a Hölder continuous transition density function $p (t, x, y)$ with respect to the Lebesgue measure on $D$ that satisfies the estimate with $\gamma=0$ and the comparison constant $C_0$ depending only on $d$, $\alpha$, $r_0$ and the constants $
\sC_k$, $k=1, \cdots, 4$ in and . In particular, this implies that the process $X$ can be refined so it can start from every point in $D$. Thus as a special case of Theorem \[t:main\], we have the following.
\[T:4.3\] Suppose that $X$ is a symmetric $\alpha$-stable-like process on a closed $d$-set $D$ in $\bR^d$. Assume $\mu\in {\bf K}_{\alpha,0}$ and $F\in {\bf J}_{\alpha, 0}$. Let $q $ be the density of the Feynman-Kac semigroup of $X$ corresponding to $A^{\mu, F}$. For any $T>0$, there exists a constant $
C_{9}>1$ such that for all $(t, x, y)\in (0, T]\times D\times D$, $$C_{9}^{-1} q(t,x,y)\le q (t, x, y) \le
C_{9} q(t,x,y).$$
\[R:4.4\] Let $n\geq 1$ be an integer and $d\in (0, n]$. In general, a Borel subset $D$ in $\bR^n$ is said to be a $d$-set if there exist a measure $\mu$ and constants $r_0>0$, $ \sC_2>
\sC_1>0$ so that $$\label{eqn:generaldset}
\sC_1\, r^d\le \mu ( B(x,r) \cap D)
\le \sC_2 \, r^d \qquad \hbox{for all }~x\in
D \hbox{ and } 0<r\leq r_0,$$ It is established in [@CK] that for every $\alpha \in (0, 2)$, a symmetric $\alpha$-stable-like process $X$ can always be constructed on any closed $d$-set $D$ in $\bR^n$ via the Dirichlet form $(\sE, \sF)$ on $L^2(D; \mu)$ defined by – but with the $d$-measure $\mu (dx)$ in place of the Lebesgue measure $dx$ there. Moreover by [@CK Theorem 1.1], the process $X$ has a jointly Hölder continuous transition density function $p (t, x, y)$ with respect to the $d$-measure $\mu$ on $D$ that satisfies the estimate with $\gamma=0$. The proof of Theorem \[t:main\] also works for such process $X$; in other words, Theorem \[T:4.3\] continues to hold for such kind of symmetric stable-like processes.
Killed symmetric $\alpha$-stable processes {#ss:stable}
------------------------------------------
A symmetric $\alpha$-stable process $X$ in $\bR^d$ is a Lévy process whose characteristic function is given by $ \E_0 \left[ \exp(i \xi \cdot X_t ) \right]
=e^{-t |\xi|^\alpha} .$ It is well-known that the process $X$ has a Lévy intensity function $J(x, y)={\cal A}(d, -\alpha)|x-y|^{-(d+\alpha)},$ where $$\label{e:cal a}
{\cal A}(d, -\alpha)= \alpha 2^{-1+\alpha} \Gamma(\frac{d+\alpha}2)
\pi^{-d/2}(\Gamma(1-\frac{\alpha}2))^{-1}.$$ Here $\Gamma$ is the Gamma function defined by $\Gamma(\lambda):=
\int^{\infty}_0 t^{\lambda-1} e^{-t}dt$ for every $\lambda > 0$. Let $X^D$ be the killed symmetric $\alpha$-stable process $X^D$ in a $C^{1,1}$ open set $D$. It follows from [@CKS] that $X^D$ satisfies the assumption of Section \[s:1\] with $\gamma=\alpha/2$. Thus as a special case of Theorem \[t:main\], we have the following.
\[t:ks\] Suppose that $X$ is a killed symmetric $\alpha$-stable process in a $C^{1,1}$ open set $D$. Assume $\mu\in {\bf K}_{\alpha,
\alpha/2}$ and $F\in {\bf J}_{\alpha, \alpha/2}$. Let $q_D$ be the density of the Feynman-Kac semigroup of $X$ corresponding to $A^{\mu,
F}$. For any $T>0$, there exists a constant $
C_{10}>1$ such that for all $(t, x, y)\in (0, T]\times D\times D$, $$C_{10}^{-1} \psi_{\alpha/2}(t, x, y)q(t,x,y)\le q_D(t, x, y) \le
C_{10}
\psi_{\alpha/2}(t, x, y)q(t,x,y).$$
Let $X^{m}$ be a relativistic $\alpha$-stable process in $\bR^d$ with mass $m>0$, i.e., $X^{m}$ is a Lévy process in $\bR^d$ with $$\E_0\left[\exp (i
\xi \cdot X^{m}_t)
\right]=\exp \left(t\left(
m^\alpha-(|\xi|^2+m^{2})^{\alpha/2} \right) \right).$$ $X^{m}$ has a Lévy intensity function $J^m(x,y)= {\cal A}(d, -\alpha) \varphi (m^{1/\alpha}|x-y|)|x-y|^{-d-\alpha}$ where $$\label{e:4.6}
\varphi (r):= 2^{-(d+\alpha)} \, \Gamma \left(
\frac{d+\alpha}{2} \right)^{-1}\, \int_0^\infty s^{\frac{d+\alpha}{
2}-1} e^{-\frac{s}{ 4} -\frac{r^2}{ s} } \, ds,$$ which is decreasing and is a smooth function of $r^2$ satisfying $\varphi(0) = 1$ and $$\label{e:4.2}
\varphi (r) \asymp e^{-r}(1+r^{(d+\alpha-1)/2} ) \qquad \hbox{on } [0,
\infty)$$ (see [@CS03b pp. 276-277] for details).
Let $X^{m, D}$ be a killed relativistic $\alpha$-stable process in a bounded $C^{1,1}$ open set. Define $$\begin{aligned}
K^m_t := \exp \left(\sum_{0<s \le t} \ln(\varphi(m^{1/\alpha}
|X^D_s-X^D_{s-}|))
+m\,(t\wedge\tau_D)\right).\end{aligned}$$ Since $\int_{\bR^d} J(x,y)-J^m(x,y)dy=m$ for all $x\in \bR^d$ (see [@Ry]), it follows from [@CS03b p.279] that $X^{m, D}$ can be obtained from the killed symmetric $\alpha$-stable process $X^D$ in $D$ through the non-local Feynman-Kac transform $K^m_t$. That is, $\E_x \big[ f(X^{m, D}_t) \big]
:= \E_x \left[ K^m_t f(X^D_t) \right]$ By , for any $M>0$, there exists a constant $c=c(d, \alpha, M,
\text{diam} (D))>0$ such that for all $m\in (0, M]$, $|\ln(\varphi(m^{1/\alpha} |x-y|))| \leq c (|x-y|^2 \wedge 1) $ and so, by Proposition \[p:newqw1\], $ F_m(x, y) :=\ln(\varphi(m^{1/\alpha} |x-y|)) \in {\bf J}_{\alpha, \alpha/2}$. The constant function $m$ is in ${\bf K}_{\alpha,
\alpha/2}$ and so $N^{\alpha, \alpha/2}_{ m, F_m}(t)$ goes to zero as $t$ goes to zero uniformly on $m \in (0,
M]$. Thus, as an application of Theorem \[t:main\], we arrive at the following result, which is the bounded open set case of a more general result recently obtained in [@CKS2] by a different method.
\[t:krs\] Suppose that $D$ is a bounded $C^{1, 1}$ open set in $\bR^d$. For any $m>0$, let $p^m_D$ be the transition density of the killed relativistic $\alpha$-stable process with weight $m$ in $D$. For any $
M>0$ and $T>0$, there exists a constant $
C_{11}>1$ such that for all $m\in (0,
M]$ and $(t, x, y)\in (0,
T]\times D\times D$, $$C_{11}^{-1} \psi_{\alpha/2}(t, x, y)q(t,x,y)\le p^m_D(t, x, y) \le
C_{11}
\psi_{\alpha/2}(t, x, y)q(t,x,y).$$
Censored stable processes
-------------------------
Fix an open set $D$ in $\bR^d$ with $d\geq 1$. Recall that ${\cal A} (d, -\alpha)$ is the constant defined in . Define a bilinear form $\sE $ on $C_c^\infty(D)$ by \[e:csdf\] (u, v):= 1[2]{} \_D \_D (u(x)-u(y))(v(x)-v(y)) dxdy , u, v C\_c\^(D). Using Fatou’s lemma, it is easy to check that the bilinear form $ (
\sE , C^\infty_c (D))$ is closable in $L^2 (D, dx)$. Let $\sF$ be the closure of $C^\infty_c(D)$ under the Hilbert inner product $\sE _1:=\sE +(\,\cdot\, ,\, \cdot\, )_{L^2(D, dx)}.$ As noted in [@BBC], $(\sE , \sF )$ is Markovian and hence a regular symmetric Dirichlet form on $L^2(D, dx)$, and therefore there is an associated symmetric Hunt process $Y=\{Y_t, t\ge 0, \P_x, x\in D\}$ taking values in $D$ (cf. Theorem 3.1.1 of [@FOT]). The process $Y$ is the censored $\alpha$-stable process in $D$ that is studied in [@BBC]. By , the jumping kernel $J(x, y)$ of the censored $\alpha$-stable process $Y$ is given by $$J(x, y)=\frac{
{\cal A}
(d, -\alpha)}{|x-y|^{d+\alpha}} \qquad \hbox{for } x, y \in
D.$$
As a particular case of a more general result established in [@CKS1 Theorem 1.1], when $\alpha \in (1, 2)$ and $D$ is a $C^{1, 1}$ open subset of $\bR^d$, the censored $\alpha$-stable process on $D$ satisfies the assumption of Section \[s:1\] with $\gamma=\alpha-1$. Thus as a special case of Theorem \[t:main\], we have the following:
\[t:cs\] Suppose that $\alpha \in (1, 2)$ and that $Y$ is a censored stable process in a $C^{1,1}$ open set $D$. Assume $\mu\in {\bf
K}_{\alpha, \alpha-1}$ and $F\in {\bf J}_{\alpha, \alpha-1}$. Let $q_D$ be the density of the Feynman-Kac semigroup of $Y$ corresponding to $A^{\mu, F}$. For any $T>0$, there exists a constant $C_{12}>1$ such that for all $(t, x, y)\in (0, T]\times D\times D$, $$C_{12}^{-1} \psi_{\alpha-1}(t, x, y)q(t,x,y)\le q_D(t, x, y) \le
C_{12}
\psi_{\alpha-1}(t, x, y)q(t,x,y).$$
Similar to [@BBC], we can define a censored relativistic $\alpha$-stable process in $D$. Alternatively, with $$K_t := \exp \left(\sum_{0<s \le t} \ln(\varphi(m^{1/\alpha} (|Y_{s-}- Y_s|))+{\cal A} (d, -\alpha)
\int^t_0 \int_D\frac{1-\varphi(m^{1/\alpha} |Y_s-y|)}{|Y_s-y|^{\alpha+d}} dy ds\right),$$ if $D$ is a bounded $C^{1,1}$ open set, a censored relativistic stable process $Y^m$ can also be obtained from the censored stable process $Y$ through the Feynman-Kac transform $K_t$. That is, $ \E_x [ f(Y^m_t)]=\E_x [ K_t f(Y_t) ]$ (see [@CKi; @CS03b]). By an argument similar to that of Subsection \[ss:stable\], one can see that $ F_m:=\ln (\varphi(m^{1/\alpha} |x-y|)) \in {\bf J}_{\alpha, \alpha/2}$. Moreover, since $$g_m(x):=\int_D (1-\varphi(m^{1/\alpha} |x-y|))|x-y|^{-\alpha-d} dy \le \int_{\bR^d} (1-\varphi(m^{1/\alpha} |x-y|))|x-y|^{-\alpha-d} dy =m,$$ $g_m \in {\bf K}_{\alpha,\alpha/2}$ and $N^{\alpha, \alpha/2}_{g_m, F_m}(t)$ goes to zero as $t$ goes to zero uniformly on $m \in (0, M]$. Thus as a particular case of Theorem \[t:cs\], we have the following.
\[t:crs\] Suppose that $\alpha \in (1, 2)$ and that $D$ is a bounded $C^{1, 1}$ open set in $\bR^d$. For any $m>0$, let $q^m_D$ be the transition density of the censored relativistic $\alpha$-stable process with weight $m$ in $D$. For any $
M>0$ and $T>0$, there exists a constant $C_{13}>1$ such that for all $m\in (0,
M]$ and $(t, x, y)\in (0,
T]\times D\times D$, $$C_{13}^{-1} \psi_{\alpha-1}(t, x, y)q(t,x,y)\le q^m_D(t, x, y) \le
C_{13}
\psi_{\alpha-1}(t, x, y)q(t,x,y).$$
In fact, Theorems \[t:cs\] and \[t:crs\] are applicable to certain class of censored stable-like processes whose Dirichlet heat kernel estimates are given in [@CKS1].
Stable processes with drifts
----------------------------
Let $\alpha \in (1, 2)$ and $d\ge 2$. In this subsection, we apply our main result to a non-symmetric process.
For $b=(b_1, \dots, b_d)$ with $b_i \in {\mathbb K}_{d, \alpha-1}$, a Feller process $Z$ on $\bR^d$ with infinitesimal generator $\sL^b:=\Delta^{\alpha/2}+ b(x)\cdot \nabla$ is constructed in [@BJ] through the fundamental solution of $\sL^b$. Let $Z^D$ be the subprocess of $Z$ killed upon leaving $D$. The following result is established in [@CKS4].
If $\alpha \in (1, 2)$, $d\ge 2$ and $D$ is a bounded $C^{1, 1}$ open set, then $Z^D$ has a jointly continuous transition density function $p_D(t, x, y)$ that satisfies with $\gamma = \alpha/2$.
Thus as a special case of Theorem \[t:main\], we also have the following:
\[t:ksnew\] Suppose that $\alpha \in (1, 2)$, $d\ge 2$, that $D$ is a bounded $C^{1, 1}$ open set and that $Z^D$ is the subprocess of $Z$ killed upon leaving $D$. Assume $\mu\in {\bf K}_{\alpha,
\alpha/2}$ and $F\in {\bf J}_{\alpha, \alpha/2}$. Let $q_D$ be the density of the Feynman-Kac semigroup of $Z^D$ corresponding to $A^{\mu,
F}$. For any $T>0$, there exists a constant $
C_{14}>1$ such that for all $(t, x, y)\in (0, T]\times D\times D$, $$C_{14}^{-1} \psi_{\alpha/2}(t, x, y)q(t,x,y)\le q_D(t, x, y) \le
C_{14}
\psi_{\alpha/2}(t, x, y)q(t,x,y).$$
[99]{}
Ph. Blanchard and Z. M. Ma. Semigroup of Schrödinger operators with potentials given by Radon measures. In [*Stochastic processes, physics and geometry*]{}, 160–195, World Sci. Publishing, Teaneck, NJ, 1990.
K. Bogdan, K. Burdzy and Z.-Q. Chen. Censored stable processes. [*Probab. Theory Related Fields*]{} [**127**]{} (2003), 89–152
K. Bogdan and T. Jakubowski. Estimates of heat kernel of fractional Laplacian perturbed by gradient operators. [*Comm. Math. Phys.*]{} [**271(1)**]{} (2007), 179–198.
Z.-Q. Chen, P. J. Fitzsimmons, K. Kuwae and T.-S. Zhang. Perturbation of symmetric Markov processes. [*Probab. Theory Relat. Fields **140***]{} (2008), 239-275.
Z.-Q. Chen, P. J. Fitzsimmons, K. Kuwae and T.-S. Zhang. On general perturbations of symmetric Markov processes. [*J. Math. Pures et Appliquées **92***]{} (2009), 363-374.
Z.-Q. Chen and P. Kim. Stability of Martin boundary under non-local Feynman-Kac perturbations. [*Probab. Theory Related Fields*]{} [**128**]{} (2004), 525–564.
Z.-Q. Chen, P. Kim, and R. Song. Heat kernel estimates for Dirichlet fractional Laplacian. [*J. European Math. Soc.*]{} [**12**]{} (2010), 1307–1329.
Z.-Q. Chen, P. Kim and R. Song. Two-sided heat kernel estimates for censored stable-like processes. [*Probab. Theory Relat. Fields*]{} [**146**]{} (2010), 361–399.
Z.-Q. Chen, P. Kim and R. Song. Sharp heat kernel estimates for relativistic stable processes in open sets. [*Ann. Probab.*]{}, to appear, 2011. arXiv:0908.1509
Z.-Q. Chen, P. Kim and R. Song. Dirichlet heat kernel estimates for fractional Laplacian under gradient perturbation. [*Ann. Probab.*]{}, to appear, 2011. arXiv:1011.3273
Z.-Q. Chen and T. Kumagai. Heat kernel estimates for stable-like processes on $d$-sets. [*Stoch. Proc. Appl.*]{} [**108**]{} (2003), 27–62.
Z.-Q. Chen and R. Song. Drift transforms and Green function estimates for discontinuous processes. [*J. Funct. Anal.*]{} [**201**]{} (2003), 262–281.
K. L. Chung. Doubly-Feller process with multiplicative functional. [*Seminar on Stochastic Processes, 1985*]{}, 63–78. Progr. Probab. Statist., [**12**]{}, Birkhäuser Boston, MA, 1986.
K. L. Chung and Z. Zhao. [*From Brownian Motion to Schrödinger’s Equation*]{}. Springer-Verlag, Berlin, 1995.
C. Doléans-Dade. Quelques applications de la formule de changement de variables pour les semimartingales. [*Z. Wahrsch. **16***]{} (1970), 181-194.
P. J. Fitzsimmons and R. K. Getoor. Smooth measures and continuous additive functionals of right Markov processes. In [*Itô’s Stochastic Calculus and Probability Theory*]{}, 31-49, Springer, Tokyo, 1996.
M. Fukushima, Y. Oshima and M. Takeda. [*Dirichlet Forms and Symmetric Markov Processes*]{}. 2nd Edition. Walter de Gruyter, 2011.
P. Kim and R. Song. Two-sided estimates on the density of Brownian motion with singular drift. [*Illinois J. Math.*]{} [**50(3)**]{} (2006), 635–688.
L. Riahi. Estimates of Green functions and their applications for parabolic operators with singular potentials. [*Colloq. Math.*]{} [**95(2)**]{} (2003), 267–283.
L. Riahi. Dirichlet Green functions for parabolic operators with singular lower-order terms. [*JIPAM. J. Inequal. Pure Appl. Math.*]{} [**8(2)**]{} (2007), Article 36.
M. Ryznar. Estimates of Green function for relativistic $\alpha$-stable process. [*Potential Anal.*]{} **17(1)** (2002), 1–23.
M. Sharpe. [*General Theory of Markov Processes*]{}. Academic Press, Boston, 1988.
B. Simon. Schrödinger semigroups. [*Bull. Amer. Math. Soc. (N.S.) **7***]{} (1982), 447–526.
R. Song. Probabilistic approach to the Dirichlet problem of perturbed stable processes. [*Probab. Theory Relat. Fields*]{} [**95**]{} (1993), 371–389.
R. Song. Feynman-Kac semigroups with discontinuous additive functionals. [*J. Theor. Probab.*]{} [**8**]{} (1995), 727-762.
R. Song. Two-sided estimates on the density of the Feynman-Kac semigroups of stable-like processes. [*Electron. J. Probab.*]{} [**11 (6)**]{} (2006), 146–161.
C. Wang. On estimates of the density of Feynman-Kac semigroups of $\alpha$-stable-like processes. [*J. Math. Anal. Appl.*]{} [**348**]{} (2008), 938–970.
Qi S. Zhang. Gaussian bounds for the fundamental solutions of $\nabla (A\nabla u)+B\nabla u-u\sb t=0$. [*Manuscripta Math.*]{} [**93**]{} (1997), 381–390.
Z. Zhao. A probabilistic principle and generalized Schrödinger perturbation. [*J. Funct. Anal.*]{} [**101**]{} (1991), 162–176.
0.3truein
[**Zhen-Qing Chen**]{}
Department of Mathematics, University of Washington, Seattle, WA 98195, USA
E-mail: `[email protected]`
[**Panki Kim**]{}
Department of Mathematics, Seoul National University, Building 27, 1 Gwanak-ro, Gwanak-gu, Seoul 151-747, Republic of Korea
E-mail: `[email protected]`
[**Renming Song**]{}
Department of Mathematics, University of Illinois, Urbana, IL 61801, USA
E-mail: `[email protected]`
[^1]: Research partially supported by NSF Grants DMS-0906743 and DMR-1035196.
[^2]: This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(0409-20110087).
[^3]: Research supported in part by a grant from the Simons Foundation (208236).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Geometrical frustration arised in spin ices leads to fascinating emergent physical properties. Nowadays there is a wide diversity of the artificial structures, mimicking spin ice at the nanoscale and demonstrating some new effects. Most of the nanoscaled spin ices are two dimensional. Ferromagnetic inverse opal-like structures (IOLS) are among inspiring examples of the three-dimensional system exhibiting spin ice behaviour. However detailed examination of its properties is not straightforward. Experimental technique which is able to unambiguously recover magnetization distribution in 3D mesoscaled structures is lacking. In this work we used an approach based on complementary exploiting of small-angle neutron diffraction technique and micromagnetic simulations. External magnetic field was applied along three main directions of the IOLS mesostructure. Comparison of the calculated and measured data allowed us to determine IOLS magnetic state. The results are in good agreement with the spin ice model. Moreover influence of the demagnetizing field and vortex states on the magnetizing process were revealed. Additionally, we speculate that this approach can be also applied to other 3D magnetic mesostructures.'
author:
- 'A.A. Mistonov$^{1}$, I.S. Dubitskiy$^{2}$, I.S. Shishkin$^{2}$, N.A. Grigoryeva$^1$, A. Heinemann$^3$, N.A. Sapoletova$^{4}$, G.A. Valkovskiy$^{1}$, S.V. Grigoriev$^{1,2}$'
bibliography:
- 'biblio.bib'
title: 'Magnetic structure of the promising candidate for three-dimensional artificial spin ice: small angle neutron diffraction and micromagnetic simulations'
---
Introduction {#sec:introduction}
============
Spatially ordered ferromagnetic nanosystems attract significant attention, since they can mimic natural objects (e.g. pyrochlore spin ices), but at the same time bring some new physics. 3D nanostructures are of considerable interest because of possible qualitatively new properties [@fernandez2017three]. The promising example is inverse opal-like structure (IOLS) based on ferromagnetic metal [@grigoriev2009structural; @grigoryeva2011magnetic]. It can be considered as the three-dimensional continuous network, consisting of cubical and tetrahedral shaped nanoscaled elements with concave faces, connected by long narrow links (“legs”) and arranged into face-centered cubic (fcc) lattice. Non-trivial geometry of the IOLS causes a complicated distribution of the magnetization, which cannot be unravelled by conventional techniques. It was shown in [@mistonov2013three; @dubitskiy2015spin], that magnetization distribution in IOLS is almost fully determined by the direction of the magnetic moments in “legs” which are arranged along four $\langle 111\rangle$-type axes of the fcc IOLS mesostructure. A strong shape anisotropy of the “legs” leads to an uniform distribution of the magnetization inside them. In other words the “legs” appear to be of Ising-type in a wide range of external fields. The magnetic structure of the IOLS is mostly governed by two “forces”: external magnetic field **H** and the ice rule. In our case, the ice rule states, that the number of “legs” with the magnetic moments pointing inward and outward each nanoobject (cube or tetrahedron) should be equal.
Therefore IOLS can be considered to belong to the family of nanostructured artificial spin ices (ASI) [@nisoli2013colloquium; @wang2006artificial]. Initially ASI were designed to mimic atomic spin ice systems firstly discovered in pyrochlore lattices [@bramwell2001spin; @den2000dipolar]. Various realizations of ASI demonstrate emergent physical properties that can be even richer than those of natural materials [@nisoli2018topology]. However, two-dimensional ASI with coplanar arrangement of magnetic moments differ from the conventional atomic spin ice systems. On the contrary, the inverse opals are three-dimensional and can be considered as a more direct mesoscale analogue of pyrochlore spin ice. Additionally, in recent years [@lee2007iron; @hsueh2011nanoporous; @Eslami2014; @mamica2016magnetoferritin; @Utke2008] the number of three-dimensional nano- and mesoscaled systems increases drastically [@fernandez2017three] owing to development of the synthesis techniques. This underlines the importance of determination of the three-dimensional magnetic structure. Nevertheless experimental investigation of the magnetization distribution inside such systems is still challenging.
It is possible to determine three-dimensional magnetic structure at atomic scale (e.g. by neutron diffraction) as well as at microscale (e.g. using tomography), while for nano and mesoscale (hundreds of nanometers) this is non-trivial. Conventional methods are limited to simple forms of samples [@Wernsdorfer1996nucl; @Biziere2013imaging], small sample sizes [@yu2010real; @Zhao03052016] or modest spatial resolution [@manke2010three]. In order to understand fine features of the magnetic structure we suggest to exploit simultaneously small-angle neutron diffraction and micromagnetic simulations.
Small-angle neutron scattering (SANS) is a powerful technique for the mesoscopic magnetic structure studies [@fitzsimmons2014neutron; @grutter2017complex; @grigoriev2010nanostructures; @gunther2014magnetic]. In principle it allows one to overcome the above mentioned limitations. However, in most cases the interpretation of the SANS data is not simple and can be inconclusive. Thus, other techniques should be involved, which can provide either two-dimensional surface (magnetic force microscopy (MFM) [@hartmann1999magnetic], Lorenz microscopy [@de20012]) or integral (superconducting quantum interference device (SQUID)-magnetometry) information about the sample. Another approach is to use micromagnetic calculations, which are able to replace or to complement actual measurements [@bertotti1998hysteresis].
In order to interpret SANS data unambiguously one has to calculate Fourier transform of the magnetization distribution in the sample under study [@gunther2014magnetic; @ott2013numerical]. In most cases the magnetization distribution is assumed to be uniform [@michels2014magnetic]. This assumption is reasonable for small or saturated nanoparticles however it is not acceptable for more complex systems. In the latter case one has to determine magnetization distribution in the system in order to find further the Fourier transform of this distribution. In doing so, the most rigorous and straightforward way is to solve micromagnetic equations [@gilbert1955lagrangian; @bertotti1998hysteresis; @aharoni2000introduction]. The apparent advantage of the micromagnetic calculations is that no adjustable parameters have to be used. The magnetization distribution depends on well known parameters of the material and its geometric arrangement. We used the scheme described above, in order to calculate the Fourier transform of the magnetization distribution in the IOLS mesostructure unit cell and its evolution during the magnetizing process.
In this paper we suggest to consider three scenarios of the interplay between the external magnetic field and the ice rule and reveal magnetic structure using SANS and micromagnetic simulations in pair. This combination was successfully applied to two-phase bulk ferromagnets [@michels2014magnetic; @michels2014micromagnetic; @erokhin2012micromagnetic]. Additionally, simulation results were recently reported for the single nanowire [@vivas2017small].
We assume, that the proposed approach can be widely used for the investigation of magnetic nanostructures and mesoscopic three-dimensional periodical systems like segmented nanowires [@Sergelius2017], nanosprings [@Nam2017], gyroids [@hsueh2011nanoporous], mesocrystals [@Sturm2016].
The paper is organized in the following way. Section \[sec:samples\] gives the essence of the sample preparation procedure. Section \[sec:SANS\] describes the small-angle neutron scattering experiment and some details of data treatment. Section \[sec:MM\_calculations\] gives the description of the micromagnetic simulation procedure. In Section \[sec:results\] the results of the SANS experiment and micromagnetic calculations are presented along with discussion. Section \[sec:conclusion\] gives concluding remarks.
Samples {#sec:samples}
=======
The inverse opal-like Co structure was prepared by electrodeposition of cobalt inside the voids of the artificial opal. The artificial opal template was synthesized by electric-field-assisted vertical deposition of polystyrene microspheres (D = 620 nm; RSD $<$ 10%) [@napolskii2010] on Si (100) wafer PVD-coated with a 200 nm thick gold layer. The cobalt electrocrystallization into the opal voids was carried out in three-electrode cell [@sapoletova2010controlled] from the electrolyte containing 0.2M CoSO$_4$ + 0.5M Na$_2$SO$_4$ + 0.3M H$_3$BO$_3$ + 3.5M C$_2$H$_5$OH at deposition potential -0.8 V versus Ag/AgCl reference electrode at room temperature. The area of the sample was 1 cm$^2$. In order to obtain the free-standing metallic structure on the support, the polystyrene microspheres were dissolved in toluene for several hours.
Scanning electron microscopy (SEM) images of the Co IOLS are given in Fig. \[ris:Co\_SEM\]. One can see, that the surface represents a hexagonal closepacked layer of spherical voids with the lattice constant of about 600 nm. An average lateral size of the structural domain exceeds 100 $\mu$m. Desorientation of different domains is less, than 3$^{\circ}$. According to the SEM data thickness of the Co IOLS is 12 $\mu$m. With the help of the SEM images one can estimate the degree of polystyrene microspheres deformation (sintering) which determines “legs” aspect ratio [@dubitskiy2015spin; @dubitskiy2017dependence]. These quantity is defined as $r/r'-1$ expressed in percents, where $r$ — microsphere radius, $r'$ — half of the distance between adjacent microspheres centers. It is equal to 2% for the considered sample.
Spatial ordering of the sample on the mesoscale was investigated by microradian x-ray diffraction technique. Details of data acquisition and treatment can be found elsewhere [@grigoriev2009structural; @grigoryeva2011magnetic; @chumakova2014periodic; @dubitskiy2017study]. It was shown, that the sample possesses fcc symmetry with lattice constant $a_{0}$ = 845$\pm$10 nm.
The method of sample fabrication allows one to synthesize structures with orientation of the \[111\] crystallographic axis of the IOLS fcc mesostructure perpendicular to the substrate (here and further we deal only with crystallographic axes related to IOLS mesostructure, not to the cobalt atomic structure). The direction of drying of the aqueous suspension of polystyrene microspheres (or meniscus moving) corresponds to the \[20$\bar{2}$\] axis. Thus, the orientation of the opal-like crystal is already established at the stage of synthesis. The main axes those determine the sample orientation are shown in Fig. \[ris:Co\_SEM\].
The appearance of the structural elements of the IOLS can be easily understood. We described them elsewhere [@mistonov2013three; @mistonov2015ice; @dubitskiy2015spin; @dubitskiy2017dependence] but outline briefly here. The tetrahedral and octahedral voids of the fcc structure transform into two types of nanoobjects during the inversion process. We denote them quasitetrahedra and quasicubes, but further “quasi” is omitted. In Fig. \[ris:Basic\_element\] the unit cell (a) and the primitive cell (b) of the IOLS are presented. The primitive cell consists of three parts: a tetrahedron, a cube and another tetrahedron. They are connected to each other by “legs” along one of four $\langle 111 \rangle$-type axes. The surfaces of the cube and tetrahedra are concave resembling the voids between the spheres. For IOLS with the period of 845 nm one can estimate that the cube has edge of about 240 nm, while tetrahedra have edges of about 165 nm . The connecting “legs” have a length of about 175 nm.
Methods and data treatment {#sec:methods}
==========================
Small-angle neutron diffraction {#sec:SANS}
-------------------------------
SANS measurements were carried out with the instrument SANS-1 (Heinz Maier Leibnitz Zentrum, Garching, Germany). Neutron beam with a mean wavelength of $\lambda$=1.7 nm and a wavelength spread of $\Delta \lambda/\lambda$ = 10% was used. The Sample - Detector distance of 21.719 m was chosen with appropriate collimation to cover scattering vectors q from 0.005 to 0.06 nm$^{-1}$. Due to the large wavelength, the intensity of the incident beam was low enough to work without a beamstop. Thus the transmitted beam as well was recorded by the detector. Since the system ordered at the mesoscale small-angle scattering was actually small-angle diffraction. An external magnetic field up to 30 kOe was applied along the \[111\], $[\bar{1}2\bar{1}]$ and \[100\] crystallographic axes of the fcc mesostructure. In first two cases, the sample was irradiated by the beam along \[111\] axis, while in latter — along \[100\] axis. All measurements were performed at room temperature. Geometrical schemes used in the experiment as well as recorded diffraction patterns are shown in Fig. \[ris:exp\_geometry\].
The intensity of non-polarized neutrons scattering consists of the following contributions: $$I_{tot}=I_N+I_M,
\label{eq:total_section}$$ $$I_N(\textbf{q})\propto{S_n(\textbf{q})|A_nF_n(\textbf{q})|^2},
\label{eq:nuclear_section}$$ $$I_M(\textbf{q})\propto{S_m(\textbf{q})|A_m|^2|\widetilde{\textbf{M}}_{\perp}(\textbf{q})|^2},
\label{eq:magnetic_section}$$
where $I_N$ and $I_M$ are intensities of the nuclear and magnetic scattering respectively, **q** — scattering vector, S$_n$, S$_m$ are nuclear and magnetic structure factors, $A_n$ and $A_m$ are amplitudes of nuclear and magnetic scattering, respectively, F$_n$ — nuclear form-factor and $\widetilde{\textbf{M}}_{\perp}(\textbf{q})$ – the magnetic form-factor component perpendicular to $\textbf{q}$. For the IOLS with the voids of radius R the F$_n$ is equal to that of the sphere of radius R [@FeiginSvergun], S$_n$ is the structure factor of the fcc lattice with the lattice constant a$_0$. $\widetilde{\textbf{M}}(\textbf{q})$ is the Fourier transform of the magnetization distribution in the IOLS unit cell and S$_m$ is determined by the correlation of the magnetization distributions between unit cells.
Both nuclear and magnetic scattering intensities contain the contribution from periodical structure (Bragg peaks) as well as from disordered elements (diffuse scattering). Since we are interested in magnetic ordering in the IOLS unit cell we should exclude diffuse scattering which is caused by magnetic domains.
The diffraction patterns presented in Fig. \[ris:exp\_geometry\] were taken from fully magnetized samples therefore the intensity of the peaks contains both nuclear and magnetic contributions. The data treatment procedure is described elsewhere [@grigoriev2009structural; @grigoryeva2011magnetic], here we summarize it shortly. In order to extract the magnetic contribution caused by periodical magnetic structure one has to subtract nuclear contribution as well as diffuse one. At first we integrated diffraction pattern azimuthally near the reflex in order to obtain I$_{tot}$(q)-dependence. Then we approximated this dependence by the sum of diffuse (squared Lorenz function) and Bragg (Gauss function) contributions and determined the integral intensity of the latter. Finally we subtracted the value, obtained for demagnetized sample (at the external magnetic field equalled to the coercive force H$_{\text{C}}$), containing mostly nuclear contribution, since magnetic scattering caused by disordered domains has been subtracted in the previous step. Thus, we obtained field dependencies of the Bragg peaks intensities.
It is well known that diffraction patterns can be obtained in the experiment only if neutron coherence length is much larger than the period of the structure [@grigor2007two; @petukhov2002high; @grigoriev2010nanostructures]. Transverse coherence length of the neutron beam $l_{\text{tr}}$ determines minimal peak width. This length depends on angular size of the neutron source [@benfield2004structure]. In our case it is equal to 3 $\mu$m . Longitudinal coherence length of the neutron beam $l_{\text{long}}$ can be found from the spectral width of the source $\Delta\lambda$ [@born1980principles]. For the experimental setup used in present work $l_{\text{long}}$ equals to 1 mm. Therefore total coherence volume can be estimated as $l_{\text{tr}}^2l_{\text{long}}$ = 9000 $\mu$m$^3$. It covers for about 25 IOLS unit cells. At the same time beam spot size is about 1 cm$^2$. As a result the diffraction patterns consist of the incoherent sum of the waves scattered by different parts of the sample. It limits the analysis of the structure factor contribution in terms of peaks shape and peaks width. Another option is to consider the integral peak intensity. In this case one has to take into account the magnetic form factor contribution to the peak intensity. This contribution appears to be the most significant and substantially varies with variation of the external field.
Micromagnetic calculations {#sec:MM_calculations}
--------------------------
Micromagetic simulations were carried out by means of Nmag [@fischbacher2007systematic; @fischbacher2009parallel] modelling package provided by University of Southampton. Nmag implements time integration of Landau-Lifshitz-Gilbert equation using “Finite Element Method/Boundary Element Method” [@fredkin1990hybrid]. Boundary element matrix was approximated by hierarchical matrices [@hackbusch1999sparse]. We used the finite element discretization scheme since IOLS unit cell contains a lot of curved facets.
We simulated magnetization distribution in the inverse opal unit cell under the field applied along \[111\], \[100\] and $[\bar{1}2\bar{1}]$ crystallographic directions of IOLS mesostructure (Fig. \[ris:Basic\_element\]) by using the same approach, as in Ref. [@dubitskiy2015spin]. The degree of sintering was chosen to be 2% according to the experimental data.
We used the following material parameters of bulk Co: the exchange coupling constant $A=3\cdot10^{-11}$ J/m and the saturation magnetization $M_S$ = 1.4$\times$10$^6$ A/m [@abo2013definition; @han2009influence]. Exchange, demagnetization and Zeeman energies were taken into account.
Wide-angle powder x-ray diffraction experiments showed that Co-based IOLS consist of polycrystalline hcp cobalt with the grain size less than 30 nm [@grigoryeva2011magnetic]. No distinct texture was found. The grains play a significant role in the magnetic behaviour of bulk polycrystalline ferromagnets [@michels2014micromagnetic; @sepehri2013high]. Nevertheless, the magnetic state of nanostructures fabricated from polycrystalline cobalt is mostly determined by the shape anisotropy and magnetocrystalline term is often neglected [@castan2014magnetic; @rodriguez2014high; @phatak2014visualization; @fernandez2009magnetization; @pathak2014experimental]. It should be noted that hysteresis loops measured for polycrystalline fcc Ni- and hcp Co-based IOLS with the same structural period do not exhibit any qualitative difference [@dubitskiy2015spin] although magnetocrystalline anisotropy of fcc nickel phase is significantly smaller than that of hcp cobalt. Moreover, numerical simulation for Ni- and Co-based IOLS shows a qualitative agreement with the experiment in spite of neglecting the magnetocrystalline anisotropy contribution [@dubitskiy2015spin]. It implies that magnetocrystalline anisotropy is not the major energy term which determines the inverse opal magnetic structure. Furthermore taking polycrystallinity into account gives rise to adjustable parameters (grain shape, size, etc.) which cannot be easily obtained from the experiment but can probably affect numerical simulations [@erokhin2017optimization]. In order to not introduce adjustable parameters in our model and restrict ourselves to IOLS geometrical characteristics measured in the experiment, we neglected magnetocrystalline anisotropy.
The linear size of the finite element (tetrahedron) was chosen to be smaller than the exchange length $l_{\text{ex}}=\sqrt{2A/{\mu_0M^2_S}}=4.9$ nm. In our model open boundary conditions (OBC) were applied instead of periodic boundary conditions (PBC), since PBC is not expected to affect significantly the results of the simulation due to the large size of the unit cell and its high porosity, while require very large computational resources. It is necessary to take into account demagnetizing field related to the sample shape (thin film). If the field is applied perpendicular to the sample surface (along \[111\] direction) then the “effective film” model introduced in [@dubitskiy2015spin] is sufficient to obtain quantitative agreement with the experiment. However this model is not applicable to two other cases of the magnetic field directions. So we used the so-called macrogeometry approach [@fangohr2009new] instead. We considered the array of 18 copies of IOLS unit cell arranged in (111) plane in order to simulate thin film shape. It was found out that the calculation results do not alter with increasing the number of copies. Full 50-points hysteresis curve simulation takes about month and a half on the eight node SMP machine. The typical value of RAM used is 60 GB.
In contrast to [@dubitskiy2015spin], where total magnetization was compared with the experimental results of SQUID-magnetometry, here we extracted the magnetic form factor $\widetilde{\textbf{M}}(\textbf{q})$ of the IOLS unit cell. It involved numerous calculations of the Fourier transform of the magnetization distribution $\textbf{M}(\textbf{r})$. Namely, we calculated the Fourier transform of $\textbf{M}(\textbf{r})$ for all values of the external magnetic fields used in the SANS experiment:
$$\widetilde{\textbf{M}}(\textbf{q})=\int\limits_{\text{unit~cell}}\textbf{M}(\textbf{r})e^{i\textbf{qr}}d\textbf{r}$$
It should be noted, that integration is taken over the whole unit cell, which is of submicron size and contains four primitive cells; thus, the obtained results contain information about the different primitive cells.
Due to local defects of the IOLS mesostructure, the “legs” reversal fields and hence the magnetic form factor values may have some dispersion [@shen2012dynamics]. At a first approximation the calculated magnetic form factor of the ideal structure can be considered as the average value of the real one.
The form factor component $\widetilde{\textbf{M}}_{\perp}(\textbf{q})$ which is perpendicular to $\widehat{\textbf{q}}=\textbf{q}/q$ was calculated for the field values up to 30 kOe according to [@squires2012introduction]:
$$|\widetilde{\textbf{M}}_{\perp}(\textbf{q})|^2=|\widehat{\textbf{q}}\times \widetilde{\textbf{M}}(\textbf{q}) \times \widehat{\textbf{q}}|^2$$
It provided the field dependence of the $\widetilde{\textbf{M}}_{\perp}(\textbf{q})$.
Results and discussion {#sec:results}
======================
In this section we interpret the results of the neutron diffraction experiment by assigning magnetic states, obtained in micromagnetic simulations to the particular points of the experimental field dependencies. The states are shown in the insets in Figs. \[ris:H\_111\]-\[ris:H\_100\] and corresponding points are marked by the same letters.
Firstly it is noteworthy to compare the two-dimensional experimental SANS patterns and the simulation results (Fig. \[ris:exp\_geometry\]). One can observe the hexagonal arrangement of the Bragg reflexes in Fig. \[ris:exp\_geometry\]a,b and four-fold symmetry in Fig. \[ris:exp\_geometry\]c which is caused by the symmetry of the IOLS and the geometry of the experiment. The brightest well-resolved maxima in all cases correspond to diffraction at the crystallographic planes of {202}-family. The respective scattering vector is $q_{202}$ = 20.9 $\pm ~0.3~\mu$m$^{-1}$, i.e. the lattice constant of the fcc structure unit cell is a$_0$ = 840 $\pm$10 nm.
Although we did not take the structure factor into account directly the calculated patterns (being cross-sections of the magnetic form factor spatial distribution) are very similar to the corresponding experimental ones, those represent the intensity distribution (i.e. the product of the form factor and the structural one) \[Fig. \[ris:exp\_geometry\]d-f\]. It means, that even four primitive cells stacked into the unit cell are sufficient to produce structural motif of the IOLS.
However, one can observe the difference between the experimental and calculated pictures. For instance, the central part of the diffraction patterns contains mainly the intensity of the direct beam, while in the calculated maps at the same place the magnetic form-factor near zero angle is plotted. In Fig. \[ris:exp\_geometry\]a one can see additional hexagonally arranged peaks at q of about 10 $\mu$m$^{-1}$. These peaks are actually the cross-sections of the Bragg rods, arising due to the finite thickness of the sample or stacking faults [@chumakova2014periodic]. The intensity of 2$\bar{2}$0, 02$\bar{2}$, 0$\bar{2}$2 and $\bar{2}$20 reflexes should be equal, when the field is applied along \[$\bar{1}2\bar{1}$\] direction \[Fig. \[ris:exp\_geometry\]b\] according to the symmetry, but since the sample was slightly tilted they do not. In Fig. \[ris:exp\_geometry\]c one can see, that the direct beam has an elliptical form. Moreover the reflexes of 202 family are different by intensity, though should be the same. This is due to the refraction of the beam, since the sample, which represents a thin film inclined to the beam at 54$^{\circ}$ \[Fig. \[ris:exp\_geometry\]i\].
It should be noted that in the saturated state the measured and calculated intensities, presented in Fig. \[ris:exp\_geometry\], are determined by the spatial structure of the inverse opal. Fig. \[ris:rem\_2D\_pattern\] shows an example of calculated and experimental patterns in the remanent state (the field, was initially applied along $[\bar{1}2\bar{1}]$ axis). The difference is much noticeable as compared with the patterns, obtained in the saturated state.
Similar discrepancy was obtained for the other field directions. This difference is caused by the absence of the magnetic structure factor in calculations. Structure factor minima are expected to suppress significantly the intensity of the observed bright spots of the form factor and *vice versa* its maxima can enhance weak form factor intensity, which manifests itself in the observed intensity pattern. Thus, the calculated form factor distribution does not look as experimental pattern, but nobody should expect it. On the other hand, one can see, that the Bragg peaks in the experimental patterns are situated at the same positions for 30 kOe \[Fig. \[ris:exp\_geometry\]b\] as well as for 0 kOe \[Fig. \[ris:rem\_2D\_pattern\]a\]. Thus, the structural motif of the fcc lattice is caught by the simulation even in the remanent state.
Relatively small coherent volume along with possible defects of the IOLS magnetic structure should lead to minor variation of the structure factor in the external field (i.e. the structure factor for the saturated and remanent state should be virtually the same; see Sec. \[sec:methods\]A). Therefore the strongest signal from the magnetic form factor in the Bragg peaks should be contained as the latter is suppressed by the structure factor in other regions of the reciprocal space.
Since namely the “legs” determine magnetic structure of the IOLS, i.e. their reorientation results in strong changes of magnetic state, further we focus on them. The typical magnetization distribution in the IOLS primitive cell is exemplified in Fig. \[ris:prim\_cell\_calc\]a. The calculated SANS intensity depends on the magnetization distribution in the whole unit cell. However for the sake of simplicity we have shown only magnetic state of one primitive cell, it is sufficient for our purposes. Moreover the complicated pictures of the magnetization distribution obtained by micromagnetic simulations were replaced by illustrative schemes, including solely “legs” magnetization \[Fig. \[ris:prim\_cell\_calc\]b\].
![(Color online) (a) Calculated magnetization distribution in the IOLS primitive cell. Color indicates the value of x-component of the normalized magnetization. (b) Corresponding simplified scheme. Different colors indicates different $\langle 111\rangle$-type directions.[]{data-label="ris:prim_cell_calc"}](Primitive_cell_transit-eps-converted-to.pdf){width="8.5cm"}
With the help of the spin ice model one can easily reveal the state, which the IOLS tends to fall in, if the external field is applied along the given direction. In this state magnetization of the “legs” should have a positive projection on the field direction. If the external magnetic field is near the saturation value then the shape anisotropy is suppressed and the “legs” magnetization tends to align along the field direction. In the lower fields the “legs” magnetization still has a non-negative projection on the field direction but already points along the “legs”. Further we denote the corresponding magnetic state of the primitive cell as “target magnetic element” (TME). Thus, TME is defined by the direction of the external field and corresponds to the configuration of the magnetic moments that is favorable for the field (i.e. leads to the lower Zeeman energy). Depending on the “convenience” of TME for the ice rule three scenarios arise.
**H $\parallel$ \[111\]: Scenario of competition.** When the field is applied along \[111\] axis of the IOLS, one “leg” in each tetrahedron is parallel to the field direction and the other “legs” have equal acute angles to the field — 72$^{\circ}$ \[Fig. \[ris:Basic\_element\]c\].
One can see that the field tries to bring the IOLS into the configuration in which all “legs” magnetic moments have a positive projection to the field direction. However this configuration leads to the ice rule violation i.e. either magnetic charge accumulation (the magnetic moments in three “legs” of tetrahedron point inward and the moment of the fourth “leg” points outward — 3-in-1-out state) and sink (3-out-1-in state) arise in tetrahedra. In other words, the field competes with the ice rule. The field decrease should lead to the ice rule fulfillment. Thus, for this geometry, ice rule is obeyed only in the restricted field region, until 3-in-1-out and 1-in-3-out configurations become more profitable in the total energy balance.
Fig. \[ris:H\_111\] shows the experimental and calculated intensities of 202-type reflexes as a function of the external magnetic field applied along \[111\] axis. One can see that the curve starts (at H = 0 kOe) and ends with low value of intensity, i.e. there is no remanent magnetization \[Fig. \[ris:H\_111\]a\]. The same results were obtained for hysteresis curves measured by SQUID [@dubitskiy2015spin]. Because of that the determination of the coercivity is challenging. The estimated value is about 0.5 kOe. Hysteresis behaviour can be observed up to 4 kOe, while complete merging of the branches occurs at 6.5 kOe.
The calculations shown, that TME indeed represents a set of 3-in-1-out and 3-out-1-in configurations (**B** in Fig. \[ris:H\_111\]b). However when the field is reduced 2-in-2-out states restore in tetrahedra. Herewith, the range of ice rule fulfillment depends on the “legs” shape anisotropy [@dubitskiy2015spin; @dubitskiy2017dependence]. The simulation results suggest that the ice rule is valid if field is less than 3.6 kOe (**C** in Fig. \[ris:H\_111\]b). In general, the appearance of the calculated curve is very similar to the experimental one. The variation of the intensity and the magnetic form factor from H = 0 to the coercive field related to fully demagnetized sample (**A** in Fig. \[ris:H\_111\]b) are quite similar. There is also a hysteresis in the low-field part of the calculated curve, but it may represent a double-loop hysteresis.
One can see also that for higher field values the experimental and calculated curves begin to diverge. Therefore real system resists ice rule violation more than the model one. This is possibly caused by demagnetizing field related to the sample shape which may be underestimated in our model.
**H $\parallel$ \[$\bar{1}2\bar{1}$\]: Scenario of independence.**
When magnetic field is applied along \[$\bar{1}2\bar{1}$\] direction two “legs” in each tetrahedron are at an angle of $62^{\circ}$ to the field, the third “leg” has smaller angle of $19.5^{\circ}$ and the fourth one (along \[111\]) is perpendicular to the field \[Fig. \[ris:Basic\_element\]c\].
TME in this case is defined in the following way: the magnetic moments of the “legs”, which are inclined to the field, have positive projections on the field direction, while the perpendicular ones due to the ice rule arrange in the same way. Therefore perpendicular to the field component of the magnetization should appear [@mistonov2013three]. It is clear that the direction of the magnetic moments in the “legs” that are perpendicular to the field does not affect Zeeman energy. This direction is determined solely by the ice rule. In this sense the ice rule and the field are independent.
The micromagnetic calculations confirm these findings and make it possible to analyze details of magnetization process and interpret SANS data [@dubitskiy2017dependence]. It is convenient to divide all 202-type reflections into two groups: $\textbf{q}_{20\bar2}$ and $\textbf{q}_{\bar202}$ which are perpendicular to the field direction \[$\bar{1}2\bar{1}$\]; and $\textbf{q}_{02\bar2}$, $\textbf{q}_{0\bar22}$, $\textbf{q}_{2\bar20}$, $\textbf{q}_{\bar220}$ which are at an angle $\alpha$ of $30^{\circ}$ or $150^{\circ}$ to \[$\bar{1}2\bar{1}$\] direction \[Fig. \[ris:exp\_geometry\]b\]. Measured intensity of these reflections groups varies differently in the external field \[Fig. \[ris:H\_121\]a,c\]. The same result has been obtained by means of micromagnetic simulation \[Fig. \[ris:H\_121\]b,d\].
The experimental curves \[Fig. \[ris:H\_121\]a,c\] start (at H = 0 Oe) with abrupt decrease of intensity, since the sample was magnetized initially in the direction, which is opposite to the applied field. The minimum at H = 0.2 Oe corresponds to the coercive force, where the sample should be totally demagnetized. Then the curves, obtained for the peaks of these two groups become different. In Fig. \[ris:H\_121\]a one can see two-stepped increase of the intensity: rapid growth until 0.45 kOe and smooth gain in higher fields. The saturation field is about 15-20 kOe. The reversal branch repeats the direct one down to 0.45 kOe, where hysteresis arises. At the same time, one can see well-pronounced maximum in the curve in Fig. \[ris:H\_121\]c. There is no hysteresis after this point, but it takes place in smaller fields. Interestingly, that the intensity drops abruptly in the zero field point.
Maximum value of the calculated intensity shown in Fig. \[ris:H\_121\]d is reached when magnetization in the “legs”, which are inclined to the field, has a positive projection on the field direction, while magnetization of the “legs” those are perpendicular to the field is oriented in the same way (i.e. magnetic configuration corresponds to TME) (state **C**). Further field increase leads to the inclination of the magnetization in all “legs” towards the field direction. Thus, the value of $\widetilde{\textbf{M}}_{\perp}(\textbf{q})$ decreases and hence peak intensity also decreases (see Eq. \[eq:magnetic\_section\]).
If one starts to decrease the external field then the magnetic moments in the “legs” which are perpendicular to the field reverse in order to minimise the whole demagnetization energy of the system. As a result perpendicular to the field magnetization component decreases. A rapid drop of this component takes place till the field of 0.6 kOe (transition from state **C** to **D**). Decreasing branch of calculated curve exhibits a kink in this field. The same but less pronounced kink is present in the experimental curve in the field of 0.4 kOe. Minimum of the intensity corresponds to fully demagnetized state (state **A**).
It is worth noting that magnetic moments of the “legs” which are at minimum angle to the external field (red in the Fig. \[ris:H\_121\] reverse in smaller field range than perpendicular ones (state **B**). This range coincides with the hysteresis region range in Fig. \[ris:H\_121\]b. Reflections of the $\textbf{q}_{20\bar2}$ group are more sensitive to the magnetization of these “legs” than $\textbf{q}_{02\bar2}$ ones.
One can conclude that remagnetization process starts with reversal of magnetic moments in the “legs” perpendicular to the field. Due to the ice rule this reversal is accompanied by tilt of the magnetic moments in “legs” that are at angle of $62^{\circ}$ to the field. After that the magnetization of the “legs” that are at angle of $19.5^{\circ}$ reverses. This model is supported by quantitative agreement between the calculated and experimental hysteresis curves.
**H $\parallel$ \[100\]: Scenario of cooperation.** When the field is applied along \[100\] direction all angles between the “legs” and the field are the same and equal to $54.7^{\circ}$ \[Fig. \[ris:Basic\_element\]c\]. In this case the state with positive projections of the magnetic moments on the field direction (TME) automatically obey the ice rule. In other words the ice rule and the field “cooperate” to minimize the total energy.
The experimental and calculated intensities of 202-type reflexes as a function of external magnetic field applied along \[100\] direction are shown in Fig. \[ris:H\_100\]. One can see, that the look of the experimental curve \[Fig. \[ris:H\_100\]a\] is similar to one, presented in Fig. \[ris:H\_121\]a. It has initial decrease of intensity, while field increasing from 0 Oe to the coercivity at 0.2 Oe, abrupt and smooth increase regions and decrease of intensity after saturation. Even the values of the critical fields are very close to each other. The calculated curve \[Fig. \[ris:H\_100\]b\] has the same form, but abrupt increase of intensity is two-stepped and decrease of intensity to the final state is also abrupt in contrast to the experimental curve. These jumps of the calculated curve are caused by “legs” magnetic moments reversals. In the real system they are less pronounced most likely because of large number of “legs” and spread of their shape anisotropy.
Because of the “cooperation”, during field decrease till the field of 0.4 kOe state **C** (TME) is stable. However one can see the fall of the calculated Fourier transform in Fig. \[ris:H\_100\](b). The same fall was obtained in the experiments (Fig. \[ris:H\_100\](a)). This drop can not be interpreted in terms of the spin ice model accounting for the reorientation of magnetic moments in the “legs”. The reason is discussed below in “Beyond spin ice model” subsection. Further field decrease leads to “legs” magnetization reversal and substantial drop of the form-factor value at 0 kOe (state **D**). Then IOLS falls in fully demagnetized state (state **A**). However the magnetization restores rapidly during field growth by two-step process. Firstly the system reaches state **B** and then in the field of 1.2 kOe the magnetic moments in all “legs” complete reversal (state **C**). The branches of the hysteresis curve merge only when the field value reaches 1.8 kOe. This discrepancy also can not be explained by the “legs” magnetization reversal model.
For all considered field directions the branches of hysteresis loops merge when the magnetic moments in the “legs” stop to reverse and fall into the final state. According to the spin ice model the branches should merge at the lowest field value if the ice rule and magnetic field “cooperate” and at the highest one otherwise. Both the experimental and calculated results seem to be in agreement with this model. The merging point was found to be 1 kOe, 3.5 kOe and 7 kOe when the field is applied along \[100\] (cooperation), \[$\bar{1}2\bar{1}$\] (independence), \[111\] (competition) directions respectively. the corresponding calculated values are 1.8 kOe (field along \[100\]), 3 kOe (\[$\bar{1}2\bar{1}$\]) and 6 kOe (\[111\]).
To make a tentative conclusion, the experimentally obtained hierarchy of the hysteresis loop merging points is indeed linked to the interplay between the magnetic field and the ice rule. Nevertheless, a rigorous interpretation of the experimental results is not possible in the frame of the spin ice model. Nonuniform states arisen in the IOLS structural elements have to be taken into account by means of the micromagnetic calculations. More detailed discussion is presented in the next subsection.
**Beyond the spin ice model.** The spin ice model allows one to explain main features of the experimental results. However, there are two important factors, which are not taken into account by this approach.
{width="0.99\linewidth"}
The first one is the shape anisotropy caused by the entire sample shape. Since the sample represents a thin film (12 $\mu$m) the strongest stray field appears, when the external field is perpendicular to the sample plane (along \[111\] axis), and the weakest one — when the field is in the sample plane (along \[$\bar{1}2\bar{1}$\] axis). Intermediate case is realized when the external field is applied along \[100\] axis. In the calculations presented here, the demagnetization effects were taken into account, that significantly improved the agreement between the simulated and experimental curves.
It is worth noting, that demagnetization could change the hierarchy of the branches merging fields observed in different geometries. However, nor in the experiment, neither in the calculations this hierarchy was not altered.
Secondly the magnetic moments reversal in the “legs” is generally accompanied by abrupt changes of the magnetic form factor. However the calculated form factor field dependencies also contain regions of smooth form factor variation.
In particular state **C** in all geometries is stable in a relatively large field range during the field decrease. In this state the spin ice rule is fulfilled therefore the external magnetic field competes solely with the “legs” shape anisotropy. Nevertheless one can notice reduction of both the measured intensity and calculated magnetization Fourier transform when the field decreases although the “legs” configuration is not changed. It was found that this reduction is caused by the magnetization rotation in tetrahedra and particularly by the vortex states arisen in the cube. The reason why the vortex states are likely to appear in the cube is that its linear size is relatively large (approximately 240 nm). The vortices were observed for all considered field directions.
The most considerable influence of the vortex on the form factor field dependence was found when the field was applied along \[100\] axis (“cooperation” case). The vortex evolution causes form factor variation up to the field of 0.4 kOe in the case of the decreasing branch of the hysteresis loop and from the field of 1.2 kOe in the case of increasing one (state **C**). Additionally, upon field increase vortex state becomes reversible only in the field of 1.8 kOe. One can suggest that in the experiment all the magnetic moments in the “legs” fall in the final state at 1 kOe and additional hysteresis loop between 1 kOe and 2.5 kOe is caused by non-uniform states in the cube. Snapshots of the vortex evolution are shown in Fig. \[ris:vortex\_100\]. Thus not only the anisotropic “legs” control the magnetic properties of the IOLS, but also non-uniform magnetic states. Therefore inverse opals have additional degrees of freedom compared to the conventional atomic spin-ices. In general, not complete agreement between the experimental and calculations can be linked to the bunch of factors: 1) local defects of IOLS structure which lead to spread in “legs” reversal fields values; 2) defects of IOLS mesostructure such as twinning and stacking faults; 3) magnetocrystalline anisotropy of cobalt grains; 4) partial account of magnetization field related to the whole sample shape; 5) exclusion of the structure factor variation from our model. Besides that, some experimental imperfections (e.g. inaccuracy of the sample alignment in the beam) can give rise to additional inconsistence. However based on the comparison of the experimental and simulated data we suggest that the magnetic structure of three-dimensional artificial mesoscaled spin ice was resolved.
\
Concluding remarks {#sec:conclusion}
==================
We successfully uncovered magnetic states of the inverse opal-like structure during the magnetizing process by using complementary small-angle neutron diffraction and micromagnetic calculations. Although we have considered only specific points of the magnetization curves, one can also reveal magnetic state in the intermediate points. Owing to the micromagnetic simulations it is possible to find magnetization distribution, while by using experimental data, one can adjust field values corresponded to this distribution.
According to the performed calculations, in the geometry, when the field is along \[111\] axis 3-in-1-out and 1-in-3-out configurations indeed arise in the tetrahedra. When the field is along \[$\bar{1}2\bar{1}$\] axis the ice rule gives rise to perpendicular to the field component of magnetization. When the field is along \[100\] axis system magnetizes in the fastest way, despite larger demagnetizing field, than in the case, when H is along \[$\bar{1}2\bar{1}$\]. The ice rule has the range of validity for the first case, while for two others it is fulfilled up to high fields, until shape anisotropy of the “legs” holds magnetization.
We have shown, that two-dimensional patterns of magnetic form-factor contribution are in good agreement with experimental diffraction maps taken at saturation state for all considered external field directions. But at low field values, the calculated and experimental results differ. Nevertheless calculated intensity dependencies of Bragg peaks on the external field are in semi-quantitative agreement with the experimental ones.
We speculate, that the suggested approach can be successfully applied for the revealing the magnetic structure of complicated three-dimensional periodic systems.
Acknowledgements
================
Authors thank A.V. Syromyatnikov for the fruitful discussions. Also authors are grateful to the staff of SANS-1 beamline at FRM-II, especially to S.-A. Siegfried for the hospitality and invaluable help during experimental time. Scanning electron microscopy studies were performed at the Research park of St.Petersburg State University Interdisciplinary Center for Nanotechnology (http://nano.spbu.ru/index.php/en) and micromagnetic calculations were performed using computational resources provided by Resource Center “Computer Center of SPbU” (http://www.cc.spbu.ru/en). Work was supported by Russian Foundation for Basic Research (pr. ofi-m 14-22-01113).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The onset of collective behavior in a population of globally coupled oscillators with randomly distributed frequencies is studied for phase dynamical models with arbitrary coupling; the effect of a stochastic temporal variation in the frequencies is also included. The Fokker-Planck equation for the coupled Langevin system is reduced to a kinetic equation for the oscillator distribution function. Instabilities of the phase-incoherent state are studied by center manifold reduction to the amplitude dynamics of the unstable modes. Depending on the coupling, the coefficients in the normal form can be singular in the limit of weak instability when the diffusive effect of the noise is neglected. A detailed analysis of these singularities to all orders in the normal form expansion is presented. Physically, the singularities are interpreted as predicting an altered scaling of the entrained component near the onset of synchronization.'
---
=-0.0in =-0.0in =8.25in =6.5in \[section\] \[section\] \[section\] \[section\]
[**[Synchronization of globally-coupled phase oscillators:\
singularities and scaling for general couplings]{}**]{}
John David Crawford\
Department of Physics and Astronomy\
University of Pittsburgh\
Pittsburgh, PA 15260
K.T.R. Davies\
Department of Physics\
Duquesne University\
Pittsburgh, Pennsylvania 15282-1503
keywords: oscillators, bifurcation, symmetry
January 9, 1997
Introduction
============
A phase dynamics description of $N$ limit cycle oscillators is feasible for sufficiently weak interaction.[@erkop; @kur; @ashwin] The uncoupled system is attracted to an N-torus and for weak coupling this attracting torus will persist; the oscillator phases $(\theta_1(t),\ldots,\theta_N(t))$ are coordinates on the perturbed torus. In this paper we study such systems where the interaction is mediated by a global coupling.[@kur; @ashwin] More precisely, we consider models of the form, $$\dot{\theta}_i=\omega_i+\frac{K}{N}\sum^{N}_{j=1} f(\theta_j-\theta_i) +
\xi_i(t)\label{eq:gcoupled}$$ with the frequencies $\omega_i$ of the uncoupled oscillators drawn from a distribution $g(\omega)$ characteristic of the population or array. For both physical and mathematical reasons, it is interesting to include in (\[eq:gcoupled\]) the effect of extrinsic white noise $\xi_i(t)$ perturbing the deterministic phase dynamics; this term is defined by the ensemble averages $<\xi_i(t)>=0$ and $<\xi_i(s)\,\xi_j(t)> = 2D\delta_{ij}\delta(s-t)$. When the oscillators are uncoupled ($K=0$), then the noise and the intrinsic variation in the frequencies spreads the population in phase. For suitably chosen coupling functions $f(\phi)$ as the strength of the coupling $K$ increases, the instantaneous frequencies $\dot{\theta}_i$ may become entrained and the phase disorder of the uncoupled system gives way to an entrained state with partial or total phase coherence.
The collective oscillations produced by such frequency-entrained states is of interest in physics, chemistry, biology and rather recently neuroscience.[@kur; @win; @win2; @stro1; @gray] The phase dynamics model (\[eq:gcoupled\]) was popularized especially by Kuramoto and his collaborators for the coupling $f(\phi)=\sin\phi$ since this choice allowed a mathematically tractable description of the onset of entrainment.[@kur] For a coupling $f(\phi)$ of arbitrary form, the model is a special case of a general normal form obtained by averaging.[@ashwin] Very recently, the averaged equations for a series array of weakly interacting Josephson junctions has been shown to have the form (\[eq:gcoupled\]) with $f(\phi)=\sin\phi$ at leading order in the coupling strength $K$.[@wcs] This result extended similar studies that had treated the junctions as identical oscillators.[@ssw; @ws]
Most research on phase oscillators with global couplings has been motivated by their appeal as theoretical models of synchronization rather than as realistic descriptions of specific experimental systems. In this literature, until very recently, the preoccupation with the simple coupling function considered by Kuramoto has been nearly univeral; see [@stro1; @kn] for discussions with many references. Nevertheless, couplings $f(\phi)$ of more general form are of interest for several reasons. First, interactions that are derived when a reduction to phase dynamics is actually carried out can easily have a more complicated structure.[@kur; @hmm] Secondly, recent results indicate that when the form of $f(\phi)$ is generalized, the emergence of the entrained state can have qualitative different features.[@dainew; @jdc3; @dainew96] More precisely, if $K_c$ is the critical coupling, above which the synchronized state appears, then the “size” of the entrained component scales like $(K-K_c)^\beta$ above threshold. In the absence of noise, Daido discovered $\beta=1$ when $f(\phi)$ contains a second harmonic such as $\sin2\phi$, but for the Kuramoto model the exponent drops to $\beta=1/2$, i.e. the entrainment for the more general coupling can be much weaker at the same coupling stength $K$.[@dainew] In this paper, we present a detailed analysis of the first transistion to a state with frequency entrainment that illuminates the interplay between the intrinsic disorder $g(\omega)$, the external noise, and the coupling in determining the exponent $\beta$. In particular, we make no further restrictions on $f(\phi)$ other than to assume it has a Fourier expansion $$f(\phi)=\sum_{n=-\infty}^{\infty}\,f_n\,e^{in\phi}.\label{eq:fexp}$$
For large oscillator populations, it is natural to analyze the system in the limit $N\rightarrow\infty$ because of the relative simplicity afforded by a continuum description of the population using a density function $\rho(\theta,\omega,t)$. A convenient definition divides out the frequency dependence of the uncoupled system so that $g(\omega)\,\rho(\theta,\omega,t)\,d\theta\,d\omega$ describes the fraction of oscillators with natural frequency in $[\omega,\omega+d\omega]$ and phase in $[\theta,\theta+d\theta]$, and $\rho(\theta,\omega,t)\,d\theta$ describes the fraction of oscillators at natural frequency $\omega$ with phase in $[\theta,\theta+d\theta]$. This latter interpretation fixes the normalization $\int\rho(\theta,\omega,t)\,d\theta=1$. The dynamics of the system (\[eq:gcoupled\]) is formulated as a kinetic equation for $\rho(\theta,\omega,t)$, $$\frac{\partial\rho}{\partial t}+\frac{\partial(\rho v)}{\partial \theta}=D
\frac{\partial^2\rho}{\partial \theta^2},\label{eq:eveqn}$$ with the deterministic part of the phase velocity (\[eq:gcoupled\]) expressed as an integral over the population $$v(\theta,\omega,t)=\omega +K \int^{2\pi}_0\,d\theta'
\int^{\infty}_{-\infty}\,d\omega'
f(\theta'-\theta)\rho(\theta',\omega',t)\,g(\omega').\label{eq:vel}$$ This equation has been applied recently to the Kuramoto model;[@sak]-[@jdc0] our application to (\[eq:gcoupled\]) for arbitrary coupling is a generalization of the results for $f(\phi)=\sin\phi$.[@sm; @jdc0] For completeness, the derivation of (\[eq:eveqn\]) - (\[eq:vel\]) is outlined in Appendix \[app:deriv\].
The kinetic equation has a special solution, $$\rho=\rho_0\equiv\frac{1}{2\pi},$$ which describes a population of oscillators spread out unpreferentially in phase. This uniform or incoherent distribution is an equilibrium since $v(\theta,\omega)=\omega+Kf_0$ for $\rho=\rho_0$ implies $\partial_t\rho=0$. The occurrence of instabilities of $\rho_0$ leading to states with nontrivial phase dependence is established by a linear stability analysis of $\rho_0$ in section \[sec:linear\] and a nonlinear analysis of the instabilities in sections \[sec:ampeqn\] - \[sec:special\]. The nonlinear analysis focuses on the evolution of the linearly unstable mode described by the flow on the unstable manifold of the equilibrium; the amplitude equation describing this flow are formulated and studied in general terms in section \[sec:ampeqn\]. The detailed calculation of the unstable manifold flow given in section \[sec:exp\] relies on expansions in the amplitude of the unstable mode. The exponent $\beta$ emerges as the scaling exponent required to balance the linear instability against the nonlinear terms in the amplitude equation. In the regime of weak linear instability, different exponents can arise because the behavior of the nonlinear terms varies considerably depending on the form of the coupling and strength of the external noise. These features of the expansion are studied in sections \[sec:p1p2\] - \[sec:special\].
Our conclusions regarding the properties of the amplitude equation are valid to all orders in the expansion. This level of generality is possible in part because the recursion relations for the expansion are considerably simplified by the symmetry of the problem. The model (\[eq:eveqn\]) - (\[eq:vel\]) has ${\rm SO}(2)$ or ${\rm O}(2)$ symmetry depending on details of the population and the coupling. The group ${\rm O}(2)$ is generated by rotations $\beta\cdot(\theta,\omega)=(\theta +\beta,\omega)$ and reflections $\kappa\cdot(\theta,\omega)=-(\theta,\omega)$ which act on functions $\rho(\theta,\omega)$ in the usual way: for any transformation $\chi\in{\rm O}(2)$, $(\chi\cdot\rho)(\theta,\omega)\equiv\rho(\chi^{-1}\cdot(\theta,\omega))$. The dynamics (\[eq:eveqn\]) - (\[eq:vel\]) is equivariant with respect to rotations for arbitrary choices of $g(\omega)$ and $f(\phi)$; if, in addition, $g(\omega)=g(-\omega)$ [*and*]{} $f(\phi)=-f(-\phi)$, then the system commutes with the reflection $\kappa$ as well. In the latter circumstance the model has ${\rm O}(2)$ symmetry, otherwise the rotational symmetry alone corresponds to ${{\rm SO}(2)}$. Note that the mean frequency of $g(\omega)$ can always be shifted to zero and we assume that this has been done.
The onset of linear instability for the incoherent state $\rho_0$ has some unusual features, especially in the limit of weak noise $D\rightarrow0$. This was first noticed in the context of the Kuramoto model,[@sm]-[@jdc0] and is emphasized in the linear analysis of section \[sec:linear\]. In the absence of noise, the unstable modes correspond to eigenvalues emerging from a neutral continuum at onset; thus, the standard framework of center manifold reduction cannot be directly applied. We proceed by first deriving the amplitude equation on the center-unstable manifold with $D>0$, in this case the diffusion term (\[eq:eveqn\]) shifts the continuous spectrum off the imaginary axis and the center manifold reduction is straightforward.[@craw5; @iv] The resulting amplitude equation is then examined by first reducing the noise $D\rightarrow0$ and then letting the linear growth rate of the mode approach zero $\gamma\rightarrow0^+$. In this way, we start with an amplitude equation on the unstable manifold and examine its features as the unstable eigenvalues are allowed to approach a neutral continuous spectrum on the imaginary axis.
Our procedure for extracting $\beta$ from the amplitude equation can be illustrated by a simple example. Let $A(t)$ denote the mode amplitude satisfying $$\dot{A}(t)=\gamma\,A+a_1(\gamma)A^3+a_2(\gamma)A^5+\cdots\label{eq:aeqnintro}$$ where the nonlinear coefficients $a_1, a_2,\ldots$ depend on the parameters of the problem such as the growth rate $\gamma$. For small $\gamma$, this equation should be viewed as a singular perturbation problem with the linear term $\gamma\,A$ representing the perturbation. If $\gamma>0$, then there is always a neighborhood of the equilibrium $A=0$ where the perturbation dominates the unperturbed system and can completely change the dynamics. As in other such singular problems, a possible strategy is to seek a (singular) change of variables which transforms (\[eq:aeqnintro\]) into a regular perturbation problem. Thus it is natural to define a new amplitude by $${A}(t)=\gamma^\beta r(\gamma^\delta t)\label{eq:newvar}$$ and rewrite (\[eq:aeqnintro\]) in terms of $r(\tau)$ $$\frac{d r}{d\tau} =\gamma^{1-\delta}\,r(\tau) + \gamma^{2\beta-\delta}
a_1(\gamma)r^3+\gamma^{4\beta-\delta}a_2(\gamma)r^5+\cdots.\label{eq:aeqnb}$$ If possible, the choice of $\beta$ and $\delta$ should be made so that each term in (\[eq:aeqnb\]) is well behaved as $\gamma\rightarrow0^+$, and furthermore so that the effect of $\gamma>0$ is a regular perturbation of the system at $\gamma=0$. The standard values are of course $\beta=1/2$ and $\delta=1$ yielding $$\frac{d r}{d\tau} =r(\tau) + a_1(\gamma)r^3
+\gamma\,a_2(\gamma)r^5+\cdots.\label{eq:aeqnc}$$ The unperturbed system $\dot{r}=r+a_1(0)r^3$ now includes the linear term and near $r=0$ the effect of small $\gamma$ is unimportant. The change of variables (\[eq:newvar\]) is singular at $\gamma=0$ as expected, but nevertheless we have determined the overall scaling $\beta=1/2$ for nonlinear solutions near the equilibrium in the regime of weak instability. In addition, since terms in (\[eq:aeqnb\]) at fifth order or higher are at least ${\cal
O}(\gamma)$, the unperturbed problem truncates to a simple balance between the linear instability and the dominant nonlinear term. Note that since $\gamma=0$ at $K=K_c$, then just above threshold $\gamma\sim(K-K_c)$, and the definition in (\[eq:newvar\]) implies $|A|\sim(K-K_c)^\beta$ for the scaling of the mode ampltude near onset.
The preceeding discussion contains the unstated assumption that the coefficients $a_1, a_2,\ldots$ have well-defined [*finite*]{} limits as $\gamma\rightarrow0$. In the phase dynamical model (\[eq:eveqn\]) - (\[eq:vel\]) this assumption is correct for any choice of $f(\phi)$ and $g(\omega)$ as long as $D>0$. However if we set $D=0$ before taking the limit $\gamma\rightarrow0^+$, then the nonlinear coefficients are singular in general although the details depend on $f(\phi)$. These singularities require a different choice for $\beta$ to accomplish the desiderata enumerated above. It turns out that when one has a coupling with only a single Fourier component such as in the Kuramoto model then the expansion coefficients are nonsingular to all orders. Thus single component couplings are special and very atypical. The most typical or generic behavior is $\beta=1$ as was found by Daido in the examples he considered.[@dainew; @dainew96] We further discuss the relation between our approach and his theory in section \[sec:daido\].
A key aspect of the singular behavior encountered in this problem is the emergence of the critical eigenvalues from a neutrally stable continuous spectrum. It has become clear recently that this characteristic is shared by a diverse set of additional examples which include the onset of linear instability in collisionless plasma[@jdc1]-[@jdcaj] and in fluid shear flows[@case2]-[@briggs], instabilities of solitary waves[@pegowein1]-[@pegwein3], and bifurcations in “mean field” descriptions of the dynamics of bubble clouds in fluids[@russo]. Aside from the present work, only for the collisionless plasma models have the resulting singularities in the amplitude expansion been fully analyzed. Our proofs in this paper are adapted from the techniques used in [@jdc2].
Linear analysis {#sec:linear}
===============
Let $\eta\equiv\rho(\theta,\omega,t)-\rho_0$ denote the deviation from $\rho_0$, and rewrite the dynamics (\[eq:eveqn\]) - (\[eq:vel\]) to describe the evolution of $\eta$: $$\frac{\partial\eta}{\partial t}={\rm \cal L}\eta +{{\rm \cal
N}({\eta})}\label{eq:dyn}$$ where the linear and nonlinear operators are defined by $$\begin{aligned}
{\rm \cal L}\eta&=& D \frac{\partial^2\eta}{\partial \theta^2} -
(\omega+K\,f_0)\frac{\partial\eta}{\partial \theta} +
\frac{K}{2\pi}\int^{2\pi}_0\,d\theta'\int^{\infty}_{-\infty}\,d\omega'
f'(\theta'-\theta)\,\eta(\theta',\omega',t)\,g(\omega')
\label{eq:lop}\\
{{\rm \cal N}({\eta})}&=&
K\int^{2\pi}_0\,d\theta'\int^{\infty}_{-\infty}\,d\omega'
\,\eta(\theta',\omega',t)\,g(\omega')
\left[\rule{0in}{0.25in}\eta(\theta,\omega,t)\,f'(\theta'-\theta)\right.
\label{eq:nop}\\
&&\hspace{3.0in}-\left.
\frac{\partial\eta}{\partial\theta}(\theta,\omega,t)\,f(\theta'-\theta)
\right].\nonumber\end{aligned}$$ In (\[eq:lop\]) - (\[eq:nop\]), $f'(\phi)\equiv df/d\phi$, and note that the normalization of $\rho$ implies $$\int^{2\pi}_0\,d\theta\,\eta(\theta,\omega,t)=0.\label{eq:nomean}$$
The equilibrium $\rho_0$ is invariant under translations $\beta$ and reflections $\kappa$; hence the dynamics for $\eta$ has the same symmetries as the original system. This can be verified explicitly by noting that the operators ${\rm \cal L}$ and ${\rm \cal N}$ commute with rotations for arbitrary choices of $g(\omega)$ and $f(\phi)$, and also commute with $\kappa$ when $g(\omega)=g(-\omega)$ and $f(\phi)=-f(-\phi)$.
Since $\rho_0$ is independent of $\theta$, the eigenfunctions ${\rm \cal
L}\Psi=\lambda\Psi$ have the form $$\Psi(\theta,\omega)=\psi(\omega)\,e^{in\theta},$$ with $\psi(\omega)$ required to satisfy the eigenvalue equation $$L_n\psi=\lambda\,\psi,\label{eq:evleqn}$$ where $$L_n\psi\equiv-in\left[(\omega+K\,f_0-inD)\psi+{K\,f_n^\ast}
\int^{\infty}_{-\infty}\,d\omega'\,g(\omega')\,\psi(\omega')\right].$$ Because the $n=0$ component of $\eta$ is zero (\[eq:nomean\]), we can restrict our consideration of (\[eq:evleqn\]) to $n\neq0$.
The spectrum of $L_n$ corresponds to values of $\lambda$ where the resolvent $R_n(\lambda)\equiv(\lambda-L_n)^{-1}$ fails to be a bounded operator. By solving the equation $(\lambda-L_n)\hat{\phi}=\phi$ for $\hat{\phi}(\omega)$ we obtain the resolvent explicitly, i.e. $\hat{\phi}=R_n(\lambda)\phi$ with $$R_n(\lambda)\phi=\frac{-i/n}{\omega+K f_0-i\lambda/n-inD}
\left[\phi(\omega)-\frac{K f_n^\ast}{\Lambda_n(i\lambda/n)}
\int_{-\infty}^\infty\frac{d\omega'\;g(\omega')\phi(\omega')}
{\omega'+K f_0-i\lambda/n-inD}\right]\label{eq:resolv}$$ where, in the second term, $\Lambda_n(z)$ denotes $${\Lambda_{n}\,({z})}\equiv 1+{K\,f_{n}^\ast}
\int^{\infty}_{-\infty}\,d\omega\frac{g(\omega)}{\omega+Kf_0-z-inD}.
\label{eq:Lam}$$ The details of a completely rigorous discussion of $R_n(\lambda)$ will depend of course on the choice of a specific function space for the problem, e.g. square integrable functions of $\omega$. For our purposes, it suffices to ignore these technicalities and discuss the spectrum of $L_n$ carefully though heuristically.
There are two ways $R_n(\lambda)\phi$ can be singular: when $i\lambda/n$ hits a zero of the function $\Lambda_{n}$ and when the denominator in (\[eq:resolv\]) vanishes, i.e. when $\lambda=-in(\omega+K
f_0-inD)$. Since $-\infty<\omega<\infty$, this latter set of values coincides with the line ${\rm Re}(\lambda)=-n^2D$, while the first set of values are typically discrete $\lambda=-inz_j$ since $\Lambda_n(z_j)=0$ will have discrete solutions $\{z_1, z_2, \ldots\}$. These two components generally correspond to the essential spectrum and the point spectrum (eigenvalues), respectively.
This distinction is clarified by solving the eigenvalue equation directly. Let $\lambda=-inz$, then (\[eq:evleqn\]) reads $$(\omega+K\,f_0-z-inD)\psi(\omega)=
-K\,f_n^\ast\,\int^{\infty}_{-\infty}\,d\omega'\,g(\omega')\,\psi(\omega').
\label{eq:evleqnb}$$ The possible solutions for ${\rm Im}(z)\neq-nD$ and ${\rm Im}(z)=-nD$ are quite different and we discuss them separately.
For ${\rm Im}(z)\neq-nD$ or ${\rm Re}(\lambda)\neq-n^2D$, we can divide by $(\omega+K\,f_0-z-inD)$ and obtain the form of the eigenfunction $$\psi(\omega)=
\frac{-K\,f_n^\ast\,\int\,d\omega'\,g(\omega')\,\psi(\omega')}
{\omega+K\,f_0-z-inD}\label{eq:form}$$ in terms of the constant $\int\,d\omega'\,g\,\psi$. A nontrivial solution in (\[eq:form\]) is found only if this constant is non-zero, in which case we can adopt the normalization $$\int^{\infty}_{-\infty}\,d\omega'\,g(\omega')\,\psi(\omega')=1.\label{eq:ncond}$$ Together (\[eq:form\]) and (\[eq:ncond\]) determine an eigenfunction $$\psi(\omega)=
\frac{-K\,f_n^\ast}
{\omega+K\,f_0-z-inD}\label{eq:efcnsoln}$$ [*provided*]{} the normalization condition (\[eq:ncond\]) for $\psi$ can be satisfied self-consistently. This requirement is precisely $\Lambda_n(z)=\Lambda_n(i\lambda/n)=0$. Thus a root $\Lambda_n(z)=0$ implies the existence of an eigenvalue $\lambda=-inz$ with eigenfunction (\[eq:efcnsoln\]).
For all choices of $g(\omega)$ and $f(\phi)$, $\Lambda_n(z)$ satisfies the identity $${\Lambda_{n}\,({z})}^\ast={\Lambda_{-n}\,({z}^\ast)};
\label{eq:Lam1}$$ in addition, with reflection symmetry $f(\phi)=-f(-\phi)$ and $g(\omega)=g(-\omega)$, a further identity holds $${\Lambda_{n}\,({z})}={\Lambda_{-n}\,(-{z})}.
\label{eq:Lam2}$$ The first identity implies that complex eigenvalues must come in conjugate pairs $(\lambda, \lambda^\ast)$ with conjugate eigenvectors $\Psi(\theta,\omega)$ and $\Psi(\theta,\omega)^\ast$, respectively. For a real eigenvalue $({\rm
Re}(z)=0)$, the first identity reduces to the second (\[eq:Lam2\]) and implies that the eigenvalue will always have (at least) two linearly independent eigenvectors, $\Psi$ and $\Psi^\ast$. When we have a complex eigenvalue for a reflection symmetric population, then (\[eq:Lam2\]) holds and again there will be two linearly independent eigenvectors $\Psi(\theta,\omega)$ and $\Psi(-\theta,-\omega)$. Without reflection symmetry, the typical case is a complex eigenvalue, but with reflection symmetry both real and complex eigenvalues can occur generically.
The actual occurrence of roots for $\Lambda_n(z)$ is a detailed question depending on the frequency distribution $g$ and the coupling; examples have been found for which $\Lambda_n(z)$ has no roots.[@sm] In the complex $z$ plane, $\Lambda_n(z)$ is an analytic function for ${\rm
Im}(z)\neq-nD$ with a discontinuity across ${\rm Im}(z)=-nD$ whenever $g({\rm Re}(z)-Kf_0)\neq0$. The discontinuity signals a branch cut for $\Lambda_n(z)$ and the magnitude of the discontinuity is given by the boundary values at $ {\rm Re}(z)=r$: $$\lim_{\epsilon\rightarrow0^+}\Lambda_n(r-inD\pm i\epsilon)=
1+{K\,f_{n}^\ast}\left[{\rm PV}
\int^{\infty}_{-\infty}\,d\omega\frac{g(\omega)}{\omega+Kf_0-r}\right]
\pm i \pi\,K f_{n}^\ast\, g(r-Kf_0).
\label{eq:bvalue}$$ As the parameters of the coupling or frequency distribution $g$ are varied, the roots of $\Lambda_n(z)$ generally change. In particular, roots can appear or disappear at the branch cut.
When $D=0$, the appearance of a root coincides with the onset of linear instability and the boundary values (\[eq:bvalue\]) provide closed form expressions for the coupling $K_c$ and frequency $\Omega_c=-{\rm Im}(\lambda)$ at criticality. Let $z_0=(\Omega+i\gamma)/n$ denote the root with $\gamma>0$, and $0=\Lambda_n((\Omega+i\gamma)/n)$. As $\gamma\rightarrow0^+$, we obtain $$0=1+{K_c\,f_{n}^\ast}\left[{\rm PV}
\int^{\infty}_{-\infty}\,d\omega\frac{g(\omega)}{\omega-\Omega'}
\right]
+ i\pi\, K_c f_{n}^\ast\, g(\Omega').
\label{eq:criticality}$$ where $\Omega'=\Omega_c/n-K_cf_0$. The real and imaginary parts of this equation must vanish separately and this implies $$\begin{aligned}
\frac{-{\rm Im}(f_n)}{\pi\,|f_n|^2\,g(\Omega')}&=&K_c\\
&&\nonumber\\
\frac{\pi\,{\rm Re}(f_n)g(\Omega')}{{\rm
Im}(f_n)}&=&\int_0^{\infty}\frac{d\Delta}{\Delta}
\left[g(\Omega'+\Delta)-g(\Omega'-\Delta)\right].\end{aligned}$$ The second equation determines $\Omega'$ and then the first equation determines $K_c$; for $n=1$ and $f_0=0$ these expressions reduce to the formulas derived by Daido in his order function analysis.[@dai93]
For ${\rm Im}(z)=-nD$ or ${\rm Re}(\lambda)=-n^2D$, the eigenvalue equation (\[eq:evleqnb\]) has no nonsingular solutions. One can introduce distributions as solutions following the closely related spectral theory for the Vlasov equation, but we shall not require this development for our study (cf. [@ch] for details).
In general, the spectrum of ${\cal L}$ has both eigenvalues and a continuous component: for each mode number $n=1,2,\ldots$, there is a line of continuous spectrum at $\mbox{\rm Re}\,\lambda=-n^2D$; in addition ${\rm \cal L}$ has eigenvalues when the function $\Lambda_{n}$ has roots. We are specifically interested in eigenvalues that cross the imaginary axis, signifying instabilities of $\rho_0$. Depending on the coupling, this can occur for any Fourier component $n=l$, provided there is a root $\Lambda_{l}\,({z_0})=0$ such that ${\rm Im}(z_0)$ becomes positive. Henceforth $l$ and $z_0$ refer to the eigenvalue $\lambda=-ilz_0$ determining the instability and the unstable eigenfunction is $$\begin{aligned}
\Psi(\theta,\omega)&=&\psi(\omega)\,e^{il\theta}\\
\psi(\omega)&=&\frac{-K\,f_l^\ast}
{\omega+K\,f_0-z_0-ilD}.\label{eq:efcn0}\end{aligned}$$ Clearly the coupling must have a Fourier component $f_l\neq0$.
We further restrict the problem with two additional assumptions. First, the root $z_0$ is simple, i.e. $\Lambda'_{l}\,({z_0})\neq0$ where $\Lambda'_{l}\equiv d\Lambda_{l}/dz;$ secondly, the center subspace at criticality is two-dimensional, in particular, all other roots of $\Lambda_{n}$ for any $n>0$ have negative imaginary parts and remain bounded away from the real axis. The first assumption ensures that there are no generalized eigenvectors; the second restricts the analysis to populations with rotation symmetry (SO(2)) and complex eigenvalues, or populations with rotation and reflection symmetry (O(2)) and a real eigenvalue.[@jdc0] This restriction excludes one possible codimension-one bifurcation, a Hopf bifurcation with O(2) symmetry, which involves a four-dimensional center subspace; a discussion of $O(2)$ Hopf for the Kuramoto model has been given elsewhere.[@jdc0]
The projection of $\eta$ along the critical mode (\[eq:efcn0\]) requires an inner product $$(A,B)\equiv\int_{0}^{2\pi}d\theta\int_{-\infty}^{\infty}
d\omega\,A(\theta,\omega)^\ast\,B(\theta,\omega),\label{eq:ip}$$ so that we can define an adjoint operator for ${\rm \cal L}$ $$({\rm \cal L}^\dagger\,A,B)=(A,{\rm \cal L}\,B).$$ In terms of the Fourier expansion $A=\sum_n\,A_n\exp (in\theta$), the inner product becomes $(A,B)=2\pi\,\sum_n\,<A_n,B_n>$ using a convenient notation for the $\omega$ integration $$<A_n,B_n>\equiv\int_{-\infty}^{\infty}
d\omega\,A_n(\omega)^\ast\,B_n(\omega).\label{eq:ipw}$$ A simple calculation shows that $${\rm \cal L}^\dagger\,A=\sum_{n=-\infty}^\infty\,
e^{in\theta}L_n^\dagger A_n(\omega),$$ with $$(L_n^\dagger A_n)(\omega)=in\left[(\omega+K\,f_0+inD)A_n(\omega)
+K\,f_n\,g(\omega)\int_{-\infty}^\infty\,d\omega'\,A_n(\omega')\right].$$
The discussion of the spectrum of $L_n^\dagger$ parallels the analysis for $L_n$, and we simply state the results we shall require. The root $\Lambda_{l}\,({z_0})=0$ implies an eigenfunction ${\rm \cal L}^\dagger\tilde{\Psi}=\lambda^\ast\tilde{\Psi}$ where $$\begin{aligned}
\tilde{\Psi}(\theta,\omega)&=&\frac{1}{2\pi}
\tilde{\psi}(\omega)\,e^{il\theta}\\
\tilde{\psi}(\omega)&=&\frac{-g(\omega)}
{\Lambda'_{{l}}\,({z_0})^\ast(\omega+Kf_0-z_0^\ast+ilD)}.\label{eq:adjef}\end{aligned}$$ This adjoint eigenfunction satisfies $(\tilde{\Psi},\Psi)=1$ and defines the projection, $\eta\rightarrow(\tilde{\Psi},\eta)\;\Psi$, from $\eta$ onto the $\Psi$ component of $\eta$. This projection is the only aspect of the adjoint theory that we need.
Before developing the nonlinear theory, we illustrate our discussion with a simple example of the instabilities we intend to analyze. Let the oscillator population be described by a Lorentzian frequency distribution, $$g(\omega)=\frac{\Delta}{\pi}\,\left[\frac{1}{\omega^2+\Delta^2}\right],
\label{eq:lorentz}$$ then for ${\rm Im}(z)\neq-lD$, $\Lambda_{l}\,({z})$ is easily evaluated by contour integration to obtain $$\Lambda_{l}\,({z})=\left\{\begin{array}{cc}
\frac{z-K(f_0+f_l^\ast)+i(lD-\Delta)}{z-Kf_0+i(lD-\Delta)}&
{\rm Im}(z)<-lD\\
&\\
\frac{z-K(f_0+f_l^\ast)+i(lD+\Delta)}{z-Kf_0+i(lD+\Delta)}&
{\rm Im}(z)>-lD.\end{array}\right.
\label{eq:loroots}$$ For ${\rm Im}(z)<-lD$, there is a root at $z_s=K(f_0+f_l^\ast)-i(lD-\Delta)$ when the coupling satisfies $K\,{\rm Im}(f_l)>\Delta.$ The eigenvalues $\lambda=-ilz_s$ have negative real parts and fall to the left of the continuous spectrum at $\mbox{\rm
Re}\,\lambda=-l^2D$. An eigenvalue sitting to the right of the continuous spectrum requires a root with ${\rm Im}(z)>-lD$; for this example $\Lambda_{l}\,({z})$ has one such root at $z_u=K(f_0+f_l^\ast)-i(lD+\Delta)$ when the coupling satisfies $K\,{\rm Im}(f_l)<-\Delta$. This root yields the eigenvalue $$\lambda=-ilz_u=\gamma-i\Omega,\label{eq:eigen}$$ where $$\begin{aligned}
\gamma&=&-l\left[ lD+\Delta+K\,{\rm Im}(f_l)\right]\\
\Omega&=&lK[{\rm Re}(f_l)+ f_0].\end{aligned}$$ These modes are linearly stable for $lD+\Delta+K\,{\rm Im}(f_l)>0$ and become linearly unstable for $K>K_c$ where $K_c=-(lD+\Delta)/{\rm Im}(f_l)$. Note, that we find eigenvalues of $L_l$ only when $|K\,{\rm Im}(f_l)|>\Delta$; in the limit $|K\,{\rm Im}(f_l)|\rightarrow\Delta$ these eigenvalues approach the continuous spectrum and are “absorbed” as the underlying root $z_u$ or $z_s$ crosses the branch cut onto a different Riemann sheet of $\Lambda_{l}\,({z})$.
The diffusion due to external noise simplifies the spectrum by displacing the lines of continuous spectrum into the left half plane. With $D>0$ the eigenvalues associated with instabilities first emerge from the continuum and then subsequently reach the imaginary axis. Thus, the critical eigenvalues are isolated on the imaginary axis, and one has a critical spectrum of standard form for bifurcation theory. If we let $D\rightarrow0$, then the unstable eigenvalues emerge directly from a continuous spectrum on the imaginary axis. One of our central concerns is the effect of this continuous spectrum for $D=0$ on the $\gamma\rightarrow0^+$ limit of the vector field on the center-unstable manifold. This is investigated below by deriving this vector field first for $D>0$ and then analyzing the effect of taking the limit $\gamma\rightarrow0^+$ when $D=0$.
Amplitude equation on the unstable manifold {#sec:ampeqn}
===========================================
With diffusion included, the appearance of an unstable mode in the spectrum can be analyzed by a straightforward application of center manfold reduction.[@craw5; @iv] This enables us to derive approximately the vector field on the center-unstable manifold of $\rho_0$, and study the resulting bifurcation. This vector field will depend on the parameters of the problem and we shall examine its features as the growth rate of the unstable mode is progressively weakened, a limit which we denote by $\gamma\rightarrow0^+$. The properties of this limit are strongly dependent on the form of the coupling $f(\phi)$ and whether the external noise $D>0$ is held fixed as $\gamma$ vanishes.
Center manifold reduction yields an amplitude equation describing the time-asymptotic behavior of the unstable mode. The mode amplitude $\alpha(t)\equiv(\tilde{\Psi},\eta)$ and describes the projection of $\eta$ onto the unstable eigenspace (\[eq:efcn0\]): $$\eta(\theta,\omega,t)=[\alpha(t)\Psi(\theta,\omega)+ {\rm c.c.}] +
S(\theta,\omega,t)\label{eq:decomp}$$ where $(\tilde{\Psi},S)=0.$ In terms of $\alpha$ and $S$, the evolution equation (\[eq:dyn\]) becomes $$\begin{aligned}
\dot{\alpha}&=&\lambda\, \alpha +(\tilde{\Psi},{{\rm \cal N}({\eta})})
\label{eq:alph1}\\
\frac{\partial S}{\partial t}&=&{\rm \cal L} S+{{\rm \cal N}({\eta})}
-\left[(\tilde{\Psi},{{\rm \cal N}({\eta})})\,\Psi + {\rm c.c.} \right].
\label{eq:S1}\end{aligned}$$ These coupled equations are equivalent to (\[eq:dyn\]); by restricting them to solutions on the center-unstable manifold we obtain an autonomous two-dimensional flow for $\alpha(t)$.
Near the equilibrium $\eta=0$, the center-unstable manifold is described by a function $H$, $$S(\theta,\omega)=H(\theta,\omega,\alpha,\alpha^\ast),\label{graph}$$ that satisfies $0=H(\theta,\omega,0,0)=\partial_\alpha\,H(\theta,\omega,0,0)
=\partial_{\alpha^\ast}\,H(\theta,\omega,0,0)$. Using this representation of the manifold, the general solution in (\[eq:decomp\]) can be adapted for solutions on the manifold, $$\eta^u(\theta,\omega,t)=[\alpha(t)\Psi(\theta,\omega)+ {\rm c.c.}] +
H(\theta,\omega,\alpha(t),\alpha(t)^\ast);\label{eq:umfd}$$ the time dependence of $\eta^u(\theta,\omega,t)$ is entirely determined by the dynamics of $\alpha(t)$. An autonomous equation for $\alpha(t)$ follows next by restricting the general equation (\[eq:alph1\]) to solutions of the special form (\[eq:umfd\]): $$\dot{\alpha}=\lambda\, \alpha +(\tilde{\Psi},{{\rm \cal N}({\eta^u})}).
\label{eq:alph1cm}$$ This is the amplitude equation we wish to derive and analyze.
Certain general features of the amplitude equation follow from the symmetry properties of the evolution equation (\[eq:dyn\]) which are, of course, shared by its expression in (\[eq:alph1\]) - (\[eq:S1\]). The action of the rotations $\beta$ and reflection $\kappa$ on the variables $(\alpha,S(\theta,\omega))$ can be determined from their definitions in (\[eq:decomp\]). The rotation $\beta\cdot(\theta,\omega)\rightarrow(\theta+\beta,\omega)$ of the phase acts by $$\beta\cdot(\alpha,S(\theta,\omega))=
(\alpha \,e^{-il\beta},S(\theta-\beta,\omega)),\label{eq:trans}$$ and $$\kappa\cdot(\alpha,S(\theta,\omega))=(\alpha^\ast,S(-\theta,-\omega))
\label{eq:ref}$$ for the reflection.
The right hand side of the amplitude equation commutes with these transformations of $\alpha$ whenever they are symmetries of (\[eq:alph1\]) - (\[eq:S1\]); in particular the amplitude equation commutes with rotations (\[eq:trans\]). There are useful standard results on the expression of such equivariant vector fields. For example, a two-dimensional vector field $\dot{\alpha}=V(\alpha,\alpha^\ast)$ that commutes with $\alpha\rightarrow \alpha \,e^{-il\beta}$ can be written as $V(\alpha,\alpha^\ast)=\alpha\,p(\sigma)$ where $\sigma\equiv|\alpha|^2$ and $p(\sigma)$ is a smooth function.[@gss] Hence we know the amplitude equation (\[eq:alph1cm\]) has the form $$\lambda\, \alpha +(\tilde{\Psi},{{\rm \cal N}({\eta^u})})=\alpha\,p(\sigma)
\label{eq:ampsym}$$ although $p(\sigma)$ must be determined from the model. In general, $p(\sigma)$ is complex, but when the model has reflection symmetry then $p(\sigma)$ must be real-valued.
Analysis of $H(\theta,\omega,\alpha,\alpha^\ast)$
-------------------------------------------------
The dynamical invariance of the center-unstable manifold implies an equation for the graph function $H$. Consistency between the time dependence of $S$ in (\[graph\]) and the evolution of $S$ described by (\[eq:alph1\])-(\[eq:S1\]) requires that $H(\theta,\omega,\alpha(t),\alpha^\ast(t))$ satisfy $$\left[\dot{\alpha}\,\frac{\partial H}{\partial {\alpha}}
+\dot{\alpha}^\ast\,\frac{\partial H}{\partial {\alpha}^\ast}\right]
={\cal{L}} H+{\cal{N}}(\eta^u)-\left[(\tilde{\Psi},{\cal{N}}(\eta^u))\,\Psi +
(\tilde{\Psi},{\cal{N}}(\eta^u))^\ast\,\Psi^\ast\right]\label{Heqn}$$ where $\dot{\alpha}$ on the left hand side refers to the amplitude equation (\[eq:alph1cm\]).
There are general constraints on $H$ due to the symmetries of (\[eq:alph1\]) - (\[eq:S1\]) because we expect the unstable manifold to be mapped onto itself by a symmetry transformation. For a rotation $\beta$, this symmetry invariance of the manifold implies $$H(\theta-\beta,\omega,\alpha,\alpha^\ast)=
H(\theta,\omega,\alpha \,e^{-il\beta},\alpha^\ast \,e^{il\beta}),
\label{eq:Htrans}$$ and if reflection symmetry holds then $H$ must satisfy $$H(-\theta,-\omega,\alpha,\alpha^\ast)=
H(\theta,\omega,\alpha^\ast ,\alpha)
\label{eq:Href}$$ as well.
These relations constrain the form of the Fourier expansion of H $$H(\theta,\omega,\alpha,\alpha^\ast)=
\sum_{n=-\infty}^{\infty}H_n(\omega,\alpha,\alpha^\ast)\,e^{in\theta};
\label{eq:scm}$$ applying (\[eq:Htrans\]) to the components $H_n$ shows that the component must vanish unless $n$ is an integer multiple of $l$. Note that the $n=0$ component of $H$ is also identically zero because of the general property $\int\,d\theta\,\eta=0$. When $n=ml>0$, then $H_{ml}$ must have the form $$\begin{aligned}
H_l(\omega,\alpha,\alpha^\ast)&=&\alpha\,\sigma h_{1}(\omega,\sigma)
\label{eq:Hl}\\
H_{ml}(\omega,\alpha,\alpha^\ast)&=&\alpha^m\,h_{m}(\omega,\sigma)
\hspace{0.5in}m>1.\label{eq:Hml}\end{aligned}$$ The functions $h_{m}(\omega,\sigma)$ are unconstrained by the rotations, but must satisfy $h_{m}(-\omega,\sigma)^\ast=h_{m}(\omega,\sigma)$ when reflection symmetry holds. In addition, the general relation $(\tilde{\Psi},S)=0$ becomes $(\tilde{\Psi},H)=0$ for solutions on the unstable manifold, and for the components of $H$ this implies $$<\tilde{\psi},h_{1}>=0.\label{eq:horthog}$$
The symmetry argument for (\[eq:Hl\]) and (\[eq:Hml\]) can be briefly summarized. Applying (\[eq:Htrans\]) to the Fourier coefficients of $H$ leads to the identity $$H_n(\omega,\alpha,\alpha^\ast)e^{-in\beta}=
H_n(\omega,\alpha e^{-il\beta},\alpha^\ast e^{il\beta}).
\label{eq:ftrans}$$ For $\beta=2\pi/l$ this immediately implies that $H_n=0$ unless $\exp(-i2\pi
n/l)=1$, hence $n$ must be an integer multiple of $l$. Recall that if a function $F(\omega,\alpha,\alpha^\ast)$ is rotation invariant under phase shifts $\alpha\rightarrow\alpha\exp-il\beta$, then it can be written as $F(\omega,\alpha,\alpha^\ast)=g(\omega,\sigma)$ where the form of $g(\omega,\sigma)$ is determined by $F$.[@gss] In this case, the invariant function $(\alpha^\ast)^mH_{ml}(\omega,\alpha,\alpha^\ast)=g(\omega,\sigma)$ must have the form $g(\omega,\sigma)=\sigma^mh(\omega,\sigma)$ to accomodate the factor $(\alpha^\ast)^m$. Dividing by $(\alpha^\ast)^m$ gives $H_{ml}(\omega,\alpha,\alpha^\ast)=\alpha^m\,h(\omega,\sigma)$ as in (\[eq:Hml\]). The general property $\partial_\alpha H(\theta,\omega,0,0)=0$ requires that $h(\omega,\sigma)$ for $H_l$ have the special form $h(\omega,\sigma)=\sigma h_1(\omega,\sigma)$ as in (\[eq:Hl\]).
The equations for $h_m(\omega,\sigma)$ follow by substituting (\[eq:Hl\]) - (\[eq:Hml\]) and (\[eq:ampsym\]) into the equation for $H$ (\[Heqn\]). The results of this straightforward calculation for $m=1, 2$ and $m>2$ are summarized below. In writing these equations two further notations prove convenient. First, define $\Gamma_m$ by $$\Gamma_{m}(\sigma)\equiv
\int_{-\infty}^\infty\,d\omega\,g(\omega)h_{m}(\omega,\sigma),\label{eq:defG}$$ and secondly let ${\rm P}_\perp$ denote the projection operator $$({\rm P}_\perp\phi)(\omega)\equiv \phi(\omega)-<\tilde{\psi},\phi>\psi(\omega)
\label{eq:projop}$$ which projects $\phi$ onto the orthogonal complement of the eigenvector $\psi$.
The $n=l$ component ($m=1$) of (\[Heqn\]) can be written $$\begin{aligned}
[L_l-2p-p^\ast]h_1-(p+p^\ast)\sigma\frac{\partial h_1}{\partial\sigma}&=&
(2\pi iKl)\;{\rm P}_\perp\left\{\rule{0.0in}{0.25in}
f_l(1+\sigma\Gamma_1^\ast)h_2
+f_{2l}^\ast\Gamma_2(\psi^\ast+\sigma h_1^\ast)\right.\nonumber\\
&&\hspace{0.75in}+f_{2l}\,\sigma\,\Gamma_2^\ast \,h_3
+f_{3l}^\ast\sigma\Gamma_3 h_2^\ast\label{eq:m=1}\\
&&\hspace{0.0in}\left.
+\sum_{m=3}^\infty\sigma^{m-1}\left[f_{ml}\Gamma_m^\ast h_{m+1}
+f_{(m+1)l}^\ast\Gamma_{m+1} h_m^\ast\right]\right\}.\nonumber\end{aligned}$$ The $n=2l$ component of (\[Heqn\]) is $$\begin{aligned}
[L_{2l}-2p]h_2-(p+p^\ast)\sigma\frac{\partial h_2}{\partial\sigma}&=&
(2\pi iK(2l))\;\left\{\rule{0.0in}{0.25in}
f_l^\ast[\psi+\sigma(h_1+\Gamma_1\psi)+\sigma^2\Gamma_1h_1]\right.
\label{eq:m=2}\\
&&\hspace{0.1in}+f_l\sigma(1+\sigma\Gamma_1^\ast)h_3
+f_{3l}^\ast\sigma[\Gamma_3\psi^\ast+\sigma\Gamma_3h_1^\ast]
\nonumber\\
&&\hspace{0.1in}\left.
+\sum_{m=2}^\infty\sigma^mf_{ml}\Gamma_m^\ast h_{m+2}
+\sum_{m=4}^\infty\sigma^{m-2}f_{ml}^\ast\Gamma_{m}h_{m-2}^\ast
\right\}.\nonumber\end{aligned}$$ For $m\geq3$, the Fourier components of (\[Heqn\]) are $$\begin{aligned}
[L_{ml}-mp]h_m-(p+p^\ast)\sigma\frac{\partial h_m}{\partial\sigma}&=&
(2\pi iK(ml))\;\left\{\rule{0.0in}{0.25in}
f_l^\ast[h_{m-1}+\sigma \Gamma_1h_{m-1}]
+f_l\sigma(1+\sigma\Gamma_1^\ast)h_{m+1}\right.\nonumber\\
&&\hspace{1.1in}
+\sum_{n=2}^{m-2}f_{(m-n)l}^\ast\Gamma_{m-n}h_n\label{eq:m>2}\\
&&\hspace{0.1in}
+f_{(m-1)l}^\ast\Gamma_{m-1}[\psi+\sigma h_1]
+f_{(m+1)l}^\ast\sigma\Gamma_{m+1}[\psi^\ast+\sigma h_1^\ast]\nonumber\\
&&\hspace{0.1in}\left.
+\sum_{n=2}^{\infty}f_{nl}\sigma^n\Gamma_{n}^\ast h_{m+n}
+\sum_{n=2}^{\infty}f_{(m+n)l}^\ast\sigma^n\Gamma_{m+n}h_n^\ast\right\}.
\nonumber\end{aligned}$$ Here and below, it is understood that sums are omitted when the lower value of the summation index exceeds the upper value, e.g. for $m=3$ the first sum in (\[eq:m>2\]) is omitted.
The symmetry constraints imply that solutions on the unstable manifold (\[eq:umfd\]) have the general form $$\begin{aligned}
\eta^u(\theta,\omega,t)&=&
\left[\rule{0.0in}{0.25in}\alpha(t)\Psi(\theta,\omega) +
\alpha(t)\,\sigma\,h_1(\omega,\sigma(t))\,e^{il\theta}\right.\label{eq:umfdb}\\
&&\hspace{1.0in}\left.+
\sum_{m=2}^\infty\alpha^m(t)\,h_m(\omega,\sigma(t))\,e^{iml\theta}\right]
+ {\rm c.c.}\nonumber\end{aligned}$$ with $\alpha(t)$ satisfying the amplitude equation (\[eq:alph1cm\]). The set of equations (\[eq:ampsym\]) and (\[eq:m=1\]) - (\[eq:m>2\]) determine the functions $p(\sigma)$ and $h_m(\omega,\sigma)$, however they cannot be solved explicitly except using the amplitude expansions introduced in the next section. What has been achieved to this point is to formulate the problem entirely in terms of functions of a single [*real*]{} variable $\sigma$. This simplification is essential to the analysis to follow.
Expansions and recursion relations {#sec:exp}
==================================
Our conclusions regarding the nonlinear development of the instability are based on amplitude expansions for $p$ and $h_m(\omega,\sigma)$, and the corresponding expansion of $\Gamma_m$ from (\[eq:defG\]): $$\begin{aligned}
p(\sigma)&=&\sum_{j=0}^\infty p_j\,\sigma^j\label{eq:pseries}\\
h_m(\omega,\sigma)&=&\sum_{j=0}^\infty h_{m,j}(\omega)\,\sigma^j\\
\Gamma_{m}(\sigma)&=&<g,h_{m}>=\sum_{j=0}^\infty
\Gamma_{m,j}\,\sigma^j\label{eq:defg}\end{aligned}$$ where $\Gamma_{m,j}\equiv\int\,d\omega\,g\,h_{m,j}$. The orthogonality condition (\[eq:horthog\]) implies $$<\tilde{\psi},h_{1,j}>=0\label{eq:orthog}$$ at each order in the expansion of $h_1$.
The coefficients $p_j$ and $h_{m,j}(\omega)$ are determined by recursively solving (\[eq:ampsym\]) and (\[eq:m=1\]) - (\[eq:m>2\]). Relations for $p_j$ follow from (\[eq:ampsym\]) using the general solution form in (\[eq:umfdb\]) and the expansions (\[eq:pseries\]) - (\[eq:defg\]). At each order in $\sigma^j$, $j\geq1$, we solve (\[eq:ampsym\]) for $p_j$ and find $$\begin{aligned}
p_j&=&-(2\pi i K l)\left\{
f_l\left[<\tilde{\psi},h_{2,j-1}>+\sum_{n=0}^{j-2}
\Gamma_{1,n}^\ast<\tilde{\psi},h_{2,j-n-2}>\right]\right.\label{eq:pj}\\
&&\hspace{0.25in}
+f_{2l}^\ast \left[\Gamma_{2,j-1}<\tilde{\psi},\psi^\ast>+\sum_{n=0}^{j-2}
\Gamma_{2,n}<\tilde{\psi},h_{1,j-n-2}^\ast>\right]
+f_{2l}\sum_{n=0}^{j-2}\Gamma_{2,n}^\ast<\tilde{\psi},h_{3,j-n-2}>\nonumber\\
&&\hspace{0.25in}\left.+\sum_{m=3}^{j+1}f_{ml}^\ast\sum_{n=0}^{j-m+1}
\Gamma_{m,n}<\tilde{\psi},h_{m-1,j-n-m+1}^\ast>
+\sum_{m=3}^{j}
f_{ml}\sum_{n=0}^{j-m}\Gamma_{m,n}^\ast<\tilde{\psi},h_{m+1,j-n-m}>
\right\}.\nonumber\end{aligned}$$
The functions $h_{m,j}(\omega)$ are determined similarly by substituting (\[eq:pseries\]) - (\[eq:defg\]) into the component equations (\[eq:m=1\]) - (\[eq:m>2\]). At every order, $h_{m,j}$ is then obtained by applying the resolvent operator (\[eq:resolv\]) to certain auxilliary functions $I_{m,j}(\omega)$ constructed from lower order terms, i.e. $$h_{m,j}=-R_{ml}(\mu_{m,j})\,I_{m,j}(\omega)\label{eq:gen}$$ where $$\mu_{m,j}\equiv(m+j)\lambda+j\lambda^\ast+\delta_{m,1}(\lambda+\lambda^\ast),
\label{eq:arg}$$ and the detailed expressions for $I_{m,j}$ are provided below.
The action of the resolvent is given in (\[eq:resolv\]) $$R_{ml}(\mu_{m,j})\,I_{m,j}=\frac{-i/ml}{\omega+K f_0-\nu_{m,j}-imlD}
\left[I_{m,j}(\omega)-\frac{K f_{ml}^\ast}{\Lambda_{ml}(\nu_{m,j})}
\int_{-\infty}^\infty\frac{d\omega'\;g(\omega')I_{m,j}(\omega')}
{\omega'+K f_0-\nu_{m,j}-imlD}\right]\label{eq:R}$$ with $\nu_{m,j}\equiv i\mu_{m,j}/ml$. It is important to note from (\[eq:arg\]) that $\nu_{m,j}$ can be rewritten $$\nu_{m,j}=z_0+\frac{2i(j+\delta_{m,1})\gamma}{ml};\label{eq:numj}$$ hence for $m\geq1$, the poles $\omega=\nu_{m,j}-K\,f_0+imlD$ in (\[eq:R\]) all sit in the upper half of the complex $\omega$ plane along the line ${\rm Re}(\omega)={\rm Re}(z_0)-K\,f_0$.
The coefficients $p_j$ are determined by the integrals $\Gamma_{m,j}$, $<\tilde{\psi},h_{m,j}>$, $<\tilde{\psi},h_{m,j}^\ast>$, and $<\tilde{\psi},\psi^\ast>$. As we are interested in the form of these coefficients near the onset of the instability ($\gamma\rightarrow0$) when the noise is neglected (i.e. $D=0$), it is important to determine the behavior of such integrals in this limit. In particular, we will need to estimate any singular behavior exhibited by each type of integral.
In the remainder of the paper, our analysis of the amplitude expansion will be facilitated by two identities. From the definition of $\Gamma_{m,j}$ in (\[eq:defg\]) and the resolvent expression for $h_{m,j}(\omega)$ in (\[eq:gen\]), the following useful relations are obtained $$\begin{aligned}
\Gamma_{m,j}&=&\frac{i/ml}{\Lambda_{ml}(\nu_{m,j})}
\int^\infty_{-\infty}\;d\omega
\frac{g(\omega)\,I_{m,j}(\omega)}{(\omega+Kf_0-\nu_{m,j}-imlD)}
\label{eq:identity1}\\
&&\nonumber\\
h_{m,j}(\omega)&=&\frac{iI_{m,j}(\omega)}{ml(\omega+Kf_0-\nu_{m,j}-imlD)}-
\frac{K f_{ml}^\ast\,\Gamma_{m,j}}{(\omega+Kf_0-\nu_{m,j}-imlD)}.
\label{eq:identity2}\end{aligned}$$ Given $I_{m,j}$, these identities allow the properties of $h_{m,j}$ and $\Gamma_{m,j}$ to be quickly determined.
The $I_{m,j}$ are in turn calculated from the following recursion relations in terms of lower order coefficients. The application of these recursion relations to calculate the coefficients $p_j$ is summarized in Table 1 and illustrated in Section \[sec:p1p2\].
Recursion relations for $I_{m,j}$
---------------------------------
The expansion of (\[eq:m=1\]) using (\[eq:pseries\]) - (\[eq:defg\]) determines $h_{1,j}$ at each order in (\[eq:R\]): $h_{1,j}=-R_{l}(\mu_{1,j})\,I_{1,j}$ where for $m=1$ and $j\geq0$, $$\begin{aligned}
I_{1,j}&=&\sum_{k=0}^{j-1} [(2+k)p_{j-k}+(1+k)p_{j-k}^\ast]\,h_{1,k}
\label{eq:i1j}\\
&&\hspace{0.2in}+(2\pi i K l)\;{\rm P}_\perp
\left\{f_{l}\left[h_{2,j}+\sum_{n=0}^{j-1}h_{2,n}\Gamma_{1,j-n-1}^\ast\right]
\right.\nonumber\\
&&\hspace{1.5in}+f_{2l}^\ast\left[\psi^\ast\Gamma_{2,j}+
\sum_{n=0}^{j-1}h_{1,n}^\ast\Gamma_{2,j-n-1}\right]
+f_{2l}\,\sum_{n=0}^{j-1}h_{3,n}\Gamma_{2,j-n-1}^\ast\nonumber\\
&&\hspace{0.5in}\left.
+\sum_{m=3}^{j+2}f_{ml}^\ast\,\sum_{n=0}^{j-m+2}\,
h_{m-1,n}^\ast\Gamma_{m,j-m-n+2}
+\sum_{m=3}^{j+1}f_{ml}\,\sum_{n=0}^{j-m+1}\,h_{m+1,n}\Gamma_{m,j-m-n+1}^\ast
\right\}.
\nonumber\end{aligned}$$ Since $<\tilde{\psi},{\rm P}_\perp\phi>=0$ by construction, this recursion relation together with the orthogonality property of $h_{1,j}$ (\[eq:orthog\]) implies a corresponding orthogonality property for $I_{1,j}$ $$<\tilde{\psi},I_{1,j}>=0.\label{eq:Iorthog}$$
For $m=2$, the expansion of (\[eq:m=2\]) yields $h_{2,j}=-R_{l}(\mu_{2,j})\,I_{2,j}$ where for $j=0$ $$\begin{aligned}
I_{2,0}&=&2\pi i K (2 l)\;f_l^\ast\;\psi(\omega),\label{eq:i20}\end{aligned}$$ and for $j\geq1$, $$\begin{aligned}
I_{2,j}&=&\sum_{k=0}^{j-1}\,[2p_{j-k}+k(p_{j-k}+p_{j-k}^\ast)]\,h_{2,k}
\label{eq:i2j}\\
&&\hspace{0.2in}+2\pi i K (2l)\left\{f_l\left[h_{3,j-1}+
\sum_{n=0}^{j-2}\,h_{3,n}\,\Gamma_{1,j-n-2}^\ast\right]\right.\nonumber\\
&&\hspace{0.5in}+
f_l^\ast\left[h_{1,j-1}+\Gamma_{1,j-1}\,\psi
+\sum_{n=0}^{j-2}\,h_{1,n}\,\Gamma_{1,j-n-2}\right]
+f_{2l}\sum_{n=0}^{j-2}\,h_{4,n}\,\Gamma_{2,j-n-2}^\ast\nonumber\\
&&\hspace{0.5in}
+f_{3l}^\ast\left[\psi^\ast\Gamma_{3,j-1}
+\sum_{n=0}^{j-2}\,h_{1,n}^\ast\,\Gamma_{3,j-n-2}\right]
+f_{3l}\,\sum_{n=0}^{j-3}h_{5,n}\,\Gamma_{3,j-n-3}^\ast\nonumber\\
&&\hspace{0.5in}\left.
+\sum_{k=4}^{j+2}
f_{kl}^\ast\,\sum_{n=0}^{j-k+2}h_{k-2,n}^\ast\,\Gamma_{k,j-n-k+2}
+\sum_{k=4}^{j}
f_{kl}\,\sum_{n=0}^{j-k}h_{k+2,n}\,\Gamma_{k,j-n-k}^\ast\right\}.\nonumber\end{aligned}$$
For $m\geq3$ and $j\geq0$ we have $h_{m,j}=-R_{ml}(\mu_{m,j})\,I_{m,j}$ with $$\begin{aligned}
I_{m,j}&=&\sum_{k=0}^{j-1}\,[mp_{j-k}+k(p_{j-k}+p_{j-k}^\ast)]\,h_{m,k}
\label{eq:imj}\\
&&\hspace{0.2in}+2\pi i K (m l)
\left\{f_l\,\left[h_{m+1,j-1}+
\sum_{k=0}^{j-2}h_{m+1,k}\Gamma_{1,j-k-2}^\ast\right]
+f_l^\ast\,\left[h_{m-1,j}+
\sum_{k=0}^{j-1}h_{m-1,k}\Gamma_{1,j-k-1}\right]\right.\nonumber\\
&&\hspace{0.5in}+f_{(m-1)l}^\ast\,
\left[\psi\Gamma_{m-1,j}+\sum_{k=0}^{j-1}h_{1,k}\Gamma_{m-1,j-k-1}\right]
+\sum_{n=2}^{m-2}f_{nl}^\ast
\sum_{k=0}^{j}h_{m-n,k}\Gamma_{n,j-k}
\nonumber\\
&&\hspace{1.0in}
+f_{(m+1)l}^\ast
\left[\psi^\ast\Gamma_{m+1,j-1}+\sum_{k=0}^{j-2}h_{1,k}^\ast\Gamma_{m+1,j-k-2}
\right]\nonumber\\
&&\hspace{1.5in}\left.
+\sum_{n=2}^{j}f_{nl}\sum_{k=0}^{j-n}
h_{m+n,k}\Gamma_{n,j-n-k}^\ast
+\sum_{n=m+2}^{m+j}f_{nl}^\ast \sum_{k=0}^{j-n+m}
h_{n-m,k}^\ast\Gamma_{n,j-n-k+m}\right\}.\nonumber\end{aligned}$$ In this final expression it is understood that a term with a negative subscript is omitted, e.g. the term $f_l\,h_{m+1,j-1}$ is dropped when $j=0$.
Pinching singularities at low order {#sec:p1p2}
===================================
At this point it is instructive to evaluate the low order coefficients in the expansion and understand how the singularities arise mathematically. We first calculate the cubic coefficient $p_1$ and show that it has a singularity when $f_{2l}\neq0$. Then we evaluate the fifth order coefficient for the special case of a coupling with $f_{2l}=0$ and no singularity in $p_1$.
Evaluation of the coefficient $p_1$ {#sec:p1}
-----------------------------------
The calculation of $p_{1}$ in (\[eq:pj\]) yields $$p_1=-2\pi iKl\left[f_l<\tilde{\psi},h_{2,0}>+
f_{2l}^\ast\Gamma_{2,0}<\tilde{\psi},\psi^\ast>\right]\label{eq:nfcoeff}$$ where the function $h_{2,0}$ is determined from (\[eq:identity2\]) and (\[eq:i20\]) $$h_{2,0}(\omega)=\frac{i\,I_{2,0}}{2l(\omega+Kf_0-z_0-i2lD)}
-\frac{K\,f_{2l}^\ast\,\Gamma_{2,0}}{(\omega+Kf_0-z_0-i2lD)}
\label{eq:cmcoeff}$$ with $\Gamma_{2,0}$ given by (\[eq:identity1\]) $$\Gamma_{2,0}=\frac{i/2l}{\Lambda_{{2l}}\,({z_0})}
\int\,d\omega\frac{g(\omega)\,I_{2,0}(\omega)}{(\omega+Kf_0-z_0-i2lD)}
=\frac{2\pi\,K\,f_{l}^\ast\,\Lambda'_{{l}}\,({z_0})}{\Lambda_{{2l}}\,({z_0})}
+{\cal O}(D).\label{eq:g20}$$ The asymptotic expression for small $D$ in (\[eq:g20\]) shows that generically $\Gamma_{2,0}\neq0$.
The most important qualitative feature of $I_{2,0}(\omega)$ and $h_{2,0}(\omega)$ is that they are meromorphic functions on the complex $\omega$ plane. Let $z_0=(\Omega+i\gamma)/l$ denote the real and imaginary parts of the root, then each function has a pole at $\omega_+=(\Omega/l-Kf_0)+i(l^2D+\gamma)/l$. In addition, $h_{2,0}(\omega)$ has a second pole at $(\Omega/l-Kf_0)+i(2l^2D+\gamma)/l$. Both poles lie in the upper half plane and although they can approach the real axis as the $\gamma\rightarrow0^+$, the integrals $<\tilde{\psi},h_{2,0}>$ and $\Gamma_{2,0}$ in (\[eq:nfcoeff\]) have well-behaved nonsingular limits even if $D$ is zero.
This observation regarding analytic structure can be extended considerably: [*all*]{} of the functions $I_{m,j}(\omega)$ and $h_{m,j}(\omega)$ are meromorphic functions in the complex $\omega$ plane and all of their poles lie along the vertical line ${\rm Re}(\omega)=-Kf_0+{\rm Re}(z_0)$. When $D=0$, all of the poles approach the real axis in the limit $\gamma\rightarrow0^+$; in general, however the poles will be found in both the upper and lower half plane and integrals over $I_{m,j}(\omega)$ or $h_{m,j}(\omega)$ will exhibit pinching singularities. These conclusions follow from the recursion relations in light of two simple observations. First, each $I_{m,j}(\omega)$ is built up as a sum of various $h_{m,j}(\omega)$ from lower order and the sum of two meromorphic functions is again meromorphic. Secondly, the application of the resolvent (\[eq:R\]) to a meromorphic $I_{m,j}(\omega)$ yields a meromorphic function for $h_{m,j}(\omega)$ in (\[eq:identity2\]).
The simplest pinching singularity is illustrated by $<\tilde{\psi},\psi^\ast>$ in $p_1$; from (\[eq:efcn0\]) and (\[eq:adjef\]) one sees that the integrand has poles $\omega_\pm=(\Omega/l-Kf_0)\pm i(l^2D+\gamma)/l$ above and below the contour along the real axis. Now the $\gamma\rightarrow0^+$ limit produces a pinching singularity when $D$ is zero, and this pinch contributes a factor $(\gamma+l^2D)$ to the denominator when the integral is evaluated: $$<\tilde{\psi},\psi^\ast>=
-\frac{lf_l\,{\rm Im}(f_l)}{(\gamma+l^2D)\,|f_l|^2\,\Lambda'_{{l}}\,({z_0})}.
\label{eq:sing}$$ This singularity will occur in the $p_1$ coefficient (\[eq:nfcoeff\]) provided $D\approx0$ and $f_{2l}\neq0$.
The derivation of (\[eq:sing\]) is simple and illustrates the key idea used below in the general analysis: make a partial fraction expansion of the integrand to isolate the singularity. In this case we have $$<\tilde{\psi},\psi^\ast>=\frac{K\,f_l}{\Lambda'_{{l}}\,({z_0})}
\int^\infty_{-\infty}\,d\omega\,
\frac{g(\omega)}{(\omega-\omega_+)(\omega-\omega_-)}.
\label{eq:p1int}$$ The partial fraction expansion yields $$\frac{1}{(\omega-\omega_+)(\omega-\omega_-)}=
\frac{-il/2}{\gamma+l^2D}
\left(\frac{1}{(\omega-\omega_+)}-\frac{1}{(\omega-\omega_-)}\right),
\label{eq:pfex}$$ so that (\[eq:p1int\]) becomes $$<\tilde{\psi},\psi^\ast>=\frac{-ilK\,f_l}{2\Lambda'_{{l}}({z_0})(\gamma+l^2D)}
\left(\frac{\Lambda_{{l}}({z_0})-1}{K\,f_l^\ast}-
\frac{\Lambda_{{l}}({z_0})^\ast-1}{K\,f_l}\right).
\label{eq:p1intb}$$ This reduces to (\[eq:sing\]) since $\Lambda_{{l}}({z_0})=0$. For $D=0$, the cubic coefficient diverges like $p_1\sim\gamma^{-1}$ as $\gamma\rightarrow0^+$ provided $f_{2l}\neq0$; this turns out to reflect a general pattern $p_j\sim\gamma^{-2j+1}$ which is proved below in Theorem \[thm:main\].
The significance of this divergence can be appreciated by considering the amplitude equation truncated at third order $$\dot{\alpha}=\lambda\alpha+p_1\,\alpha|\alpha|^2,$$ and re-examining the determination of $\beta$ in (\[eq:aeqnintro\]). Introducing polar variables $\alpha=A(t)\exp(-i\psi(t))$, the amplitude equation reads $$\begin{aligned}
\dot{A}&=&\gamma A+{\rm Re}(p_1)A^3\hspace{1.0in}
\dot{\psi}=\Omega -{\rm Im}(p_1)A^2\label{eq:polar}\end{aligned}$$ where $\lambda=\gamma-i\Omega.$ Let $p_1=a_1\gamma^{-\nu}$ denote the asymptotic behavior of the cubic coefficient with $\nu\geq0$, then in terms of the rescaled amplitude (\[eq:newvar\]) the equations become $$\begin{aligned}
\frac{d r}{d\tau}&=&\gamma^{1-\delta} r+\gamma^{2\beta-\nu-\delta}{\rm
Re}(a_1)r^3\label{eq:radial}
\hspace{1.0in}
\dot{\psi}=\Omega -\gamma^{2\beta-\nu}{\rm Im}(a_1)r^2.\end{aligned}$$ In order to remove the singular perturbation due to the linear term we again take $\delta=1$ and then solve for $\beta=(1+\nu)/2$ to balance the cubic nonlinearity against the linear instability. This makes it clear that a divergence in the cubic coefficient will increase $\beta$ and for $\nu=1$ we predict $\beta=1$. This of course neglects possible effects from higher order nonlinear terms, in particular singularities at higher order could turn out to be more important than the cubic singularity. The effects of higher order singularities are systematically studied in section \[sec:sing\].
The fifth order coefficient $p_2$ {#sec:p2}
---------------------------------
When $f_{2l}=0$ the cubic coefficient is nonsingular, and we briefly consider what occurs at fifth order in the amplitude expansion. Setting $f_{2l}=0$, the fifth order coefficient is given by $$\begin{aligned}
p_2&=&-(2\pi i K l)\left\{f_l\left[<\tilde{\psi},h_{2,1}>+
\Gamma_{1,0}^\ast<\tilde{\psi},h_{2,0}>\right] +
f_{3l}^\ast\Gamma_{3,0}<\tilde{\psi},h_{2,0}^\ast>\right\}.\label{eq:p2p}\end{aligned}$$ where $$\begin{aligned}
I_{1,0}(\omega)&=&(2\pi i K l)\;{\rm P}_\perp
\left\{f_{l}h_{2,0}\right\}=(2\pi i K l)\,f_{l}\;
\left\{h_{2,0}(\omega)-\psi(\omega)<\tilde{\psi},h_{2,0}>\right\}
\label{eq:I10}\\
I_{3,0}(\omega)&=&2\pi i K (3l)\,
\;f_l^\ast\; h_{2,0}(\omega)\label{eq:I30}\\
I_{2,1}(\omega)&=&2p_1h_{2,0}(\omega)+2\pi i K (2l)\,
\left[f_lh_{3,0}(\omega)+f_l^\ast(h_{1,0}(\omega)+\Gamma_{1,0}\psi(\omega))
\right.\\
&&\hspace{2.5in}\left.+
f_{3l}^\ast\Gamma_{3,0}\psi^\ast(\omega)\right].\nonumber\end{aligned}$$
There are five integrals to consider in (\[eq:p2p\]). Since $h_{2,0}$ has poles only in the upper plane, the integral $<\tilde{\psi},h_{2,0}>$ is not singular. From (\[eq:I30\]), we note that $I_{3,0}$ similarly has poles only in the upper half plane hence $\Gamma_{3,0}$ has no pinching singularity. In addition, $I_{1,0}$ in (\[eq:I10\]) has poles in the upper plane only and this implies that $\Gamma_{1,0}$ is nonsingular, but the calculation is more subtle. From (\[eq:identity1\]) we have $$\Gamma_{1,0}=\frac{i/l}{\Lambda_{l}(\nu_{1,0})}
\int^\infty_{-\infty}\;d\omega
\frac{g(\omega)\,I_{1,0}(\omega)}{(\omega+Kf_0-\nu_{1,0}-ilD)}.
\label{eq:identity10}$$ Since $\Lambda_{l}(\nu_{1,0})=\Lambda_{l}(z_0+2\gamma/l)$ from Eq. (\[eq:numj\]), the first factor in (\[eq:identity10\]) is singular as $\gamma\rightarrow0^+$. However the integral is nonsingular in this limit and because of the orthogonality relation (\[eq:Iorthog\]) the leading term is zero so the integral is actually ${\cal
O}(\gamma)$. The vanishing of the integral as $\gamma\rightarrow0^+$ cancels the singularity from $\Lambda_{l}(\nu_{1,0})$ and $\Gamma_{1,0}$ is well-behaved.
The remaining integrals do exhibit pinching singularities which can be evaluated with the partial fraction procedure illustrated above. Rewriting $p_2$ to isolate these singular terms we find for $D=0$: $$\begin{aligned}
p_2&=&-(2\pi i K l)\left\{f_l\,<\tilde{\psi},h_{2,1}>
+f_{3l}^\ast\Gamma_{3,0}<\tilde{\psi},h_{2,0}^\ast>+\;\;{\rm nonsingular
\;\;terms}\right\}
\label{eq:p2psing}\end{aligned}$$ with $$\begin{aligned}
<\tilde{\psi},h_{2,1}>&=&-2\pi K\,f_{3l}^\ast\Gamma_{3,0}\,
\int_{-\infty}^\infty\,
\frac{d\omega\,\tilde{\psi}(\omega)^\ast\,{\psi}(\omega)^\ast}
{\omega+K\,f_0-(z_0+i\gamma/l)}+\;\;{\rm nonsingular \;\;terms}
\label{eq:t1}\\
&&\nonumber\\
<\tilde{\psi},h_{2,0}^\ast>&=&
-2\pi K\,f_{l}\,
\int_{-\infty}^\infty\,
\frac{d\omega\,\tilde{\psi}(\omega)^\ast\,{\psi}(\omega)^\ast}
{\omega+K\,f_0-z_0^\ast}.\label{eq:t2}\end{aligned}$$ A partial fraction expansion of each of these integrals allows an explicit evaluation of the singular behavior $$\begin{aligned}
\int_{-\infty}^\infty\,
\frac{d\omega\,\tilde{\psi}(\omega)^\ast\,{\psi}(\omega)^\ast}
{\omega+K\,f_0-(z_0+i\gamma/l)}&=&
-\frac{il^2f_l\;[{\rm Im}(f_l)+{\cal O}(\gamma)]}
{3\,|f_l|^2\Lambda'_{l}(z_0)\,\gamma^2}\label{eq:sing1}\\
&&\nonumber\\
\int_{-\infty}^\infty\,
\frac{d\omega\,\tilde{\psi}(\omega)^\ast\,{\psi}(\omega)^\ast}
{\omega+K\,f_0-z_0^\ast}&=&
\frac{il^2f_l\;[{\rm Im}(f_l)+{\cal O}(\gamma)]}
{2\,|f_l|^2\Lambda'_{l}(z_0)\,\gamma^2};\label{eq:sing2}\end{aligned}$$ each integral diverges like $\gamma^{-2}$.
The resulting evaluation of $p_2$ yields $$\begin{aligned}
p_2&=&-\frac{2l(\pi\, l\,K\,f_l)^2\,f_{3l}^\ast\,\Gamma_{3,0}\,
[{\rm Im}(f_l)+{\cal O}(\gamma)]}
{3\,|f_l|^2\,\Lambda'_{l}(z_0)\,\gamma^2}+\;\;{\rm nonsingular \;\;terms};
\label{eq:p2final}\end{aligned}$$ finally, the $\Gamma_{3,0}$ factor is generically nonzero and can be evaluated from (\[eq:identity1\]) as $$\Gamma_{3,0}=
-2(\pi\,K\,f_l^\ast)^2\,\frac{\Lambda''_{l}(z_0)}{\Lambda_{3l}(z_0)}.$$ Hence $p_2\sim\gamma^{-2}$ for the special case $f_{2l}=0$ when $p_1$ is nonsingular. However, if $f_{2l}$ [*and*]{} $f_{3l}$ both vanish, then the divergent terms in (\[eq:p2final\]) disappear and both $p_1$ and $p_2$ are nonsingular.
These conclusions are generalized below. In Proposition \[prop:f2l0\] we show that when $f_{2l}=0$ but $f_{3l}\neq0$, then $p_j\sim\gamma^{-2j+2}$ describes the strongest possible singularity of $p_j$. In Proposition \[prop:kur\] we find that the coefficients $p_j$ are nonsingular at every order if $f_{nl}=0$ for all $n>1$.
The implications of the $p_2$ divergence for the scaling of the amplitude, $\gamma^\beta\,r(\gamma t)$, can be seen by rewriting the radial equation (\[eq:radial\]) to allow for both the third and fifth order terms: $$\frac{d r}{d\tau}=\gamma r+ {\rm Re}(p_1) \gamma^{2\beta-1}r^3
+ {\rm Re}(p_2) \gamma^{4\beta-1}r^3.\label{eq:fifth}$$ When $p_1$ is nonsingular and $p_2\sim\gamma^{-2}$, then we must take $\beta=3/4$ to obtain an expansion that is finite through fifth order terms. In this case however the general singularity result $p_j\sim\gamma^{-2j+2}$ in Proposition \[prop:f2l0\] does not exclude higher order terms that would require a still larger value of $\beta$; this point is discussed further in Section \[sec:newscaling\] below.
Singularity structure of the expansion to all orders {#sec:sing}
====================================================
Our goal is a systematic understanding of the singularities in $p_j$ to all orders in the amplitude expansion. These singularities are smoothed by the noise (cf. the singularity in (\[eq:sing\])), and in the discussion below we always set $D=0$. In this analysis, it is useful to introduce an index which keeps track of the worst case singularity for each coefficient.
Definition of the index
-----------------------
The pinching singularities at every order occur in integrands of a standard form. Let $$D_n(\alpha,\omega)\equiv\frac{1}{(\omega-\alpha_1)(\omega-\alpha_2)
\cdots(\omega-\alpha_n)}
\label{eq:Ddef}$$ for $n>0$ where $\alpha\equiv(\alpha_1,\ldots,\alpha_n)$ and define $D_0(\alpha,\omega)\equiv1$. Evaluating $p_j$ for $j\geq1$ involves integrals of the form $\int d\omega\,\phi(\omega){\cal G}(\omega)$ where $${\cal G}(\omega)=D_m(\beta,\omega)^\ast\,D_n(\alpha,\omega)
\label{eq:indatom}$$ with $m+n\geq1$. Here $\phi(\omega)$ denotes a smooth function assumed to be well behaved as $\gamma\rightarrow0$ and to vanish for $|\omega|\rightarrow\infty$. For example, $\phi(\omega)=g(\omega)$ is often the smooth function of interest. The poles in (\[eq:indatom\]) are given by $$\begin{aligned}
\alpha_j&=&z_0-Kf_0+i\gamma\delta_j\hspace{0.5in}j=
1,\ldots,n\label{eq:poles1}\\
\beta^\ast_j&=&
z_0^\ast-Kf_0-i\gamma\zeta_j\hspace{0.5in}j=1,\ldots,m;\label{eq:poles2}\end{aligned}$$ hence they lie along the vertical line $\omega={\rm Re }(z_0)-Kf_0$ at locations determined by the numbers $\delta_j\geq0$ and $\zeta_j\geq0$ which are assumed to be independent of $\gamma=l\,{\rm Im}(z_0)$.
The [*index*]{} of ${\cal G}(\omega)$ in (\[eq:indatom\]) is defined to be $${\mbox{\rm Ind }} [{\cal G}]\equiv m+n-1.\label{eq:inddef}$$ We extend this definition to assign an index if ${\cal G}(\omega)$ is multiplied by a smooth function $\phi(\omega)$ as described above $${\mbox{\rm Ind }} [\phi(\omega){\cal G}(\omega)]
\equiv{\mbox{\rm Ind }} [{\cal G}(\omega)].\label{eq:inddefb}$$
Since we assume $m+n\geq1$ in (\[eq:indatom\]), ${\mbox{\rm Ind }}
[\phi\,{\cal G}]\geq0$. If $mn\neq0$, then as ${\gamma\rightarrow0^+}$, the integral $\int d\omega\,\phi(\omega){\cal G}(\omega)$ may diverge due to a pinching singularity at $\omega={\rm Re }(z_0)-Kf_0$. The smooth functions $\phi(\omega)$ are always assumed to be sufficiently well behaved at large $|\omega|$ that the only possible divergence is due to the pinching singularity. The index of $\phi\,{\cal G}$ is simply related to the maximum possible strength of this divergence.
\[prop:sing\] For ${\cal G}(\omega)$ in with $m+n\geq1$ and $\phi(\omega)$ a smooth function as described above, the integral of $\phi\,{\cal G}$ satisfies $$\lim_{{\gamma\rightarrow0^+}}\; \left|\gamma^{J}\;
\int^\infty_{-\infty}\,d\omega\,\phi(\omega)\;{\cal
G}(\omega)\right|<\infty\label{eq:sing0}$$ with $J={\mbox{\rm Ind }} [\phi\,{\cal G}]$. If $J$ is replaced by $J-1$, then the limit diverges in general unless $mn=0$ in which case the limit is zero for any $J>0$.
[*[**Proof**]{}.*]{}
> See Appendix \[app:proof\]. [**$\Box$**]{}
Our application of Prop. \[prop:sing\] to the recursion relations for $I_{m,j}$ requires a further generalization of the index to allow for sums of functions with well-defined indices and products of $\phi\,{\cal G}$ with singular functions of $\gamma$. In each case, the generalized index is defined so that Eq. (\[eq:sing0\]) remains true, i.e. the index for the composite function gives the [*maximal*]{} possible divergence of its integral.
First, let $G_1(\omega)$ and $G_2(\omega)$ denote functions with well-defined indices, ${\mbox{\rm Ind }} [{ G}_1]\geq{\mbox{\rm Ind }} [{ G}_2]$. The index of their sum is defined as the largest of the two individual indices: $${\mbox{\rm Ind }} [{ G}_1+{G}_2]\equiv{\mbox{\rm Ind }} [{ G}_1].\label{eq:sum}$$ Clearly, ${\mbox{\rm Ind }} [{ G}_1]$ gives the maximal possible divergence of $\int\,d\omega\,(G_1+G_2)$. Secondly, let $q(\gamma)$ denote a function of $\gamma$ with the asymptotic behavior $$\lim_{\gamma\rightarrow0^+}\;q(\gamma)\sim\gamma^{-\nu},$$ then the index of $q(\gamma)\,{ G_1}(\omega)$ is defined by $${\mbox{\rm Ind }} [q\,{ G_1}]\equiv {\mbox{\rm Ind }}
[{G_1}]+\nu.\label{eq:prod}$$ This completes the definition of the index.
By applying (\[eq:sum\]) and (\[eq:prod\]) the indices of $I_{m,j}$ and $h_{m,j}$ are determined from the recursion relations; some simple examples are provided in section \[sec:examples\]. We stress again that when applied to a composite function $G(\omega)$ the estimate in (\[eq:sing0\]) does not necessarily determine the true singularity, but only an upper bound on the possible divergence of the integral. In many examples considered below this upper bound is actually realized, but not always.
There are two immediate implications of the index definition worth noting. First, complex conjugation doesn’t change the index $${\mbox{\rm Ind }} [{ G}]={\mbox{\rm Ind }} [{ G}^\ast].\label{eq:cc}$$ Secondly, if $G(\omega)$ has a well defined index, then dividing $G$ by $(\omega-\alpha)$ or $(\omega-\beta^\ast)$ simply increases the index of $G$ by one: $${\mbox{\rm Ind }} [{ G}/(\omega-\alpha)]=
{\mbox{\rm Ind }} [{ G}/(\omega-\beta^\ast)]=
{\mbox{\rm Ind }} [{ G}]+1.\label{eq:divide}$$ Here $\alpha, \beta$ introduce additional poles as defined in (\[eq:poles1\]) - (\[eq:poles2\]) and thereby increase the index of every term in $G$ by one; this implies (\[eq:divide\]).
The second observation (\[eq:divide\]) provides a useful relation between the divergence of $\Gamma_{m,j}$ and the divergence of $<\tilde{\psi},h_{m,j}>$. Suppose that $\gamma^{-\nu}$, $\nu>0$, denotes the actual divergence of $\Gamma_{m,j}$. Then, from (\[eq:identity1\]) this must be due to certain terms in $I_{m,j}$; in addition when $m=1$ then $\Lambda_{l}(\nu_{1,j})$ is of order $\gamma$ for small $\gamma$, which increases the divergence by one. From (\[eq:identity2\]), $<\tilde{\psi},h_{m,j}>$ is given by $$\begin{aligned}
<\tilde{\psi},h_{m,j}>&=&\frac{-i}{ml\,\Lambda'_{l}(z_0)}
\int^\infty_{-\infty}\;
\frac{d\omega\,g(\omega)\,I_{m,j}(\omega)}
{(\omega+Kf_0-z_0)(\omega+Kf_0-\nu_{m,j})}\label{eq:int2}\\
&&\nonumber\\
&&\hspace{0.25in}+
\frac{K f_{ml}^\ast\,\Gamma_{m,j}}{\Lambda'_{l}(z_0)}
\int^\infty_{-\infty}\;
\frac{d\omega\,g(\omega)}{(\omega+Kf_0-z_0)(\omega+Kf_0-\nu_{m,j})},
\nonumber\end{aligned}$$ and we conclude that its divergence cannot be worse than $\gamma^{-(\nu+1-\delta{m,1})}$. The second integral in (\[eq:int2\]) has no pinching singularity and the singularity of the second term is determined by $\Gamma_{m,j}\sim\gamma^{-\nu}$. The first term in (\[eq:int2\]) involves an integral whose integrand differs from the integrand in (\[eq:identity2\]) only by the extra factor $(\omega+Kf_0-z_0)$ in the denominator. From (\[eq:divide\]) this factor simply increases the index of every term in $I_{m,j}(\omega)/{(\omega+Kf_0-\nu_{m,j})}$. In particular, the terms that determined the singularity of $\Gamma_{m,j}\sim\gamma^{-\nu}$ in the first place have their index increased by 1. Thus we can infer the divergence of the first term in (\[eq:int2\]) by adding one to $\nu$ and subtracting off the effect of $\Lambda_{ml}(\nu_{m,j})$ if $m=1$; this proves that the singularity of $<\tilde{\psi},h_{m,j}>$ is no worse than $\gamma^{-(\nu+1-\delta{m,1})}$. Note that if $\Gamma_{m,j}$ is not singular ($\nu=0$) then the same reasoning shows that $<\tilde{\psi},h_{m,j}>$ will also be nonsingular. In summary, we have proven
\[prop:prop2\] If the asymptotic behavior of $\Gamma_{m,j}$ satisfies $$\lim_{\gamma\rightarrow0^+}\gamma^{\nu}|\Gamma_{m,j}|<\infty
\label{eq:prop2a}$$ for an integer $\nu>0$, then $$\lim_{\gamma\rightarrow0^+}\gamma^{\nu'}|<\tilde{\psi},h_{m,j}>|<\infty
\label{eq:pjint}$$ with $${\nu'}=\left\{\begin{array}{cc}
\nu& m=1\\
\nu+1&m>1.\end{array}\right.\label{eq:prop2c}$$ If $\nu=0$ in , then $$\lim_{\gamma\rightarrow0^+}|<\tilde{\psi},h_{m,j}>|<\infty.\label{eq:prop2e}$$
This result is useful because $\nu'$ in (\[eq:prop2c\]) often turns out to be less than the index of the integrand of $<\tilde{\psi},h_{m,j}>$ and thus the Proposition can provide sharper estimates of the actual divergences.
It is very important to recognize that the conclusions in (\[eq:pjint\]) and (\[eq:prop2e\]) do not in general apply to $<\tilde{\psi},h_{m,j}^\ast>$. Even though $h_{m,j}$ and $h_{m,j}^\ast$ have the same index, the most singular terms in $<\tilde{\psi},h_{m,j}>$ are not necessarily the most singular terms in $<\tilde{\psi},h_{m,j}^\ast>$. For $<\tilde{\psi},h_{m,j}^\ast>$, if we wish to avoid explicitly evaluating the integral, our only means of estimating the divergence is by calculating the index of the integrand.
Calculation of the index {#sec:examples}
------------------------
These definitions allow the index of $I_{m,j}$ and $h_{m,j}$ to be calculated recursively to all orders. We illustrate this index calculation for the first few levels of Table I; our most general results are stated in Theorem \[thm:main\] and Proposition \[prop:f2l0\].
At the top of Table I we have the eigenfunctions defined by linear theory, these have a single pole in the upper half plane so $${\mbox{\rm Ind }} [\psi(\omega)]=
{\mbox{\rm Ind }} [\tilde{\psi}(\omega)]=0.\label{eq:efcni}$$ Next we consider $I_{2,0}$ and $h_{2,0}$; from (\[eq:i20\]) and (\[eq:cmcoeff\]) we calculate $${\mbox{\rm Ind }} [I_{2,0}(\omega)]=0\hspace{1.0in}
{\mbox{\rm Ind }} [h_{2,0}(\omega)]=1.\label{eq:linind}$$ The index of $I_{2,0}$ follows from (\[eq:efcni\]), while the expression for $h_{2,0}$ has two terms; the first has index 1 and the second has index 0, hence the index of $h_{2,0}$ is 1.
Since $h_{2,0}$ only has poles in the upper half plane, the integrals $\Gamma_{2,0}$ and $<\tilde{\psi},h_{2,0}>$ have no pinching singularities and remain finite as $\gamma\rightarrow0^+$: $$\begin{aligned}
\Gamma_{2,0}&\sim&\gamma^{0}\hspace{1.0in}
<\tilde{\psi},h_{2,0}>\sim\gamma^{0}.\label{eq:exsing}\end{aligned}$$ Note that the integrand in (\[eq:exsing\]) has index equal to 2, but there is no divergence, this illustrates the second part of Proposition \[prop:prop2\]. By contrast the integrand of $<\tilde{\psi},\psi^\ast>$ has a smaller index than $<\tilde{\psi},h_{2,0}>$ but yields a singular integral (\[eq:sing\]) as $\gamma\rightarrow0^+$: $$<\tilde{\psi},\psi^\ast>\sim\gamma^{-1}.\label{eq:cubdiv}$$
The fifth order coefficient $p_2$ provides an instructive exercise in applying the index to estimate the singularity of a more complicated expression. The general form of the coefficient (not assuming $f_{2l}=0$) is $$\begin{aligned}
p_2&=&-(2\pi i K l)\left\{f_l\left[<\tilde{\psi},h_{2,1}>+
\Gamma_{1,0}^\ast<\tilde{\psi},h_{2,0}>\right] +
f_{2l}\,\Gamma_{2,0}^\ast<\tilde{\psi},h_{3,0}> \right.\\
&&\hspace{0.75in}\left.+f_{2l}^\ast
\left[\Gamma_{2,1}<\tilde{\psi},\psi^\ast>+
\Gamma_{2,0}<\tilde{\psi},h_{1,0}^\ast>\right]+
f_{3l}^\ast\Gamma_{3,0}<\tilde{\psi},h_{2,0}^\ast>\right\}\nonumber\end{aligned}$$ where $$\begin{aligned}
I_{3,0}&=&2\pi i K (3l)
\;[f_l^\ast\; h_{2,0}+f_{2l}^\ast\,\psi\,\Gamma_{2,0}]\label{eq:I30b}\\
I_{1,0}&=&(2\pi i K l)\;{\rm P}_\perp
\left\{f_{l}h_{2,0}+f_{2l}^\ast\,\psi^\ast\,\Gamma_{2,0}\right\}
\label{eq:I10b}\\
I_{2,1}&=&2p_1h_{2,0}+2\pi i K (2l)\,
\left[f_lh_{3,0}+f_l^\ast(h_{1,0}+\Gamma_{1,0}\psi)+
f_{3l}^\ast\Gamma_{3,0}\psi^\ast\right].\end{aligned}$$
We can immediately estimate the integral $<\tilde{\psi},h_{2,0}^\ast>$ from the fact that the integrand has index 2 $$<\tilde{\psi},h_{2,0}^\ast>\sim\gamma^{-2}.$$
The two terms in $I_{3,0}$ have indices of 1 and 0, respectively, and the poles of $I_{3,0}$ are all in the upper half plane. Writing out the expression in (\[eq:I10b\]) $${\rm P}_\perp
\left\{f_{l}h_{2,0}+f_{2l}^\ast\,\psi^\ast\,\Gamma_{2,0}\right\}=
f_{l}h_{2,0}+f_{2l}^\ast\,\psi^\ast\,\Gamma_{2,0}-
\psi\,<\tilde{\psi},[f_{l}h_{2,0}+f_{2l}^\ast\,\psi^\ast\,\Gamma_{2,0}]>,$$ we find terms of index 1, 0, and 1, respectively. The third term is already singular due to $<\tilde{\psi},\psi^\ast>$ and the second term has a pole in the lower half plane. These observations imply $$\begin{aligned}
{\mbox{\rm Ind }} [I_{3,0}(\omega)]&=&1\hspace{1.0in}
{\mbox{\rm Ind }} [I_{1,0}(\omega)]\leq1,\label{eq:g10}\end{aligned}$$ and from (\[eq:identity1\]) $$\Gamma_{1,0}\sim\gamma^{-2}\hspace{1.0in}
\Gamma_{3,0}\sim\gamma^{0}.$$ The inequality in (\[eq:g10\]) reflects the fact that there are two terms of index 1 and we have not ruled out the possibility of a cancellation. The indices of $h_{3,0}$ and $h_{1,0}$ are now easily evaluated from (\[eq:identity2\]) $$\begin{aligned}
{\mbox{\rm Ind }} [h_{3,0}(\omega)]&=&2\hspace{1.0in}
{\mbox{\rm Ind }} [h_{1,0}(\omega)]\leq2,\label{eq:h10}\end{aligned}$$ and the corresponding integrals are estimated following (\[eq:pjint\]) $$\begin{aligned}
<\tilde{\psi},h_{3,0}>&\sim&\gamma^{0}\hspace{1.0in}
<\tilde{\psi},h_{1,0}>\sim\gamma^{-2}.\label{eq:h10int}\end{aligned}$$ For $p_2$ we also need to estimate $<\tilde{\psi},h_{1,0}^\ast>$ and this is done by noting from (\[eq:h10\]) that the integrand has an index of (at most) 3 so $$<\tilde{\psi},h_{1,0}^\ast>\sim\gamma^{-3}.$$
The five terms in $I_{2,1}$ all have index 2 except $\Gamma_{3,0}\psi^\ast$ which has index 0, hence $${\mbox{\rm Ind }} [I_{2,1}(\omega)]\leq2.$$ The terms involving $h_{1,0}$ and $\psi^\ast$ have poles in the lower half plane and therefore produce pinching singularities in $\Gamma_{2,1}$. By comparing (\[eq:identity1\]) for $\Gamma_{2,1}$ to (\[eq:cubdiv\]) and (\[eq:h10int\]) we can estimate a maximum singularity of $$\Gamma_{2,1}\sim\gamma^{-2}$$ which in turn implies ${\mbox{\rm Ind }} [h_{2,1}(\omega)]\leq3,$ and $$<\tilde{\psi},h_{2,1}>\sim\gamma^{-3}\label{eq:term1}$$ from (\[eq:identity2\]) and (\[eq:pjint\]), respectively.
The result in (\[eq:term1\]) gives the divergence of the first term in $p_2$, the asymptotic behavior of the remaining five terms can be evaluated from the foregoing estimates: $$\begin{aligned}
\Gamma_{1,0}^\ast<\tilde{\psi},h_{2,0}>&\sim&\gamma^{-2}\\
\Gamma_{2,0}^\ast<\tilde{\psi},h_{3,0}>&\sim&\gamma^{0}\\
\Gamma_{2,1}<\tilde{\psi},\psi^\ast>&\sim&\gamma^{-3}\\
\Gamma_{2,0}<\tilde{\psi},h_{1,0}^\ast>&\sim&\gamma^{-3}\\
\Gamma_{3,0}<\tilde{\psi},h_{2,0}^\ast>&\sim&\gamma^{-2}.\end{aligned}$$ Hence barring some accidental cancellation among the most divergent terms we expect $p_2\sim\gamma^{-3}$ for the general case when $f_{2l}\neq0$; more precisely we have proved $$\lim_{\gamma\rightarrow0^+}\gamma^3|p_2|<\infty.\label{eq:p2gen}$$
The main result
---------------
Our main result on the singularity structure of the amplitude expansions can now be proved. In the preceeding sections, we have calculated the indices of $I_{m,j}$ and $h_{m,j}$ through the first two levels of Table I, and determined the singular behavior of the cubic and fifth order coefficients. This information is now inferred for the entire theory to all orders.
\[thm:main\] For $j\geq1$, the singularities of the coefficients in the amplitude expansion satisfy $$\lim_{{\gamma\rightarrow0^+}}\;
\gamma^{2j-1}\;\left|p_j\right|<\infty.\label{eq:pjasy}$$ For $m\geq1$ and $j\geq0$, the indices of $I_{m,j}$ and $h_{m,j}$ obey $$\begin{aligned}
{\mbox{\rm Ind }} \left[I_{m,j}\right]&\leq&
J_{m,j}+2\delta_{m,1}.\label{eq:keyind}\\
{\mbox{\rm Ind }} \left[h_{m,j}\right]&\leq&
J_{m,j}+1+2\delta_{m,1}\label{eq:hind}\end{aligned}$$ where $J_{m,j}\equiv m+2j-2$, and the integrals in [(\[eq:pj\])]{} satisfy $$\begin{aligned}
\lim_{{\gamma\rightarrow0^+}}\;\gamma^{J_{m,j}+3\delta_{m,1}}\;
\left|\Gamma_{m,j}
\right|&<&\infty\label{eq:gmklasy}\\
\lim_{{\gamma\rightarrow0^+}}\;\gamma^{J_{m,j}+1+2\delta_{m,1}}\;
\left|<\tilde{\psi},h_{m,j}>\right|
&<&\infty.\label{eq:intest}\end{aligned}$$
We are only able to prove the results in (\[eq:keyind\]) - (\[eq:hind\]) as upper bounds on the index, but we expect them to hold as equalities in most cases. Note that the bounds in (\[eq:gmklasy\]) - (\[eq:intest\]) give weaker divergences than one would estimate simply from the index of the integrand as calculated using (\[eq:keyind\]) - (\[eq:hind\]). For example, from (\[eq:hind\]) the integrand of $<\tilde{\psi},h_{m,j}>$ has index $J_{m,j}+2+2\delta_{m,1}$, but the bound in (\[eq:intest\]) assures us that the divergence of this integral will always be at least one power of $\gamma$ less. There is no contradiction since the index only gives a strict upper limit on divergence exhibited by a given integral.
[*[**Proof.**]{}*]{}
> 1. The relations (\[eq:pjasy\]) - (\[eq:intest\]) have been checked explicitly for $I_{2,0}$, $h_{2,0}$, $\Gamma_{2,0}$, and $p_1$ in the analysis of section \[sec:p1\]. In addition, in section \[sec:examples\] they have been verified for $I_{1,0}$, $h_{1,0}$, $\Gamma_{1,0}$, $I_{3,0}$, $h_{3,0}$, $\Gamma_{3,0}$, $I_{2,1}$, $h_{2,1}$, $\Gamma_{2,1}$, and $p_2$. This proves the theorem for the first two levels of Table I.
>
> 2. We extend (\[eq:pjasy\]) - (\[eq:intest\]) to all $p_j$, and $\{h_{m,j}$, $I_{m,j}\}$ by induction using the recursion relations. Assume that (\[eq:pjasy\]) - (\[eq:intest\]) are true down to some arbitrary level of Table I, and consider what the recursion relations imply for the coefficients evaluated at the next level. Let $h_{m',j'}$ and $I_{m',j'}$ denote coefficients at the next level that can be evaluated from lower order quantities which satisfy (\[eq:pjasy\]) - (\[eq:intest\]). We first consider the various recursion relations for $h_{m',j'}$ and $I_{m',j'}$, and prove that if (\[eq:pjasy\]) - (\[eq:intest\]) hold on the right hand side of these relations, then (\[eq:keyind\]) - (\[eq:intest\]) will also hold for the coefficients obtained on the left. We organize this part of the proof by noting the importance of establishing two relations for $I_{m',j'}$:
>
> 1. the index identity (\[eq:keyind\]) $${\mbox{\rm Ind }} \left[I_{m',j'}\right]\leq
> J_{m',j'}+2\delta_{m',1}.\label{eq:keyind2}$$
>
> 2. the estimate $$\lim_{{\gamma\rightarrow0^+}}\;\gamma^{J_{m',j'}+2\delta_{m',1}}
> \left|\int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,I_{m',j'}(\omega)}{(\omega-\alpha)}\right|<\infty
> \label{eq:indstep}$$ where $\alpha$ is any pole of the form described in (\[eq:poles1\]) (in the upper half plane). Note that (\[eq:keyind2\]) implies that the integrand in (\[eq:indstep\]) has an index of $J_{m',j'}+2\delta_{m',1}+1$; thus (\[eq:indstep\]) gives a divergence which is less by one power of $(1/\gamma)$ than would be “naively” predicted using the index of $I_{m',j'}$ in (\[eq:keyind2\]).
>
> If we can prove (\[eq:keyind2\]) and (\[eq:indstep\]) for $I_{m',j'}$, then the remaining properties (\[eq:hind\]) - (\[eq:intest\]) are easily obtained as follows. First, (\[eq:gmklasy\]) follows from (\[eq:indstep\]) and the identity (\[eq:identity1\]). Next (\[eq:hind\]) follows by applying (\[eq:keyind2\]) and (\[eq:gmklasy\]) to the identity (\[eq:identity2\]). Finally (\[eq:intest\]) follows from Proposition \[prop:prop2\] and (\[eq:gmklasy\]). Thus the crux of the matter is to verify (\[eq:keyind2\]) and (\[eq:indstep\]); this is done using the recursion relations for $I_{m',j'}$, i.e. using (\[eq:i1j\]) - (\[eq:imj\]).
>
> 3. [*Verification of [(\[eq:keyind2\])]{}*]{}.
>
> The index of each term on the right hand side of the recursion relations for $I_{m',j'}(\omega)$ can be evaluated by applying (\[eq:pjasy\]) - (\[eq:intest\]); this exercise shows that all of these terms have an index less than or equal to $J_{m',j'}+2\delta_{m',1}$. This establishes (\[eq:keyind2\]) for $I_{m',j'}$. A few examples from (\[eq:i1j\]) illustrate these index calculations.
>
> 1. The terms in $I_{1,j}$ that depend on $p_j$ have the form $[(2+k)p_{j-k}+(1+k)p_{j-k}^\ast]\,h_{1,k}$; from (\[eq:hind\]) we have ${\mbox{\rm Ind }} \left[h_{1,k}\right]\leq2k+2$ and from (\[eq:pjasy\]) the singularity of $p_{j-k}$ is determined, hence $${\mbox{\rm Ind }} \left[p_{j-k}\,h_{1,k}\right]\leq2(j-k)-1+(2k+2)=2j+1
> \label{eq:index1}$$ which is consistent with (\[eq:keyind2\]) for $(m',j')=(1,j)$.
>
> 2. Next consider the terms $$\sum_{m=3}^{j+2}f_{ml}^\ast\,\sum_{n=0}^{j-m+2}\,{\rm P}_\perp
> [h_{m-1,n}^\ast\Gamma_{m,j-m-n+2}]
> \label{eq:secondex}$$ which expand to $${\rm P}_\perp
> [h_{m-1,n}^\ast\Gamma_{m,j-m-n+2}]=
> \Gamma_{m,j-m-n+2}\,[h_{m-1,n}^\ast-
> <\tilde{\psi},h_{m-1,n}^\ast>\,\psi].\label{eq:index3}$$ The index of each term can be evaluated $$\begin{aligned}
> {\mbox{\rm Ind }} \left[h_{m-1,n}^\ast\Gamma_{m,j-m-n+2}\right]&\leq&2j
> \label{eq:index3a}\\
> {\mbox{\rm Ind }}
> \left[<\tilde{\psi},h_{m-1,n}^\ast>\,\Gamma_{m,j-m-n+2}\,\psi\right]
> \label{eq:index3b}
> &\leq&2j+1;\end{aligned}$$ thus ${\mbox{\rm Ind }} \left[{\rm P}_\perp
> [h_{m-1,n}^\ast\Gamma_{m,j-m-n+2}]\right]\leq2j+1$ which is consistent with (\[eq:keyind2\]) for $(m',j')=(1,j)$. In this example, both indicies are obtained from (\[eq:hind\]) and (\[eq:gmklasy\]). The divergence of $\Gamma_{m,j-m-n+2}\sim\gamma^{-(2j-m-2n+2)}$ is estimated from (\[eq:gmklasy\]), and (\[eq:hind\]) gives ${\mbox{\rm Ind }} \left[h_{m-1,n}^\ast\right]\leq
> m+2n-2$. Hence the integrand of $<\tilde{\psi},h_{m-1,n}^\ast>$ has an index no greater than $m+2n-1$, and we conclude that the worst possible divergence is $<\tilde{\psi},h_{m-1,n}^\ast>\sim\gamma^{-(m+2n-1)}$. These estimates immediately imply (\[eq:index3a\]) - (\[eq:index3b\]).
>
> 3. A more subtle example is the term ${\rm P}_\perp\,h_{2,j}$ in (\[eq:i1j\]); from (\[eq:projop\]) we have that the index of ${\rm P}_\perp\,h_{2,j}=h_{2,j}-\psi<\tilde{\psi},h_{2,j}>$ is equal to the largest index obtained from $h_{2,j}$ and $<\tilde{\psi},h_{2,j}>\,\psi(\omega)$. These terms have indices $${\mbox{\rm Ind }} \left[h_{2,j}\right]\leq2j+1\hspace{0.5in}
> {\mbox{\rm Ind }} \left[<\tilde{\psi},h_{2,j}>\,\psi(\omega)\right]\leq2j+1,
> \label{eq:index2}$$ so ${\mbox{\rm Ind }} \left[{\rm P}_\perp\,h_{2,j}\right]\leq2j+1$ which is consistent with (\[eq:keyind2\]) for $(m',j')=(1,j)$. The first index in (\[eq:index2\]) comes from (\[eq:hind\]) and the second follows from (\[eq:intest\]) and (\[eq:linind\]). This example is delicate because (\[eq:index2\]) requires that we use (\[eq:intest\]) to estimate $<\tilde{\psi},h_{2,j}>\sim\gamma^{-(2j+1)}$ rather than (\[eq:hind\]) as in the previous calculation (\[eq:index3b\]).
>
> Proceeding in this way the index of every term in the recursion relations (\[eq:i1j\]) - (\[eq:imj\]) can be calculated and (\[eq:keyind2\]) verified.
>
> 4. [*Verification of [(\[eq:indstep\])]{}*]{}.
>
> We prove (\[eq:indstep\]) by integrating the recursion relations for $I_{m',j'}(\omega)$ and examining the resulting recursion relations for $$\int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,I_{m',j'}(\omega)}{(\omega-\alpha)}.
> \label{eq:crux}$$ The right hand side of the recursion relations (\[eq:i1j\]) - (\[eq:imj\]) involve products of $\psi$, $\psi^\ast$, $h_{m,j}$, or $h_{m,j}^\ast$ with integrals whose singularities can be estimated by applying (\[eq:pjasy\]) - (\[eq:intest\]) as illustrated above. Now when we divide by $(\omega-\alpha)$ and integrate as in (\[eq:crux\]) we get additional integrals of the following types $$\begin{aligned}
> \int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,\psi(\omega)}{(\omega-\alpha)}&\hspace{0.3in}&
> \int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,\psi^\ast(\omega)}{(\omega-\alpha)}\label{eq:inta}\\
> &&\nonumber\\
> \int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,h_{m,j}^\ast(\omega)}{(\omega-\alpha)}
> &\hspace{0.3in}&\int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,h_{m,j}(\omega)}{(\omega-\alpha)}.\label{eq:intb}\end{aligned}$$ The first integral in (\[eq:inta\]) has no pinching singularity and the second has index 1 and diverges as $\gamma^{-1}$; the singularity of the integrals in (\[eq:intb\]) can be estimated from (\[eq:hind\]) and (\[eq:intest\]), respectively: $$\begin{aligned}
> \lim_{{\gamma\rightarrow0^+}}\;\gamma^{J_{m,j}+2+2\delta_{m,1}}
> \left|\int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,h_{m,j}^\ast(\omega)}{(\omega-\alpha)}\right|&<&\infty
> \label{eq:intd}\\
> \lim_{{\gamma\rightarrow0^+}}\;\gamma^{J_{m,j}+1 +2\delta_{m,1}}
> \left|\int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,h_{m,j}(\omega)}{(\omega-\alpha)}\right|&<&\infty.
> \label{eq:intc}\end{aligned}$$ With these estimates the terms contributing to (\[eq:crux\]) can be shown to diverge at worst like $$\left(\frac{1}{\gamma}\right)^{J_{m',j'}+2\delta_{m',1}}.$$ This establishes (\[eq:indstep\]) for $I_{m',j'}$, and completes the induction argument that (\[eq:keyind\]) - (\[eq:intest\]) are propagated by the recursion relations.
>
> 5. Now we finish the proof of the theorem by verifying that the estimates for the integrals in (\[eq:gmklasy\]) and (\[eq:intest\]) imply the estimate of $p_j$ in (\[eq:pjasy\]); this is a straightforward application of (\[eq:gmklasy\]) and (\[eq:intest\]) to the right hand side of the recursion relations (\[eq:pj\]). A few examples from (\[eq:pj\]) suffice to illustrate how this is done.
>
> 1. Consider the term $<\tilde{\psi},h_{2,j-1}>$; from (\[eq:intest\]) the singularity of this integral is at most $$\left|<\tilde{\psi},h_{2,j-1}>\right|\sim
> \left(\frac{1}{\gamma}\right)^{J_{2,j-1}+1}=
> \left(\frac{1}{\gamma}\right)^{2j-1}
> \label{eq:ex1}$$ which is consistent with (\[eq:pjasy\]). This example is another case where the divergence of the integral is best characterized with (\[eq:intest\]) rather than inferred from the index of the integrand which is $2j$.
>
> 2. Other terms involve products of integrals, for example, $\Gamma_{2,j-1}<\tilde{\psi},\psi^\ast>$. Using (\[eq:gmklasy\]) and (\[eq:cubdiv\]) we determine that the singularity of this product is at most $$\left|\Gamma_{2,j-1}<\tilde{\psi},\psi^\ast>\right|\sim
> \left(\frac{1}{\gamma}\right)^{J_{2,j-1}+1}
> \label{eq:ex2}$$ which is the same as in (\[eq:ex1\]).
>
> 3. The terms $$\sum_{m=3}^{j+1}f_{ml}^\ast\sum_{n=0}^{j-m+1}
> \Gamma_{m,n}<\tilde{\psi},h_{m-1,j-n-m+1}^\ast>;
> \label{eq:exfinal}$$ are a final example. Applying (\[eq:hind\]) and (\[eq:gmklasy\]) shows a maximum singularity of $$\left|\Gamma_{m,n}<\tilde{\psi},h_{m-1,j-n-m+1}^\ast>\right|\sim
> \left(\frac{1}{\gamma}\right)^{J_{m,n}+J_{m-1,j-n-m+1}+2}=
> \left(\frac{1}{\gamma}\right)^{2j-1}
> \label{eq:ex3}$$ which is also consistent with (\[eq:pjasy\]).
>
> In this manner every term in the recursion relation for $p_j$ is shown to satisfy the general estimate in (\[eq:pjasy\]).
>
> [**$\Box$**]{}
Implications of Theorem \[thm:main\]
------------------------------------
Our discussion in section \[sec:p1\] of the significance of the coefficient singularities did not consider terms in the amplitude equation beyond third order. If all such terms are included then the amplitude equation in polar variables (\[eq:polar\]) becomes $$\begin{aligned}
\dot{A}&=&\gamma A+\sum_{j=1}^\infty{\rm Re}(p_j)A^{2j+1}\\
\dot{\psi}&=&\Omega -\sum_{j=1}^\infty{\rm Im}(p_j)A^{2j}.\end{aligned}$$ With the asymptotic behavior of the coefficients is given by Theorem \[thm:main\] $$p_j=\frac{[a_j+{\cal O}(\gamma)]}{\gamma^{2j-1}},$$ we re-express the dynamics in terms of the rescaled amplitude (\[eq:newvar\]) ${A}(t)=\gamma^\beta r(\gamma^\delta t)$ $$\begin{aligned}
\frac{d r}{d\tau}&=&\gamma^{1-\delta}
r+\sum_{j=1}^\infty\gamma^{2j(\beta-1)-\delta+1}{\rm Re}(a_j)[1+{\cal
O}(\gamma)] r^{2j+1}\label{eq:rdot}\\
\frac{d \psi}{d t}&=&\Omega -\sum_{j=1}^\infty\gamma^{2j(\beta-1)+1}{\rm
Im}(a_j)[1+{\cal O}(\gamma)]r^{2j}.\end{aligned}$$ Now we see that the singularities of $p_j$ are such that the choices $\delta=1$ and $\beta=1$ remove them to [*all*]{} orders, leaving an asymptotically finite system $$\begin{aligned}
\frac{d r}{d\tau}&=& r+\sum_{j=1}^\infty{\rm Re}(a_j)[1+{\cal O}(\gamma)]
r^{2j+1}\label{eq:reqn}\\
\frac{d \psi}{d t}&=&\Omega -\gamma\sum_{j=1}^\infty{\rm Im}(a_j)[1+{\cal
O}(\gamma)]r^{2j}.\label{eq:psieqn}\end{aligned}$$
In (\[eq:reqn\]) there is a further remarkable consequence of the singularities, namely that the terms at higher order in $r$ no longer appear to be negligible even for small $\gamma$. This conclusion is provisional in the sense that we have not proven $a_j\neq0$ at high order. Nevertheless, our analysis of $p_1$ and $p_2$ shows that in general $a_1$ and $a_2$ are non-zero, and inspection of the recursion relations shows no evidence that the series will truncate. This feature does not affect the phase evolution (\[eq:psieqn\]) quite so dramatically. If we assume a time-independent solution $r=R_{rw}$ of (\[eq:reqn\]) then the nonlinear frequency of this state $$\Omega_{rw}=\Omega -\gamma\sum_{j=1}^\infty{\rm Im}(a_j)[1+{\cal
O}(\gamma)]R_{rw}^{2j}\label{eq:rwave}$$ is very close to the linear frequency $\Omega$ for small $\gamma$. Such solutions are called “rotating wave” states as they describe a solution (\[eq:umfdb\]) $$\begin{aligned}
\eta^u(\theta,\omega,t)&=&
\left[\rule{0.0in}{0.25in}\gamma^\beta\,R_{rw}e^{-i\Omega_{rw}t}
\Psi(\theta,\omega) +
\gamma^{3\beta}\,R_{rw}^3e^{-i\Omega_{rw}t}\,
h_1(\omega,\gamma^{2\beta}\,R_{rw}^2)\,e^{il\theta}\right.\label{eq:rwsoln}\\
&&\hspace{1.0in}\left.+
\sum_{m=2}^\infty\gamma^{m\beta}\,R_{rw}^me^{-im\Omega_{rw}t}\,
h_m(\omega,\gamma^{2\beta}\,R_{rw}^2)\,e^{iml\theta}\right]
+ {\rm c.c.}\nonumber\end{aligned}$$ that appears time independent in a rotating frame $\theta'=\theta-\Omega_{rw}t/l$.
The asymptotic behavior of $\Gamma_{m,j}$ is given by Theorem \[thm:main\] as $$\Gamma_{m,j}=\frac{[b_{m,j}+{\cal O}(\gamma)]}{\gamma^{J_{m,j}+3\delta_{m,1}}}$$ where the constants $b_{m,j}$ are independent of $\gamma$. The expansion of $\Gamma_m$ thus has the asymptotic form $$\Gamma_{m}(\sigma)=\sum_{j=0}^\infty
\Gamma_{m,j}\,\sigma^j=\left(\frac{1}{\gamma}\right)^{m-2+3\delta_{m,1}}
\sum_{j=0}^\infty[b_{m,j}+{\cal O}(\gamma)]r^{2j}.
\label{eq:asym}$$ This motivates the definition of the nonsingular coefficient $\hat{\Gamma}_{m}$ $$\Gamma_{m}(\sigma)\equiv\left(\frac{1}{\gamma}\right)^{m-2+3\delta_{m,1}}
\hat{\Gamma}_{m}(r^{2});
\label{eq:ghat}$$ $\hat{\Gamma}_{m}$ will play a role shortly in our discussion of Daido’s order parameter.
Special Cases: $f_{2l}=0$ {#sec:special}
=========================
The general analysis shows that if $f_{2l}\neq0$, then the resulting singularities require a rescaling of the amplitude by a factor of $\gamma$. In this section, we consider some of the possiblities that can occur for couplings with $f_{2l}=0$. First, we note that the general recursion relations simplify somewhat when $f_{2l}=0$; the terms proportional to $f_{2l}$ are explicitly shown in (\[eq:pj\]) and (\[eq:i1j\]) - (\[eq:imj\]), and we simply drop them.
Absence of the cubic singularity: $f_{2l}=0$ and $f_{3l}\neq0$
--------------------------------------------------------------
When $f_{2l}=0$ so that the cubic coefficient is non-singular, then if $f_{3l}\neq0$ the first singularity appears in the fifth order coefficient $p_2\sim\gamma^{-2}$. The general structure of the amplitude expansion for this situation is characterized by the following proposition.
\[prop:f2l0\] Suppose that $f_{2l}=0$ and $f_{3l}\neq0$, then for $j\geq1$, the singularities of the coefficients in the amplitude expansion satisfy $$\lim_{{\gamma\rightarrow0^+}}\;
\gamma^{2j-2}\;\left|p_j\right|<\infty.\label{eq:pjasysc}$$ For $m\geq1$ and $j\geq0$, the indices of $I_{m,j}$ and $h_{m,j}$ obey $$\begin{aligned}
{\mbox{\rm Ind }} \left[I_{m,j}\right]&\leq&
J_{m,j}+2\delta_{m,1}\label{eq:keyindsc}\\
{\mbox{\rm Ind }} \left[h_{m,j}\right]&\leq&
J_{m,j}+1+2\delta_{m,1}\label{eq:hindsc}\end{aligned}$$ where $J_{m,j}\equiv m+2j-2$, and the integrals in [(\[eq:pj\])]{} satisfy $$\begin{aligned}
\lim_{{\gamma\rightarrow0^+}}\;
\gamma^{J_{m,j}+3\delta_{m,1}-(1-\delta_{m,2}\delta_{j,0})}\;
\left|\Gamma_{m,j}\right|&<&\infty\label{eq:gmklasysc}\\
\lim_{{\gamma\rightarrow0^+}}\;\gamma^{J_{m,j}+2\delta_{m,1}}\;
\left|<\tilde{\psi},h_{m,j}>\right|
&<&\infty.\label{eq:intestsc}\end{aligned}$$
Compared to the conclusions of Theorem \[thm:main\] the bounds on the indices are the same so there is nothing to prove. The estimates of the singularities in (\[eq:pjasysc\]) and (\[eq:gmklasysc\]) - (\[eq:intestsc\]) are softened by one factor of $\gamma$. Since $J_{2,0}=0$, there is no singularity in $\Gamma_{2,0}$ even in the general case and the odd correction term $\delta_{m,2}\delta_{j,0}$ in (\[eq:gmklasysc\]) adjusts the exponent to allow for this. In fact, $\Gamma_{m,0}$, $m>1$ is nonsingular in general, although we have not used this.
[*[**Proof.**]{}*]{}
> 1. The results in (\[eq:pjasysc\]) - (\[eq:intestsc\]) have been established by the calculations in sections \[sec:p1\] and \[sec:p2\]; they are extended to all orders by induction. The argument is entirely analogous to the proof given for Theorem \[thm:main\] and we indicate only the necessary modifications and omit the details.
>
> 2. The only adjustment required is that the second crucial identity (\[eq:indstep\]) in the general proof is now replaced by the estimate $$\lim_{{\gamma\rightarrow0^+}}\;\gamma^{J_{m',j'}+2\delta_{m',1}
> -(1-\delta_{m',2}\delta_{j',0})}
> \left|\int^\infty_{-\infty}\;d\omega
> \frac{g(\omega)\,I_{m',j'}(\omega)}{(\omega-\alpha)}\right|<\infty
> \label{eq:indstepsc}$$ where $\alpha$ is any pole of the form described in (\[eq:poles1\]) (in the upper half plane). The first index identity (\[eq:keyind2\]) is still appropriate: $${\mbox{\rm Ind }} \left[I_{m',j'}\right]\leq
> J_{m',j'}+2\delta_{m',1}.\label{eq:keyind2sc}$$ With this adjustment the repetition of the previous argument is straightforward.
>
> [**$\Box$**]{}
Implications for scaling {#sec:newscaling}
------------------------
The consequences of this Theorem for the scaling of the amplitude are ambiguous. If we rewrite the radial equation from (\[eq:rdot\]) with $\delta=1$ $$\frac{d r}{d\tau}= r+\sum_{j=1}^\infty\gamma^{2j\beta-1}{\rm Re}(p_j)
r^{2j+1},\label{eq:rnewa}$$ and then insert the estimated singularity $p_j\sim a_j\gamma^{-2j+2}$ we find $$\frac{d r}{d\tau}= r+\sum_{j=1}^\infty\gamma^{2j(\beta-1)+1}{\rm Re}(a_j)
r^{2j+1}.\label{eq:rnewb}$$ Now the requirement that all the nonlinear terms should be nonsingular, $2j(\beta-1)+1\geq0$, results in a $j$-dependent estimate on the exponent: $\beta\geq1-1/2j$. We know the fifth order singularity occurs so from $j=2$ we have $\beta=3/4$ as discussed previously in Section \[sec:p2\]. However our analysis of the recursion relations is not sharp enough to determine if the leading singularities are actually present at every order. Consequently, we have a range of possible values of $\beta$ $$\frac{3}{4}\leq\beta\leq 1.\label{eq:range}$$
This range can be reduced to $\beta=1$ if we require that the rescaling also yield a finite expansion for $\Gamma$ with coefficient singularities as estimated in (\[eq:gmklasysc\]). However it must be borne in mind that these singularities are upper bounds. Our firmest prediction remains of the range of values in (\[eq:range\]) since the bound $\beta\geq3/4$ is conclusively required by our fifth order calculations and the the bound $\beta\leq1$ follows from our analysis of the general case.
Single component couplings
--------------------------
An additional case of interest is the circumstance which holds for the Kuramoto model $f(\phi)=\sin\phi$; a coupling that lacks all harmonics of the Fourier component $f_l$ driving the instability. In this exceptional case the amplitude expansion is entirely free of singularities.
If $f_{nl}=0$ for $n>1$, then the coefficients $p_j$ in [(\[eq:pj\])]{} are all nonsingular:\[prop:kur\] $$\lim_{\gamma\rightarrow0^+}\,|p_j|<\infty$$ for $j\geq0$.
[*[**Proof.**]{}*]{}
> The functions $I_{m,j}(\omega)$ and $h_{m,j}(\omega)$ have poles in the upper half plane only and consequently the integrals appearing in the coefficients $p_j$ are all free of pinching singularities in the limit $\gamma\rightarrow0^+$. Also $\Gamma_{m,j}$ is nonsingular from (\[eq:identity1\]) in light of the discussion following (\[eq:identity10\]). Hence there are no singularities.
>
> [**$\Box$**]{}
The absence of singularities results immediately in the conclusion $\beta=1/2$ for the scaling exponent.
Evaluation of Daido’s order function {#sec:daido}
====================================
Daido’s definition of the order function $H(\theta)$ is a direct generalization of the familiar order parameter of Kuramoto and is motivated by the mean field form of the phase dynamics (\[eq:gcoupled\]).[@dainew; @dainew96; @dai2] He assumes there is a component of the population collectively entrained with frequency $\Omega_e$ and introduces the shifted phases $\psi_j\equiv\theta_j-\Omega_e t$, then phase equations (\[eq:gcoupled\]) can be rewritten $$\dot{\psi}_i=\Delta_i-H({\psi}_i,t) +
\xi_i(t)\label{eq:gcoupledmf}$$ where $\Delta_i\equiv\omega_i-\Omega_e$ and $H(\theta,t)$ is the mean field $$H(\theta,t)\equiv -\sum_{n=-\infty}^{\infty}\,f_n\,e^{-in\theta}
\left[\frac{1}{N}\sum_{j=1}^{N}\,e^{in(\theta_j(t)-\Omega_e t)}\right].
\label{eq:mf}$$ The additive noise term $\xi_i(t)$ in (\[eq:gcoupled\])is omitted in the analysis, and for large $N$, Daido assumes $\Omega_e$ can be chosen so that the limit $$H(\theta)\equiv \lim_{t\rightarrow\infty}H(\theta,t)\label{eq:daido}$$ exists. The time-asymptotic mean field then defines the order function $H(\theta)$. A norm of $H$, $$\|H\|^2=\int_0^{2\pi}d\theta\,H(\theta)^2/2\pi,$$ serves as an order parameter for models with multi-component couplings.
It is interesting to evaluate $H$ using our continuum description. The average over the population in (\[eq:mf\]) can be expressed in terms of $\rho(\theta,\omega,t)$: $$\lim_{N\rightarrow\infty}
\frac{1}{N}\sum_{j=1}^{N}\,e^{in(\theta_j(t)-\Omega_e t)}
=\int_0^{2\pi}\,d\theta\int_{-\infty}^{\infty}\,d\omega\,g(\omega)
\rho(\theta,\omega,t) e^{in(\theta-\Omega_e t)},$$ and this yields a general expression for the order function $$H(\theta)\equiv -\lim_{t\rightarrow\infty}
\sum_{n=-\infty}^{\infty}\,f_n\,e^{-in\theta}
\int_0^{2\pi}\,d\theta'\int_{-\infty}^{\infty}\,d\omega'\,g(\omega')
\rho(\theta',\omega',t) e^{in(\theta'-\Omega_e t)}.\label{eq:ofcn}$$ Let $H^u$ denote the order function obtained from solutions on the unstable manifold (\[eq:umfdb\]); a straightforward substitution from (\[eq:umfdb\]) then yields $$\begin{aligned}
H^u(\theta)/2\pi&=& -\lim_{t\rightarrow\infty}
\left\{f_l\,e^{-il(\theta+\Omega_et)}\left[\alpha(t)^\ast+
\alpha(t)^\ast\sigma(t)\int_{-\infty}^{\infty}\,d\omega \,g(\omega)
h_1(\omega,\sigma(t))^\ast\right]\right.\label{eq:ofcnum}\\
&&\hspace{0.25in}\left.
+\sum_{m=2}^\infty\,f_{ml}\,e^{-iml(\theta+\Omega_et)}
(\alpha(t)^\ast)^m\int_{-\infty}^{\infty}\,d\omega \,g(\omega)
h_m(\omega,\sigma(t))^\ast)+{\rm cc}\right\};\nonumber\end{aligned}$$ note that we have set $f_0=0$ for simplicity.
Daido’s assumption of time-independent behavior (\[eq:daido\]) is naturally satisfied when the system is described by a “rotating wave” solution of the form (cf. (\[eq:rwsoln\])) $$\alpha_{rw}(t)=\gamma^\beta\,R_{rw}\,e^{-i\phi_0}\,e^{-i\Omega_{rw}t}
\label{eq:rw}$$ with $R_{rw}$ and $\Omega_{rw}$ the time-independent amplitude and frequency, respectively. For these asymptotic states by choosing $\Omega_e=\Omega_{rw}/l$, the limit in (\[eq:ofcnum\]) exists and the order function is $$\frac{H^u_{rw}}{2\pi}= \left[
f_l^\ast\,\gamma^\beta
R_{rw}\,e^{-i\phi_0}\,(1+\gamma^{2\beta}R_{rw}^2\Gamma_1)
e^{il\theta}+\sum_{m=2}^\infty\,
f_{ml}^\ast\,\gamma^{m\beta}R_{rw}^{m}\,e^{-im\phi_0}
\,\Gamma_m\,e^{iml\theta}\right]
+{\rm cc}$$ with norm given by $$\|H\|^2=(2\pi)^2\gamma^{2\beta}R_{rw}^2\,\left[|f_l|^2\,
(1+\gamma^{2\beta}R_{rw}^2\Gamma_1)^2
+\sum_{m=2}^\infty\,|f_{ml}|^2\,
\gamma^{(2m-2)\beta}R_{rw}^{2m-2}\,\Gamma_m^2\right].\label{eq:rwnorm}$$
Our results on the possible values of $\beta$ imply corresponding predictions for the scaling of $\|H\|$ near onset. There are several different cases depending on the coupling and whether or not noise is included; a summary appears in Table II.
1. Suppose that either $D>0$, or $D=0$ and we have a “single component” coupling: $f_l\neq0$ and $f_{ml}=0$ for $m>1$. Then as $\gamma\rightarrow0^+$, there are no singularities and (\[eq:rwnorm\]) gives $$\|H\|=2\pi\gamma^{\beta}R_{rw}\,|f_l|\,\left[1+{\cal O}(\gamma^{2\beta})\right]
\sim(K-K_c)^\beta\label{eq:rwnormkur}$$ with $\beta=1/2$. In each of these circumstances, the prediction $\|H\|\sim(K-K_c)^{1/2}$ is in agreement with Daido’s analysis of $\|H\|$.[@dainew96]
2. Suppose that $D=0$ and we have the general case with $f_l\neq0$ and $f_{2l}\neq0$. The coefficients of the amplitude expansion have the generic singularities (\[eq:pjasy\]) and the coefficients $\Gamma_m$ are also singular (\[eq:ghat\]). Substituting (\[eq:ghat\]) into (\[eq:rwnorm\]) we find $$\|H\|^2=(2\pi)^2\gamma^{2\beta}R_{rw}^2\,\left[|f_l|^2\,
(1+\gamma^{2(\beta-1)}R_{rw}^2\hat{\Gamma}_1(R_{rw}^2))^2
+\sum_{m=2}^\infty\,|f_{ml}|^2\,\gamma^{2(m-1)(\beta-1)}
R_{rw}^{2m-2}\,\hat{\Gamma}_m^2\right];\label{eq:rwnormsing}$$ in this case we have determined that $\beta=1$ which yields $$\|H\|=2\pi\gamma\, R_{rw}\,
\left[|f_l|^2\,(1+R_{rw}^2\hat{\Gamma}_1)^2
+\sum_{m=2}^\infty\,|f_{ml}|^2\,R_{rw}^{2m-2}\,\hat{\Gamma}_m^2\right]^{1/2}
\sim(K-K_c)^1.\label{eq:rwnormfinal}$$ This is the scaling for $\|H\|$ Daido found for transitions at $l=1$ when there was no noise and the coupling had components at $l=2$. We obtain this result for a transition at arbitrary $l$ if the coupling satisfies $f_{2l}\neq0$ and $D=0$.
3. Suppose that $D=0$ and we have the special case with $f_l\neq0$, $f_{2l}=0$, and $f_{3l}\neq0$. In this case, our analysis of the recursion relations leaves considerable uncertainty in the value of $\beta$ as indicated by the range (\[eq:range\]).
4. Suppose that $D=0$ and we have the special case with O(2) symmetry and also a coupling that satisfies $f(\phi)=-f(\phi+\pi/l)$. This requires $f_n=0$ unless $n$ is an odd multiple of $l$ and the non-zero components must be imaginary. In this case again one has typically $f_l\neq0$ and $f_{3l}\neq0$, and our analysis of the recursion relations specifies $\beta$ only within the range (\[eq:range\]). However this range nevertheless [*excludes*]{} the value $\beta=1/2$ Daido obtains for such symmetric systems.[@dainew; @dai2; @dai93b] The reason for this apparent discrepancy has not been found.
Discussion
==========
Our analysis shares a key conclusion with Daido’s study of scaling based on the order function. In the absence of noise, the presence of a coupling component $f_{2l}$ at the first harmonic of the critical mode number $l$ will slow the onset of synchronized behavior so that the phase-locked component of the population scales like $(K-K_c)$ rather than $(K-K_c)^{1/2}$ as in the single-component coupling models. However, both approaches suffer a common deficiency; neither provides a simple explanation of [*why*]{} the onset of synchronization is slower if the coupling has a second harmonic component ($f_{2l}\neq0$). This shift seems to be a true many-body effect requiring large $N$, but a dynamical mechanism has not been identified. Perhaps further numerical studies will shed some light on this issue.
For a general coupling with $f_{2l}\neq0$, our calculations show two effects of adding noise to the phase dynamics. The diffusion term tends to suppress the linear instability so that $K_c$ increases, but in addition the noise prevents the poles from actually reaching the real axis and therefore removes the coefficient singularities; this modifies the scaling $(K-K_c)^\beta$ of the synchronized state changing $\beta$ from $1$ to $1/2$. The interaction of these two effects raises the intriguing possiblity that under some circumstances, adding noise could actually yield an enhancement in the level of synchronization. In some recent studies of mean field oscillator models it has been noted that noise can enchance a collective oscillation.[@hr94]
The linear stability analysis of section \[sec:linear\] clearly shows that for multi-component couplings one can expect multiple linear instabilities and this will allow for studies of the interaction between two different synchronizing transitions. In other settings the study of simultaneous linear instabilities, or codimension-two mode interactions, has been one of the keys to obtaining an analytical theory capable of describing the system beyond the threshold of the first instability. Such a theory can identify possibilities of secondary bifurcation and predict the appearance of states with nontrivial time-dependence. Despite the widespread interest in synchronization, and the large literature on the Kuramoto model, there has been no analysis of the codimension-two bifurcations that arise when more than one mode of synchronization is possible.
Although the justification for a phase model varies from one context to another, in the setting of coupled oscillators it requires two essential assumptions: weak coupling and a frequency spread $g(\omega)$ that is “sufficiently” small.[@erkop; @kur; @ashwin] Moving beyond these assumptions necessitates restoring the dynamics of the oscillator amplitudes $r_i(t)$, and this adds an additional term to the kinetic equation of the schematic form $$\frac{\partial}{\partial r}\cdot\left(\dot{r}\rho\right)$$ where $r=(r_1,\ldots,r_N)$ are the oscillator amplitudes and $\dot{r}$ is the non-stochastic component of the amplitude dynamics. In this situation, $\rho$ describes the joint distribution of oscillator phases and amplitudes. The effect of such a term on the synchronizing transition can be studied by analyzing the effect on the singularities of the nonlinear coefficients.
Acknowledgements
================
This work supported by NSF grant PHY-9423583.
Proof of Proposition \[prop:sing\] {#app:proof}
==================================
Let $z_0=(\Omega+i\gamma)/l$ denote the root in (\[eq:poles1\]) - (\[eq:poles2\]). If $mn=0$ then there is no pinching singularity and the integral in (\[eq:sing0\]) has a finite limit; hence the limit is zero if $J>0$.
Assume $mn>0$, then for integrals with $m+n=2$ the limit can be evaluated from the Plemej formula[@musk] after a partial fraction expansion. The expansion isolates the singularity at $\gamma=0$, $$\int^\infty_{-\infty}\,\frac{d\omega\,\phi(\omega)}{(\omega-\alpha_1)
(\omega-\beta_1^\ast)}=\frac{-i/\gamma}{\delta_1+\zeta_1+2/l}
\left[\int^\infty_{-\infty}\,\frac{d\omega\,\phi(\omega)}{(\omega-\alpha_1)}-
\int^\infty_{-\infty}\,\frac{d\omega\,\phi(\omega)}{(\omega-\beta_1^\ast)}
\right],\label{eq:pfexp}$$ and the limit (\[eq:sing0\]) yields $$\begin{aligned}
\lim_{\gamma\rightarrow0^+}\gamma
\left|\int^\infty_{-\infty}\,\frac{d\omega\,\phi(\omega)}{(\omega-\alpha_1)
(\omega-\beta_1^\ast)}\right|&=&
\frac{1}{(\delta_1+\zeta_1+2/l)}\lim_{\gamma\rightarrow0^+}\left|
\int^\infty_{-\infty}\,\frac{d\omega\,\phi(\omega)}{(\omega-\alpha_1)}-
\int^\infty_{-\infty}\,\frac{d\omega\,\phi(\omega)}{(\omega-\beta_1^\ast)}
\right|\nonumber\\
&&\nonumber\\
&=&\frac{2\pi|\phi(\Omega/l-Kf_0)|}{(\delta_1+\zeta_1+2/l)}.\end{aligned}$$
The partial fraction expansion in (\[eq:pfexp\]) expresses an integrand of index 1 in terms of integrands of index 0. More generally, for integrands $$\int^\infty_{-\infty}\,d\omega\,\phi(\omega)\;{\cal G}(\omega)$$ with index $m+n-1>1$, by expanding the integrand in partial fractions, they can be re-expressed as a sum of two integrands with index $m+n-2$ that are multiplied by an overall factor of $\gamma^{-1}$. Thus, if the partial fraction integrands of index $m+n-2$ satisfy (\[eq:sing0\]) with $J=m+n-2$, then the original integrand will satisfy (\[eq:sing0\]) with $J=m+n-1$; the value of $J$ must be incremented by one to allow for the factor of $\gamma^{-1}$.
A simple induction argument along these lines establishes (\[eq:sing0\]) for all $m+n\geq1$. In the limit in (\[eq:sing0\]) is found to be non-zero in general so, barring an accidental cancellation, if $J$ is replaced by $J-1$ then the modified limit will diverge as $\gamma^{-1}$.
Derivation of the model {#app:deriv}
=======================
Basically, the kinetic equation arises as the first member of a coupled hierarchy of equations quite similar to the BBGKY hierarchy well known in the kinetic theory of gases and this is probably the most elementary systematic approach. There are alternative derivations using path integral methods.[@bonilla]
The solution of the system of $N$ stochastic differential equations in (\[eq:gcoupled\]) is a Markov process with transition probability ${\cal
P}({\bf\theta},{\bf\omega},t|{\bf\theta}',{\bf\omega}',t')$ where ${\bf\theta}=(\theta_1,\ldots,\theta_N)$ and ${\bf\omega}=(\omega_1,\ldots,\omega_N)$.[@doer] This transition probability satisfies a linear Fokker-Planck equation $$\frac{\partial {\cal P}}{\partial t}+\frac{\partial}{\partial {\bf\theta}}\cdot
\left(\dot{{\bf\theta}}{\cal P}\right)=
D\frac{\partial}{\partial
{\bf\theta}}\cdot\frac{\partial{\cal P}}{\partial {\bf\theta}}\label{eq:fp}$$ subject to the initial condition $${\cal P}({\bf\theta},{\bf\omega},t'|{\bf\theta}',{\bf\omega}',t')
=\delta^N({\bf\theta}-{\bf\theta}')\,\delta^N({\bf\omega}-{\bf\omega}').$$ The notation $\dot{{\bf\theta}}=(\dot{\theta_1},\ldots,\dot{\theta}_N)$ in (\[eq:fp\]) refers only to the non-stochastic part of the dynamics $$\dot{\theta}_i=\omega_i+\frac{K}{N}\sum^{N}_{j=1} f(\theta_j-\theta_i).$$
Given an initial normalized ensemble of populations ${\cal
E}_0({\bf\theta},{\bf\omega})$, the ensemble for $t>t_0$ is determined by $${\cal E}({\bf\theta},{\bf\omega},t)=\int\, d{\bf\omega}'\;\int
\,d{\bf\theta}'\;
{\cal P}({\bf\theta},{\bf\omega},t|{\bf\theta}',{\bf\omega}',t_0)\;
{\cal E}_0({\bf\theta}',{\bf\omega}');$$ equivalently, this is can be written as an evolution equation for ${\cal
E}({\bf\theta},{\bf\omega},t)$ $$\frac{\partial {\cal E}}{\partial t}+\frac{\partial}{\partial {\bf\theta}}\cdot
\left(\dot{{\bf\theta}}{\cal E}\right)=
D\frac{\partial}{\partial
{\bf\theta}}\cdot\frac{\partial{\cal E}}{\partial {\bf\theta}}\label{eq:fp2}$$ with initial condition ${\cal E}_0$. The original phase dynamics (\[eq:gcoupled\]) is unchanged if we exchange a pair of oscillators $(\theta_i,\omega_i)
\leftrightarrow(\theta_j,\omega_j)$ and it is natural to select initial ensembles ${\cal E}_0$ with this same invariance under pairwise exchange. In addition, for our problem, the values of the frequencies $\omega_i$ are assumed to obey a fixed distribution $g(\omega)$. This requirement necessitates a suitable choice of initial ensemble: $${\cal E}_0({\bf\theta},{\bf\omega})=
G_{0}({\bf\theta},{\bf\omega})\prod_{j=1}^{N} g(\omega_j)
\label{eq:ens}$$ where the normalization of ${\cal E}_0({\bf\theta},{\bf\omega})$ implies $$1=\int\, d{\bf\theta}\;G_{0}({\bf\theta},{\bf\omega}).$$
A reduced distribution $\rho_s$ for the ensemble is defined in the usual way $$\rho_s(\theta_1,\ldots,\theta_s,\omega_1,\ldots,\omega_s,t)\equiv
\int\,d\theta_{s+1}\cdots\,d\theta_{N}\int\,d\omega_{s+1}\cdots\,d\omega_{N}\,
{\cal E}({\bf\theta},{\bf\omega},t),\label{eq:rdist}$$ and one can derive a hierarchy of coupled equations expressing the evolution of $\rho_s$ in terms of $\rho_s$ and $\rho_{s+1}$ by performing the integration in (\[eq:rdist\]) on the evolution equation (\[eq:fp2\]). For our purposes it is sufficient to find the equation for $\rho_1$; integrating (\[eq:fp2\]) over $(\theta_2,\ldots,\theta_N)$ and $(\omega_2,\ldots,\omega_N)$ yields $$\begin{aligned}
\lefteqn{\frac{\partial \rho_1}{\partial t}
+\frac{\partial}{\partial\theta_1}
\left[\rule{0in}{0.25in}(\omega_1+{Kf(0)}/{N})\rho_1
+\frac{K(N-1)}{N}\int d\theta_2\int d\omega_2 f(\theta_2-\theta_1)
\rho_2(\theta_1,\theta_2,\omega_1,\omega_2,t)\right]}
\hspace{0.5in}\nonumber\\
&&\hspace{4.25in}=D\frac{\partial^2 \rho_1 }{\partial \theta_1^2}.\end{aligned}$$ In terms of the two-oscillator correlation function, $$C(\theta_1,\theta_2,\omega_1,\omega_2,t)\equiv
\rho_1(\theta_1,\omega_1,t)\,\rho_1(\theta_2,\omega_2,t)
-\rho_2(\theta_1,\theta_2,\omega_1,\omega_2,t),$$ this becomes $$\begin{aligned}
\lefteqn{\frac{\partial \rho_1}{\partial t}
+\frac{\partial}{\partial\theta_1}
\left[\rule{0in}{0.25in}(\omega_1+{Kf(0)}/{N})\rho_1
+\frac{K(N-1)\rho_1}{N}
\int d\theta_2\int d\omega_2 f(\theta_2-\theta_1)
\,\rho_1(\theta_2,\omega_2,t)\right]=}\hspace{0.25in}\nonumber\\
&&\hspace{0.25in}D\frac{\partial^2 \rho_1 }{\partial
\theta_1^2}+\frac{\partial}{\partial\theta_1}
\left[\rule{0in}{0.25in}\frac{K(N-1)}{N}\int d\theta_2\int d\omega_2
f(\theta_2-\theta_1)
C(\theta_1,\theta_2,\omega_1,\omega_2,t)\right].\end{aligned}$$
We obtain an autonomous equation for $\rho_1$ by letting $N\rightarrow\infty$ and discarding the correlation term $$\begin{aligned}
\frac{\partial \rho_1}{\partial t}
+\frac{\partial}{\partial\theta_1}
\left[\rule{0in}{0.25in}\omega_1\,\rho_1
+{K\,\rho_1}\,
\int d\theta_2\int d\omega_2 f(\theta_2-\theta_1)
\,\rho_1(\theta_2,\omega_2,t)\right]&=&D\frac{\partial^2 \rho_1 }{\partial
\theta_1^2}.\label{eq:kinetic}\end{aligned}$$ In some cases, one can rigorously prove that with a mean field interaction the effects of correlations are negligible in the limit $N\rightarrow\infty$.[@zwan; @daw; @bonilla]) Whether this kind of result holds for the models considered here appears to be an unsettled issue. Finally, the ensemble choice in (\[eq:ens\]) implies that $$\int_0^{2\pi}\,d\theta_1\,\rho_1(\theta_1,\omega_1,t_0)=g(\omega_1),$$ so it is natural to define $\rho(\theta_1,\omega_1,t)$ by $$\rho_1(\theta_1,\omega_1,t)\equiv g(\omega_1)
\rho(\theta_1,\omega_1,t).$$ Substituting this expression for $\rho_1$ into (\[eq:kinetic\]) yields the kinetic equation (\[eq:eveqn\]) - (\[eq:vel\]).
It is interesting to note that the choice of initial ensemble (\[eq:ens\]) can be adapted to treat other forms of quenched randomness.[@bonilla2] For example, we can allow for random variation in the coupling strength or even in the sign of the coupling between different pairs of phases.
[99]{}
G.B. Ermentrout and N. Kopell, Frequency plateaus in a chain of weakly coupled oscillators, I, [*SIAM J. Math Anal.*]{} [**15**]{} (1984) 215.
Y. Kuramoto, [**Chemical Oscillations, Waves, and Turbulence**]{}, Springer-Verlag, New York (1984).
P. Ashwin and J.W. Swift, The dynamics of $n$ weakly coupled identical oscillators, [*J. Nonlinear Sci.*]{} [**2**]{} 69.
A.T. Winfree, Biological rhythms and the behavior of populations of coupled oscillators, [ J. Theor. Biol.]{} [**16**]{} (1967) 15-42.
A.T. Winfree, [**Geometry of Biological Time**]{}, Springer-Verlag, New York (1990).
S.H. Strogatz, Norbert Wiener’s Brain Waves, in Lecture Notes in Biomathematics, Vol. 100, ed. S. Levin, (Springer Verlag, New York, 1994).
C.M. Gray, P. Konig, A.K. Engel and W. Singer, Oscillatory responses in cat visual cortex exhibit inter-columnar synchronization which reflects global stimulus properties, [Nature]{} [**338**]{} (1989) 334-337.
K. Wiesenfeld, P. Colet, S.H. Strogatz, Synchronization transitions in a disordered Josephson series array, [Phys. Rev. Lett.]{} [**76**]{} (1996) 404.
J.W. Swift, S.H. Strogatz, and K. Wiesenfeld, Averaging of globally coupled oscillators, [Phys. Rev. E]{} [**51**]{} (1995) 1020.
K. Wiesenfeld and J.W. Swift, Averaged equations for Josephson junction series arrays, [Physica D]{} [**55**]{} (1992) 239.
Y. Kuramoto and I. Nishikawa, Statistical macrodynamics of large dynamical systems. Case of a phase transition in oscillator communities, [*J. Stat. Phy.*]{} [**49**]{} (1987) 569-605.
D Hansel, G. Mato, and C. Meunier, Phase dynamics for weakly coupled Hodgkin-Huxley neurons, [ Europhysics. Lett.]{} [**23**]{} (1993) 367-372.
H. Daido, Generic scaling at the onset of macroscopic mutual entrainment in limit-cycle oscillators with uniform all-to-all coupling, [ Phys. Rev. Lett.]{} [**73**]{} (1994) 760-763.
J.D. Crawford, Scaling and singularities in the entrainment of globally coupled oscillators, [ Phys. Rev. Lett.]{} [**74**]{} (1995) 4341-4344.
H. Daido, Onset of cooperative entrainment in limit-cycle oscillators with uniform all-to-all interactions: bifurcation of the order function, [Physica D]{} [**91**]{} (1996) 24-66.
H. Sakaguchi, Cooperative phenomena in coupled oscillator systems under external fields, [ Prog. Theor. Phys.]{} [**79**]{} (1988) 39-46.
S.H. Strogatz and R. Mirollo, Stability of incoherence in a population of coupled oscillators, [ J. Stat. Phys.]{} [**63**]{} (1991) 613-635.
S.H. Strogatz, R. Mirollo and P.C. Matthews, Coupled nonlinear oscillators below the synchronization threshold: relaxation by generalized Landau damping, [ Phys. Rev. Lett.]{} [**68**]{} (1992) 2730-2733.
J.D. Crawford, Amplitude expansions for instabilities in populations of globally-coupled oscillators, [ J. Stat. Phys]{} [**74**]{} (1994) 1047-1084.
J.D. Crawford, Introduction to bifurcation theory, [*Rev. Mod. Phys.*]{} [**63**]{} (1991) 991-1037.
A. Vanderbauwhede and G. Iooss, 1992, Centre manifold theory in infinite dimensions, [**Dynamics Reported**]{}, Vol. 1, Springer-Verlag, New York, 125-163.
J.D. Crawford, Universal trapping scaling on the unstable manifold of a collisionless electrostatic mode, [ Phys. Rev. Lett.]{} [**73**]{} (1994) 656-659.
J.D. Crawford, Amplitude equations for electrostatic waves: universal singular behavior in the limit of weak instability, [ Phys. Plasmas]{} [**2**]{} 97-128 (1995).
J.D. Crawford and A. Jayaraman, Nonlinear saturation of electrostatic waves: mobile ions modify trapping scaling, [ Phys. Rev. Lett.]{} [**77**]{} 3549 (1996).
K. Case, Stability of inviscid plane Couette flow, [ Phys. Fl.]{} [**3**]{} 143 (1960).
R.J. Briggs, J.D. Daugherty, and R.H. Levy, Role of Landau damping in crossed-field electron beams and inviscid shear flow, [ Phys. Fl.]{} [**13**]{} 421-432 (1970).
R. Pego and M.I. Weinstein, [ Phil. Trans. R. Soc. Lond. A]{} [**340**]{} (1992) 47-94.
R. Pego and M.I. Weinstein, in [*Differential Equations with Applications to Mathematical Physics*]{}, W.F. Ames, E.M. Harrell II, and J.V. Herod, eds., Academic Press, Orlando, 1993. pp. 273-286.
R. Pego, P. Smereka, and M.I. Weinstein, Oscillatory instability of solitary waves in a continuum model of lattice vibrations, [Nonlinearity]{}, [**8**]{} 921-941 (1995).
G. Russo and P. Smereka, Kinetic theory for bubbly flow I: collisionless case, [SIAM J. Appl. Math.]{}, [**56**]{} (1996).
H. Daido, Critical conditions of macroscopic mutual entrainment in uniformly coupled limit-cycle oscillators, [ Prog. Theor. Phys.]{} [**89**]{} (1993) 929-934.
J.D. Crawford and P.D. Hislop, Application of the method of spectral deformation to the Vlasov-Poisson system, [*Ann. Phys.*]{} [**189**]{} (1989) 265-317.
M. Golubitsky, I. Stewart, and D.G. Schaeffer, [**Singularities and Groups in Bifurcation Theory**]{}: Vol. II, Appl. Math. Sci. [**69**]{}, Springer-Verlag, New York (1988).
H. Daido, Order function and macroscopic mutual entrainment in uniformly coupled limit-cycle oscillators, [ Prog. Theor. Phys.]{} [**88**]{} (1992) 1213-1218.
H. Daido, A solvable model of coupled limit-cycle oscillators exhibiting partial perfect synchrony and novel frequency spectra,, [Physica D]{} [**69**]{} (1993) 394-403.
V. Hakim and W.-J. Rappel, Noise-induced periodic behavior in the globally coupled complex Ginzburg-Landau equation, [*Europhys. Lett.*]{} [**27**]{} (1994) 637.
N. I. Muskhelishvili, [**Singular Integral Equations**]{} (Noordhoff, Groningen, The Netherlands, 1953), pp. 56-61.
L.L. Bonilla, Stable nonequilibrium probability densities and phase transitions for mean-field models in the thermodynamic limit, [*J. Stat. Phys.*]{} [**46**]{} (1987) 659.
C.R. Doering, Modelling complex systems: stochastic processes, stochastic differential equations, and Fokker-Planck equations, in [**Lectures in Complex Systems**]{}, Eds. L. Nadel and D. Stein, (Addison Wesley, New York, 1991), pp. 3-51.
Rashmi C. Desai and Robert Zwanzig, Statistical mechanics of a nonlinear stochastic model, [*J. Stat. Phy.*]{} [**19**]{} (1978) 1-24.
D.A. Dawson, Critical dynamics and fluctuations for a mean-field model of cooperative behavior, [*J. Stat. Phy.*]{} [**31**]{} (1983) 29.
L.L. Bonilla, Glassy synchronization in a population of coupled oscillators, [*J. Stat. Phys.*]{} [**70**]{} (1993) 921.
[l|ccccccc]{} &$m=1$&$m=2$&$m=3$&$m=4$&$m=5$&$m=6$&$\cdots$\
\
$p_0$& $\psi(\omega)$ & & & & & &\
$p_1$& - &$h_{2,0}$& & & & &\
&$h_{1,0}$& &$h_{3,0}$& & & &\
$p_2$& &$h_{2,1}$& & & & &\
&$h_{1,1}$& & &$h_{4,0}$& & &\
$p_3$& &$h_{2,2}$&$h_{3,1}$& & & &\
&$h_{1,2}$& & & &$h_{5,0}$& &\
$p_4$& &$h_{2,3}$& $h_{3,2}$&$h_{4,1}$ && &\
&$h_{1,3}$& & & &&$h_{6,0}$ &\
$p_5$& &$h_{2,4}$& $h_{3,3}$&$h_{4,2}$ &$h_{5,1}$& &\
&& & & && &\
$\vdots$&&& & && &\
\[table1\]
[lccccc]{} Noise&Symmetry&Coupling&$\beta$&$\beta_D $ (for $l=1$)&Comment\
\
$D>0$&SO(2) or O(2)& arbitrary & 1/2 & - & generic case with noise\
\
$D=0$&SO(2) or O(2)& $f_{2l}\neq0$ & 1 & 1 & generic case without noise\
\
$D=0$&SO(2) or O(2)& $f_{2l}=0$, $f_{3l}\neq0$ & \[3/4,1\] & - &\
\
$D=0$&O(2)& $f(\phi)=-f(\phi+\pi/l)$ & \[3/4,1\] & 1/2& $\beta$ and $\beta_D$ differ\
\
$D=0$&SO(2) or O(2)& $f_{ml}=0,\,m>1$ & 1/2 & 1/2 & single component case\
\
\[table2\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper introduces nucleus clustering in Voronoï tessellations of plane surfaces with applications in the geometry of digital images. A *nucleus cluster* is a collection of Voronoï regions that are adjacent to a Voronoï region called the cluster nucleus. Nucleus clustering is a carried out in a strong proximity space. Of particular interest is the presence of maximal nucleus clusters in a tessellation. Among all of the possible nucleus clusters in a Voronoï tessellation, clusters with the highest number of adjacent polygons are called *maximal nucleus clusters*. The main results in this paper are that strongly near nucleus clusters are strongly descriptively near and every collection of Voronoï regions in a tessellation of a plane surface is a Zelins’kyi-Soltan-Kay-Womble convexity structure.'
address:
- 'Computational Intelligence Laboratory, University of Manitoba, WPG, MB, R3T 5V6, Canada and Department of Mathematics, Faculty of Arts and Sciences, Adiyaman University, 02040 Adiyaman, Turkey'
- ' Department of Mathematics, Faculty of Arts and Sciences, Adiyaman University, 02040 Adiyaman, Turkey and Computational Intelligence Laboratory, University of Manitoba, WPG, MB, R3T 5V6, Canada'
author:
- 'J.F. Peters$^{\alpha}$'
- 'E. İnan$^{\beta}$'
title: Strongly Near Voronoï Nucleus Clusters
---
[^1]
Introduction
============
This paper introduces nucleus clustering in Voronoï tessellations of surfaces in Euclidean space $\mathbb{R}^d, d\geq 2$. In this article, nucleus clustering is restricted to plane surfaces with applications in the geometry of digital images. A *nucleus cluster* is a collection of Voronoï regions that are adjacent to a Voronoï region called the cluster nucleus, which is a variation of the notion of a Harer-Edelsbrunner nerve [@Edelsbrunner2010compTop §III.2, p. 59].
[R]{}[0.35]{}
{width="25mm"}
Every Voronoï region of a site $s$ is a convex polygon containing all points that are nearer $s$ than to any other site in a Voronoï tessellation of a surface. Voronoï regions are strongly near, provided the regions have points in common. This form of clustering leads to the introduction of what are known as nucleus-clusters. A *nucleus cluster* is a collection of Voronoï regions that are strongly near a central Voronoï region called the cluster nucleus in a Voronoï tessellation. A *maximal nucleus cluster* is a collection of a maximal number of Voronoï regions that are strongly near the mesh nucleus. Maximal nucleus clusters serve as indicators of high object concentration in a tessellated image. This form of clustering leads to object recognition in many forms of application images.
Preliminaries
=============
This section introduces strongly near proximity and Voronoï tessellation of a plane surface based on recent work on computational proximity [@Peters2016ComputationalProximity], computational geometry [@Edelsbrunner1999; @Edelsbrunner2001; @Edelsbrunner2014; @Edelsbrunner2010compTop]. Strong proximities were introduced in [@Peters2015AMSJmanifolds], elaborated in [@Peters2016ComputationalProximity] (see, also, [@Inan2015]) and are a direct result of earlier work on proximities [@DiConcilio2006; @DiConcilio2013mcs; @Naimpally1970; @Naimpally2009; @Naimpally2013]. Nonempty sets $A$ and $B$ have strong proximity (denoted $A\ {\mathop{\delta}\limits^{\doublewedge}}\ B$), provided $A$ and $B$ have points in common. Let $E$ be the Euclidean plane, $S\subset E$ (set of mesh generating points), $s\in S$. A Voronoï region (denoted by $V(s)$) is defined by $$V(s) = \left\{x\in E: {\left\|x - s\right\|}\leq {\left\|x - q\right\|}, \mbox{for all}\ q\in S\right\}\ \mbox{(Vorono\"{i} region)}.$$
A partial view of a Voronoï tessellation of a plane surface is shown in Fig. \[fig:nucleusCluster\]. The Voronoï region $N$ in this tessellation is the nucleus of a mesh clustering containing all of those polygons adjacent to $N$. Let $X$ be a collection of Voronoï regions containing $N$, endowed with the strong proximity ${\mathop{\delta}\limits^{\doublewedge}}$. Briefly, a proximity relation is strong, provided $A\ {\mathop{\delta}\limits^{\doublewedge}}\ B, A,B\in X$ have points in common. Then the nucleus mesh cluster (denoted by ${\mbox{\large$\mathfrak{C}$}}\ N$) in this sample tessellation is defined by $${\mbox{\large$\mathfrak{C}$}}\ N = \left\{A\in X: {\mathop{\mbox{cl}}\limits^{\doublewedge}}\ A\ {\mathop{\delta}\limits^{\doublewedge}}\ N\right\}\ \mbox{(Vorono\"{i} mesh nucleus cluster)}.\\$$ That is, a nucleus mesh cluster ${\mbox{\large$\mathfrak{C}$}}\ N$ is a collection of nonempty sets $A$ whose closure is strongly near the cluster nucleus $N$ (in that case, each $A\in {\mbox{\large$\mathfrak{C}$}}\ N$ has points in common with $N$). For example, the set of points in the convex polygon $N$ in Fig. \[fig:nucleusCluster\] has points in common with each of the adjacent polygons, [*i.e.*]{}, each polygon adjacent to $N$ has an edge in common with $N$. Let $B$ be a polygon adjacent to $N$. $B\ {\mathop{\delta}\limits^{\doublewedge}}\ N$, since $B$ and $N$ have in edge in common.
A *concrete* (*physical*) set $A$ of points $p$ that are described by their location and physical characteristics, [*e.g.*]{}, gradient orientation (angle of the tangent to $p$. Let $\varphi(p)$ be the gradient orientation of $p$. For example, each point $p$ with coordinates $(x,y)$ in the concrete subset $A$ in the Euclidean plane is described by a feature vector of the form $(x,y,\varphi(p(x,y))$. Nonempty concrete sets $A$ and $B$ have descriptive strong proximity (denoted $A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$), provided $A$ and $B$ have points with matching descriptions. In a region-based, descriptive proximity extends to both abstract and concrete sets [@Peters2016ComputationalProximity §1.2]. For example, every subset $A$ in the Euclidean plane has features such as area and diameter. Let $(x,y)$ be the coordinates of the centroid $m$ of $A$. Then $A$ is described by feature vector of the form $(x,y,area, diameter)$. Then regions $A,B$ have descriptive proximity (denoted $A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$), provided $A$ and $B$ have matching descriptions.
The notion of strongly proximal regions extends to convex sets. A nonempty set $A$ is a *convex set* (denoted ${\mbox{conv}}A$), provided, for any pair of points $x,y\in A$, the line segment $\overline{xy}$ is also in $A$. The empty set $\emptyset$ and a one-element set $\left\{x\right\}$ are convex by definition. Let $\mathscr{F}$ be a family of convex sets. From the fact that the intersection of any two convex sets is convex [@Edelsbrunner2014 §2.1, Lemma A], it follows that $$\mathop{\bigcap}\limits_{A\in\mathscr{F}} A\ \mbox{is a convex set}.$$ Convex sets ${\mbox{conv}}A, {\mbox{conv}}B$ are strongly proximal (denote ${\mbox{conv}}A {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{conv}}B$), provided ${\mbox{conv}}A, {\mbox{conv}}B$ have points in common. Convex sets ${\mbox{conv}}A, {\mbox{conv}}B$ are descriptively strongly proximal (denoted ${\mbox{conv}}A {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ {\mbox{conv}}B$), provided ${\mbox{conv}}A, {\mbox{conv}}B$ have matching descriptions.
Let $X$ be a Voronoï tessellation of a plane surface equipped with the strong proximity ${\mathop{\delta}\limits^{\doublewedge}}$ and descriptive strong proximity ${\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}$ and let $A,N\in X$ be Voronoï regions. The pair $\left(X,\left\{{\mathop{\delta}\limits^{\doublewedge}},{\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\right\}\right)$ is an example of a proximal relator space [@Peters2016relator]. The two forms of nucleus clusters (ordinary nucleus cluster denoted by ${\mbox{\large$\mathfrak{C}$}}$) and descriptive nucleus clusters are examples of mesh nerves [@Peters2016ComputationalProximity §1.10, pp. 29ff], defined by $$\begin{aligned}
{\mbox{\large$\mathfrak{C}$}}N &= \left\{A\in X: A\ {\mathop{\delta}\limits^{\doublewedge}}\ N\right\}\ \mbox{(nucleus cluster)}.\\
{\mathfrak{C}_{\Phi}}N &= \left\{A\in X: A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ N\right\}\ \mbox{(descriptive nucleus cluster)}.\end{aligned}$$
A nucleus cluster is *maximal* (denoted by ${\mbox{max}\mathfrak{C}}N$), provided $N$ has the highest number of adjacent polygons in a tessellated surface (more than one maximal cluster in the same mesh is possible). Similarly, a descriptive nucleus cluster is maximal (denoted by ${\mbox{max}\mathfrak{C}_{\Phi}}N$), provided $N$ has the highest number of polygons in a tessellated surface descriptively near $N$, [*i.e.*]{}, the description of each $A\in {\mbox{max}\mathfrak{C}_{\Phi}}N$ matches the description of nucleus $N$ and the number of polygons descriptively near $N$ is maximal (again, more than one ${\mbox{max}\mathfrak{C}_{\Phi}}N$ is possible in a Voronoï tessellation).\
![ ${\mbox{\large$\mathfrak{C}$}}\ N_1\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}\ N_2\ \mbox{and}\ {\mbox{\large$\mathfrak{C}$}}\ N_2\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}\ N_3$[]{data-label="fig:N1N2N3maxClusters"}](nucleusClustersN1N2N3){width="65mm"}
Let $X$ the collection of Voronoï regions shown in Fig. \[fig:N1N2N3maxClusters\] with $N_1,N_2,N_3\in X$. In addition, let $2^X$ be the family of all subsets of Voronoï regions in $X$. Then ${\mbox{\large$\mathfrak{C}$}}N_1,{\mbox{\large$\mathfrak{C}$}}N_2,{\mbox{\large$\mathfrak{C}$}}N_3\in 2^X$ nucleus clusters in the tessellation. In this sample plane surface tessellation, ${\mbox{\large$\mathfrak{C}$}}N_1\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}N_2$, since $A\ {\mathop{\delta}\limits^{\doublewedge}}\ B$ for some $A\in {\mbox{\large$\mathfrak{C}$}}N_1, B\in {\mbox{\large$\mathfrak{C}$}}N_2$. Similarly, ${\mbox{\large$\mathfrak{C}$}}N_2\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}N_3$. In addition, nucleus clusters ${\mbox{\large$\mathfrak{C}$}}N_2,{\mbox{\large$\mathfrak{C}$}}N_3$ are maximal (denoted by ${\mbox{max}\mathfrak{C}}N_2,{\mbox{max}\mathfrak{C}}N_3$), since nuclei $N_2,N_3$ in the tessellation have the maximal number of adjacent Voronoï regions, namely, 10 adjacent regions. Let the description of a nucleus cluster in the Euclidean plane be described by its number of sides of its nucleus. Then ${\mbox{max}\mathfrak{C}}N_2\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ {\mbox{max}\mathfrak{C}}N_3$, since $N_2\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ N_3$, [*i.e.*]{}, the description of $N_2$ strongly matches the description of $N_3$ inasmuch as the description of the one nucleus is contained in the description of the other nucleus. In a more complete description, we would also consider the gradient orientation of the nucleus edges. In the case, $N_2\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ N_3$, provided each nucleus has at least one edge with a gradient orientation that matches the gradient orientation of an edge in the other nucleus.
\[thm:nucleusCluster\] Let $X$ be a set of Voronoï regions in the tessellation of a plane surface, endowed with the proximities ${\mathop{\delta}\limits^{\doublewedge}},{\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}},{\delta_{\Phi}}$ with $A,N\in X$. In addition, let $2^X$ be the family of all subsets of Voronoï regions in $X$. Then
$A\in X$ implies $A\in {\mbox{\large$\mathfrak{C}$}}N$ for some $N\in X$.
${\mbox{\large$\mathfrak{C}$}}N \in 2^X$ implies $A\ {\mathop{\delta}\limits^{\doublewedge}}\ N$ for some $A\in X$.
The union of all Voronoï nucleus clusters cover a plane surface, [*i.e.*]{}, $$X = \mathop{\bigcup}\limits_{N\in X} {\mbox{\large$\mathfrak{C}$}}N\ \mbox{(Nucleus cluster covering property)}.$$
$N,N'\in X$ implies ${\mbox{\large$\mathfrak{C}$}}N,{\mbox{\large$\mathfrak{C}$}}N'\in 2^X$.
Let the description of $N\in X$ be the number of edges on the polygon $N$. Then\
${\mbox{max}\mathfrak{C}}N, {\mbox{max}\mathfrak{C}}N'\in 2^X$ implies $N\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ N'$ for $N,N' \in X$.
\[prop:sn\] ${\mbox{\large$\mathfrak{C}$}}N\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}N'$, if and only $A\ {\mathop{\delta}\limits^{\doublewedge}}\ B$ for some $A\in {\mbox{\large$\mathfrak{C}$}}N, B\in {\mbox{\large$\mathfrak{C}$}}N'$.
\[prop:snd\] ${\mbox{\large$\mathfrak{C}$}}N\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}N'$, if and only if $A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$ for some $A\in {\mbox{\large$\mathfrak{C}$}}N, B\in {\mbox{\large$\mathfrak{C}$}}N'$.
$A\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}\ B$, for $A\in X, {\mbox{\large$\mathfrak{C}$}}\ B\in 2^X$ implies ${\mbox{\large$\mathfrak{C}$}}N\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}B$ for some $N\in X$, where $A\ {\mathop{\delta}\limits^{\doublewedge}}\ N$.
${\mbox{\large$\mathfrak{C}$}}N\ \cap\ {\mbox{\large$\mathfrak{C}$}}N'\neq \emptyset$ implies $A\ {\mathop{\delta}\limits^{\doublewedge}}\ B$ for some $A\in {\mbox{\large$\mathfrak{C}$}}N, B\in {\mbox{\large$\mathfrak{C}$}}N'$.
Let ${\mbox{\large$\mathfrak{C}$}}N\ {\mathop{\cap}\limits_{\Phi}}\ {\mbox{\large$\mathfrak{C}$}}N' = \left\{A\in {\mbox{\large$\mathfrak{C}$}}N \cup {\mbox{\large$\mathfrak{C}$}}N': A\in \Phi({\mbox{\large$\mathfrak{C}$}}N)\ \&\ A\in \Phi({\mbox{\large$\mathfrak{C}$}}N')\right\}$ (descriptive intersection of nucleus clusters). Then ${\mbox{\large$\mathfrak{C}$}}N\ {\mathop{\cap}\limits_{\Phi}}\ {\mbox{\large$\mathfrak{C}$}}N'\neq \emptyset$ implies $A\ {\delta_{\Phi}}\ B$ for some $A\in {\mbox{\large$\mathfrak{C}$}}N, B\in {\mbox{\large$\mathfrak{C}$}}N'$.
We prove only \[prop:sn\]$^o$ and \[prop:snd\]$^o$. The proof of the remaining parts are direct consequences of the definitions.\
\[prop:sn\]$^o$: $A\ {\mathop{\delta}\limits^{\doublewedge}}\ B$ ($A$ and $B$ have a common edge) for some $A\in {\mbox{\large$\mathfrak{C}$}}N, B\in {\mbox{\large$\mathfrak{C}$}}N'$, if and only if ${\mbox{\large$\mathfrak{C}$}}N, {\mbox{\large$\mathfrak{C}$}}N'$ are adjacent, if and only if ${\mbox{\large$\mathfrak{C}$}}N\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}N'$.\
\[prop:snd\]$^o$: $A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$ for some $A\in {\mbox{\large$\mathfrak{C}$}}N, B\in {\mbox{\large$\mathfrak{C}$}}N'$, if and only if the description of $A$ matches the description of $B$ ($A,B$ can be either adjacent or non-adjacent), if and only if ${\mbox{\large$\mathfrak{C}$}}N\ \cap\ {\mbox{\large$\mathfrak{C}$}}N'\neq \emptyset$, if and only if, $A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$.
Let $X$ be a set of Voronoï regions in the tessellation of a plane surface, endowed with the proximities ${\mathop{\delta}\limits^{\doublewedge}},{\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}},{\delta_{\Phi}}$ with $A,N\in X$. From Theorem \[thm:nucleusCluster\].\[prop:sn\], the nuclei in adjacent Voronoï nucleus clusters have a strong affinity in the sense that each of the clusters contains a Voronoï region that is strongly near a Voronoï region in an adjacent cluster. For example, in Fig. \[fig:N1N2N3maxClusters\], clusters ${\mbox{\large$\mathfrak{C}$}}N_1,{\mbox{\large$\mathfrak{C}$}}N_2$ share a pair of adjacent polygons. The nuclei in adjacent Voronoï nucleus clusters have a strong descriptive affinity, provided the nuclei have matching descriptions. It also the case that Voronoï regions $V(s),V(s')\in X$ are descriptively near, provided $s\ {\delta_{\Phi}}\ s'$, [*i.e.*]{}, the description of $s$ matches the description of $s'$. Hence, from Theorem \[thm:nucleusCluster\].\[prop:snd\], ${\mbox{\large$\mathfrak{C}$}}V(s)\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}V(s')$.
Main Results
============
\[lem:sn\] $A\ {\mathop{\delta}\limits^{\doublewedge}}\ B \Rightarrow A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$.
$A\ {\mathop{\delta}\limits^{\doublewedge}}\ B$ implies that $A$ and $B$ have points in common. Hence, there are points in $A$ and $B$ with the same descripitons, [*i.e.*]{}, $A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$
${\mbox{\large$\mathfrak{C}$}}\ N\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}\ M \Rightarrow {\mbox{\large$\mathfrak{C}$}}\ N\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}\ M$.
Immediate from Lemma \[lem:sn\] and Theorem \[thm:nucleusCluster\].\[prop:snd\].
The descriptive intersection [@Peters2014book §1.9, p. 43] of nonempty sets $A,B$ (denoted $A\ \ {\mathop{\cap}\limits_{\Phi}}\ B$) in an $n$-dimensional Euclidean space $\mathbb{R}^n$ is defined in the following way.
[**($\boldsymbol{\Phi}$)**]{}
: $\Phi(A) = \left\{\Phi(x)\in\mathbb{R}^n: x\in A\right\}$, set of feature vectors.
[**($\boldsymbol{{\mathop{\cap}\limits_{\Phi}}}$)**]{}
: $A\ {\mathop{\cap}\limits_{\Phi}}\ B = \left\{x\in A\cup B: \Phi(x)\in \Phi(A)\ \&\ \Phi(x)\in \Phi(B)\right\}$.
That is, the descriptive intersection of $A$ and $B$ contains all $a\in A,b\in B$ that are descriptively near each other.
${\mathfrak{C}_{\Phi}}N\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mathfrak{C}_{\Phi}}M \Leftrightarrow A\ {\mathop{\cap}\limits_{\Phi}}\ B\neq \emptyset$ for some $A\in {{\mathop{\mathfrak{C}}\limits^{\doublewedge}}_{\Phi}}\ N, B\in {{\mathop{\mathfrak{C}}\limits^{\doublewedge}}_{\Phi}}\ M$.
${\mathfrak{C}_{\Phi}}N\ {\mathop{\delta}\limits^{\doublewedge}}\ {\mathfrak{C}_{\Phi}}M\Leftrightarrow {\mbox{\large$\mathfrak{C}$}}N\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ {\mbox{\large$\mathfrak{C}$}}M\ \mbox{(from the definition of ${\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}$)}\ \Leftrightarrow A\ {\mathop{\delta_{_{\Phi}}}\limits^{\doublewedge}}\ B$ for some $A\in {\mbox{\large$\mathfrak{C}$}}N, B\in {\mbox{\large$\mathfrak{C}$}}M$ (from Theorem \[thm:nucleusCluster\].\[prop:snd\]), if and only if $A\ {\mathop{\cap}\limits_{\Phi}}\ B\neq \emptyset$.
\[def:Convexity\][Zelins’kyi-Soltan-Kay-Womble Convexity Structure]{}[[@Soltan1984convexity; @Zelinskii2015convexity; @Kay1971convexitySpace]]{}. Let $\mathscr{F} = 2^X$ be the family of all subsets of a nonempty set $X$ and let subfamilies $\mathscr{A},\mathscr{B}\in\mathscr{F}$. The family $\mathscr{F}$ on $X$ is called a *Zelins’kyi-Soltan-Kay-Womble convexity structure*, provided it satisfies the following axioms.
[(C0)]{}
: $\emptyset$ and $X$ belong to $\mathscr{F}$.
[(C1)]{}
: $\mathscr{A}\ \cap\ \mathscr{B} \in \mathscr{F}$ for all subfamilies $\mathscr{A},\mathscr{B}\in\mathscr{F}$.
The pair $\left(X,\mathscr{F}\right)$ is a *Zelins’kyi-Soltan-Kay-Womble convexity space*.
\[thm:FamilyOfSetsConvexity\][[@Peters2016ComputationalProximity]]{} The family of all subsets $\mathscr{F} = 2^X$ of a nonempty set $X$ is a Zelins’kyi-Soltan-Kay-Womble convexity structure.
Let $\mathcal{A}\in \mathscr{F}$. $X$ and $\emptyset$ are in $\mathscr{F}$. In addition, $\mathop{\bigcap}\limits_{A\in\mathcal{A}} A\in \mathscr{F}$. Hence, $\mathscr{F}$ is a Zelins’kyi-Soltan-Kay-Womble convexity structure.
\[thm:VoronoiConvexityStructure\] Let $X$ be a collection of Voronoï regions in the tessellation of a plane surface, $2^X$ the family of all subsets of $X$, ${\mbox{\large$\mathfrak{C}$}}N,{\mbox{\large$\mathfrak{C}$}}M\in 2^X$ such that ${\mbox{\large$\mathfrak{C}$}}N\cap{\mbox{\large$\mathfrak{C}$}}M\neq \emptyset$. The family $2^X$ is a Zelins’kyi-Soltan-Kay-Womble convexity structure.
For a nonempty $X$, both $\emptyset$ and $X$ are subsets in $2^X$ (Axiom (C0)). Let ${\mbox{\large$\mathfrak{C}$}}N,{\mbox{\large$\mathfrak{C}$}}M$ be subcollections in $2^X$. ${\mbox{\large$\mathfrak{C}$}}N\cap{\mbox{\large$\mathfrak{C}$}}M\neq \emptyset$ implies that ${\mbox{\large$\mathfrak{C}$}}N,{\mbox{\large$\mathfrak{C}$}}M$ share at least one Voronoï region. Consequently, ${\mbox{\large$\mathfrak{C}$}}N\cap{\mbox{\large$\mathfrak{C}$}}M\in 2^X$ (Axiom (C1)). Hence, from Theorem \[thm:FamilyOfSetsConvexity\], $2^X$ is a Zelins’kyi-Soltan-Kay-Womble convexity structure.
From Theorem \[thm:VoronoiConvexityStructure\], the collection of Voronoï regions $\left\{{\mbox{\large$\mathfrak{C}$}}N_2,{\mbox{\large$\mathfrak{C}$}}N_3\right\}$ in the tessellation shown in Fig. \[fig:N1N2N3maxClusters\] is a convexity structure, since ${\mbox{\large$\mathfrak{C}$}}N_2,{\mbox{\large$\mathfrak{C}$}}N_3$ have a Voronoï region in common.
Applications
============
Several applications arise from the introduction of Voronoï clustering.
[**Satellite Images**]{}
: Detecting surface objects and locations of sharp differences in terrain. Surface objects are revealed by one or more occurrences of maximal nucleus clusters.
[**FMRI Images**]{}
: High cortical activity corresponds to maximal nucleus clusters in brain tissue. The leads to the detection and classification of cortical activity associated with the tessellation of fMRI images. For example, the Voronoï mesh in Fig. \[fig:N1N2N3maxClusters\] has been extracted from the tessellation of an fMRI image of the brain. Mesh nucleus clustering is directly related to recent studies of fMRI images [@Tozzi2016CogNeuro4Dbrain; @Tozzi2016JNeuroSciSymmetries].
[**Tomography Images**]{}
: High concentration of fossils correspond to the presence and distribution of maximal nucleus clusters in 3D tomography images derived from drill core samples.
[99]{} A. Di Concilio, G. Gerla, *Quasi-metric spaces and point-free geometry*, Math. Structures Comput. Sci. [**16**]{} (2006), no. 1, 115–137, MR2220893.
A. Di Concilio, *Point-free geometries: [P]{}roximities and quasi-metrics*, Math. in Comp. Sci. [**7**]{} (2013), no. 1, 31-42, MR3043916. H. Edelsbrunner, *Computational Topology. Advances in discrete and computational geometry*. Contemp. Math., 223, Amer. Math. Soc., Providence, RI, 1999, MR1661380.
H. Edelsbrunner, *Geometry and Topology for Mesh Generation*. Cambridge University Press, Cambridge, UK, 2001, 2006. xii+177 pp. ISBN: 978-0-521-68207-7; 0-521-68207-X, MR2223897.
H. Edelsbrunner, *A Short Course in Computational Geometry and Topology*. Springer Briefs in Applied Sciences and Technology. Springer, Cham, 2014. x+110 pp. ISBN: 978-3-319-05956-3; 978-3-319-05957-0, MR3328629.
H. Edelsbrunner, *Computational Topology. An Introduction*. Amer. Math. Soc., Providence, RI, 2010. xii+241 pp. ISBN: 978-0-8218-4925-5, MR2572029.
C. Guadagni, *Bornological Convergences on Local Proximity Spaces and $\omega_{\mu}$-Metric Spaces*, Ph.D thesis, Università degli Studi di Salerno, Dipartimento di Matematica, supervisor: A. Di Concilio, 2015, 72pp.
E. İnan, *Algebraic Structures on Nearness Approximation Spaces*, Ph.D thesis, İnönü University, Department of Mathematics, supervisors: S. Keleş and M.A. Öztürk, 2015, vii+113pp.
D. Kay, E. Womble, [*Automatic convexity theory and relationships between the carath‘eodory, helly and radon numbers*]{}, Pacific Journal of Math. 38 (1971), no. 2, 471–485.
S.A. Naimpally, B.D. Warrack, *Proximity spaces*, Cambridge University Press, Cambridge Tract in Mathematics and Mathematical Physics 59, Cambridge, UK, 1970, ISBN 978-0-521-09183-1, MR2573941. S.A. Naimpally, *Proximity Approach to Problems in Topology and Analysis*, Oldenbourg Verlag, Munich, Germany, 2009, 73 pp., ISBN 978-3-486-58917-7, MR2526304.
S.A. Naimpally, J.F. Peters, *Topology with applications. Topological spaces via near and far. With a foreword by Iskander A. Taimanov*. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2013. xvi+277 pp. ISBN: 978-981-4407-65-6, MR3075111.
J.F. Peters, *Topology of Digital Images. Visual Pattern Discovery in Proximity Spaces*. Intelligent Systems Reference Library 63, Springer (2014). DOI 10.1007/978-3-642-53845-2. URL http://dx.doi.org/10.1007/978-3-642-53845-2. XV + 411 pp. Zentralblatt MATH Zbl 1295 68010.
J.F. Peters, *Computational Proximity. Excursions in the Topology of Digital Images*. Springer, Intelligent Systems Reference Library, Berlin, 2016, *in press*.
J.F. Peters, *Visibility in proximal [D]{}elaunay meshes*, Advances in Math. 4 (2015), no. 1, 41-47.
J.F. Peters, C. Guadagni, *Strong proximities on smooth manifolds and Vorono¨ı diagrams*, Advances in Math. 4 (2015), no. 2, 97-107.
J.F. Peters, *Proximal relator spaces*, Filomat (2016), *in press*.
V.P. Soltan, *Introduction to the axiomatic theory of convexity \[[R]{}ussian With [E]{}nglish and [F]{}rench summariess*, Shtiintsa, Kishinev, 1984. 224 pp., MR0779643.
A. Tozzi, J.F. Peters, *Towards a fourth spatial dimension of brain activity*, Cognitive Neurodynamics (2016), 1-11, DOI 10.1007/s11571-016-9379-z.
A. Tozzi, J.F. Peters, *A topological approach unveils system invariances and broken symmetries in the brain*, J. of Neuroscience Research (2016), 1-12, *in press*.
Y. Zelins’ky, [*Generalized convex envelopes of sets and the problem of shadow*]{}, Journal of Mathematical Sciences, 211 (2015), no. 5, 710-717.
[^1]: The research has been supported by the Scientific and Technological Research Council of Turkey (TÜBİTAK) Scientific Human Resources Development (BIDEB) under grant no: 2221-1059B211402463 and the Natural Sciences & Engineering Research Council of Canada (NSERC) discovery grant 185986.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider $N$ events that are defined on a common probability space. Those events shell have a common probability function that is symmetric with respect to interchanging the events. We ask for the probability distribution of the number of events that occur. If the probability of a single event is proportional to $1/N$ the resulting count probability is Poisson distributed in the limit of $N\rightarrow \infty$ for independent events. In this paper we calculate the characteristic function of the limiting count probability distribution for events that are correlated up to an arbitrary but finite order.'
address: 'Institut für Physik, Universität Greifswald, Felix-Hausdorff-Str. 6, D-17489 Greifswald, Germany'
author:
- Rüdiger Kürsten
date: 'October 29, 2019'
title: On the Count Probability of Many Correlated Symmetric Events
---
Introduction
============
We might distribute $N$ grains of rice randomly in a room and ask for the number of grains that lie on a marked subset of the ground like e.g. a circle drawn on it. If we assume that each grain lies on each position equally likely and furthermore that the positions of all grains are independent, the resulting number of grains in the marked area will be binomial distributed. If we increase the amount of rice grains and the area of the room such that the ration between them remains constant, the binomial distribution will converge to a Poisson distribution in the limit $N \rightarrow \infty$ [@Poisson1837]. In this paper we generalize this result for the case that the positions of the grains are not independent but still symmetric with respect to interchanges of the grains. That means it is impossible to distinguish the different grains statistically. We consider correlations of arbitrary but finite order. Later on we define precisely what is meant by correlation order.
Clearly, the probabilistic analysis of the problem is not restricted to rice grain but it can be applied to any large set of randomly positioned objects. The study of the distribution of large sets of points in space is known as spatial point pattern analysis [@Diggle83; @IPSS08]. It has applications in divers fields such as for example ecology [@VMGMW16], astronomy [@KSRBBGMPW97] or statistics of crimes [@MSBST11]. Furthermore, large sets of points appear as the positions of molecules in statistical physics.
For our purposes, the exact position of each point is not important because we only care if it lies within the marked area or not. Hence we can consider the event that a given particle lies within the marked space. In that way we obtain $N$ events from the positions of $N$ particles. In statistical physics, those events are usually correlated due to interactions between the particles and also other objects are often correlated. Formally, we consider $N$ events that are correlated and statistically indistinguishable. In that formulation it is not important whether those events are related to positions of particles in space.
Our main result is an explicit formula for the characteristic function of the number of events that occur in the limit $N\rightarrow \infty$. If correlations are limited to some order $l_{\text{max}}$, the characteristic function will depend on exactly $l_{\text{max}}$ parameters. The formula was already given in [@KSZI19] for the particular application to spatial point distributions, however, without proof. Ref. [@KSZI19] also gives an efficient Monte Carlo algorithm to sample the parameters of the distribution in the case of spatial point patterns. Furthermore, [@KSZI19] investigates not only the number of particles within a marked set in space but also the number of neighbors of a randomly chosen particle. If the neighborhood is defined by some spatial relation, e.g. if particles are considered to be neighbors if they lie within some given distance, the problem will be equivalent to the number of particles within an arbitrarily placed circle for independent homogeneously distributed particles. However, for correlated particles the two problems are not identical, but related. In fact the number of neighbor distribution can be obtained from our main result and depends on $l_{\text{max}}$ additional parameters, cf. [@KSZI19].
The paper is organized as follows. In Section \[sec:definitions\] we give some basic definitions introducing for example the common probability function of $N$ events, their correlation function of order $k$, and the count probability function that gives the probability that exactly $s$ of the $N$ events occur. Furthermore, we give some Lemma that are useful later on. In Section \[sec:result\] we give our main result, Theorem \[theorem:main\], together with its proof. In Section \[sec:discussion\] we conclude with a short discussion of the result.
Basic Definitions and some Lemma\[sec:definitions\]
===================================================
Instead of explicitly writing the intersections, unions or complements of the considered events we describe all relevant events using indicator functions.
Let $E$ be an event defined on a probability space $(\Omega, \Sigma, P)$. We define its $\mathbbm{1}_E: \Omega \rightarrow \{0,1 \}$ by $$\begin{aligned}
\mathbbm{1}_E(\omega) := \begin{cases} 1 \text{ if } \omega \in E \\ 0 \text{ if } \omega \notin E. \end{cases}
\label{eq:indicator}
\end{aligned}$$
\[def:symmetricsequenceofevents\] We call a finite sequence of events $E_1, \cdots, E_N$ that are defined on a common probability space , if for each finite sequence $r_1, \cdots, r_N$, with $r_i\in \{0,1\}$ for $i \in \{1,\cdots, N\}$ it holds $$\begin{aligned}
P(\mathbbm{1}_{E_1}=r_1, \cdots, \mathbbm{1}_{E_N}=r_N)=P(\mathbbm{1}_{E_1}=r_{\sigma(1)}, \cdots, \mathbbm{1}_{E_N}=r_{\sigma(N)})
\end{aligned}$$ for all permutations $\sigma$ of the elements $\{1, \cdots, N\}$.
That means it is impossible to distinguish the events statistically.
\[def:probabilityfunction\] Given a symmetric sequence of events $E_1, \cdots, E_N$, we call the function $P_k: \{0,1\}^k \rightarrow [0,1]$ defined by $$\begin{aligned}
P_{k}(r_1, r_2, \cdots, r_k):= \sum_{r_l \in \{0,1\} \text{ for }l \in \{k+1,\cdots, n\}}P(\mathbbm{1}_{E_1}=r_1, \cdots, \mathbbm{1}_{E_N}=r_N)
\label{eq:probabilityfunction}
\end{aligned}$$ the , where $1\le k \le N$ and $r_i \in \{0,1\}$.
\[remark:symmetryoftheprobabilityfunction\] The probability functions are also symmetric, that means $$\begin{aligned}
P_k(r_1, \dots, r_k)=P_k(r_{\sigma(1)}, \dots, r_{\sigma(k)})
\label{eq:symmetryoftheprobabilityfunction}
\end{aligned}$$ for all permutations $\sigma$ of the elements $\{1, \dots, k\}$ and for all $k\in \{1,\dots, N\}$, which follows immediately from the Definitions \[def:symmetricsequenceofevents\] and \[def:probabilityfunction\].
\[def:correlationfunction\] Given a symmetric sequence of events $E_1, \cdots, E_N$, we define the $G_1:\{0,1\} \rightarrow [0,1]$ as $G_1:\equiv P_1$ and $G_k:\{0,1\}^k \rightarrow \mathbb{R}$ for $1<k\le N$ recursively by $$\begin{aligned}
G_k(r_1, \cdots, r_k):= &P_k(r_1, \cdots, r_k)
\label{eq:correlationfunction}
\\
&- \sum_{\sigma}\sum_{l=1}^{k-1} \frac{1}{(l-1)!} \frac{1}{(k-l)!}G_{l}(r_1, r_{\sigma(2)}, \cdots, r_{\sigma(l)}) P_{k-l}(r_{\sigma(l+1)}, \cdots, r_{\sigma(k)}),
\notag
\end{aligned}$$ where the sum over $\sigma$ denotes the sum over all permutations of the elements $\{2, \cdots, k\}$.
The idea of a decomposition in different correlation orders is due to Ursell who essentially introduced the expansion given by Definition \[def:correlationfunction\], however, not properly normalized to serve as a probability [@Ursell27]. Mayer and Montroll used exactly the expansion of Definition \[def:correlationfunction\] [@MM41].
\[remark:correlationfunction\] We can rewrite Eq. of Definition \[def:correlationfunction\] as $P_l(\dots) = G_l(\dots) + \dots$ and insert it recursively for all $P_l$ of order $l<k$ into Eq. and eventually replace $P_1$ by $G_1$. Performing such an expansion, only $P_k$ and $G$-functions remain on the right hand side of Eq. . It follows inductively from Eq. that the indexes of all correlation functions on the right hand side are ordered, that is for each term $G_l(r_{i_1}, \dots, r_{i_l})$ appearing in the expansion of Eq. it holds $i_1<i_2<\dots<i_l$. Thus, instead of Eq. we might alternatively write $$\begin{aligned}
&G_k(r_1, \dots, r_k)=P_k(r_1, \dots, r_k) - \sum \{ \text{over all possible products of $G$-functions such that}
\notag
\\
&\text{each of the arguments $r_1, \dots, r_k$ appears exactly once and for each $G$-function}
\notag
\\
&\text{the arguments are ordered.}\}
\label{eq:alternativedefinitioncorrelationfunction}
\end{aligned}$$
Two examples of the alternative formulation of Definition \[def:correlationfunction\] given in the Remark \[remark:correlationfunction\] are the three- and four-event correlation functions $G_3$ and $G_4$ that are defined by $$\begin{aligned}
G_3(r_1, r_2, r_3)=&P_3(r_1, r_2, r_3)-G_1(r_1)G_1(r_2)G_1(r_3)-G_1(r_1)G_2(r_2, r_3)
\notag
\\
&-G_1(r_2)G_2(r_1, r_3)-G_1(r_3)G_2(r_1, r_2),
\label{eq:g3}
\\
G_4(r_1, r_2, r_3, r_4)=&P_4(r_1, r_2, r_3, r_4)-G_1(r_1)G_1(r_2)G_1(r_3)G_1(r_4)
\nonumber
\\
&-G_2(r_1,r_2)G_1(r_3)G_1(r_4)-G_2(r_1,r_3)G_1(r_2)G_1(r_4)
\nonumber
\\
&-G_2(r_1,r_4)G_1(r_2)G_1(r_3)-G_2(r_2,r_3)G_1(r_1)G_1(r_4)
\nonumber
\\
&-G_2(r_2,r_4)G_1(r_1)G_1(r_3)-G_2(r_3,r_4)G_1(r_1)G_1(r_2)
\nonumber
\\
&-G_2(r_1,r_2)G_2(r_3, r_4)-G_2(r_1,r_3)G_2(r_2, r_4)-G_2(r_1,r_4)G_2(r_2, r_3)
\nonumber
\\
&-G_3(r_1, r_2, r_3)G_1(r_4)-G_3(r_1, r_2, r_4)G_1(r_3)
\nonumber
\\
&-G_3(r_1, r_3, r_4)G_1(r_2)-G_3(r_2, r_3, r_4)G_1(r_1).
\end{aligned}$$
\[lemma:symmetryofthecorrelationfunction\] Let $E_1, \dots, E_N$ be a symmetric sequence of correlated events, then the corresponding correlation functions are symmetric, that is $$\begin{aligned}
G_k(r_1, \dots, r_k) = G_k(r_{\sigma(1)}, \dots, r_{\sigma(k)})
\label{eq:symmetryofthecorrelationfunction}
\end{aligned}$$ for all permutations $\sigma$ of the elements $\{1, \dots, k\}$, where $1\le k \le N$.
Obviously, the statement is satisfied for $k=1$ as the identity is the only permutation of one element. We assume that the statement is true for all $k$ that satisfy $1\le k < l$ and aim to show that it also holds for $l$. Considering the definition of $G_l$ given in the formulation of Eq. we realize that $P_l$ on the right hand side is symmetric due to Remark \[remark:symmetryoftheprobabilityfunction\]. Thus it remains to show that the remaining terms (without $P_l$) on the right hand side of Eq. are symmetric. We notice that those terms only consist of products of correlation functions $G_k$ with $k<l$ such that we can apply the induction hypothesis to them. Therefore we identify the terms $G_k(r_{i_\sigma(1)},\dots, r_{i_{\sigma(k)}} )$ for all permutations $\sigma$ of the elements $\{1, \dots, k\}$, where $i_j$ are some indexes.
Then, for any permutation $\sigma$ of the elements $\{1, \dots, l \}$, we can assign for each term in the sum in Eq. $G_{k_1}(r_{i_1^{k_1}}, \dots, r_{i_{k_1}^{k_1}}) G_{k_2}(r_{i_1^{k_2}}, \dots, r_{i_{k_2}^{k_2}}) \dots G_{k_s}(r_{i_1^{k_s}}, \dots, r_{i_{k_s}^{k_s}})$ the term $G_{k_1}(r_{\sigma(i_1^{k_1})}, \dots, r_{\sigma(i_{k_1}^{k_1})}) G_{k_2}(r_{\sigma(i_1^{k_2})}, \dots, r_{\sigma(i_{k_2}^{k_2})}) \dots G_{k_s}(r_{\sigma(i_1^{k_s})}, \dots, r_{\sigma(i_{k_s}^{k_s})})$ which is due to the above identification also a term in the sum in Eq. . Here, $i_j^k$ are some indexes. Hence the permutation $\sigma$ only exchanges terms in the sum in Eq. which does not affect the result of the sum.
\[def:correlationcoefficient\] Given a symmetric sequence of events $E_1, \cdots, E_N$ with correlation function of order $k$, $G_k$, with $1\le k \le N$ we call $C_k\in \mathbb{R}$ defined by $$\begin{aligned}
C_k:= N^k G_k(1, \cdots, 1)
\label{eq:correlationcoefficient}
\end{aligned}$$ the associated with the symmetric sequence of events.
\[lemma1\] Given a symmetric sequence of correlated events it holds for any correlation function $G_k$ of order $k\ge 2$ $$\begin{aligned}
G_k(1, r_2, \dots, r_k)=-G_k(0, r_2, \dots, r_k)
\label{eq:correlationlemma}
\end{aligned}$$ for all values of $r_i\in \{0,1\}$ for $i\in \{2, \dots, k\}$.
This result was already given by Mayer and Montroll [@MM41].
By Definition \[def:correlationfunction\] we have for the two-event correlation function $G_2(r_1, r_2) = P_2(r_1, r_2) - P_1(r_1) P_2(r_2)$. Summing this equation over $r_1=1,0$ we obtain the result for $k=2$. For $k>2$ we obtain the result by induction immediately from the Definition \[def:correlationfunction\] when summing over $r_1=1,0$. The first term on the right gives $P_{k-1}(r_2, \cdots, r_k)$ and the $l=1$-term from the sum gives $-P_{k-1}(r_2, \cdots, r_k)$. All other terms of the sum are zero by induction hypothesis.
\[lemma:reduction\_of\_correalation\_function\] Given a symmetric sequence of events $E_1, \cdots, E_N$ with correlation coefficient of order $k$, $C_k$, with $2\le k\le N$ it holds for the corresponding correlation function $$\begin{aligned}
G_k(r_1, \cdots, r_k) = N^{-k} C_k \prod_{l=1}^{^k} (-1)^{r_l+1}.
\label{eq:propcorrelationfunction}
\end{aligned}$$
By Lemma \[lemma1\] the ’flipping’ of $r_1$ in $G_k(r_1, r_2, \cdots, r_k)$ changes the value of the function by a factor of $-1$. Due to the symmetry given by Lemma \[lemma:symmetryofthecorrelationfunction\] we obtain a factor of $-1$ also when flipping one of the other arguments of $G_k$. Thus, the claim follows directly from the Definition \[def:correlationcoefficient\].
We call a finite symmetric sequence of events $E_1, \cdots, E_N$ for $2\le k \le N$, if the corresponding correlation parameter is zero, that is $$\begin{aligned}
C_k=0.
\label{eq:correlation_free}
\end{aligned}$$
We call a finite symmetric sequence of events $E_1, \cdots, E_N$ , if it is not correlation free of order $k$, that is if $$\begin{aligned}
C_k\neq 0.
\label{eq:correlation_parameter}
\end{aligned}$$
We call a finite symmetric sequence of events $E_1, \cdots, E_N$ , if $C_k=0$ for all $k\ge 2$ and with $k\ge 2$ if $C_k\neq 0$ and $C_j=0$ for all $j>k$.
\[def:count\_probability\] Given a symmetric sequence of events $E_1, \cdots, E_N$, we call the function $p_N:\mathbb{N}_{0}\rightarrow [0,1]$ defined by $$\begin{aligned}
p_{N}(s):= \langle \delta(s-\sum_{l=1}^{^N}\mathbbm{1}_{E_l}) \rangle,
\label{eq:countprobability}
\end{aligned}$$ the of the sequence of symmetric events, where $\langle \cdot \rangle$ denotes the expectation value and $$\begin{aligned}
\delta(k) = \begin{cases} 1 \text{ if } k=0 \\
0 \text{ else.}\end{cases}
\label{eq:delta}
\end{aligned}$$
Main Result\[sec:result\]
=========================
\[theorem:main\] Let $(E_1^N, E_2^N, \cdots, E_N^N)_{N \in \mathbb{N}, N \ge N_0}$ be a sequence of symmetric sequences of events such that for all $N\ge N_0$ the finite symmetric sequences of events $E^N_1, \cdots, E^N_N$ are correlated up to order $l_{\text{max}}$ with the same correlation parameters $C_1, \cdots, C_{l_{\text{max}}}$. Let furthermore $p_N(s)$ be the count probability of the symmetric sequence of events $E^N_1, \cdots, E^N_N$, then $$\begin{aligned}
\lim_{N\rightarrow \infty} p_N(s) = p_{\infty}(s),
\label{eq:theorem}
\end{aligned}$$ where the characteristic function of the limiting distribution $p_{\infty}$ is given by $$\begin{aligned}
\chi(u)= \sum_{s=0}^\infty \exp(ius)p(s) = \exp\bigg[ \sum_{l=1}^{l_{\text{max}}} \sum_{t=0}^l (-1)^{l-t} \frac{C_l}{l!} \binom{l}{t} \exp(itu) \bigg].
\label{eq:characteristicfunction}
\end{aligned}$$
We start with Definition \[def:count\_probability\] which can be written as $$\begin{aligned}
p_{N}(s)= \sum_{r_i\in\{0,1\} } P_{N}(r_1, \cdots, r_N) \delta\bigg(s-\sum_{i=1}^N r_i \bigg).
\label{eq:definition_count_probability2}
\end{aligned}$$ For simplicity we first investigate $p_N(N)$. In that case, only the term with all $r_i=1$ in Eq. (\[eq:definition\_count\_probability2\]) contributes to the sum, as only for this term the $\delta$-function has the value one. Next, we replace the probability function $P_N$ by correlation functions according to Definition \[def:correlationfunction\]. In particular we insert Eq. (\[eq:alternativedefinitioncorrelationfunction\]) of remark \[remark:correlationfunction\]. In the resulting sum, all summands containing correlation functions of order greater than $l_{\text{max}}$ are zero because we assumed that those correlation functions are zero. From the remaining summands, many are equal due to the symmetry given by Lemma \[lemma:symmetryofthecorrelationfunction\]. Each summand can be characterized by the numbers $k_l$ of factors of $G_l(1, 1, \cdots, 1)$ for $l=1, 2, \cdots, N$. The number of equal summands, that is the number of summands characterized by the same set of numbers $k_l$ is given by the combinatorial factor $$\begin{aligned}
\prod_{l=1}^N \bigg( \frac{1}{l!}\bigg)^{k_l} M^{n_l}_{l, k_l},
\label{eq:combinatorialfactor1}
\end{aligned}$$ where $$\begin{aligned}
n_l= N - \sum_{l'=1}^{l-1} k_{l'}.
\label{eq:combinatorialfactor2}
\end{aligned}$$ The above sum is meant to have no terms if the upper bound is smaller than the lower. The combinatorial factor $M^{n_l}_{l, k_l}$ denotes the number of possibilities to choose $l$ ordered $k_l$-plets from $n_l$ elements. It can be calculated by $$\begin{aligned}
M^{n_l}_{l, k_l}= \binom{n_l}{l k_l} \binom{l k_l}{k_l} \binom{l(k_l-1)}{k_l} \cdots \binom{k_l}{k_l} (k_l!)^l \frac{1}{k_l!}.
\label{eq:combinatorialfactor3}\end{aligned}$$ The first factor gives the number of possibilities to choose the $l k_l$ arguments of the $l$ $k_l$-plets. The second factor gives the number of choices of the argument of the first $k_l$-plet, the next factor the number of possible choices for the second $k_l$-plet and so on. The factor $(k_l!)^l$ gives the number of possible orders of the arguments of the $l$ $k_l$-plets and the last factor takes care of the fact that the $l$ $k_l$-plets are not ordered. The above formula, Eq. (\[eq:combinatorialfactor3\]), simplifies to $$\begin{aligned}
M^{n_l}_{l, k_l}= \frac{n_l!}{(n_l-l k_l)!} \frac{1}{k_l!}.
\label{eq:combinatorialfactor4}\end{aligned}$$
![Illustration of the combinatorial factor $\bigg(\frac{1}{l!}\bigg)^{k_l}M^{n_l}_{l, k_l}$. \[fig:1\]](figure1.pdf){width="90.00000%"}
The construction of the combinatorial factors $\bigg( \frac{1}{l!}\bigg)^{k_l} M^{n_l}_{l, k_l}$ in Eq. (\[eq:combinatorialfactor1\]) is illustrated in Figure \[fig:1\]. There are $M^{n_l}_{l, k_l}$ possibilities to choose $l$ ordered $k_l$-plets. Than the arguments for each factor $G_l$ are chosen as the columns displayed in Figure \[fig:1\]. In that way we obtain all possible permutations of the arguments for each $G_l$-factor. However, in Eq. (\[eq:alternativedefinitioncorrelationfunction\]) the arguments of all correlation factors are ordered. This is compensated by the factor of $\bigg( \frac{1}{l!}\bigg)^{k_l}$ in Eq. (\[eq:combinatorialfactor1\]).
Putting all together and collecting all the equal summands in $P_{N}(N)$ we obtain $$\begin{aligned}
p_N(N)=\prod_{l=1}^{l_{\text{max}}} \Bigg[ \sum_{k_l=0}^\infty \bigg(\frac{C_l}{N} \bigg)^{k_l} \bigg( \frac{1}{l!}\bigg)^{k_l} M^{n_l}_{l, k_l}\Bigg]\delta\bigg( N - \sum_{l=1}^{l_{\text{max}}}k_l \bigg),
\label{eq:pnn}\end{aligned}$$ where we inserted $C_l/N$ for $G_l(1, 1, \cdots, 1)$ according to Definition \[def:correlationcoefficient\]. The delta-function takes care of the fact that the total number of arguments of the correlation functions in each summand equals $N$.
For general arguments $s$ of $p_N(s)$ we go along the same lines. We insert $P_N$ from Eq. (\[eq:alternativedefinitioncorrelationfunction\]) into Eq. (\[eq:definition\_count\_probability2\]) and collect equal summands from the sum over the values of the arguments $r_i$ and over the different products of correlation functions. In analogy to Eq. (\[eq:pnn\]) we find $$\begin{aligned}
p_{N}(s) =\sum_{k_{1,0}}^{\infty} \Bigg[ \bigg(\frac{C_1}{N} \bigg)^{k_{1,0}} M^{n_{1,0}}_{1, k_{1,0}} \Bigg] \times \sum_{k_{1,1}}^{\infty} \Bigg[ \bigg(1- \frac{C_1}{N} \bigg)^{k_{1,1}} M^{n_{1,1}}_{1, k_{1,1}} \Bigg]
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{q=0}^{l} \Bigg[ \sum_{k_{l,q}=0}^{\infty} \bigg( (-1)^q \frac{C_l}{N^l} \bigg)^{k_{l,q}} M^{n_{l,q}}_{l, k_{l,q}} \bigg( \frac{1}{q!} \frac{1}{(l-q)!} \bigg)^{k_{l,q}}\Bigg] \Bigg\}
\nonumber
\\
\times \delta\bigg(s-\sum_{l=1}^{l_{\text{max}}}\sum_{q=0}^l (l-q)k_{l,q} \bigg) \delta\bigg(N - \sum_{l=1}^{l_{\text{max}}}\sum_{q=0}^l l k_{l,q} \bigg),
\label{eq:pns1}\end{aligned}$$ where $$\begin{aligned}
n_{l,q} = N - \sum_{l'=1}^{l-1}\sum_{q'=0}^{l'} k_{l', q'} - \sum_{q'=0}^{q-1}k_{l, q'}.
\label{eq:nlq}\end{aligned}$$ Here, $k_{l, q}$ denotes the number of factors $G_l$ with $q$ arguments that have the value zero and $l-q$ arguments that have the value one. That is $l$ denotes the correlation order and $q$ denotes the number of arguments that are zero. The factors $\frac{C_1}{N}$ and $1-\frac{C_1}{N}$ are, according to Definition \[def:correlationcoefficient\], the values of the correlation function $G(1)$ and $G(0)$, respectively. The factor $(-1)^q\frac{C_l}{N^l}$ gives the value of $G_l$ with $q$ arguments with the value zero and $l-q$ arguments with the value one according to Lemma \[lemma:reduction\_of\_correalation\_function\]. The factor $\frac{1}{q!} \frac{1}{(l-q)!}$ is the analog of $\frac{1}{l!}$ in Eq. (\[eq:pnn\]). However, here we do not divide by the number of possible permutations of the arguments of $G_l$ but by the number of possible permutation of only that arguments that have the value one and only that arguments that have the value zero. The first delta-function ensures that the total number of arguments of the correlation functions in each summand that have the value one equals $s$ and the second delta-function ensures that the total number of arguments of the correlation functions in each summand equals $N$.
Inserting the combinatorial factor $M^{n_{l,q}}_{l, k_{l,q}}$ given by Eq. (\[eq:combinatorialfactor4\]) into Eq. (\[eq:pns1\]) the product of the first factor in Eq. (\[eq:combinatorialfactor4\]) results in a factor $N!$ as most of the terms cancel. Eventually we obtain from Eq. (\[eq:combinatorialfactor4\]) $$\begin{aligned}
p_{N}(s) =N! \sum_{k_{1,0}}^{\infty} \Bigg[ \bigg(\frac{C_1}{N} \bigg)^{k_{1,0}} \frac{1}{k_{1,0}!} \Bigg] \times \sum_{k_{1,1}}^{\infty} \Bigg[ \bigg(1- \frac{C_1}{N} \bigg)^{k_{1,1}} \frac{1}{k_{1,1}!} \Bigg]
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{q=0}^{l} \Bigg[ \sum_{k_{l,q}=0}^{\infty} \bigg( (-1)^q \frac{C_l}{N^l} \bigg)^{k_{l,q}} \frac{1}{k_{l, q}!} \bigg( \frac{1}{q!} \frac{1}{(l-q)!} \bigg)^{k_{l,q}}\Bigg] \Bigg\}
\nonumber
\\
\times \delta\bigg(s-\sum_{l=1}^{l_{\text{max}}}\sum_{q=0}^l (l-q)k_{l,q} \bigg) \delta\bigg(N - \sum_{l=1}^{l_{\text{max}}}\sum_{q=0}^l l k_{l,q} \bigg).
\label{eq:pns2}\end{aligned}$$ Evaluating the sum over $k_{1,1}$ taking into account the second delta-function we obtain $$\begin{aligned}
p_{N}(s) =\frac{N!}{(N-\sum_{l=2}^{l_{\text{max}}}\sum_{q=0}^l l k_{l,q} - k_{1,0})!} \sum_{k_{1,0}}^{\infty} \Bigg[ \bigg(\frac{C_1}{N} \bigg)^{k_{1,0}} \frac{1}{k_{1,0}!} \Bigg]
\nonumber
\\
\times \bigg(1- \frac{C_1}{N} \bigg)^{N-\sum_{l=2}^{l_{\text{max}}}\sum_{q=0}^l l k_{l,q} -k_{1,0}}
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{q=0}^{l} \Bigg[ \sum_{k_{l,q}=0}^{\infty} \bigg( (-1)^q \frac{C_l}{N^l} \bigg)^{k_{l,q}} \frac{1}{k_{l, q}!} \bigg( \frac{1}{q!} \frac{1}{(l-q)!} \bigg)^{k_{l,q}}\Bigg] \Bigg\}
\nonumber
\\
\times \delta\bigg(s-\sum_{l=1}^{l_{\text{max}}}\sum_{q=0}^l (l-q)k_{l,q} \bigg).
\label{eq:pns3}\end{aligned}$$ The first factor can be written as $$\begin{aligned}
\frac{N!}{(N-\sum_{l=2}^{l_{\text{max}}}\sum_{q=0}^l l k_{l,q} - k_{1,0})!} = N^{\sum_{l=2}^{l_{\text{max}}}\sum_{q=0}^l l k_{l,q} + k_{1,0}}[1+ R(N)],
\label{eq:largeN}\end{aligned}$$ where the remainder $R(N)$ goes to zero as $\propto 1/N$ for $N\rightarrow \infty$. Inserting Eq. (\[eq:largeN\]) into Eq. (\[eq:pns3\]) almost all powers of $N$ cancel and we obtain $$\begin{aligned}
p_{N}(s) = [1+R(N)]\sum_{k_{1,0}}^{\infty} \Bigg[ \frac{C_1^{k_{1,0}}}{k_{1,0}!} \Bigg]
\bigg(1- \frac{C_1}{N} \bigg)^{N-\sum_{l=2}^{l_{\text{max}}}\sum_{q=0}^l l k_{l,q} -k_{1,0}}
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{q=0}^{l} \Bigg[ \sum_{k_{l,q}=0}^{\infty} \bigg( (-1)^q C_l \bigg)^{k_{l,q}} \frac{1}{k_{l, q}!} \bigg( \frac{1}{q!} \frac{1}{(l-q)!} \bigg)^{k_{l,q}}\Bigg] \Bigg\}
\nonumber
\\
\times \delta\bigg(s-\sum_{l=1}^{l_{\text{max}}}\sum_{q=0}^l (l-q)k_{l,q} \bigg).
\label{eq:pns4}\end{aligned}$$ Performing the limit $N\rightarrow \infty$ yields $$\begin{aligned}
p_{\infty}(s) = \sum_{k_{1,0}}^{\infty} \Bigg[ \frac{C_1^{k_{1,0}}}{k_{1,0}!} \Bigg] \exp(-C_1)
\nonumber
\\
\times \prod_{l=2}^{l_{max}} \Bigg\{ \prod_{q=0}^{l} \Bigg[ \sum_{k_{l,q}=0}^{\infty} \bigg( (-1)^q C_l \bigg)^{k_{l,q}} \frac{1}{k_{l, q}!} \bigg( \frac{1}{q!} \frac{1}{(l-q)!} \bigg)^{k_{l,q}}\Bigg] \Bigg\}
\nonumber
\\
\times \delta\bigg(s-\sum_{l=1}^{l_{\text{max}}}\sum_{q=0}^l (l-q)k_{l,q} \bigg).
\label{eq:pns5}\end{aligned}$$ Now, we perform the sum over $k_{1,0}$ taking the delta-function into account to obtain $$\begin{aligned}
p_{\infty}(s) = C_1^s \exp(-C_1)\frac{1}{(s-\sum_{l=2}^{l_{\text{max}}}\sum_{q=0}^l (l-q)k_{l,q} )!}
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{q=0}^{l} \Bigg[ \sum_{k_{l,q}=0}^{\infty} \bigg(\frac{(-1)^q C_l (C_1)^{q-l}}{q!(l-q)!} \bigg)^{k_{l,q}} \frac{1}{k_{l, q}!} \Bigg] \Bigg\}.
\label{eq:pns6}\end{aligned}$$ The first line is independent of $k_{l,l}$. Hence we can perform the sums over $k_{l,l}$ that result in exponential factors yielding $$\begin{aligned}
p_{\infty}(s) = C_1^s \exp\bigg(\sum_{l=1}^{l_{\text{max}}} (-1)^l \frac{C_l}{l!}\bigg)\frac{1}{(s-\sum_{l=2}^{l_{\text{max}}}\sum_{q=0}^{l-1} (l-q)k_{l,q} )!}
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{q=0}^{l-1} \Bigg[ \sum_{k_{l,q}=0}^{\infty} \bigg(\frac{(-1)^q C_l (C_1)^{q-l}}{q!(l-q)!} \bigg)^{k_{l,q}} \frac{1}{k_{l, q}!} \Bigg] \Bigg\}.
\label{eq:pns7}\end{aligned}$$ Next, we substitute $q$ by $t=l-q$ and calculate the characteristic function resulting in $$\begin{aligned}
\chi(u)=\sum_{s=0}^\infty p_{\infty}(s)\exp(ius)
\nonumber
\\
= \sum_{s=0}^\infty \exp(ius) C_1^s \exp\bigg(\sum_{l=1}^{l_{\text{max}}} (-1)^l \frac{C_l}{l!}\bigg)\frac{1}{(s-\sum_{l=2}^{l_{\text{max}}}\sum_{t=1}^{l} t k_{l,t} )!}
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{t=1}^{l} \Bigg[ \sum_{k_{l,t}=0}^{\infty} \bigg(\frac{(-1)^{l-t} C_l (C_1)^{-t}}{t!(l-t)!} \bigg)^{k_{l,t}} \frac{1}{k_{l, t}!} \Bigg] \Bigg\}.
\label{eq:characteristic_function}\end{aligned}$$ We can evaluate the sum over $s$ using $$\begin{aligned}
\sum_{s=0}^\infty \frac{\bigg[C_1 \exp(iu)\bigg]^s}{(s-a)!}=\sum_{s=a}^\infty \frac{\bigg(C_1 \exp(iu)\bigg)^s}{(s-a)!}
\nonumber
\\
=[C_1 \exp(iu)]^a \sum_{\tilde{s}=0}^\infty \frac{ \bigg(C_1 \exp(iu)\bigg) ^{\tilde{s}} }{\tilde{s}!}=[C_1 \exp(iu)]^a\exp\Bigg[C_1 \exp(iu) \Bigg],
\label{eq:sumovers}\end{aligned}$$ where we used that $1/(s-a)!$ is considered to be zero if $s-a<0$ in the first equation. Furthermore we used the substitution $\tilde{s}=s-a$. With $a=\sum_{l=2}^{l_{\text{max}}}\sum_{t=1}^{l} t k_{l,t}$ we obtain from Eq. (\[eq:characteristic\_function\]) with Eq. (\[eq:sumovers\]) $$\begin{aligned}
\chi(u)= \exp\Bigg[C_1 \exp(iu) \Bigg] \exp\bigg(\sum_{l=1}^{l_{\text{max}}} (-1)^l \frac{C_l}{l!}\bigg)
\nonumber
\\
\times \prod_{l=2}^{l_{\text{max}}} \Bigg\{ \prod_{t=1}^{l} \Bigg[ \sum_{k_{l,t}=0}^{\infty} \bigg(\frac{(-1)^{l-t} C_l \exp(iul) }{t!(l-t)!} \bigg)^{k_{l,t}} \frac{1}{k_{l, t}!} \Bigg] \Bigg\}.
\label{eq:characteristic_function2}\end{aligned}$$ Eventually we can evaluate the remaining sums over $k_{l,t}$ which result in exponential functions. Collecting all terms we obtain Eq. (\[eq:characteristicfunction\]).
It is remarkable that one could also evaluate the sums over $k_{l, l-1}$ directly in Eq. (\[eq:pns7\]). However, evaluating the remaining sums directly seems to be not doable.
Summary and Discussion\[sec:discussion\]
========================================
We calculated the probability distribution of the number of occurring events from a set of $N$ correlated events in the limit $N \rightarrow \infty$ under the assumption that the events are statistically indistinguishable and correlations are limited to an arbitrary order $l_{\text{max}}$. The limiting distribution is is given by Eq. (\[eq:pns7\]) that can be further simplified evaluating the sums over $k_{l,l-1}$, however, this expression contains still infinite sums. Therefore it is not suitable for practical evaluations. We calculate the characteristic function of the limiting distribution, Eq. (\[eq:characteristicfunction\]), which has a surprisingly simple form containing only finite sums. Setting all correlation parameters $C_k=0$ for $k>1$ we recover the characteristic function of the Poisson distribution.
[1]{}
S. D. Poisson. . Bachelier, Paris, France, 1837.
P. J. Diggle. Academic Press, 1983.
J. Illian, A. Penttinen, H. Stoyan, and D. Stoyan. , volume 70. John Wiley & Sons, 2008.
E. Vel[á]{}zquez, I. Mart[í]{}nez, S. Getzin, K. A. Moloney, and T. Wiegand. An evaluation of the state of spatial point pattern analysis in ecology. , 39(11):1042–1055, 2016.
M. Kerscher, J. Schmalzing, J. Retzlaff, S. Borgani, T. Buchert, S. Gottlöber, V. Müller, M. Plionis, and H. Wagner. . , 284(1):73–84, 01 1997.
G. O. Mohler, M. B. Short, P. J. Brantingham, F. P. Schoenberg, and G. E. Tita. Self-exciting point process modeling of crime. , 106(493):100–108, 2011.
R. Kürsten, S. Stroteich, M. Zumaya Hérnandez, and T. Ihle. Multiple particle correlation analysis of many-particle systems: Formalism and application to active matter, arXiv:1910.04244, 2019.
H. D. Ursell. The evaluation of Gibbs’ phase-integral for imperfect gases. , 23(6):685–697, 1927.
J. E. Mayer and E. Montroll. Molecular distribution. , 9(1):2–16, 1941.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $A$ be a Noetherian ring, $J\subseteq A$ an ideal and $C$ a finitely generated $A$-module. In this note we would like to prove the following statement. Let $\{I_n\}_{n\geq 0}$ be a collection of ideals satisfying : (i) $I_n\supseteq J^n$, for all $n$, (ii) $J^s\cdot I_s \subseteq I_{r+s}$, for all $r,s\geq 0$ and (iii) $I_n\subseteq I_m$, whenever $m\leq n$. Then ${\operatorname{Ass}}_A(I_nC/J^nC)$ is independent of $n$, for $n$ sufficiently large. Note that the set of prime ideals $\cup_{n\geq 1} {\operatorname{Ass}}_A(I_nC/J^nC)$ is finite, so the issue at hand is the realization that the primes in ${\operatorname{Ass}}_A(I_nC/J^nC)$ *do not* behave periodically, as one might have expected, say if $\bigoplus _{n\geq 0}I_n$ were a Noetherian $A$-algebra generated in degrees greater than one. We also give a multigraded version of our results.'
address:
- 'Department of Mathematics, University of Kansas, Lawrence, KS 66045'
- 'Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai 400 076, India'
author:
- Daniel Katz
- 'Tony J. Puthenpurakal'
title: 'Quasi-finite modules and asymptotic prime divisors'
---
Introduction
============
Let $A$ be a Noetherian ring, $J\subseteq A$ an ideal and $C$ a finitely generated $A$-module. Then, by a well known theorem of Brodmann (see [@B]), ${\operatorname{Ass}}_A(C/J^nC)$ is a stable set of prime ideals for $n$ large. Brodmann’s result has many applications and has been generalized in various forms. For example, see [@KMR], [@KS], [@EW], [@KW], [@KR] and [@H], among others.
In this note we are motivated by the following question. Given an ideal $J\subseteq A$, a finitely generated $A$-module $C$ and a filtration of ideals $\{I_n\}_{n\geq 0}$ with $J^n\subseteq I_n$ for all $n$, when is the set of associated primes ${\operatorname{Ass}}_A(I_nC/J^nC)$ a stable set of prime ideals ? It turns out that the desired stabilty holds under very mild conditions on the filtration $\{I_n\}_{n\geq 0}$. It is important to note that the set of prime ideals $\cup_{n\geq 1} {\operatorname{Ass}}_A(I_nC/J^nC)$ is well known to be a finite set and the issue at hand is the realization that the primes in ${\operatorname{Ass}}_A(I_nC/J^nC)$ *do not* behave periodically, as one might have expected, say if $\bigoplus _{n\geq 0}I_n$ were a Noetherian $A$-algebra generated in degrees greater than one (see [@EW]).
Not surprisingly, our approach is through graded modules defined over finitely generated $A$-algebras. To elaborate, let $R = \bigoplus_{n \geq 0}R_n$ be a finitely standard graded $A$-algebra, i.e., $R = A[R_1]$ and $R_1$ is a finite $A$-module. Let $M = \bigoplus_{n\in {\mathbb{Z} }}M_n$ be a graded $R$-module with $M_n = 0$ for all $n$ sufficiently small. We will assume throughout that each $M_n$ is a finitely generated $A$-module. Note that we do *not* assume that $M$ is a finitely generated $R$-module. Throughout, we set $L := H^0_{R_+}(M) = \bigoplus_{n\in {\mathbb{Z} }} L_n$. We say that $M$ is *quasi-finite* if, in addition, $L_n = 0$ ${\textrm{for \textit{n} sufficiently large}}$. It turns out that for the theory of asymptotic prime divisors, it is the quasi-finite property that is crucial, and not the finite generation of the module $M$. Indeed, almost all of the results we give in the context of graded modules are already known in the finite case (see [@M], [@EW] and [@H]). Of course, finite modules are quasi-finite. For an example of a quasi-finite module which is not finite, we will see below that $A[X]/A[JX]$ as a $A[JX]$-module, where $X$ is an indeterminate and $J\subseteq A$ is a proper ideal with positive grade. The notion of quasi-finite modules was introduced in [@JPV].
It turns out that our results are not much harder to come by if we consider multigraded rings, i.e., standard ${\mathbb{N}_0^d}$-graded Noetherian $A$-algebras and multigraded modules over them. In section two, we define the types of modules we are interested in and, in particular, we extend the definition of quasi-finite module to the multigraded case. We then prove our basic results concerning asymptotic prime divisors of quasi-finite multigraded modules. In particular, we note that we can achieve the standard stability result known for finite modules (see Theorem \[main\]). In section three we give some specific examples and applications of the results in section two, especially in the singly graded case. Our problems would be simpler to solve but less interesting if the graded (or multigraded) modules under consideration were always quasi-finite. To deal with modules $M$ which are not quasi-finite, we look at the quasi-finite module $M/L$. In section four, we consider what happens in the case of multigraded modules that are not necessarily quasi-finite. It turns out that we can isolate the precise obstruction to stability of asymptotic prime divisors in the general case (see Theorem \[genstability\]). Finally, in section five we present a multi-ideal version of the result alluded to in the abstract. In particular, we prove the following theorem.
\[introthm\] Let $A$ be a noetherian ring, $C$ a finitely generated $R$-module and ${J_1, \ldots, J_d }\subseteq A$ finitely many ideals. Suppose that for each $1\leq i\leq d$, $\{I_{i,n_i}\}_{n_i\geq 0}$ is a filtration of ideals satisfying : (i) $I_{i,0} = A$, (ii) $J_i^{n_i}\subseteq I_{i,n_i}$, for all $n_i\geq 0$, (iii) For all $m_i\leq n_i\in {\mathbb{N}_0 }$, $I_{i,n_i}\subseteq I_{i,m_i}$ and (iv) $J_{r_i}\cdot I_{i,s_i} \subseteq I_{i,r_i+s_i}$, for all $r_i$ and $s_i$. Then there exists $k = (k_1, \ldots, k_d) \in {\mathbb{N}_0^d}$ such that for all $n = (n_1, \ldots, n_d)\geq k$, $${\operatorname{Ass}}_A(I_{1,n_1}\cdots I_{d,n_d}C/J_1^{n_1} \cdots
J_d^{n_d}C) = {\operatorname{Ass}}_A(I_{1,k_1}\cdots I_{d,k_d}C/J_1^{k_1} \cdots
J_d^{k_d}C).$$
Quasi-finite multigraded modules
================================
Throughout this section $R$ will denote a standard ${\mathbb{N}_0^d}$-graded Noetherian, commutative ring with identity, where ${\mathbb{N}_0 }$ denote the set of non-negative integers. We denote the degree $(0,\ldots, 0)$ component of $R$ by $A$. Here, we use the term ‘standard’ in the sense of Stanley, i.e., a standard ${\mathbb{N}_0^d}$-graded ring is one which is generated in total degree one. Rather than use excessive notation, we will simplify our notation and use $n\in {\mathbb{N}_0^d}$ to indicate $d$-tuples. We will use subscripts to denote components of $d$-tuples. Thus $n_i$ means the $i^{\textrm{th}}$ component of $n = (n_1,\ldots, n_d) \in
{\mathbb{N}_0^d}$. Superscripts will be used to indicate lists of $d$-tuples. Given $n,m\in {\mathbb{N}_0^d}$, we will write $n\geq m$ if $n_i\geq m_i$, for all $1\leq i\leq d$. Finally, we extend all of this notation in the obvious way to ${\mathbb{Z} }^d$.
**Notation.** Let $R = \bigoplus_{n\in {\mathbb{N}_0^d}}R_d$ be a Noetherian standard ${\mathbb{N}_0^d}$-graded ring as above.
1. We will write $R_+$ for the ideal consisting of all sums of homogeneous elements $x_n \in R_n$ such that $n_i\geq 1$, for all $1\leq i\leq d$. In other words, $R_+$ denotes the ideal of $R$ generated by $R_{(1,\ldots, 1)}$.
2. *Throughout this paper*, by a multigraded $R$-module we mean a ${\mathbb{Z} }^d$-graded $R$-module $M = \bigoplus_{n\in {\mathbb{Z} }^d} M_n$ such that :
1. Each component $M_n$ of $M$ is a finitely generated $A$-module.
2. There exists $a \in {\mathbb{Z} }^d$ such that $M_n = 0$ for $n \leq a$.
**Observation.**\[observation\] In the notation above, suppose that $x\in M$ and $R_c\cdot x = 0$, for some $c\in {\mathbb{N}_0^d}$ with $c \not = (0,\ldots,0)$. Then $x\in H^0_{R_+}(M)$. To see this, suppose that $t\in {\mathbb{N}_0 }$ is the largest component of $c$, so $t > 0$. Then $R_{(t-c_1,\ldots, t-c_d)}\cdot
R_c\cdot x = 0$. In other words, $R_{(t,\ldots, t)}\cdot x = 0$. Since $R$ is standard graded, $R_+^t \cdot x = 0$, therefore $x\in H^0_{R_+}(M)$.
We now define multigraded quasi-finite modules.
\[quasfinitedef\] Let $M = \bigoplus_{n\in {\mathbb{Z} }^d} M_n$ as above be an multigraded $R$-module. We say $M$ is a *quasi-finite* $R$-module if there exists $b \in {\mathbb{Z} }^d$ such that $H^0_{R_+}(M)_n = 0$ for $n \geq b$.
\[multiquasiexample\] Notice that if $M$ is an ${\mathbb{Z} }^d$-graded $R$-module as above, then it follows immediately from the definition that $M/H^0_{R_+}(M)$ is quasi-finite as a ${\mathbb{Z} }^d$-graded $R$-module. More generally, the next propostion shows that a wide range of multigraded modules are quasi-finite.
\[quasigradepositive\] Let $U\subseteq V$ be multigraded $R$-modules such that $U$ is finitely generated over $R$ and ${\operatorname{grade}}(R_+,V) > 0$. Then $M := U/V$ is a quasi-finite $R$-module.
Consider the exact sequence $$0 {\rightarrow}U {\rightarrow}V {\rightarrow}M {\rightarrow}0.$$ Taking local cohomology [with respect to]{} $R_+$ and using that $H^0_{R_+}(V) = 0$ we get $$0 {\rightarrow}H^0_{R_+}(M ) {\rightarrow}H^1_{R_+}(U).$$ But, $U$ is a finitely generated $R$-module, so there exists $b \in {\mathbb{Z} }^d$ such that $H^1_{R_+}(U)_n = 0$, for $n\in {\mathbb{N}_0^d}$ with $n\geq b$. It follows that $M$ is quasi-finite.
For example, if ${\mathcal{R}}$ is the Rees algebra determined by an ideal of positive grade and $V$ is the polynomial ring containing ${\mathcal{R}}$, then, with $U = {\mathcal{R}}$, $M := V/U$ is an infinitely generated quasi-finite ${\mathcal{R}}$-module. (See section three.)
The following proposition is well known in the case of finitely generated graded or multigraded modules. Since $R$ is Noetherian, the proofs are the same even if $M$ is not finitely generated over $R$.
\[primesofM\] Let $R$ be a not necessarily standard ${\mathbb{N}_0^d}$-graded Noetherian ring and $M$ as above be a multigraded $R$-module. Then $P\in {\operatorname{Ass}}_A(M)$ if and only if $P\in {\operatorname{Ass}}_A(M_n)$ for some $n\in {\mathbb{Z} }^d$. Moreover, $P \in {\operatorname{Ass}}_A(M_n)$ for some $n\in {\mathbb{Z} }^d$ if and only if there exists a prime $Q\in {\operatorname{Ass}}_R(M)$ with $Q\cap A = P$. Consequently, the following statements hold :
1. $\bigcup _{n\in {\mathbb{Z} }^d} {\operatorname{Ass}}_A(M_n)$ is finite if and only if ${\operatorname{Ass}}_A(M)$ is finite.
2. If ${\operatorname{Ass}}_R(M)$ is finite, then $\bigcup _{n\in {\mathbb{Z} }^d} {\operatorname{Ass}}_A(M_n)$ is finite.
Note also that for any multiplicatively closed subset $S\subseteq A$, $R_S$ is a standard ${\mathbb{N}_0^d}$-graded Noetherian $A_S$-algebra and $M_S$ is a multigraded $R_S$-module with the original gradings preserved since the elements of $S$ have degree zero. It follows easily from this that if $P\subseteq A$ is disjoint from $S$ then $P$ is the annihilator of an element of degree $n$ in $M$ if and only if $P_S$ is the annhilator of an element of degree $n$ in $M_S$. We will use this observation freely throughout this paper.
Our first theorem shows that the quasi-finite notion is sufficient to guarantee asymptotic stability of prime divisors.
\[main\] Let $R$ be a standard ${\mathbb{N}_0^d}$-graded Noetherian ring and $M$ as above be a quasi-finite $R$-module. Then there exists $b\in {\mathbb{Z} }^d$ such that for all $n, m\in {\mathbb{Z} }^d$ with $b\leq n\leq m$, ${\operatorname{Ass}}_A(M_n) \subseteq {\operatorname{Ass}}_A(M_m)$. Moreover, if ${\operatorname{Ass}}_A (M)$ is a finite set, then there exists $k\in {\mathbb{Z} }^d$ such that for all $n \in {\mathbb{Z} }^d$ with $n\geq k$, ${\operatorname{Ass}}_A( M_n) = {\operatorname{Ass}}_A (M_k)$.
Now, let $b\in {\mathbb{Z} }^d$ be such that $H^0_{R_+}(M)_b = 0 $ for all $n\in {\mathbb{Z} }^d$ with $n\geq b$. If $M_n = 0$ for all $n\in {\mathbb{Z} }^d$ with $n\geq b$, then ${\operatorname{Ass}}_A(M_n) = \emptyset$ for all such $n$ and the conclusions of the theorem readily follow. Otherwise, $M_n \not = 0$ for some $n\geq b$. Without loss of generality, we may take $n = b$ (by increasing $b$ if necessary) and assume that $M_b \not = 0$. Note that since $M_b\not \subseteq H^0_{R_+}(M)$, it follows from Observation \[observation\] that for all $n\geq b$, $R_{n-b}\cdot M_b \not = 0$. In particular, $M_n \not = 0$.
Now, take $c < h\in {\mathbb{N}_0^d}$. We first show that ${\operatorname{Ass}}_A (M_{b+c}) \subseteq {\operatorname{Ass}}_A (M_{b+h})$. Let $P \in {\operatorname{Ass}}_A (M_{b+c})$. Without loss generality, we may assume that $A$ is local at $P$ and $P = (0:u)$, for $0 \not = u \in M_{b+c}$. We now note that $R_{h-c}\cdot u \not = 0$. Indeed, suppose $R_{h-c}\cdot u = 0$. Then by Observation \[observation\] above, $u \in H^0_{R_+}(M)$. Since $u \not = 0$, this contradicts our choice of $b$. Thus, $R_{h-c}\cdot u \not = 0$. Therefore, $xu \not = 0$, for some $x\in R_{h-c}$. Since $P\cdot xu =0$, $P\in {\operatorname{Ass}}_A (M_{b+h})$, as required.
Now, suppose ${\operatorname{Ass}}_A(M)$ is finite. Then by Proposition \[primesofM\], $\bigcup_{n\in {\mathbb{Z} }^d} {\operatorname{Ass}}_A(M_n)$ is a finite set. Now, let $\{P_1, \ldots, P_r\}$ denote the prime ideals $\bigcup_{n\geq b}{\operatorname{Ass}}_A (M_n)$. We can write each $P_j = (0:_Au_j)$, where $u_j \in M_{h^j}$, for $h^1, \ldots, h^r \in {\mathbb{Z} }^d$, with each $h^j\geq b$. Choose $k\in {\mathbb{Z} }^d$ such that $h^j\leq k$, for all $1\leq j\leq r$. Then, by the paragraph above, $P_j\in {\operatorname{Ass}}_A (M_k)$ for all $j$ and hence, $P_j\in {\operatorname{Ass}}_A M_n$, for all $k\leq n$. On the other hand, if $n\geq k$, then $n\geq b$. Thus, if $P\in {\operatorname{Ass}}_A (M_n)$, $P = P_j$, for some $1\leq j\leq r$. Thus, for all $n\in {\mathbb{Z} }^d$ with $n\geq k$, ${\operatorname{Ass}}_A(M_n) = \{P_1,\ldots, P_r\} = {\operatorname{Ass}}_A (M_k)$, and this completes the proof of the proposition.
First applications
==================
Suppose we have a homogeneous inclusion of singly graded Noetherian $R$-algebras $R \subseteq S$. In other words, $S = \bigoplus_{n\geq 0} S_n$ is a Noetherian ring with $S_0 = R_0 = A$ and $S_n \supseteq R_n$ for all $n \geq 0$. We assume $R$ is standard graded, but **do not** assume that $S$ is standard graded. Consider the $R$-module $E = S/R$. Notice that $$E= \bigoplus_{n = 0}^{\infty} \frac{S_n}{R_n} \quad \text{as an $A$-module}.$$ Note that $E$ need not be a finitely generated $R$-module. We however have the following :
\[Rees\] ${\operatorname{Ass}}_A (E)$ is a finite set.
We use a technique due to Rees [@Rees1]. Let $I$ denote the ideal of $S$ generated by $R_1$. Then, as graded ideals in $S$, for each $p\geq 0$, $I^p = \bigoplus_{n\geq 0}R_pS_n$ and $I^{p+1} =
\bigoplus_{n\geq 0} R_{p+1}S_n$. Thus, as a graded $S$-module, $I^p/I^{p+1} =
\bigoplus_{n\geq 0} R_pS_n/R_{p+1}S_{n-1}$. If we now write $G$ for the associated graded ring of $S$ with respect to $I$, we then have $$G = \bigoplus_{n,p\geq 0}R_pS_n/R_{p+1}S_{n-1},$$ and as such, $G$ can be viewed as a finitely generated (not necessarily standard) bigraded $A$-algebra. Thus, ${\operatorname{Ass}}_G(G)$ is finite, so by Proposition \[primesofM\], ${\operatorname{Ass}}_A(G)$ is finite.
The advantage of Rees’s technique is due to the following observation. For $t\geq 1$, consider the filtration of $A$-modules $$R_t \subseteq R_{t-1} S_1 \subseteq R_{t-2} S_2\subseteq \ldots \subseteq R_{1}S_{t-1} \subseteq S_{t}.$$ By breaking this filtration into short exact sequences, it follows that
$${\operatorname{Ass}}_A(S_t/R_t) \subseteq \bigcup_ {j =0}^t {\operatorname{Ass}}_A(R_{t-j}S_j/R_{t-j+1}S_{j-1}) \subseteq {\operatorname{Ass}}_A(G).$$ It follows that $\bigcup_{t\geq 1} {\operatorname{Ass}}_A(R_t/S_t)$ is finite, which gives what we want.
The next proposition is just a variation on Proposition \[multiquasiexample\] and indicates when $E$ as above is quasi-finite.
\[criterion\] For $R, S$ and $E$ as above, the following are equivalent
1. $E = S/R$ is a quasi-finite $R$-module.
2. $S$ is a quasi-finite $R$-module.
In particular if ${\operatorname{grade}}(R_+, S) > 0$ then $S/R$ is a quasi-finite module.
This follows from taking local cohomology (with respect to $R_+$) of the short exact sequence $$0 {\longrightarrow}R {\longrightarrow}S {\longrightarrow}S/R {\longrightarrow}0$$ and noting that for each $i \geq 0$ we have $H^i_{R_+}(R)_j = 0$ for all $j \gg 0$.
The following corollary is an immediate consequence of Theorem \[main\], Lemma \[Rees\], and Proposition \[criterion\]. It recovers from our perspective a special case of Theorem 1.1 from [@H], though in our case, we do not need to assume that $S$ is standard graded.
\[stablegraded\] Let $R\subseteq S$ be as in the previous proposition. If ${\operatorname{grade}}(R_+,S) > 0$, then ${\operatorname{Ass}}_A(S_n/R_n)$ is independent of $n$, for $n$ sufficiently large.
\[hpvexample\] It should be noted that the conclusion of Corollary \[stablegraded\] can fail if ${\operatorname{grade}}(R_+,S) = 0$. Indeed, in [@JPV], Example 3.4, one has $S_n/R_n = 0$ for $n$ odd and $S_n/R_n \not = 0$ for $n$ even. Thus, for this example, ${\operatorname{Ass}}_A(S_n/R_n)$ is not stable for $n$ large.
In the example below, we use a result due to Herzog, Hibi and Trung from their paper [@HHT].
Let $A = K[X_1,\ldots,X_d]$ be a polynomial ring over a field $K$. Let $I_1, \ldots, I_r$ be monomial ideals in $A$. Then there exists $k\in {\mathbb{N}_0 }$ such that for all $n\geq k$, $${\operatorname{Ass}}_A \frac{ \bigcap_{i=1}^{r} I_i^n}{\left(\bigcap_{i=1}^{r} I_i \right)^n} = {\operatorname{Ass}}_A \frac{ \bigcap_{i=1}^{r} I_i^{k}}{\left(\bigcap_{i=1}^{r} I_i \right)^{k}}.$$
Consider the algebra $S = \bigoplus_{n \geq 0} \bigcap_{i=1}^{r} I_i^n$. By [@HHT Corollary 1.3] we get that $S$ is a finitely generated $A$-algebra. Notice that $R = \bigoplus_{n \geq 0}\left(\bigcap_{i=1}^{r} I_i \right)^n$ is a standard graded $A$-subalgebra of $S$. Since $S$ is a domain we have that ${\operatorname{grade}}(R_+, S) > 0$. The result now follows by Corollary \[stablegraded\].
We continue with our applications in the singly graded case, by way of illustrating the strength of the quasi-finite condition. We begin by letting $J\subseteq A$ be an ideal and $\{ I_n \}_{n\geq 0}$ a filtration of ideals in $A$ satisfying the following properties : $I_0 = A$, $J^n\subseteq I_n$ for all $n \geq 0$ and $J_r\cdot I_s \subseteq I_{r+s}$, for all $r$ and $s$. We set ${\mathcal{R}}:= \bigoplus_{n\geq 0} J^n$, the Rees algebra of $A$ with respect to $J$. For a finitely generated $A$-module $C$, we write $U:= \bigoplus_{n\geq 0} I_nC$, a not necessarily finite ${\mathcal{R}}$-module. Let $V := \bigoplus_{n\geq 0}J^nC$ be the Rees module of $C$ with respect to $J$. We set $M_C := U/V$. Notice that since $\bigcup_{n \geq 1} {\operatorname{Ass}}_A (C/J^nC)$ is a finite set, it follows that ${\operatorname{Ass}}_A M_C$ is a finite set. The following application is a special case of our main result in section five.
\[gradePositive\] Maintain the notation established in the paragraph above. Suppose ${\operatorname{grade}}(J,C) > 0$. Then $M_C$ is quasi-finite. In particular, ${\operatorname{Ass}}_A (I_nC/J^nC)$ is stable for all $n \gg 0$.
If ${\operatorname{grade}}(J,C) > 0$, then ${\operatorname{grade}}({\mathcal{R}}_+, V) > 0$, so $M_C$ is a quasi-finite ${\mathcal{R}}$-module, by Proposition \[quasigradepositive\]. For the second statement, ${\operatorname{Ass}}_A (M_C)$ is a finite set, by our comment in the previous paragraph, so the result now follows from Theorem \[main\].
The non-quasi-finite case
=========================
We now return to the notation of section two. That is, $R$ denotes a Noetherian standard ${\mathbb{N}_0^d}$-graded $A$-algebra and $M$ is a (not necessarily finitely generated) multigraded $R$-module satisfying our standard hypotheses. Throughout we set $L := H^0_{R_+}(M)$. Notice that $M/L$ is quasi-finite, since $H^{0}_{R_+}(M/L) = 0$.
We begin with the following proposition.
\[union\] Let $R$ be a Noetherian standard ${\mathbb{N}_0^d}$-graded $A$-algebra and $M$ a multigraded $R$-module. Then, $${\operatorname{Ass}}_A (M) = {\operatorname{Ass}}_A (L) \cup {\operatorname{Ass}}_A (M/L).$$ In particular, for any $q\in {\mathbb{Z} }^d$, if $M_q/L_q \not = 0$, there exists $s\in {\mathbb{N}_0^d}$ such that ${\operatorname{Ass}}_A(M_q/L_q) \subseteq
{\operatorname{Ass}}_A(M_{q+s})$. Moreover, if ${\operatorname{Ass}}_A M $ is a finite set, then there exists $k \in {\mathbb{N}_0^d}$ such that ${\operatorname{Ass}}_A(M_n/L_n) = {\operatorname{Ass}}_A(M_k/L_k)$ for all $n\in {\mathbb{N}_0^d}$ with $n\geq k$.
Because ${\operatorname{Ass}}_A(M)\subseteq {\operatorname{Ass}}_A(L)\cup {\operatorname{Ass}}_A(M/L)$, for the first statement it suffices to prove that ${\operatorname{Ass}}_A(M/L)$ is contained in ${\operatorname{Ass}}_A(M)$. But for this, the second statement suffices. Let $P\in {\operatorname{Ass}}_A(M_q/L_q)$. Then we may write $P = (L_q :_A u)$, for $u\in M_q\backslash L_q$. Thus, $P\cdot u\subseteq L_q$, so there exists $t > 0$ such that $R_{_+}^t \cdot Pu = 0$. In particular, $P\cdot R_su = 0$, for $s = (t,\ldots, t)$. Thus, $P\subseteq (0 :_A R_su)$. Suppose $a\in A$ and $a\cdot R_s u = 0$. Then, $R_{_+}^t\cdot au = 0$, so $au \in L_q$. Thus, $a \in P$. It follows that $P = (0:_AR_su)$. Since $R_su$ is a finite $A$-module, it follows that $P$ annihilates an element of $R_su$, and hence $P\in {\operatorname{Ass}}_A(M_{q+s})$. Since we may choose the same $t$ for all $P\in {\operatorname{Ass}}_A(M_q/L_q)$ (since $M_q/L_q$ is a finite $A$-module), it follows that ${\operatorname{Ass}}_A(M_q/L_q) \subseteq {\operatorname{Ass}}_A(M_{q+s})$. This proves our second assertion. For the final statement, if ${\operatorname{Ass}}_A (M) $ is a finite set, it follows from the first statement of this proposition that ${\operatorname{Ass}}_A (M/L) $ is a finite set. However, $M/L$ is quasi-finite, so our last assertion follows from Theorem \[main\].
\[multiremark\] In the preceding proof, note that since $u\not \in L$, $R_+^t\cdot u \not \subseteq L$, so that if $P\in {\operatorname{Ass}}(M/L)$, $P$ is the annihilator of an element of $M\backslash L$.
We now set $L':= (0 \colon_{M} R_+)$. Clearly ${\operatorname{Ass}}_A (L') \subseteq {\operatorname{Ass}}_A (L)$. The following proposition supplies the reverse containment.
\[L’equalsL\] If $P \in {\operatorname{Ass}}_A (L_n)$ for some $n\in {\mathbb{N}_0^d}$, then $P \in {\operatorname{Ass}}_A (L'_{n+s})$ for some $s \in {\mathbb{N}_0^d}$. In particular, ${\operatorname{Ass}}_A(L') = {\operatorname{Ass}}_A(L)$.
Only the first statement requires proof. Let $P \in {\operatorname{Ass}}_A (L_n)$. By localizing at $P$ we may assume that $A$ is local and $P = (0 \colon u)$ where $u \in L_n$. Now, for some $t\geq 1$, $R_+^t\cdot u=0$. If $t=1$, then $u \in L'_n$ and $P \in {\operatorname{Ass}}(L'_n)$, as required. Otherwise, we choose $t > 1$ least so that $R_+^{t-1}\cdot u \neq 0$. Notice $P\cdot R_+^{t-1}\cdot u = 0$. Set $s := (t-1,\ldots, t-1)$. Then there is $x \in R_s$ be such that $x\cdot u\not = 0$. We have $P\cdot xu = 0$. Also $R_+ \cdot(x u) = 0$. Thus, $x u \in L'_{n+s}$. It follows that $P \in {\operatorname{Ass}}_A (L'_{n+s})$.
Here is the main result of this section. It illustrates an obstruction to the asymptotic stability of ${\operatorname{Ass}}_A(M_n)$ in case $M$ is not quasi-finite.
\[genstability\] Let $R$ as above be a Noetherian standard ${\mathbb{N}_0^d}$-graded $A$-algebra and $M$ a multigraded $R$-module as above. Assume that ${\operatorname{Ass}}_A(M)$ is finite. Consider the following statements :
1. There exists $k\in {\mathbb{Z} }^d$ such that ${\operatorname{Ass}}_A(L'_n) = {\operatorname{Ass}}_A(L'_k)$, for all $n\in {\mathbb{Z} }^d$ with $n\geq k$.
2. There exists $l \in {\mathbb{Z} }^d$ such that ${\operatorname{Ass}}_A(L_n) = {\operatorname{Ass}}_A(L_l)$, for all $n\in {\mathbb{Z} }^d$ with $n\geq l$.
3. There exists $h\in {\mathbb{Z} }^d$ such that ${\operatorname{Ass}}_A(M_n) = {\operatorname{Ass}}_A(M_h)$, for all $n\in {\mathbb{Z} }^d$ with $n\geq h$.
Then (i) implies (ii) and (ii) implies (iii).
We first note that if there exists $t\in {\mathbb{Z} }^d$ with $L'_n = 0$ for all $n\geq t$, then $L_n = 0$, for all $n\geq t$. Indeed, if some $L_n \not = 0$ for $n\geq t$, then there exists $P\in {\operatorname{Ass}}_A(L_n)$. But by Proposition \[L’equalsL\], $P\in {\operatorname{Ass}}_A(L'_{n+s})$, for some $s\in {\mathbb{N}_0^d}$, and therefore, $L'_{n+s} \not = 0$, a contradiction. Thus, if ${\operatorname{Ass}}_A(L'_n) = \emptyset$, for all $n\geq k$, then the same holds for ${\operatorname{Ass}}_A(L_n)$. Now suppose $\emptyset \not = {\operatorname{Ass}}_A L'_{n} = {\operatorname{Ass}}_A(L'_k)$, for all $n\geq k$. Fix $n \in {\mathbb{N}_0^d}$ with $n\geq k$. Take $P\in {\operatorname{Ass}}_A(L_k)$. Then, by Proposition \[L’equalsL\], there exists $s\in {\mathbb{N}_0^s}$ with $P\in {\operatorname{Ass}}_A(L'_{k+s})$. Thus, $P\in {\operatorname{Ass}}_A (L'_k)$, by choice of $k$. Thus, $P\in {\operatorname{Ass}}_A(L'_n)\subseteq {\operatorname{Ass}}_A(L_n)$. Conversely, if $P\in {\operatorname{Ass}}_A(L_n)$, then $P\in {\operatorname{Ass}}_A(L'_{n+r})$, for some $r\in {\mathbb{N}_0^d}$. Thus, $P\in {\operatorname{Ass}}_A(L'_k)\subseteq {\operatorname{Ass}}_A(L_k)$. Thus, ${\operatorname{Ass}}_A(L_n) = {\operatorname{Ass}}_A(L_k)$, for all $n\in {\mathbb{N}_0^d}$ with $n\geq k$. Therefore (i) implies (ii).
Now suppose that (ii) holds. We have two cases. If ${\operatorname{Ass}}_A(L_n) = \emptyset$ for all $n\geq l$, then $L_n = 0$, for all $n\geq l$, and therefore $M$ is quasi-finite. Thus, the conclusion of (iii) follows from Theorem \[main\]. If ${\operatorname{Ass}}_A(L_n) \not = \emptyset$ for all $n\geq l$, then $L_n \not = 0$, for all $n\geq l$, and thus, $M_n\not = 0$, for all $n\geq l$.
To continue, first note that since ${\operatorname{Ass}}_A(M)$ is finite, Proposition \[union\] implies that ${\operatorname{Ass}}_A(M/L)$ is also finite. By Theorem \[main\], there exists $q\in {\mathbb{Z} }^d$ such that ${\operatorname{Ass}}_A (M/L)_n = {\operatorname{Ass}}_A(M/L)_q$ for all $n\in {\mathbb{Z} }^d$ with $n\geq q$, since $M/L$ is quasi-finite. If the stable value of ${\operatorname{Ass}}_A(M_n/L_n) = \emptyset$, then $M_n = L_n$ for all $n\geq q$, and thus ${\operatorname{Ass}}_A(M_n) = {\operatorname{Ass}}_A(M_l)$, for all $n\in {\mathbb{N}_0^d}$ with $n\geq h$, for any $h\in {\mathbb{Z} }^d$ with $h\geq q$ and $h\geq l$, which gives what we want. Otherwise, if $M_q/L_q \not = 0$, then by Proposition \[union\], there exists $q'\in {\mathbb{Z} }^d$ with $q' \geq q$ such that ${\operatorname{Ass}}_A( M_q/L_q)\subseteq {\operatorname{Ass}}_A (M_{q'})$.
Take $h\in {\mathbb{Z} }^d$ such that $h\geq l$ and $h\geq q'$. Let $P \in {\operatorname{Ass}}_A(M_h)$. We will show $P\in {\operatorname{Ass}}_A(M_n)$ for all $n\geq h$, and for this, we may assume that $A$ is local at $P$. If $P\in {\operatorname{Ass}}_A(L_h)$, then by the choice of $l$, $P \in {\operatorname{Ass}}_A(L_n)$ for all $n\in {\mathbb{Z} }^d$ with $n\geq h$. Thus $P\in {\operatorname{Ass}}_A(M_n)$ for all $n\geq h$. If $P\not \in {\operatorname{Ass}}_A(L_h)$, then $P = (0:_A u)$ for $u \in M_h\backslash L_h$. Thus, by Observation \[observation\] above, $R_{n-h}\cdot u \not = 0$, for all $n\in {\mathbb{Z} }^d$ with $n\geq h$. Note $n-h \geq 0$. Since $P$ annihilates $R_{n-h}\cdot u$ and $R_{n-h}\cdot u\subseteq M_n$, we have $P\in {\operatorname{Ass}}_A(M_n)$, for all $n\geq h$.
Now suppose $n\in {\mathbb{Z} }^d$ and $n\geq h$. Take $P\in {\operatorname{Ass}}_A(M_n)$. If $P\in {\operatorname{Ass}}_A(L_n)$, then $P\in {\operatorname{Ass}}_A(L_h)$, by our choice of $h$ and thus, $P\in {\operatorname{Ass}}_A(M_h)$. Otherwise, $P\in {\operatorname{Ass}}_A(M_n/L_n)$. Thus, $P\in {\operatorname{Ass}}_A(M_q/L_q)$, by the definition of $q$. Therefore, by the definition of $q'$, $P\in {\operatorname{Ass}}_A(M_{q'})$, and by Remark \[multiremark\] above, we may assume that $P$ is the annihilator of an element of $M_{q'}\backslash L_{q'}$. Thus, as in the second paragraph of this proof, we may move $P$ forward so that $P\in
{\operatorname{Ass}}_A(M_h)$, since $h\geq q'$. We now have ${\operatorname{Ass}}_A(M_h) = {\operatorname{Ass}}_A(M_n)$, for all $n\geq h$, which gives (iii).
\(a) Clearly, in light of Proposition \[primesofM\], even in the singly graded case we cannot expect stability of ${\operatorname{Ass}}_A (M_n)$ without a finiteness condition on ${\operatorname{Ass}}_A (M)$. Moreover, in the presence of this condition, we cannot do better than Theorem \[genstability\]. Indeed, notice that $L'$ is just a direct sum of $A$-modules. Thus, for example, let $\gamma$ be an irrational number with a binary expansion $\gamma := a_1a_2a_3\cdots$ and take two distinct primes $P _1$ and $P _2$ in $A$. If we set $R = A$ and $M := \bigoplus_{n\geq 0} M_n$, where $M_n := A/P _1$, for $a_n=0$ and $M_n := A/P _2$, for $a_n=1$, then $L' = M$ and ${\operatorname{Ass}}_A (M)$ is finite, but ${\operatorname{Ass}}_A (M_n)$ is neither stable nor periodic.
\(b) Neither of the implications in Theorem \[genstability\] can be reversed. To see this, let $R = k[X]$ denote the polynomial ring in one variable over the field $k$. Let $T$ be the graded $R$-module $R/X^2R$. Thus, $T = T_0\oplus T_1$, with $T_0$ and $T_1$ one-dimensional vector spaces over $k$. Moreover, $X^2\cdot T _0 = 0$, $X\cdot T_0 \not = 0$ and $X\cdot T_1 = 0$. Let $M = \bigoplus_{n\geq 0} M_n$, where $M_n = T_0$ if $n$ is even and $M_n = T_1$ if $n$ is odd. Then $M = L := H^0_{R_+}(M)$ and ${\operatorname{Ass}}_k(M_n) = (0)$, for all $n$, so ${\operatorname{Ass}}_A(L_n)$ is stable for all $n$. On the other hand, setting $L' := (0:_R R_+)$, it follows that $L'_n = 0$ for $n$ even and $L'_n \not = 0$, for $n$ odd. Thus, ${\operatorname{Ass}}_k(L'_n)$ is not stable for $n$ large. Therefore, statement (ii) in Theorem \[genstability\] does not imply statement (i). A similar example can be constructed to show that (iii) does not imply (ii) in the statement of Theorem \[genstability\].
Second Applications
===================
Let $C$ be a finitely generated $A$-module, $J_1, \ldots, J_d$ be a family of ideals and for each $1\leq i\leq d$ let $\{I_{i,n_i}\}_{n_i\geq 0}$ be a filtration of ideals satisfying : (i) $I_{i,0} = A$, (ii) $J_i^{n_i}\subseteq I_{i,n_i}$, for all $n_i\geq 0$, (iii) $I_{i,n_i}\subseteq I_{i,m_i}$, whenever $n_i\geq m_i$, and (iv) $J_{r_i}\cdot I_{i,s_i} \subseteq I_{i,r_i+s_i}$, for all $r_i$ and $s_i$. An application of the main result of this section shows that if $C$ is a finitely generated $A$-module, there exists $k = (k_1, \ldots, k_d) \in {\mathbb{N}_0^d}$ such that for all $n = (n_1, \ldots, n_d)\geq k$, $${\operatorname{Ass}}_A(I_{1,n_1}\cdots I_{d,n_d}C/J_1^{n_1} \cdots
J_d^{n_d}C) = {\operatorname{Ass}}_A(I_{1,k_1}\cdots I_{d,k_d}C/J_1^{k_1} \cdots
J_d^{k_d}C).$$
Not suprisingly, we accomplish our goal by recasting the given data in terms of multigraded rings and modules. Throughout the rest of this section, we fix ideals $J_1, \ldots, J_d \subseteq A$. Keeping the notational conventions from the previous section, we will write $J^n$ for $J_1^{n_1}\cdots J_d^{n_d}$ for all $n = (n_1, \ldots, n_d) \in {\mathbb{N}_0^d}$.
We will need the following definition.
\[multifil\] For $J_1, \ldots, J_d$ as above, we call a collection of ideals $\{I_n\}_{n\in {\mathbb{N}_0^d}}$ a *multi-filtration* with respect to $J_1, \ldots, J_d$ if the following conditions holds :
1. $I_{(0,\ldots,0)} = A$.
2. $J^n\subseteq I_n$, for all $n\in {\mathbb{N}_0^d}$.
3. For all $n\leq m\in {\mathbb{N}_0^d}$, $I_m\subseteq I_n$.
4. For all $n, h\in {\mathbb{N}_0^d}$, $J^n\cdot I_h \subseteq I_{n+h}$.
Given a multi-filtration with respect to $J_1, \ldots, J_d$, it is now a simple matter to create a multigraded set-up to which we can apply the results of the previous section.
**Notation.** Let $\{I_n\}_{n\in {\mathbb{N}_0^d}}$ be a multi-filtration with respect to ${J_1, \ldots, J_d }$. Let $C$ be a finitely generated $A$-module. We set ${\mathcal{R}}:= \bigoplus _{n\in {\mathbb{N}_0^d}} J^n$, the multigraded Rees ring of $A$ with respect to ${J_1, \ldots, J_d }$. We also set $U :=\bigoplus_{n \in {\mathbb{N}_0^d}} I_nC$, $V := \bigoplus_{n\in {\mathbb{N}_0^d}} J^nC$ and $M := U/V$. Note that ${\mathcal{R}}$ is a standard ${\mathbb{N}_0^d}$-graded Noetherian ring, and $U, V$ and $M$ are multigraded ${\mathcal{R}}$-modules satisfying the standard hypotheses from the previous section.
Our goal now is to show that, with the notation just established, there exists $k\in {\mathbb{N}_0^d}$ such that for all $n\in {\mathbb{N}_0^d}$ with $n\geq k$, ${\operatorname{Ass}}_A (M_n) = {\operatorname{Ass}}_A (M_k)$. We first note that ${\operatorname{Ass}}_A (M)$ is finite, since on the one hand, $M\subseteq \bigoplus_{n\in {\mathbb{N}_0^d}} C/J^nC$ while on the other hand, it is well known that $\bigcup_{n\in {\mathbb{N}_0^d}} {\operatorname{Ass}}_A (C/J^nC)$ is finite (in fact, ultimately stable), and therefore, ${\operatorname{Ass}}_A (\bigoplus_{n\in {\mathbb{N}_0^d}} C/J^nC)$ is finite.
We need the following lemma which follows from [@KS], Lemma 1.3 (see also [@KMR], Proposition 1.4).
\[cancellation\] Let ${J_1, \ldots, J_d }\subseteq A$ be as above and suppose that $C$ is a finitely generated $A$-module. Suppose that each $J_i$ contains a nonzero divisor on $C$. Then there exists $k\in {\mathbb{N}_0^d}$ such that for all $n\in {\mathbb{N}_0^d}$ with $n \geq k$, $(J^{n+r}C :_C J^r) = J^nC$, for all $r\in {\mathbb{N}_0^d}$.
We are now ready for the principal result of this section, which immediately yields the main result of this paper.
\[multimainthm\] Let $A$ be a Noetherian ring, $C$ a finitely generated $A$-module and ${J_1, \ldots, J_d }\subseteq A$ finitely many ideals. Suppose that $\{I_n\}_{n\in {\mathbb{N}_0^d}}$ is a multi-filtration with respect to ${J_1, \ldots, J_d }$. Then there exists $k\in {\mathbb{N}_0^d}$ such that for all $n\in {\mathbb{N}_0^d}$ with $n\geq k$, ${\operatorname{Ass}}_A (I_nC/J^nC) = {\operatorname{Ass}}_A(I_kC/J_kC)$.
Set ${\mathcal{R}}:= \bigoplus_{n\in Nd} J^n$ and $M := \bigoplus_{n\in {\mathbb{N}_0^d}} I_nC/J^nC$, so that $M$ is a not necessarily finitely generated multigraded ${\mathcal{R}}$-module. Following the notation of the previous section, we set $L := H^0_{R_+}(M)$. Since ${\operatorname{Ass}}_A(M)$ is finite, by Theorem \[genstability\] it suffices to show that there exists $l\in {\mathbb{N}_0^d}$ such that ${\operatorname{Ass}}_A(L_n) = {\operatorname{Ass}}_A(L_l)$, for all $n \in {\mathbb{N}_0^d}$ with $n\geq l$.
We now calculate an expression for $L_n$, for $n\in {\mathbb{N}_0^d}$. Suppose $u\in L_n$. Then there exists $t\in {\mathbb{N}_0 }$ such that $R_{_+}^t\cdot u = 0$ in $L$, for all $u\in L_n$ (since $L_n$ is a finitely generated $A$-module). It follows that if we set $r = (t,\ldots, t) \in {\mathbb{N}_0^d}$, then $J^r\cdot u\subseteq J^{n+r}C$ in $C$. Thus, $L_n = (J^{n+r}C :_C J^r)\cap I_n/J^n$. Note, $r$ depends upon $n$. Now, set $T := H^0_J(C)$, where $J = J_1\cdots J_d$. Thus, each $J_j$ contains a non-zerodivisor on $C/T$. It follows from Lemma \[cancellation\] that there exists $k\in {\mathbb{N}_0^d}$ such that for all $n\in {\mathbb{N}_0^d}$ with $n\geq k$ and all $r\in {\mathbb{N}_0^d}$, $(J^{n+r}C :_C J^r) \subseteq J^nC +T$.
Now, suppose $n\in {\mathbb{N}_0^d}$ and $n\geq k$. We may choose $r$ as in the previous paragraph so that $$T = (0 :_C J^r)\quad \textrm{and} \quad L_n = (J^{n+r}C :_C J^r)\cap I_n/J^n.$$ In particular, $(J^{n+r}C :_C J^r) = J^nC +T$. Increasing $k$ if necessary, we may further assume that $T\cap J^nC = 0$, by using the multigraded form of the Artin-Rees Lemma. Under these conditions $$L_n = (J^nC + T)\cap I_n/J^nC = (J^nC + I_nC\cap T)/J^nC = I_nC\cap T.$$ The path to the end of the proof is now clear. If we let $P_1, \ldots, P_g$ denote $$\bigcup_{n\geq k}{\operatorname{Ass}}_A(L_n) = \bigcup_{n\geq k} {\operatorname{Ass}}_A(I_nC\cap T),$$ then each $P_i$ is an associated prime of some $L_{h^i}$ with $h^i\in {\mathbb{N}_0^d}$, $h_i\geq k$, $1\leq i\leq r$. We now just use the fact that for all $m,n\in {\mathbb{N}_0^d}$ with $m\geq n\geq k$, $L_m\subseteq L_n$. Now, if there exists $c\in {\mathbb{N}_0^d}$ such that $L_n = 0$ for all $n\geq c$, then we take $l = c$ and note that ${\operatorname{Ass}}_A(L_n) = \emptyset$, for all $n\geq c$. Otherwise, we order the set of primes $P_1, \ldots, P_g$ so that for each $P_1, \ldots, P_s$, if $P_i$ is among these primes, there exists $a^i\in {\mathbb{N}_0^d}$ with $P\not \in {\operatorname{Ass}}_A(L_{a^i})$ and no such $a^i$ exists if $s+1\leq i\leq g$. It follows that if $l\in {\mathbb{N}_0^d}$ and $l\geq a^i$ for $1\leq i\leq s$, then for all $n\geq l$, $P_i \not \in {\operatorname{Ass}}_A(L_n)$, if $1\leq i\leq s$ and $P_j\in {\operatorname{Ass}}_A(L_n)$, for $s+1\leq j\leq g$. In particular, ${\operatorname{Ass}}_A(L_n) = {\operatorname{Ass}}_A(L_l)$, for all $n\geq l$, and the proof of the theorem is complete.
We now provide the result stated at the beginning of this section.
Let $A$ be a noetherian ring, $C$ a finitely generated $R$-module and ${J_1, \ldots, J_d }\subseteq A$ finitely many ideals. Suppose that for each $1\leq i\leq d$, $\{I_{i,n_i}\}_{n_i\geq 0}$ is a filtration of ideals satisfying :
1. $I_{i,0} = A$.
2. $J_i^{n_i}\subseteq I_{i,n_i}$, for all $n_i\geq 0$.
3. $I_{i,n_i}\subseteq I_{i,m_i}$, for all $m_i\leq n_i\in {\mathbb{N}_0 }$.
4. $J_{r_i}\cdot I_{i,s_i} \subseteq I_{i,r_i+s_i}$, for all $r_i$ and $s_i$.
Then there exists $k = (k_1, \ldots, k_d) \in {\mathbb{N}_0^d}$ such that for all $n = (n_1, \ldots, n_d)\geq k$, $${\operatorname{Ass}}_A(I_{1,n_1}\cdots I_{d,n_d}C/J_1^{n_1} \cdots
J_d^{n_d}C) = {\operatorname{Ass}}_A(I_{1,k_1}\cdots I_{d,k_d}C/J_1^{k_1} \cdots
J_d^{k_d}C).$$
For $n = (n_1, \ldots, n_d)$ in ${\mathbb{N}_0^d}$, set $I_n := I_{1,n_1}\cdots I_{d,n_d}$. The result follows immediately from Theorem \[multimainthm\] since the collection of ideals $\{I_n\}_{n\in {\mathbb{N}_0^d}}$ is a multi-graded filtration with respect to ${J_1, \ldots, J_d }$.
\[distinction\] It is important to note that in the preceding theorem and corollary, we do not need to make any assumption that yields quasi-finiteness. This should be contrasted with Corollary \[stablegraded\] and Remark \[hpvexample\]. Indeed, if we take $d = 1$ in the preceding corollary, and set $R := \bigoplus_{n\geq 0} J^n$ and $S :=\bigoplus_{n\geq 0}I_n$, then we have ${\operatorname{Ass}}_A(S_n/R_n)$ is stable for large $n$, even if ${\operatorname{grade}}(R_+,S) = 0$.
We now give two examples illustrating certain aspects of the theorem.
\[symbolic-powers\] Let ${J_1, \ldots, J_d }\subseteq A$ be finitely many ideals, let $K\subseteq A$ be any ideal and set $I_n := J^n \colon K^\infty$, for all $n\in {\mathbb{N}_0^d}$. Then $\{J^n\colon K^{\infty}\}$ is a multi-filtration with respect to ${J_1, \ldots, J_d }$, so by Theorem \[multimainthm\], there exists $k\in {\mathbb{N}_0^d}$ such that ${\operatorname{Ass}}_A (J^n\colon K^{\infty})/J^n$ is stable for all $n \geq k$. Note, that in general, $\bigoplus_{n\in {\mathbb{N}_0^d}} (J^n\colon K^{\infty})$ need not be a Noetherian $A$-algebra. Note also that, as the proof of Theorem \[multimainthm\] shows, the stability of ${\operatorname{Ass}}(M_n) = {\operatorname{Ass}}(J^n\colon K^{\infty})/J^n$ depends upon the stability of ${\operatorname{Ass}}(L_n)$, for $L := H^0_{R_+}(M)$. We determine this latter set of primes. It follows from the proof of Theorem \[multimainthm\] that there exists $c \in {\mathbb{N}_0^d}$ such that $L_n = (J^n \colon K^\infty)\cap H^0_J(A)$, for all $n\in {\mathbb{N}_0^d}$ with $n\geq c$. (Recall $J = J_1\cdots J_d$.) We now note that $$(J^n \colon K^\infty)\cap H^0_J(A) = H^0_K(A)\cap H^0_J(A),$$ for $n$ sufficiently large (in ${\mathbb{N}_0^d}$). To see this, clearly $H^0_K(A)\cap H^0_J(A) \subseteq (J^n \colon K^\infty)\cap H^0_J(A)$. For the reverse inclusion, let $u \in (J^n \colon K^\infty)\cap H^0_J(A)$. Then $uK^r \subseteq J^n$ for some $r \geq 1$. Notice $uK^r \subseteq H^0_J(A)$. So $uK^r \subseteq H^0_J(A)\cap J^n = 0$ for $n$ large in ${\mathbb{N}_0^d}$, by the multigraded version of the Artin-Rees lemma. Thus $u \in H^0_K(A)$, as required. It follows that in the present case, ${\operatorname{Ass}}(H^0_K(A)\cap H^0_J(A))$ is the stable value of ${\operatorname{Ass}}(L_n)$. Finally, suppose $H^0_J(A) \cap H^0_K(A) \not = 0$ and $P\in {\operatorname{Ass}}(H^0_J(A)\cap H^0_K(A))$. Then $P\in {\operatorname{Ass}}(A)$ and $P$ must contain both $J$ and $K$, else $H^0_K(A)_P\cap H^0_J(A)_P = 0$. Conversely, suppose $P\in {\operatorname{Ass}}(A)$ and $P$ contains $J$ and $K$. Then $P = (0\colon a)$ and thus $J\cdot a = 0$, so $a \in H^0_J(A)$. Similarly, $a\in H^0_K(A)$. It follows that ${\operatorname{Ass}}_A(H^0_J(A)\cap H^0_K(A)) = {\operatorname{Ass}}(A)\cap V(J+K)$. Therefore, for all $n$ sufficiently large in ${\mathbb{N}_0^d}$, ${\operatorname{Ass}}(L_n) = {\operatorname{Ass}}(A)\cap V(J + K)$.
Let $J\subseteq A$ be an ideal and take $I\subseteq A$ any ideal containing $J$ and set $I_n := {\overline}{I^n}$. Then, by Theorem \[multimainthm\], ${\operatorname{Ass}}_A (I_n/J^n)$ is stable for all $n \gg 0$. Let $M = \bigoplus_{n\geq 1}{\overline}{I^n}/J^n$. Then by the proof of Theorem \[multimainthm\], $L_n = H^0_J(A) \cap {\overline}{I^n}$, for all $n\gg 0$, for $L = H^0_{R_+}(M)$. As in the previous example, we calculate the stable value of ${\operatorname{Ass}}_A(L_n)$. We now make two claims.
Claim 1 : ${\operatorname{Ass}}_A (H^0_J(A) \cap {\overline}{I^n}) = \{ P \in {\operatorname{Ass}}_A {\overline}{I^n} \mid P \supseteq J \}$.
To see this, note that if $P \nsupseteq J$ then $(H^0_J(A))_P = 0$, so $P \notin {\operatorname{Ass}}_A (H^0_J(A) \cap {\overline}{I^n})$. Thus, ${\operatorname{Ass}}_A (H^0_J(A)\cap {\overline}{I^n})$ is contained in $\{ P \in {\operatorname{Ass}}_A {\overline}{I^n} \mid P \supseteq J \}$. On the other hand, if $P \supseteq J$ and $P\in {\operatorname{Ass}}_A ({\overline}{I^n})$, then we localize at $P$ and assume $A$ is local at $P$ and $J \not = A$. Say $P = (0 \colon x)$ for $x \in {\overline}{I^n}$. Then $P\cdot x = 0$. So $x \in (0\colon P) \subseteq (0 \colon J^\infty)$. Thus $x \in H^0_J(A) \cap {\overline}{I^n}$. Therefore $P \in {\operatorname{Ass}}_A (H^0_J(A) \cap {\overline}{I^n})$. This proves the first claim.
Now, set $Z = $ nil-radical of $A$. Clearly we have the following exact sequence $$0 {\rightarrow}Z {\rightarrow}{\overline}{I^n} {\rightarrow}{\overline}{I^n}/Z {\rightarrow}0.$$ We make a second claim,
Claim 2 : For $n$ sufficiently large, ${\operatorname{Ass}}_A ({\overline}{I^n}) = ({\operatorname{Ass}}_A (Z))\cup \{ P \mid P \in {\operatorname{Min}}(A) \ \textrm{and} \ I\nsubseteq P\}$. (Note, here we regard $Z$ as an $A$-module.)
To see, we first note that ${\operatorname{Ass}}_A ({\overline}{I^n}/Z) = \{ P \mid P \in {\operatorname{Min}}(A) \ \textrm{and} \ I\nsubseteq P\}$, for $n$ large. Clearly any minimal prime $P$ not containing $I$ belongs to ${\operatorname{Ass}}_A ({\overline}{I^n}/Z)$ (just localize at $P$). On the other hand, if $P \in {\operatorname{Ass}}_A ({\overline}{I^n}/Z)$, then $P \in {\operatorname{Ass}}_A (A/Z)$, so $P$ is a minimal prime. But if $P$ contains $I$, then $I_{P}$ is nilpotent. Thus, $({\overline}{I^n})_{P} = Z_{{\mathfrak{p}}}$ for $n$ large, so ${\overline}{I^n}_{P}/Z_{P} = 0$, Thus, $P \not \in {\operatorname{Ass}}_A ({\overline}{I^n}/Z)$, contradiction. Therefore, for $n$ sufficiently large, $${\operatorname{Ass}}_A ({\overline}{I^n}/Z) = \{ P \mid P \in {\operatorname{Min}}(A) \ \textrm{and} \ I\nsubseteq P\}.$$ It follows immediately from this and (5.8.1) that $${\operatorname{Ass}}_A ({\overline}{I^n}) \subseteq ({\operatorname{Ass}}_A (Z))\cup \{ P \mid P \in {\operatorname{Min}}(A) \ \textrm{and}\ I\nsubseteq P\},$$ for all large $n$. For the reverse containment, clearly ${\operatorname{Ass}}_A (Z) \subseteq {\operatorname{Ass}}_A ({\overline}{I^n})$. Moreover, if $P \in {\operatorname{Min}}(A)$ and $P \nsupseteq I $ then by localizing at $P$ one sees that $P \in {\operatorname{Ass}}_A ({\overline}{I^n})$. Thus, Claim 2 holds.
Now, as before, setting $L := H^0_{R_+}(M)$, we have $L_n = H^0_J(A)\cap {\overline}{I^n}$, for all large $n$ and by Claims 1 and 2, $$[({\operatorname{Ass}}_A (Z))\cup \{ P \mid P \in {\operatorname{Min}}(A) \ \textrm{and} \ I\nsubseteq P\}]\cap V(J),$$ is the stable value of ${\operatorname{Ass}}_A(L_n)$, for $n$ sufficiently large.
Regarding Claim 2, we have an analogue for powers of an ideal. We note that for $n$ sufficiently large, $${\operatorname{Ass}}_A (I^n) = \{ P \mid P \in {\operatorname{Ass}}A \ \textrm{and} \ P \nsupseteq I \}.$$ Indeed, if $P \in {\operatorname{Ass}}_A (A)$ and $P \nsupseteq I$ then $I^n_P = R_P$. So $P \in {\operatorname{Ass}}_A (I^n)$ for all $ n\geq 1$. On the other hand suppose $P \in {\operatorname{Ass}}_A (I^n)$ and $P \supseteq I$, say $P = (0 \colon c_n)$ where $c_n \in I^n$. So for $n$ sufficiently large we have $$c_n \in (0\colon P)\cap I^n = I^{n-n_0}(I^{n_0}\cap (0\colon P )) = 0,$$ a contradiction. Thus, for sufficiently large $n$, if $P\in {\operatorname{Ass}}_A(I^n)$, $P\in {\operatorname{Ass}}_A(A)$, and $I\not \subseteq P$.
Finally, we note the following, which should be clear from our results above. Let $C$ be a finitely generated $A$-module and $J_1, \ldots, J_d$ finitely many ideals. Assuming that $\bigcup_{n\in {\mathbb{N}_0^d}} C/J^nC$ is finite, asymptotic stability of ${\operatorname{Ass}}_A(C/J^nC)$, $n\in {\mathbb{N}_0^d}$ follows from Theorem \[multimainthm\] by taking the multi-filtration with respect to ${J_1, \ldots, J_d }$ to be $I_n = A$ for all $n\in {\mathbb{N}_0^d}$. The finiteness of $\bigcup_{n\in {\mathbb{N}_0^d}} C/J^nC$ follows fairly readily from basic properties of prime divisors of regular elements on a module by using extended Rees algebras. For example, see [@KS].
[99]{}
M. Brodmann, *Asymptotic stability of ${\operatorname{Ass}}(M/I^n M)$*, Proc. Amer. Math. Soc. **74** (1979), 16-18.
F. Hayaska, *Asymptotic stability of primes assoicated to homogenous components of multigraded modules*, Journal of Algebra, **306** (2006), 535-543.
J. Herzog, T. Hibi and N. V. Trung, *Symbolic powers of monomial ideals and vertex cover algebras*, Adv. Math. **210** (2007), 304–322.
J. Herzog, T. J. Puthenpurakal and J. K. Verma, *Hilbert polynomials and powers of ideals*, Math. Proc. Camb. Phil. Soc, Vol **145** (2008), 623-642.
D. Katz, S. McAdam, and L.J. Ratliff, Jr., *Prime divisors and divisorial ideals*, Journal of Pure and Applied Algebra, Vol.**58** (1989), 179-186.
D. Katz and G. Rice, *Asymptotic prime divisors of torsion-free symmetric powers of modules*, Journal of Algebra, Vol. **319** (2008), 2209-2234.
D. Katz and E. West, *A linear function associated to asymptotic primes*, Proc. AMS, Vol **132**, No. 6 (2004), 1589-1597
A. K. Kingsbury and R. Y. Sharp, *Asymptotic behaviour of certain sets of prime ideals*, Proc. Amer. Math. Soc. **124** (1996), 1703-1711.
S. McAdam, Asymptotic Prime Divisors, Lecture Notes in Math., vol. 1023, Springer-Verlag, New York, 1983.
D. Rees, $A$-transforms of local rings and a theorem on multiplicities of ideals, Proc. Cambridge Philos Soc. (2) **57** (1961), 8–17.
E. West, *Primes associated to multigraded modules*, J. Algebra **271** (2004), no. 2, 427�453.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present a long (116 ks) observation of the Broad-Line Radio Galaxy (BLRG) c acquired in April 2007. A BAT spectrum in 15–200 keV from the 58-month survey is also analyzed, together with an archival EPIC exposure of 20 ks obtained one year after . Our main result is the finding with of a broad Fe K line with a relativistic profile consistent with emission from an accretion disk at tens of gravitational radii from the central black hole. The XIS data indicate emission from highly ionized iron and allow us to set tight, albeit model-dependent, constraints on the inner and outer radii of the disk reflecting region, $r_{in} \simeq 10 r_g$ and $r_{out} \simeq 20 r_g$, respectively, and on the disk inclination, $i\simeq 30$. Two ionized reflection components are possibly observed, with similar contributions of $\sim$10% to the total continuum. A highly ionized one, with log$\xi$$\simeq$3 erg s$^{-1}$ cm, which successfully models the relativistic line and a mildly ionized one, with log$\xi$$\simeq$1.5 erg s$^{-1}$ cm, which models the narrow Fe K$\alpha$ and high energy hump. When both these components are included, there is no further requirement for an additional black body soft excess below 2 keV. The *Suzaku* data confirm the presence of a warm absorber previously known from grating studies. After accounting for all the spectral features, the intrinsic photon index of the X-ray continuum is $\Gamma_x\simeq1.8$ with a cutoff energy at $\sim
200$ keV, consistent with Comptonization models and excluding jet-related emission up to these energies. Comparison of the X-ray properties of c and other BLRGs to Seyferts recently observed with and BAT confirms the idea that the distinction between radio-loud and radio-quiet AGN at X-rays is blurred. The two classes form a continuum distribution in terms of X-ray photon index, reflection strength, and Fe K line width (related to the disk emission radius), with BLRGs clustered at one end of the distribution. This points to a common structure of the central engine, with only a few fundamental parameter(s) responsible for the radio-loud/radio-quiet division. The black hole spin, and in particular its rotation compared to the disk’s, may be a key one.
author:
- 'R. M. Sambruna $^{1}$, F. Tombesi $^{2,3}$, J. N. Reeves $^{4}$, V. Braito $^{5}$, L. Ballo $^{6}$, M. Gliozzi $^{1}$ and C. S. Reynolds $^{3}$'
title: The Suzaku view of 3C 382
---
[*Subject Headings:*]{} [Galaxies: active — galaxies: radio – galaxies: individual — X-rays: galaxies]{}
Introduction
============
It is by now widely accepted that Active Galactic Nuclei (AGN), including radio-loud sources, are powered by accretion of the host galaxy gas onto a central super-massive black hole (e.g., Blandford 1985). The rich variety of AGN phenomenology has been explained in terms of orientation with respect to the axis of an obscuring equatorial thick matter surrounding the nucleus (Antonucci 1993; Urry & Padovani 1995), yielding type-1 sources for face-on views and type-2 objects for edge-on views. In addition, in radio-loud sources a relativistic jet is present, connecting the innermost regions near the black hole to the radio galaxy perifery. The jet angle – defined as the angle between the jet axis and the line of sight – increases from blazars to Broad- and Narrow-Line Radio Galaxies, roughly corresponding to an increase of obscuration degree. As the jet angle increases, the importance of its emission decreases, due to beaming effects.
X-ray spectroscopy is an effective tool to investigate the central engines of AGN. Indeed, previous X-ray observations of BLRGs in the 1990s with , , and established that these sources exhibit Seyfert-like spectra with subtle but significant differences, i.e., flatter X-ray continua and weaker reflection features than in radio-quiet (see Sambruna, Eracleous, & Mushotzky 2002; Ballantyne 2007, and references therein, for a review). In particular, the emission line around $\sim$6.4 keV was observed to be narrower and of lower EW than in Seyferts, and generally unresolved at the limited resolution of the pre- detectors. No clear evidence for a relativistic accretion disk profile was found for the line in any BLRGs, or other RL AGN. These results raised the currently held scenario that some intrinsic, fundamental difference must exist in the structure of the accretion flow between the two classes of AGN, with RQ being dominated by standard optically thick, geometrically thin disks and RL having radiatively inefficient, ADAF-like, inner disks (see discussion in Eracleous et al. 2000; Ballantyne 2007). Alternatively or in addition, a non-thermal jet contribution was also possible.
The advent of with its broad-band coverage in 0.3–100 keV and improved sensitivity especially in the critical Fe K region, 6–7 keV, has provided a golden opportunity to deepen our understanding of the central engines of RL AGN. Our group has secured GO observations of all the bright, nearby ($z<0.1$) BLRGs with relatively deep, 100 ks, exposures. The observations of our program targets are described in separate papers that concentrate on one source at the time, as each classical BLRG presents somewhat unique X-ray and multi-wavelength properties (3c 390.3, Sambruna et al. 2009, S09 hereafter; 3C 445, Braito et al. 2011; Reeves et al. 2010; 3C 111, Ballo et al. in prep.). In this paper, we focus on c, a giant lobe-dominated radio galaxy well-known for its soft excess and rapid X-ray variability from previous observations (see § 2). Besides its data, we also discuss its EPIC data and a BAT spectrum (see below), both unpublished.
The observation of c confirms the complexity of its spectrum and its similarity to Seyferts. Remarkably, we find compelling evidence in the XIS data for a [*relativistic line profile*]{}, with excellent constraints on the disk inclination and inner/outer radii (§ 5.3). This is the first time that a such a detailed profile is detected in a BLRG, and indeed in a radio-loud AGN. The implications for the structure of the central engine and models of jet formation are far reaching (§ 8).
The paper is organized as follows. After describing the source properties and previous observations in § 2, in § 3 the data reduction of the new observations is presented. The timing analysis is reported in § 4. We describe the spectral fits to the various datasets in § 5 for the combined and BAT and in § 6 for [*XMM-Newton*]{}. In § 7 we report the summary of the results and the discussion following in § 8. Finally, in Appendix A we report the background and calibration tests for the XIS cameras. Throughout this paper, a concordance cosmology with H$_0=71$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\Lambda}$=0.73, and $\Omega_m$=0.27 (Spergel et al. 2003) is adopted. The energy spectral index, $\alpha$, is defined such that $F_{\nu} \propto \nu^{-\alpha}$. The photon index is $\Gamma=\alpha+1$.
c and Previous X-ray Observations
=================================
The lobe-dominated, Fanaroff-Riley II radio galaxy c ($z=0.0579$) has a double-lobe structure, with a clear jet in the northern lobe ending in a hotspot. While a counter-jet is not clear, a hotspot is also detected in the southern lobe, with a total size between hotspots at 3.89 GHz of 179 (Hardcastle et al. 1998). From this size, a lower limit to the jet inclination of $\theta=15$ is inferred (Eracleous & Halpern 1998). While it is common to assume that in BLRGs the jet axis is parallel to the disk axis, so that $\theta \sim
i$, with $i$ the disk inclination with respect to the line of sight, in this paper we will leave $i$ free to vary during the spectral fits to the X-ray data whenever possible.
Optically, c is identified with a disturbed elliptical galaxy dominated by a very bright and unresolved nucleus (Matthews, Morgan & Schmidt 1964; Martel et al. 1999), located in a moderately rich environment (Longair & Seldner 1979). The optical spectra show a strong continuum and prominent broad lines photo-ionized by a power-law type of spectrum (Saunders et al. 1989; Tadhunter, Fosbury & Quinn 1989), with FWHM $\sim 12,000$ km/s for the H$\alpha$ line (Eracleous & Halpern 1994). A recent estimate using the luminosity of the host galaxy places the mass of the central black hole around M$_{BH} \sim 10^9$ M$_{\sun}$ within 40% (Marchesini et al. 2004).
c is a variable source at X-ray (Dower et al. 1980; Barr & Giommi 1992), radio (Strom, Willis & Willis 1978), optical, and UV frequencies (Puschell 1981; Tadhunter, Perez & Fosbury 1986). At X-rays it is well studied, and was observed by all the major X-ray observatories before . Flux and spectral variability is observed on time-scales shorter than a day, with a trend of spectral softening for increasing intensity (Gliozzi et al. 2007).
Previous low-sensitivity observations of c indicated that its X-ray spectrum was remarkably similar to Seyferts (Reynolds 1997). Aside from the variable soft excess below 2 keV, c exhibits a warm absorber with ionization parameter log$\xi \sim 2.5$ erg cm s$^{-1}$, column density N$_H^W \sim 1-3 \times 10^{21}$ , and outflow velocity $\sim$1,000 km/s, clearly detected in recent high-resolution dispersion spectra (Reeves et al. 2009; Torresi et al. 2010). The X-ray continuum can be described by a power law with the “canonical” photon index $\Gamma \sim 1.8$ (e.g., Sambruna et al. 1999, S99 hereafter) and weak reflection features, including a Compton hump above 10 keV with reflection fraction $R \sim 0.2-0.6$ (Gliozzi et al. 2007) and an line. The profile of the Fe K line has remained, so far, ambiguous due to its inherent complexity, as we show here (§ 5.1), and the poor sensitivity of the pre- detectors. For example, using an unusually broad (Gaussian width $\sigma_G$$\sim$1 keV) line was inferred (S99; Reynolds 1997) while only a narrow component with $\sigma_G$$\sim$100 eV was detected with , , and the HETGS (Eracleous et al. 2000; Grandi et al. 2001; Gliozzi et al. 2007).
The source exhibits a strong, variable X-ray excess over the extrapolation of the power-law continuum from the hard X-rays below 2 keV. Extended soft X-ray emission was detected with the HRI in 0.3–2.4 keV (Prieto 2000), later confirmed with (Gliozzi et al. 2007). The contribution to the spectral soft excess from the extended emission, however, is negligible as shown by the point-like ACIS-S image in 0.5–2 keV (Gliozzi et al. 2007).
Finally, while other BLRGs have been detected at GeV gamma-rays (3C 111; Hartman et al. 2008), no signal has been reported so far from c with (Abdo et al. 2010). The lack of gamma-ray flux and the double-lobe radio structure are strong indicators that the jet in this source is misaligned and most likely does not contaminate the nuclear emission, as supported by the Seyfert-like X-ray spectrum from the and era.
New X-ray Observations: Data Reduction
======================================
The log of the X-ray observations is reported in Table 1. The exposures are after data screening and the source count rates after background subtraction, according to the procedures described below.
XIS
----
observed c on April 27th, 2007 for a net exposure time after screening of $\sim$116 ks. We used the cleaned event files obtained from version 2 of the pipeline processing. Standard screening criteria were used, namely, only events outside the South Atlantic Anomaly (SAA) as well as with an Earth elevation angle (ELV) $ > 5\ensuremath{{}^{\circ }}$ were retained, and Earth day-time elevation angles (DYE\_ELV) $ > 20\ensuremath
{{}^{\circ}}$. Furthermore, data within 256 s of the SAA passage were excluded and a cut-off rigidity of $ >6 \,\mathrm{GV}$ was adopted. The main parameters of the XIS observations are reported in Table 1.
The XIS spectra of c were extracted from a circular region of 2.9$'$ radius centered on the source. Background spectra were extracted from four circular regions offset from the main target and avoiding the serendipitous source and the calibration sources. The combined area of these four background regions is twice the area of the main target region. The XIS response (rmf) and ancillary response (arf) files were produced, using the latest calibration files available, with the *ftools* tasks *xisrmfgen* and *xissimarfgen*, respectively. The source spectra from the front illuminated (FI) CCDs were summed, and fitted jointly with the back illuminated (BI), the XIS1, spectrum after verifying that the data from the FI and BI CCDs were consistent with each other. In fact, we checked that the 2–10 keV continuum slope and fluxes in the two cases are consistent within the 2%. The net XIS background-subtracted source spectra were binned with a minimum of 25 counts per energy bin to ensure that the $\chi^2$ goodness of fit can be applied to the spectral analysis. This binning is maintained throughout the spectral analysis and only in some figures it was increased for clarity, when specified. Consistent results were found by varying the binning in a range up to 100 counts/bin.
HXD
----
For the HXD-PIN data reduction and analysis we followed the latest data reduction guide (the ABC guide Version 2). We used the rev2 data, which include all four cluster units, and the best background available, which account for the instrumental background (Kokubun et al. 2007). The source and background spectra were extracted within the common good time interval and the source spectrum was corrected for the detector dead-time. The net exposure time after screening was 113 ks.
The contribution of the diffuse cosmic X-ray background counts was simulated using the spectral form of Boldt (1987), assuming the response matrix for diffuse emission, and then added to the instrumental one. Two instrumental background files are available; background A or “quick” background, and background D or “tuned” background. We adopted the latter which is the latest release and which suffers lower systematic uncertainties of about 1.3%, corresponding to about half the uncertainty of the first release of the Non X-ray Background. With this choice of background, c is detected up to 70 keV with the PIN at a level of $\sim$20% above the background. The count rate in 10–30 keV is 0.135 $\pm$ 0.003 c/s. The HXD PIN spectrum was binned in order to have a signal-to-noise ratio greater then 10 in each bin, and the latest response file released by the instrumental team was used.
For a comparison with the *Swift* BAT spectrum, we considered the PIN spectrum in the overlapping energy band of the two detectors, E$=$15–70 keV (see § 3.4). In this band, the PIN spectrum can be fitted with a single power law with $\Gamma_{PIN}=1.95\pm0.14$, yielding a flux of $6.37^{+0.19}_{-1.23}\times 10^{-11}$ , with 90% errors.
observed c on April 28th, 2008, for a net exposure after screening of 20 ks. The analysis of the RGS spectrum has already been performed by Torresi et al. (2010) and is not reported here.
We downloaded and reduced the data of the three EPIC cameras, pn, MOS1 and MOS2. The data reduction was performed following the standard procedure with the XMM-SAS v. 8.0.1 package. We checked the observation for high background contamination, looking for flares in the light curves at energies greater than 10 keV. We excluded these bad intervals from the successive analysis, and extracted the source photons from a circular region of 40, while the background ones were collected from an adjacent source free circular region of the same size. Only single and double events were selected.
Using the SAS task *epatplot* we checked that the pile-up fraction of the EPIC pn is negligible, less than 1%. Instead, given the X-ray brightness of the source, the pile-up fraction of the EPIC MOS observations is significant, $\sim$50%. The main result of pile-up is to introduce an overall distortion of the observed continuum, as it basically consists in counting two or more low-energy photons as a single high-energy photon. In fact, we find that the 2–10 keV power-law continuum of the pn has a $\Gamma$$\simeq$1.74, while for the MOS $\Gamma$$\simeq$1.6. Therefore, given the significant pileup of the MOS and and the $\la$25% lower S/N ratio of the MOS data, due to the short exposure and lower effective area, we limited the analysis of only to data from the EPIC pn camera.
Similarly to the XIS, the net EPIC pn background-subtracted source spectrum was binned with a minimum of 25 counts per energy bin to ensure that the $\chi^2$ goodness of fit can be applied to the spectral analysis This binning is maintained throughout the successive spectral analysis and increased only in some figures for clarity of presentation, when specified. Consistent results were found by varying the binning in a range up to 100 counts/bin.
BAT Observations
-----------------
The BAT spectrum was derived from the 58-month hard X-ray survey. The data reduction and extraction procedure of the 8-channel spectrum is described in Baumgartner et al. (2010). To fit the BAT spectrum, we used the latest calibration response `diagonal_8.rsp` and background files as of December 2010.
The source is clearly detected in the energy band E$=$15–200 keV with a S/N$=$35 and a count rate $(1.35\pm0.05) \times 10^{-3}$ c/s. A fit to the mean 58-month BAT spectrum in the 15–70 keV band with a single power law model yields a photon index $\Gamma_{BAT}=2.12\pm0.20$ and a flux of $5.10^{+0.16}_{-1.93} \times 10^{-11}$ , at 90% uncertainty. These values are consistent with those derived for the HXD PIN in the same energy range (see §3.2), justifying joint fits using both instruments. This is especially important in view of the extended sensitivity to higher energies provided by the BAT, which will allow us to set better constraints on reflection models.
Timing Analysis
===============
Figure 1 shows the XIS0-3 light curves of c at soft and hard X-rays, respectively, with the flux in bins of 5760 sec (1 orbit). The data from the three detectors are plotted on separate panels for comparison. Weak flux variations are seen in both energy bands for all three detectors, with small-amplitude variations of the order of few percent. When the ratio of the counts in the two energy bands is plotted against time or total count rate, there is no evidence for spectral variability. We note the presence of two possible short time-scale flares of $\sim$5% amplitude in the XIS 3. We ascribe them to spurious instrumental fluctuations. However, as discussed in the Appendix, they have a negligible influence on the spectral analysis of the average spectrum reported here. Aside from the small apparent flares, the overall shape of the light curves of the different XIS detectors is consistent, with each showing a gradual increase in counts over the Suzaku observation.
The variability analysis was repeated for the EPIC data, despite the short exposure affected by the large background flares. No evidence for flux changes was found. Table 2 reports the soft (0.5–2 keV) and hard (2–10 keV) fluxes of c from the and observations. The reported fluxes are both observed and corrected for any intervening absorption and were derived from the best-fit models to the 0.6–200 keV +BAT and 0.5–10 keV EPIC pn spectra (see § 5.4 and § 6). Observed and absorption corrected luminosities are reported in Table 2 as well. Comparing both fluxes and luminosities in all energy bands, it is clear that the source was in similar intensity states at the epoch of the and observations.
The Broad-Band Spectrum
========================
The lack of significant spectral variability derived from a model-independent timing analysis of the XIS data (§ 4) indicates that the spectral properties of c can be safely investigated using a time-averaged spectrum accumulated over the entire 116 ks exposure. This spectrum provides the largest signal-to-noise ratio available from this observation, allowing the most detailed spectral analysis of the Fe K line region and broad-band continuum, the main goals of our project.
All spectral fits were performed using `XSPEC` v.12.6.0. The significance of adding free parameters to the model was evaluated with the F-test. We included only model components with a statistical probability $P_F$$>$99%. All uncertainties quoted are 1$\sigma$ for one parameter of interest, unless otherwise stated. The energy of the lines is reported in the source rest-frame, if not otherwise stated. The data from the XIS CCDs were fitted jointly, excluding the energy range E$=$1.5–2 keV, around the Si K edge, which is known to be affected by calibration issues (Koyama et al. 2007). When fitting the broad-band spectrum, a constant factor was introduced in the model to take into account the cross-normalization between the different instruments. In particular, a constant factor of $\simeq$1.16 was adopted for the XIS/PIN and of $\simeq$0.9 for the PIN/BAT. Throughout the spectral analysis we always include a Galactic absorption towards c of $N_H$$=$$7.4\times10^{20}$ (e.g., Dickey & Lockman 1990; Kalberla et al. 2005).
The plan of the analysis is as follows. First, we consider the combined broad-band and BAT spectra in the E$=$0.6–200 keV, excluding the Fe K band, and derive an estimate of the underlying X-ray continuum. Then, we consider also the Fe K band and perform a search for additional emission/absorption features in the spectrum. Subsequently, we perform fits to the full broad-band spectrum with a physical model including all the spectral components identified in the previous steps plus an additional relativistic line and smeared ionized reflection from the putative accretion disk. Finally, we test the simultaneous alternative modeling of the soft excess, Fe K lines and cold reflection with a second lowly ionized reflection component.
The Continuum Determination
---------------------------
The combination of the and BAT instruments provides an unprecedentedly wide and sensitive spectral coverage from 0.6 keV up to 200 keV. This is fundamental for a correct determination of the underlying X-ray continuum. From this analysis, we exclude the Fe K energy range, E$=$5.5–7.5 keV, because, as discussed in the next section, the presence of spectral complexities in this interval can affect the best-fit continuum model.
In the upper panel of Figure 2 we show the ratio of the combined XIS03, XIS1, PIN, and BAT spectra with respect to a Galactic absorbed power-law continuum. As seen, this simple model does not provide a sufficient parameterization of the data ($\chi^2/dof$$=$5027/3166) and strong deviations are present at both low and high energies. The resultant power-law photon index is $\Gamma$$\sim$1.88.
At energies lower than $\sim$2 keV there is evidence for a soft excess component. This is expected also from previous studies of c, see § 2. Following previous authors, we initially model the soft excess with a phenomenological blackbody with best-fit temperature $kT$$\simeq$100 eV. This is highly required, with a $\Delta\chi^2$$\simeq$1407 for two additional model parameters. The alternative phenomenological modeling with a power-law provides a comparable fit. However, the physical interpretation of the resultant very steep photon index of $\Gamma$$\sim$5 is more ambiguous and therefore in the following we keep the blackbody parameterization.
The excess flux in the upper panel of Figure 2 between 10–40 keV is instead indicative of a reflection component. This is also required on a physical basis by the presence of a narrow line (e.g., Lightman et al. 1988), as discussed in the next section. Therefore, we added a cold reflection component using *pexrav* in `XSPEC`. We assume an inclination angle of $i$$=$30$^\circ$, typical for BLRGs, and standard solar abundances. This component is highly required, with a $\Delta\chi^2$$\simeq$66 for two model parameters, significant at $>$99.99%. The power-law continuum slope is now $\Gamma$$\simeq$1.85, the high energy cut-off E$_c$$\sim$200–300 keV, and the reflection fraction from *pexrav* is R$\simeq$0.36. The final fit is good, with a reduced chi-squared very close to unity, $\chi^2/dof$$=$3554/3162. The ratios of the spectra with respect to this best-fit continuum model, including the Fe K band, are shown in the middle panel of Figure 2.
In conclusion, the broad-band X-ray continuum of c is well parameterized by a Galactic absorbed power-law, plus a blackbody soft excess and a reflection component. We use this as the baseline continuum model for the successive search of additional emission/absorption spectral features.
Search for Emission/Absorption Spectral Features
------------------------------------------------
Looking at the spectral ratios in the middle panel of Figure 2, weak dips indicating absorption can be seen at E$<$2 keV, as well as further spectral complexities in the Fe K region. The BLRG c is known to exhibit absorption features at energies below 2 keV due to a warm absorber (Reynolds 1997), recently studied with the and gratings (Reeves et al. 2009; Torresi et al. 2010). Despite the XIS lower resolution compared to the gratings, absorption features are visible in Figure 2 (middle panel) around 0.8–1 keV, possibly due to ionized O and Fe L. We thus include a warm absorber component in our fit, modeled by an `XSTAR` grid (Kallman & Bautista 2001). Given the limited S/N and energy resolution available with respect to the gratings studies, we fix the gas outflow velocity to 1,000 km/s, consistent with what found by Reeves et al. (2009) and Torresi et al. (2010). We assumed an ionizing X-ray power-law with photon index $\Gamma$$=$2 and turbulence velocity of 500 km/s, similar to what assumed by Reeves et al. (2009).
The inclusion of the warm absorber component is highly significant, $\Delta\chi^2$$=$30 for two additional parameters, corresponding to a detection probability $\ga$99.99%. We estimate a column density N$_H$$\simeq$$6\times 10^{20}$ cm$^{-2}$ and an ionization parameter log$\xi$$\simeq$$2.44$ erg s$^{-1}$ cm.
We now reintroduce the E$=$5.5–7.5 keV energy band in the fits. Up to now the fit statistics is $\chi^2/dof = 3524/3160$. Including the Fe K band, this changes to $\chi^2/dof = 4760/4177$. The middle panel of Figure 2 and upper panel of Figure 3 show the presence of spectral complexities in the Fe K band. In particular, a prominent emission line is visible at rest-frame energy $\sim$6.4 keV, together with several other emission lines in the energy range $\sim$5.5–7.5 keV.
As a first step, we modeled these emission features with Gaussians. The most prominent line at E$=6.41\pm0.01$ keV is clearly identified with the from lowly ionized, fluorescent material. The line is resolved, with a width of $\sigma_G$$=$$118^{+20}_{-13}$ eV, equivalent width EW$=$$68\pm9$ eV, and flux I$=$$(3.3\pm0.2)\times
10^{-5}$ ph s$^{-1}$ cm$^{-2}$. The inclusion of this feature yields a $\Delta\chi^2$$=$233 for three additional model parameters, which corresponds to a detection probability $\gg$99.99%.
Next, we added the associated component. In the fit, the energy was fixed to the expected value of E$=$7.06 keV, the width was set equal to that of the and its intensity to $\simeq$13% of the (e.g., Molendi, Bianchi & Matt 2003 and references therein). The addition of the improves the fit by $\Delta\chi^2$=13.
Three additional emission features are still present in the spectral residuals at energies of E$\sim$5.8, 6.9, and 7.5 keV. To account for these, we added three more Gaussians to the model. The two lines at higher energies are not resolved. Their energies are at E$=$$6.91\pm0.02$ keV and E$=$$7.51\pm0.03$ keV and have equivalent widths of $17\pm6$ eV and $18\pm7$ eV, respectively. The $\Delta\chi^2$ is 21 and 19, for two additional model parameters, which correspond to detection probabilities of $\simeq$99.99%. The lower energy line at E$=$$5.70\pm0.09$ keV is instead broader and resolved, with a width of $\sigma$$=385^{+77}_{-57}$ eV and equivalent width of $55\pm 12$ eV. Its significance is higher, $\Delta\chi^2$$=$58 for three additional parameters, corresponding to a detection probability $>$99.99%.
The fit with the five Gaussian lines is purely phenomenological, in order to have a first representation of the Fe K complex and a determination of the lines parameters and their significance. All the Gaussian lines are detected with a probability $\ga$99.99%. Concerning the identification of the three additional emission lines, the feature at 7.5 keV may well be identified with emission from lowly ionized Ni K$\alpha$, whose atomic transition energy is expected at E$\simeq$7.5 keV (e.g., Molendi, Bianchi & Matt 2003 and references therein). The two remaining lines at 6.9 keV and 5.8 keV are at energies not directly consistent with any significant atomic transitions and therefore have a less obvious physical interpretation. The closest atomic lines to the 6.9 keV line are the Fe XXV He$\alpha$ (1s$^2$–1s2p) at E$\simeq$6.7 keV and Fe XXVI Ly$\alpha$ (1s–2p) at E$\simeq$6.97 keV (Kallman et al. 2004). However, the identification with one of these lines would require a blue/red-shift of $\simeq$0.03c or $\simeq$0.01c, respectively. The 5.8 keV line is even more mysterious, being out of the range expected for Fe K lines, from 6.4 keV up to 7 keV, depending on the ionization. If identified with Fe K emission, the large width and red-shift could suggest strong relativistic effects and an origin close to the central black hole (see next section).
The inclusion of the warm absorber and the emission lines in the Fe K band on the continuum model provides a very good fit, with a $\chi^2/dof$$=$4416/4167. Therefore, the final best-fit model of the combined broad-band and BAT spectrum of c is rather complex and is composed by a Galactic absorber power-law continuum, a black body soft excess, a cold reflection component, a warm absorber and five Gaussian emission lines in the Fe K band. The best-fit parameters are reported in Table 3. The spectral ratios with respect to this model are shown in the lower panel of Fiure. 2.
A Relativistic Fe K Emission Line in 3C 382?
--------------------------------------------
Inspection of Figure 3 (upper panel) suggests the intriguing possibility that the 5.7 keV and 6.9 keV emission features represent two components of a single, broader feature, i.e. a relativistic Fe K line. More precisely, they would correspond to the broad red-wing and sharp blue-peak components of such a line, respectively (e.g., Fabian et al. 1989). As an initial test to this hypothesis we replaced these two features with a single relativistic line component, *diskline* in `XSPEC` (Fabian et al. 1989), in the model discussed in the previous section and reported in Table 3. Therefore, now the Fe K band is composed by three narrow emission lines, an and associated Fe K$\beta$ and Ni K$\alpha$, and a broad relativistic line (see middle panel of Figure 3).
The *diskline* model considers a non-rotating Schwarzschild black hole, with spin parameter $a$$=$0. Given that the fit is not sensitive to the power-law emissivity profile, we assumed the typical value of $\beta$$\simeq$$-2.5$ (e.g., Nandra et al. 2007; de la Calle Perez et al. 2010; Patrick et al. 2010). However, we checked that a different choice of $\beta$ in the wide range from $-5$ to 0 does not significanly change the parameters estimates reported below, being always consistent at the 90% level.
Interestingly, if the energy of the line is left free to vary, the fit yields a value larger than the expected 6.4 keV for neutral emission. The formal best-fit centroid is at E$=$6.69 keV, ranging between 6.55 and 7.10 keV at 90% confidence. This indicates that the relativistic Fe K emission originates from ionized material, preferentially Fe XXV He$\alpha$ at 6.7 keV rest-frame, but we can not clearly exclude also a contribution from Fe XXVI Ly$\alpha$ at 6.97 keV. This supports earlier models claiming that the accretion disk in BLRGs is ionized (Ballantyne et al. 2002), and adds to previous evidence from the observations of two other classical BLRGs, 3C 390.3 (Sambruna et al. 2009) and 3C 120 (Kataoka et al. 2007).
Thus, we performed fits assuming, in turn, either emission from Fe XXV or Fe XXVI. In both cases the fit is highly improved, $\Delta\chi^2$$=$60 for 4 additional parameters, corresponding to a detection significance $>$99.99%. The middle panel of Figure 3 illustrates the best-fit model for the three Gaussians and the *diskline* profile. Assuming Fe XXV (Fe XXVI) emission, we obtain good constraints on the disk inclination $i=30^\circ \pm 1^\circ$ ($26^\circ \pm 2^\circ$), on the inner and outer radii of the emitting region, $r_{in} =12 \pm 2$ $r_g$ ($10\pm2$ $r_g$) and $r_{out}=23 \pm 3$ $r_g$ ($18\pm3$ $r_g$). Indeed, a tight estimate of $i$ was expected from the 6.9 keV peak sharp fall to the blue (see Figure 3, middle panel), while the limited extension of the red “wing” indicates that the inner radius can not extend too further in close to the black hole. The equivalent width of the line is EW$=$$80\pm25$ eV and the intensity I$=$$(3.2\pm0.4)\times
10^{-5}$ ph s$^{-1}$ cm$^{-2}$.
We stress that the inclusion of the warm absorber has a negligible effect on the Fe K parameters (e.g., Reeves et al. 2004 and references therein). This can be explained by the low level of ionization and small column density of the warm absorber, which implies that only the light elements (C, O, Mg, Ne) are contributing to the absorption at energies below 2 keV.
The final fit with the relativistic line profile is very good, $\chi^2/dof=$4435/4168 and the data/model ratio with respect to this model are shown in the lower panel of Figure 3. Given the high quality of the XIS data in 4–10 keV and good determination of the underlying continuum in the wide 0.6–200 keV band, we regard the detection of a broad, relativistic Fe K line in c as robust. In support of this, we performed extensive calibration and background tests, which are described in the Appendix.
Detailed Modeling of the Relativistic Line and Ionized Reflection
-----------------------------------------------------------------
Here we discuss in more details the modeling of the broad Fe K line and possible ionized reflection component in the broad-band 0.6–200 keV spectrum. We first examine the broad profile in light of the possibility that it could be due to a rotating (either prograde or retrograde) black hole. This is possible by using the relativistic line code *relline* of Dauser et al. (2010), which allows also the black hole spin $a$ to vary between $-$0.998 and $+$0.998. We use the final best-fit model discussed in § 5.2 and reported in Table 3, substituting the 5.7 keV and 6.9 keV emission lines with a single broad relativistic line.
The main parameters of *relline* are the rest-frame energy of the line, E, the power-law emissivity profile, $\beta$, the black hole spin, $a$, the inner and outer radii of the reflecting region, $r_{in}$ and $r_{out}$, the disk inclination, $i$, and the intensity of the emission line, $I$. At first, all the above parameters were free to vary during the fit. Only $r_{in}$, $r_{out}$, $i$, $I$ were constrained. We explored the parameter space in more detail, checking the effects of assuming different values. However, given the limited S/N of the present data and the complexity of the model, we are unable to constrain the spin, emissivity and energy all together. Nevertheless, good fits are obtained for $|\beta|$$\simeq$2–3. The line energy is E$\sim$6.6–7 keV at 90%, confirming ionized emission. Similar fits are obtained assuming Fe XXV He$\alpha$ at E$=$6.7 keV or Fe XXVI Ly$\alpha$ at E$=$6.97 keV, with slightly smaller $r_{in}$ and $r_{out}$ in the latter case.
We also explored the dependence of the line shape on the black hole spin by imposing that the inner radius coincide with the Innermost Stable Circular Orbit (ISCO) of the disk. This corresponds to assuming that the beginning of the reflection region is coincident with the inner radius of the disk and that the latter extends down to the ISCO. The ISCO assumes the following values depending on $a$: $r_{ISCO}$$\simeq$1.23$r_g$, 6$r_g$ and 9$r_g$ for $a$$=$$+0.998$, $0$ and $-0.998$, respectively. A series of fits were performed with the inner radius linked to the ISCO parameter in the `relline` model in `XSPEC`, and leaving $a$ free to vary. However, also in this case we can not still constrain all the parameters simultaneously and find a clear deep minimum in the statistical distribution. For instance, given the values in the intervals $\beta$$=$1–3, E$=$6.7–6.97 keV and $a$$=$$-0.998$–$+0.998$, there are several minima, all within a difference of a few in $\Delta\chi^2$, and the fit does not reach a stable convergence.
Therefore, despite the good quality of the present data and the broad-band coverage up to 200 keV, no sensitive constraint was possible on the value of the black hole spin, i.e., we obtain similar fits for $a=-0.998$, $a=0$, and $a=+0.998$. Detailed simulations show that much larger exposures are needed, $\sim$ 300 ks, to achieve the goal of measuring $a$ with sufficient accuracy to distinguish among a static, prograde or retrograde spinning hole.
Given the results described so far, in the remaining fits to the 0.6–200 keV spectrum of c the broad Fe K line energy was fixed to E$=$6.7 keV, $a$$=$0, and $\beta=-2.5$, an intermediate value consistent with Seyferts (Nandra et al. 1997, Patrick et al. 2010, de la Calle Perez et al. 2010). Including this component, the fit is improved by $\Delta\chi^2$$=$60 for 4 additional parameters, corresponding to a high detection probability $>$99.99%. This fit yields $r_{in}$$=$$12\pm2$ $r_g$ and $r_{out}$$=$$23^{+4}_{-2}$ $r_g$, with $i$$=$$30^{\circ}\pm 1^{\circ}$. The line intensity is $I$$=$$(3.0\pm0.4)\times 10^{-5}$ ph s$^{-1}$ cm$^{-2}$ and EW$=$$76\pm18$ eV. This indicates that the broad, ionized Fe K line is produced from a narrow annulus of $r_{in}$$\sim$10$r_g$ and $r_{out}$$\sim$20$r_g$, i.e., at some distance from the black hole. We checked that a different choice of the $\beta$ in the range $-2$ to $-3$ does not change significantly the parameters estimates, being always consistent within the 90% errors.
As discussed above, the presence of ionized Fe K emission lines in the energy range 6–7 keV provides strong motivation for including an additional ionized reflection component to our baseline continuum model. Sophisticated models are known in the literature that include the proper number of emission lines expected for the fitted ionization parameter, $\xi$ (e.g., Ross, Fabian & Young 1999; Ross & Fabian 2005; Garc[í]{}a & Kallman 2010). Two such models, *reflionx* (Ross & Fabian 2005) and the more recent *xillver* (Garc[í]{}a & Kallman 2010), are available in `XSPEC` for detailed fits. We choose to use the latter, convolved with the *relconv* model of Dauser et al. (2010), which takes into account line blurring due to relativistic effects. We use a *xillver* table with log$\xi$$=$0.8-3.8 erg s$^{-1}$ cm, standard solar abundances, and an incident ionizing power-law with $\Gamma$$=$2. Consistent results were obtained using *reflionx*.
The free parameters of the *relconv* component are the power-law emissivity profile, inner/outer radii on the disk surface, spin, and inclination angle. The free parameters of *xillver* are instead the ionization level of the material and the normalization. During the first fits to the 0.6–200 keV spectrum, these parameters were left free to vary. The fits yielded $r_{in}$$\simeq$10 $r_g$, $r_{out}$$\simeq$25 $r_g$, inclination $i$$\simeq$30$^{\circ}$, and ionization parameter log$\xi$$\simeq$3 erg s$^{-1}$ cm, while the disk emissivity $\beta$ was less constrained and the spin $a$ was unconstrained. After performing a series of tests, we fixed $\beta$$=$$-2.5$, as no significant changes in the line and continuum parameters were observed within 90% uncertainties if the emissivity profile was left free to vary in the most probable range $|\beta|$$=$2–3. We then linked $r_{in}$$=$$r_{ISCO}$ in *relconv*, and performed three separate fits for $a=+0.998$, $0$, and $-0.998$, to explore the dependence on the spin. A slightly better fit was obtained for the case of negative spin, with $\Delta\chi^2$$\simeq$2 and $\simeq$5 higher with respect to the cases with zero or positive rotation. Thus, we conclude again that the quality of the current data of c does not allow us to constrain the sign of the black hole spin. Future deeper observations ($\ga$300 ks) are needed to this end. However, given that the fit is slightly better for a spin of $-0.998$ and that the overall model parameters are still consistent at the 90% whatever spin is assumed, in the following fits and in Table 4 we fix the spin to this value. The inclusion of the ionized reflection component improves the fit by $\Delta\chi^2=65$ for four additional model parameters, which correspond to a detection probability of $\gg$99.99%.
The estimated ionization of the disk reflecting material is log$\xi=3.04^{+0.03}_{-0.04}$ erg s$^{-1}$ cm, consistent with the Fe ion population being dominated by Fe XXV, with the He$\alpha$ at E$\simeq$6.7 keV being the most intense associated resonant line (e.g., Ross, Fabian & Young 1999; Ross & Fabian 2005; Garc[í]{}a & Kallman 2010). The fitted outer radius is $r_{out}$$=$$26^{+6}_{-4}$ $r_g$ and the inclination $i$$=$$29^{\circ}\pm2^{\circ}$, consistent with the more simplistic fits using *diskline* or *relline*. The fit is very good, with a reduced chi-squared very close to unity, $\chi^2/dof = 4430/4168$. The best-fit parameters of this broad-band fit are reported in Table 4, while the data, best-fit model and spectral ratios are shown in the left panels of Figure 4.
Two ionized reflection components in 3C 382?
--------------------------------------------
It should be noted that the modeling of the soft excess with a simple black body component in the previous sections, besides providing a good representation of the data, might be difficult to explain from a physical point of view. In particular, from a systematic X-ray spectral analysis of PG quasars, Gierlinski & Done (2004) demonstrated that if the soft excess is modeled with a simple black body component, the resultant temperature is in the narrow range $kT$$\sim$100–200 eV, despite the large range in Eddington luminosities and black hole masses of the sources. The temperature is also too high to be directly associated with standard Shakura-Sunyaev accretion disks. This suggests a possible atomic origin for the soft excess and a promising alternative physical explanation is in terms of a blurred ionized reflection component (e.g., Crummy et al. 2006). However, the modeling of the soft excess is still required after the inclusion of the ionized disk reflection component in the previous section. When parameterized with a black body, it still provides a high improvement of the fit, $\Delta\chi^2$$=$538 for two model parameters, which corresponds to a detection probability $\gg$99.99%.
The presence of an additional mildly ionized reflection component is required to self consistently model the Fe K$\alpha$/K$\beta$, Ni K$\alpha$ emission lines and reflection hump. In particular, the energy of the Fe K$\alpha$ is $\sim$6.4 keV for a wide range of ionization, from Fe I to Fe XVII (Kallman et al. 2004). This gives rise to the possibility of simultaneously modeling the soft excess, Fe K lines and reflection hump with an additional reflection component. Therefore, we performed a test replacing the previous *pexrav*, black body and Fe K Gaussian emission lines with a cutoff power-law continuum and an additional *xillver* reflection component. To account for the velocity broadening, we convolved the latter with a Gaussian profile with the same width of the Fe K$\alpha$ emission line, i.e., $\sigma$$=$90 eV, which corresponds to FWHM$\sim$10,000 km/s (see Table 4).
The best-fit values are reported in Table 5. The power-law continuum, warm absorber and relativistic highly ionized reflection parameters are essentially unchanged. The ionization parameter of the mildly ionized reflection is log$\xi$$=$$1.5\pm0.03$ erg s$^{-1}$ cm. This is consistent with Fe XIII–XVII being the most abundant Fe ions (Kallman et al. 2004; Garc[í]{}a, Kallman & Mushotzky 2011). As shown in the right panels of Figure 4, this provides a very good fit, with $\chi^2$$=$4467 for 4174 degrees of freedom. The fit is only slightly worse than that with the model in Table 4. The reduced $\chi^2$ is 1.06 and 1.07 using the previous and present model, respectively. However, the present model has the strong advantage to physically self-consistently explain all the main spectral features in the broad-band spectrum with the lowest number of model components. This result is remarkably similar to what observed for local Seyfert galaxies (e.g., Nandra et al. 2007; de la Calle Perez et al. 2010; Patrick et al. 2010).
The reflection fraction $R$ is conventionally used to estimate the relative contribution of the reflection component with respect to the power-law continuum emission. This parameter is directly derived when using the *pexrav* model in `XSPEC`, as in Table 3 and Table 4. This is related to the solid angle covered by the reflecting gas as $R$$=$$\Omega/2\pi$. However, anisotropies, obscuration and relativistic effects close to the black hole can all influence this simple conventional relation for AGN, to the point that it can become physically meaningless (Crummy et al. 2006; Garc[í]{}a, Kallman & Mushotzky 2011). Therefore, we define a phenomenological *reflection flux fraction*, $R_F$, as the ratio between the unabsorbed 0.5–100 keV flux of the reflection component and continuum. Its value is still comparable to the conventional parameter, $R_F$$\sim$$R$ (Garc[í]{}a, Kallman & Mushotzky 2011). As reported in Table 5, this fraction is about 0.1 for both highly and mildly ionized reflection components, which means that each of them contributes $\sim$10% on the overall broad-band spectrum. In particular, the mildly (highly) ionized component contributes $\sim$20% ($\sim$30%) in the 0.5–2 keV, $\sim$5% ($\sim$10%) in the 2–10 keV and $\sim$10% ($\sim$5%) in the 10–200 keV with respect to the power-law continuum (see Figure 4 right, upper panel).
Due to the wide-band coverage allowed by the BAT, very good constraints can be obtained on the power-law continuum energy cutoff. The latter value of $E_c$$\simeq$180 keV is completely consistent for both models presented in this section and § 5.4, see Table 4 and Table 5. This clearly suggests a thermal disk-corona origin for the primary power-law. Thus, it is unlikely that the X-ray power-law continuum of c is due to a jet.
In conclusion, the broad-band 0.6–200 keV spectrum of c can be described by a very complex model composed by a Galactic absorbed power-law continuum, a black body soft excess, a warm absorber, an Fe K$\alpha$ emission line and related Fe K$\beta$, a Ni K$\alpha$ emission line, a mildly ionized reflection component and a relativistic smeared ionized reflection component. Alternatively, it can be well represented also by a Galactic absorbed cutoff power-law continuum, a warm absorber and two ionized reflection components: a mildly ionized one which simultaneously reproduces the narrow Fe K$\alpha$ emission line, soft excess and high energy hump and a highly ionized one which reproduces the broad relativistic Fe K line. These two models are indistinguishable from a statistical point of view, but the latter provides a more physically self-consistent representation of the data, with the lowest number of model components.
The Spectrum
=============
The spectral analysis of the EPIC pn observations was carried out using the *heasoft* v. 6.5.1 package. As shown in Figure 5, the overall 0.6-10 keV pn spectrum is very similar to the XIS one, albeit at lower S/N due to the lower exposure (factor 6) which compensates for the larger EPIC effective area compared to the XIS cameras. The ratios against a Galactic absorbed power-law continuum are reported in the middle panel of Figure 5. As we can see, the fit is very poor ($\chi^2/dof = 4149/1454$) and the power-law photon index is $\Gamma \simeq 2$. Intense emission residuals are present at low and high energes, similar to the ones in the upper panel of Figure 2. These deviations indicate the presence of a soft excess and a reflection components.
Because of the lower quality of the data and the spectral similarities, and because the source was in similar intensity states at both epochs (Table 2), we assumed for the EPIC pn the best-fit model previously derived for the data from the fits to the broad-band spectrum discussed in § 5.2 and reported in Table 3. We assumed only a cold reflection component modeled with *pexrav* and fixed the cut-off energy and reflection fraction to their best-fit values. We also assumed only the presence of the narrow Fe K$\alpha$/K$\beta$ emission lines, as indicated by the data.
The fit to the EPIC pn data with this model is good, $\chi^2/dof$$=$1567/1448. The best-fit model and spectral ratios are reported in the upper and lower panels of Figure 5, respectively. The power-law continuum photon index is $\Gamma$$=$$1.87\pm0.01$. The EPIC data confirm the presence of a strong soft excess, with a blackbody temperature k$T$$=$$142\pm5$ eV and normalization $(1.72\pm0.03)\times
10^{-4}$. The warm absorber is also detected, with column density $N_H$$=$$1.8^{+0.5}_{-0.2}\times 10^{21}$ cm$^{-2}$ and ionization parameter log$\xi$$=$$2.35\pm0.05$ erg s$^{-1}$ cm, consistent with (§ 5.2).
An unresolved ($\sigma$$=$100 eV) line is detected at E$=$$6.4\pm0.03$ keV, with a flux $I$$=$$(2.5\pm0.4)\times
10^{-5}$ ph s$^{-1}$ cm$^{-2}$ and EW$=$$54^{+22}_{-14}$ eV. Due to the lower S/N of the pn data, there is no clear evidence for the associated and Ni lines present in the XIS data, or for a broad, relativistic Fe K line at 6.67 keV. The 90% upper limit for the latter is EW$<$130 eV.
In summary, an archival 20 ks EPIC pn observation of c was analyzed and found consistent with the best-fit model to the XIS data. The observed and intrinsic (corrected from any intervening absorption) fluxes and luminosities in the 0.5–2 keV and 2–10 keV bands from the EPIC pn data are reported in Table 2.
Summary of Observational Results
================================
This paper presented a 116 ks observation of the Broad-Line Radio Galaxy c ($z$=0.057). A shorter (20 ks), unpublished EPIC exposure was discussed, as well as a BAT spectrum from integrating 5-years of survey observations. The main results of the data analysis are:
[*The soft X-ray spectrum -* ]{} Below 2 keV, the XIS data indicate the presence of a warm absorber, with parameters consistent with those derived from the analysis of a previous HETGS spectrum. The soft excess previously observed is confirmed by both and with an increase in flux between the two epochs of $\sim$10%. This component can be well modeled by a phenomenological black body with $kT$$\sim$100 keV, but a more physical parameterization suggests a possible origin from mildly ionized reflecting material.
[*The Fe K Region -* ]{} An intense narrow Fe K$\alpha$ emission line is detected at E$\simeq$6.4 keV, together with the associated Fe K$\beta$ component. An additional weak unresolved emission line at E$\simeq$7.5 keV is tentatively observed, possibly identifiable with Ni K$\alpha$. However, more interestingly, a prominent Fe K line with a relativistic profile is found; its center energy is consistent at 90% confidence with emission from ionized Fe, either or both FeXXV/FeXXVI in the energy range 6.55–7.10 keV. The profile is consistent with emission from a disk annulus between inner and outer radii $r_{in}
\sim 10$ $r_g$ and $r_{out} \sim 20$ $r_g$, respectively, for an assumed disk emissivity $\beta=-2.5$. An ionized disk reflection model was included to parameterize this component in the broad-band spectrum.
[*The spectrum at energies $\ga$10 keV -* ]{} The data at medium and hard X-rays are well described by a power law with $\Gamma\simeq1.8$, cutting off around E$_c\sim 200$ keV. Above 10 keV, contributions from both mildly and highly ionized reflection are present. Each of them contribute $\sim$10% on the overall broad-band continuum, but the mildly ionized one has an higher impact at E$>$10 keV.
Discussion
==========
This paper presents new results from the analysis of our as well as archival and BAT observations of c, one of the “classical”, best-studied BLRGs from optical and radio samples. The combination of the and BAT data provides an unprecedented sensitive coverage in the energy range 0.6–200 keV, useful to disentangle the various contributions to the total nuclear X-ray emission. At first sight, the broad-band X-ray spectrum of c is remarkably Seyfert-like with most, if not all, the characteristics of a typical nearby radio-quiet AGN: a $\Gamma \sim
1.8$ power-law continuum, a warm absorber, soft excess, relativistic line, and a reflection bump above 10 keV. While typically radio-loud in other wavelengths, with giant radio lobes and a radio-to-optical Spectral Energy Distribution typical of other BLRGs (e.g., Grandi et al. 2001), c appears exclusively radio-quiet from an X-ray spectroscopy perspective.
The intrinsic photon index we measure for c after accounting for the reflection components is $\Gamma$$\sim$1.8, similar to most radio-quiet Seyfert 1s observed with (e.g., Patrick et al. 2010). Previous observations of c at X-rays (§ 2) established that the photon index varies in unison with the flux, in the sense of a softer spectrum for increasing X-ray intensity (Gliozzi et al. 2007 and references therein), a trend c shares with radio-quiet, non-jetted AGN. Together with the constraints on the high-energy cutoff of the primary power law, this implies that thermal Comptonization dominates the emission below 100 keV in c, supporting the idea that the bulk of the X-ray continuum does not originate in a jet. At present, only an upper limit is reported to the gamma-ray GeV emission at the location of c in the Fermi 15-months catalog (Abdo et al. 2010). This suggests that the contribution of a jet to the higher energies is weak, or occurs with a small duty cycle. Future monitoring will be necessary to shed light on the jet activity at high energies from the core of c.
The high quality and low background of the XIS data allows us to model with high fidelity the Fe-K emission. We confirm the presence of a narrow line at an energy of $\sim$6.4 keV, identified with the narrow component of the line. A similar narrow component, consistent with emission from the Broad Line Region (BLR), was detected in 3C 390.3 (S09) and 3C 120 (Kataoka et al. 2007), two of a handful of well-known, nearby BLRGs observed with and is regularly observed in Seyfert 1s (Patrick et al. 2010; Nandra et al. 2007). This line can be modeled with a mildly ionized reflection component with log$\xi$$\simeq$ 1.5 erg s$^{-1}$ cm (Table 5), which can also simultaneously explain the observed soft excess and part of the reflection hump. This contributes $\sim$10% on the power-law continuum.
From the width of the line of FWHM$\sim$10,000 km/s and assuming that the material is in Keplerian motion in the gravitational potential well of the super-massive black hole, with estimated mass M$_{BH} \sim 10^9$ M$_{\sun}$ (Marchesini et al. 2004), we can derive a typical distance of $\sim$0.3 pc. The similarity of the Fe K$\alpha$ line width with that of the broad H$\alpha$ line in the optical spectrum (Eracleous & Halpern 1994) and the distance on sub-pc scales possibly suggest an origin of this reflecting material from the BLR. Using the distance estimate in the definition of the ionization parameter $\xi=L_{ion}/nr^2$ (Tarter, Tucker & Salpeter 1969), where $n$ is the average absorber number density and $L_{ion}$ is the source X-ray ionizing luminosity integrated between 1 Ryd and 1000 Ryd (1 Ryd=13.6 eV), we can derive the density of the material at that location. The absorption corrected ionizing luminosity of c is $L_{ion}$$=$$8\times 10^{44}$ erg s$^{-1}$. Substituting this value, we can estimate the density of the material of $\sim$$5\times 10^7$ cm$^{-3}$.
Moreover, we also detected a weak emission line at E$\simeq$7.5 keV, most probably ascribable to Ni K$\alpha$ fluorescence emission line from the same material of the Fe K$\alpha$. In particular, the is expected to be more intense than the Ni K$\alpha$ due to the higher cosmic abundance of iron with respect to nickel, about $\sim$20 times (Molendi, Bianchi & Matt 2003; Yaqoob & Murphy 2011). However, the ratios of the intensities of the two emission lines suggest a possible slight overabundance of nickel and/or an underabundance of iron of a factor of $\sim$2 with respect to the standard solar values (e.g., Anders & Grevesse 1989; Grevesse et al. 1996). Unfortunately, we note that the treatment of Ni lines is not included in the present version of the *xillver* code (March 2010). However, the possible slight underabundance of iron or overabundance of Ni has negligible effects, within the 1$\sigma$ errors, on the parameters of the mildly ionized reflection component.
Our most interesting result from the Fe K region modeling is the detection of a relativistic profile for the broad component of the line, confirming a previous tentative result (Reynolds 1997; S99). The data quality is sufficient to provide tight constraints on the disk parameters, albeit in a model-dependent way, in particular on the inner and outer radii of the emitting region and the disk inclination. We find that the bulk of the reprocessing in c occurs in a small disk annulus, $\Delta r \sim 10
r_g$, located at a distance $r_{in} \sim 10 r_g$ from the central black hole and at an inclination $i$$\sim$25–30. Substituting the average distance of $r$$\sim$15$r_g$ in the ionization parameter equation and the estimated value of log$\xi$$\simeq$3 erg s$^{-1}$ cm (Table 5), we can derive a density for the associated highly ionized reflection component of $\sim$$10^{11}$ cm$^{-3}$. This is a realistic estimate for the inner parts of an accretion disk in AGN (Garc[í]{}a, Kallman & Mushotzky 2011).
A similar finding - that the innermost disk regions do not contribute to the Fe K emission – was previously reported for other broad-lined radio-loud AGN, the BLRGs 3C 120 and 3C 390.3 (Kataoka et al. 2007, S09), and the quasar/BLRG 4C+74.26 (Larsson et al. 2008). In these sources the inner radius of the Fe K emission was determined to be around 10–50$r_g$ from fits to the data. Thus, a pattern is starting to emerge, whereas the innermost regions of the accretion disk in broad-lined radio-loud AGN do not contribute to the Fe K emission.
A number of plausible physical scenarios have been put forth to explain the above lack of Fe K reprocessing in the innermost radii of the BLRG disks (see S09 and discussion therein). These include a highly ionized ion torus/ADAF occupying the inner disk, obscuration by the base of a jet, or lack of illumination of the inner radii of the disk by the primary power law, due to a mildly beamed source of continuum. An additional possibility is offered in the context of the “flux-trap” scenario (Reynolds et al. 2006; Garofalo 2009).
It is generally believed that jet power is, at least in part, linked to the spin of the central black hole. Radio-loud AGNs have been claimed to harbor rapidly spinning black holes (e.g., Nemmen et al. 2007, and references therein). A scenario for jet power production is provided by the “flux trap” model (Reynolds et al. 2006; Garofalo 2009), which suggests that the plunging region between accretion disks and black holes is fundamental in producing strong, spin dependent, horizon-threading magnetic fields, and therefore powerful jets. Noteworthy, it has been demonstrated that the Blandford & Znajek (1977; BZ) mechanisms is maximized for black hole spin $a \sim
-1$ (retrograde) and stronger jets/outflows are expected in this case (Garofalo 2009). This is a direct consequence of the fact that the radius of marginal stability for retrograde black holes, $r_{ISCO}\sim
9$r$_g$ for $a=-0.998$, is much larger than in prograde, $r_{ISCO}\sim
1.23$r$_g$ for $a=+0.998$, and therefore the plunging region is more extendend and the flux-trapping effect is magnified. In principle, this could be directly tested with X-ray spectroscopy because in the two cases the relativistic Fe K line profile changes significantly (Garofalo, Evans, & Sambruna 2010).
As discussed in § 5.3 and § 5.4, we tested the retrograde/prograde scenario by performing spectral fits to the data with a model that allows the black hole spin to assume [*negative*]{} values (Dauser et al. 2010). While the fit is statistically satisfactory, unfortunately the spin parameter is unconstrained due to the insufficient S/N ratio of the data. However, because of the shape of its Fe K line, c qualifies as the best candidate so far for determining the black hole spin in a radio-loud AGN through Fe K line spectroscopy. Longer exposures with are required.
c shares with two other BLRGs (3C 390.3 and 3C 120) the evidence in its X-ray spectrum for emission from an ionized disk, with similar ionization parameters $\xi \sim 2300-2700$. This supports the previous suggestion by Ballantyne et al. (2002) that ionized disks play a significant role for the X-ray emission of powerful radio-loud AGN. Evidence for ionized material in the nuclear regions of radio-loud AGN is also provided by the detection of extremely ionized outflows/winds with velocities $\ga$10,000 km/s, and properties similar to radio-quiet AGN outflows (Tombesi et al. 2010a,b; see also below).
Finally, we compare the X-ray properties of c and other BLRGs observed with to Seyfert 1s. We started this comparison in our previous paper on 3C 390.3 (S09), where we outlined the average BLRG properties – X-ray photon index, Fe K line widths, and cold reflection strengths – with the then available most current X-ray data for Seyferts from EPIC observations (Nandra et al. 2007). We noted how the data for radio-quiet AGN suggested a much larger spread of reflection fractions and line widths for Seyferts than in the and era, and how this spread lead to a blur in the distinction between radio-loud and radio-quiet sources. In particular, in BLRGs X-ray reflection and continuum properties occupy one end of the distribution for Seyferts, with significant overlap. We concluded that the division between radio-loud and radio-quiet AGN was blurred, as far as their X-ray spectral properties were concerned (S09).
A more proper comparison is now possible thanks to the recent publication of and BAT observations for a relatively large (6 sources) sample of Seyfert 1s (Patrick et al. 2010). From an accurate modeling of the 0.6-100 keV spectra using both cold and ionized reflection models, Patrick et al. (2010) found broad relativistic profiles for the Fe K emission lines in most Seyferts, with average equivalent width for the sample of EW=119$\pm$19 eV and ionized emission lines at 6.7 and 6.97 keV being relatively common. Intriguingly, the Fe K emission seems to arise from tens of r$_g$ from the central black hole (see also Nandra et al. 2007); fits with models allowing the black hole spin to vary, although in the 0–1 range only, yield typically intermediate values, $a\sim 0.7$. Both cold and ionized reflection is required to describe the data above 10 keV; the cold reflection fraction, however, varies from source to source from less than 0.4 up to 2.5 (see Table 4 in Patrick et al. 2010). The intrinsic X-ray photon index is $\Gamma_X \sim
1.8-2.2$, consistent with Comptonization models. However, the ionization parameter of the ionized disk reflection seems to be lower in the Seyferts, consistent with neutral or lowly ionized Fe, while interestingly so far the BLRGs seem to suggest high ionization disk lines (He and H-like iron), although based on small number statistics so far.
Thus, comparison of the [*broad-band 0.6–200 keV*]{} spectra of BLRGs and Seyfert 1s obtained with and the BAT confirm and reinforce the conclusion (S09) that there is a continuum of X-ray properties between the two AGN subclasses, with radio-loud sources clustered at one end of the distribution. An important consideration, however, is that in both subclasses – aside for a few egregious examples – the innermost $\sim 10$r$_g$ of the accretion disk do not contribute significantly to Fe K emission. The reason for this may be different in the two subclasses, and would remain of central importance for the radio-loud/radio-quiet AGN division.
Another major progress since S09 is the discovery that radio-loud AGNs have ionized, outflowing material from the central nuclei, similar to Seyferts (e.g., Tombesi et al. 2010a,b), contrary to what stated in S09 based on the then available evidence. Indeed, our observations of 5 BLRGs provided evidence for ultra-fast disk outflows in 3/5 sources (Tombesi et al. 2010b) with velocities $v_{out} \sim 0.1$c, kinetic power $\sim 10^{43-44}$ , and mass outflow rates comparable to the accretion rates. As mentioned above and in § 2, c itself presents evidence for a highly ionized, kpc-scale outflow through a series of soft X-ray absorption features detected with the and gratings (Reeves et al. 2009; Torresi et al. 2010), although with a much lower velocity of $\sim$1,000 km/s. The presence of ionized material along the line of sight is puzzling and its relation with the relativistic jet must be studied in much detail. As already discussed in Tombesi et al. (2010b), we did not detect any significant highly ionized Fe K absorption line in the present observation of c.
The overlapping X-ray continuum and reflection properties of BLRGs and Seyferts point to a common accretion flow structure, with similar radiative processes in place that account broadly for the X-ray emission. The reason why powerful, relativistic jets are present in BLRGs but [*not*]{} in Seyferts with similar engines could very well be identified in a process unrelated to the accretion dynamics – i.e., the black hole spin. Several theoretical studies along this line are underway, and the wealth of high-quality data that continues to provide is essential to test the model predictions in the X-ray band, paving the way for the upcoming high-precision spectroscopy with the [*Astro-H*]{} observatory.
This research has made use of data obtained from the High Energy Astrophysics Science Archive Research Center (HEASARC), provided by NASA’s Goddard Space Flight Center, and 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. RMS and FT acknowledge support from NASA through the and ADAP/LTSA programs. FT thank T. Dauser for help in using the *relline* model. FT thank J. Garc[í]{}a for help in using the *xillver* model. LB acknowledges financial support from the Spanish Ministry of Science and Innovation through a “Juan de la Cierva” fellowship, research grant AYA2009-08059. The authors thank the referee for suggestions that led to important improvements in the paper.
Background and Calibration tests for the XIS
============================================
We performed detailed background and calibration tests for the three XIS detectors to exclude possible issues affecting the Fe K region, and thus the results of our modeling described above. Background contamination could alter the profile of the Fe line; however, for the c observation analyzed here, we estimate that the background contribution to 0.6–10 keV is negligible, amounting to $\la$4–8% of the source counts.
It is well known that the XIS CCDs contain a $^{55}$Fe calibration source, located on two corners of the XIS chips, which produces a characteristic X-ray line from Mn K$\alpha$ at E$=$5.895 keV. Its intrinsic width is $\sigma$$\la$50 eV and it is used to calibrate the energy and width of the iron line (Koyama et al. 2007). We can clearly exclude any possible significant contamination of the broad line detected at the rest-frame energy of E$\simeq$5.8 keV from this calibration source on board the XIS. In fact, we checked the images of the three separate XIS cameras and find that the source is located in the center of the CCD, far from the calibration sources, which are in two corners. Moreover, the observed energy of the broad line is at E$=$5.45 keV, which is not consistent with that expected from the calibration source, and the calibration line is intrinsically narrower compared with the $\simeq$400 eV of the c feature.
As already discussed in § 3.1 the combined XIS-FI and XIS-BI fit is supported by the consistency of the power-law slope and 2–10 keV flux within $\sim$2%. Moreover, as already reported in § 4, we checked that the presence of two low amplitude ($\sim$5%) spikes in the XIS 3 light curve, see Figure 1, have negligible effects on the overall spectral analysis. This is expected from the fact that we focused on the average spectrum and the weak short time-scale spikes involved less than 20% of the exposure. These are most probably ascribable to weak instrumental fluctuations. We directly analyzed the three XIS cameras separately and ensured that there is an agreement at $\sim$1% for both the power-law continuum slope and 2–10 keV flux between the XIS 0 and XIS 3 and at $\sim$2% between the front illuminated XIS 0-3 and back illuminated XIS 1.
Finally, the spectra of the individual XIS camera were fitted separately for a consistency check of the emission line parameters. Unfortunately, the quality of the individual camera spectra does not allow us to leave all line parameters free to vary during the fit. Using the best-fit model determined from the analysis of the joint spectra (see Table 3), and freezing the lines energies and widths, we find that the EWs of the emission lines are always consistent at 90% among the separate XIS instruments and the combined XIS-FI and XIS-BI fit.
Abdo, A. A., et al. 2010, , 720, 912 Anders, E., & Grevesse, N. 1989, , 53, 197 Antonucci, R. 1993, , 31, 473 Ballantyne, D.R., Ross, R.R., & Fabian, A. C. 2002, MNRAS, 332, L45 Ballantyne, D.R. 2007, Mod. Phys. Lett. A, 22, 2397 Barr, P., & Giommi, P. 1992, , 255, 495 Baumgartner et al. 2010, ApJS, submitted Blandford, R. D., & Znajek, R. L. 1977, , 179, 433 Blandford, R.D. 1985, in Active Galactic Nuclei, ed. J.E.Dyson, Manchester Univ. Press: Manchester, p. 281 Boldt, E. 1987, PhR, 146, 215 Braito, V., Reeves, J. N., Sambruna, R. M., & Gofford, J. 2011, arXiv:1102.5234 Crummy, J., Fabian, A. C., Gallo, L., & Ross, R. R. 2006, , 365, 1067 Dauser, T., Wilms, J., Reynolds, C. S., & Brenneman, L. W. 2010, , 409, 1534 de La Calle P[é]{}rez, I., et al. 2010, , 524, A50 Dickey, J. M., & Lockman, F. J. 1990, , 28, 215 Dower, R. G., Bradt, H. V., Doxsey, R. E., Johnston, M. D., & Griffiths, R. E. 1980, , 235, 355 Eracleous, M. & Halpern, J.P. 1994, ApJS, 90, 1 Eracleous, M. & Halpern, J.P. 1998, ApJ, 505, 577 Eracleous, M., Sambruna, R.M., & Mushotzky, R.F. 2000, ApJ, 537, 654 Fabian, A. C., Rees, M. J., Stella, L., & White, N. E. 1989, , 238, 729 Garc[í]{}a, J., & Kallman, T. R. 2010, , 718, 695 Garc[í]{}a, J., Kallman, T. R., & Mushotzky, R. F. 2011, arXiv:1101.1115 Garofalo, D. 2009, , 699, 400 Garofalo, D., Evans, D. A., & Sambruna, R. M. 2010, , 406, 975 Gierli[ń]{}ski, M., & Done, C. 2004, , 349, L7 Gliozzi, M., Sambruna, R. M., Eracleous, M., & Yaqoob, T. 2007, , 664, 88 Grandi, P., Maraschi, L., Urry, C. M., & Matt, G. 2001, , 556, 35 Grevesse, N., Noels, A., & Sauval, A. J. 1996, Cosmic Abundances, 99, 117 Hardcastle, M. J., Alexander, P., Pooley, G. G., & Riley, J. M. 1998, , 296, 445 Hartman, R. C., Kadler, M., & Tueller, J. 2008, , 688, 852 Kalberla, P. M. W., Burton, W. B., Hartmann, D., Arnal, E. M., Bajaja, E., Morras, R., Pöppel, W. G. L. 2005, , 440, 775 Kallman, T. R., Palmeri, P., Bautista, M. A., Mendoza, C., & Krolik, J. H. 2004, , 155, 675 Kataoka, J. et al. 2007, PASJ, 59, 279 Koyama, K., et al. 2007, , 59, 23 Kokubun, M. et al. 2007, PASJ, 59, 53 Larsson, J., Fabian, A. C., Ballantyne, D. R., & Miniutti, G. 2008, , 388, 1037 Lightman, A. P., & White, T. R. 1988, , 335, 57 Longair, M. S., & Seldner, M. 1979, , 189, 433 Marchesini, D., Celotti, A., & Ferrarese, L. 2004, , 351, 733 Martel, A. R., et al. 1999, , 122, 81 Matthews, T. A., Morgan, W. W., & Schmidt, M. 1964, , 140, 35 Molendi, S., Bianchi, S., & Matt, G. 2003, , 343, L1 Nandra, K., O’Neill, P. M., George, I. M., & Reeves, J. N. 2007, , 382, 194 Nemmen, R. S., Bower, R. G., Babul, A., & Storchi-Bergmann, T. 2007, , 377, 1652 Patrick, A. R., Reeves, J. N., Porquet, D., Markowitz, A. G., Lobban, A. P., & Terashima, Y. 2010, arXiv:1010.2080 Prieto, M. A. 2000, , 316, 442 Puschell, J. J. 1981, , 86, 16 Reynolds, C. 1997, MNRAS, 286, 513 Reynolds, C. S., Garofalo, D., & Begelman, M. C. 2006, , 651, 1023 Reeves, J. N., Nandra, K., George, I. M., Pounds, K. A., Turner, T. J., & Yaqoob, T. 2004, , 602, 648 Reeves, J. N., Sambruna, R. M., Braito, V., & Eracleous, M. 2009, , 702, L187 Ross, R. R., Fabian, A. C., & Young, A. J. 1999, , 306, 461 Ross, R. R., & Fabian, A. C. 2005, , 358, 211 Sambruna, R.M., Eracleous, M., & Mushotzky, R. 1999, ApJ, 526, 60 (S99) Sambruna, R.M., Eracleous, M., & Mushotzky, R.F. 2002, NewAR, 46, 215 Sambruna, R. M., et al. 2009, , 700, 1473 (S09) Saunders, R., Baldwin, J. E., Rawlings, S., Warner, P. J., & Miller, L. 1989, , 238, 777 Spergel, D. N., et al. 2003, ApJS, 148, 175 Strom, R. G., Willis, A. G., & Wilson, A. S. 1978, , 68, 367 Tadhunter, C. N., Perez, E., & Fosbury, R. A. E. 1986, , 219, 555 Tadhunter, C. N., Fosbury, R. A. E., & Quinn, P. J. 1989, , 240, 225 Tarter, C. B., Tucker, W. H., & Salpeter, E. E. 1969, , 156, 943 Tombesi, F., Cappi, M., Reeves, J. N., Palumbo, G. G. C., Yaqoob, T., Braito, V., & Dadina, M. 2010, , 521, A57 (Tombesi et al. 2010a) Tombesi, F., Sambruna, R. M., Reeves, J. N., Braito, V., Ballo, L., Gofford, J., Cappi, M., & Mushotzky, R. F. 2010, , 719, 700 (Tombesi et al. 2010b) Torresi, E., Grandi, P., Longinotti, A. L., Guainazzi, M., Palumbo, G. G. C., Tombesi, F., & Nucita, A. 2010, , 401, L10 Urry, C.M. & Padovani, P. 1995, PASP, 107, 803 Yaqoob, T., & Murphy, K. D. 2011, , 18
{width="8cm" height="9.5cm"} {width="8cm" height="9.5cm"}
{width="8cm" height="12cm"}
{width="8cm" height="12cm"}
{width="8cm" height="9cm"} {width="8cm" height="9cm"}
{width="13cm" height="10cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A class of multiwavelength Fabry-Pérot lasers is introduced where the spectrum is tailored through a non-periodic patterning of the cavity effective index. The cavity geometry is obtained using an inverse scattering approach and can be designed such that the spacing of discrete Fabry-Pérot lasing modes is limited only by the bandwidth of the inverted gain medium. A specific two-color semiconductor laser with a mode spacing in the THz regime is designed, and measurements are presented demonstrating the simultaneous oscillation of the two wavelengths. The extension of the Fabry-Pérot laser concept described presents significant new possibilities in laser cavity design.'
author:
- 'S. O’Brien$^{1}$, S. Osborne$^{1}$, K. Buckley$^{1}$, R. Fehse$^{1}$, A. Amann$^{1}$, E. P. O’Reilly$^{1}$, L. P. Barry$^{2}$, P. Anandarajah$^{2}$, J. Patchell$^{3}$ and J. O’Gorman$^{3}$'
title: 'Two-Color Fabry-Pérot Laser Diode with THz Primary Mode Spacing'
---
The most familiar laser cavity geometry is the Fabry-Pérot (FP) laser, which comprises an active gain medium and two external mirrors providing feedback for oscillation. In this geometry the longitudinal lasing mode wavelengths are determined by the half wave resonance condition: $\lambda^{m} = 2 n L_{c}/m$, where $m = 1,2,...$, $\lambda^{m}$ is the free space wavelength of the $m\mbox{th}$ mode, $L_{c}$ is the cavity length, and $n$ is the cavity refractive index [@agrawal].
A fundamental limitation of the basic FP geometry is the lack of any frequency selectivity other than that provided by the gain medium. Because the gain bandwidth is much larger than the FP mode spacing in typical semiconductor lasers, more complex laser cavity geometries have been conceived in order to control and manipulate semiconductor laser spectra. For example, one dimensional systems such as the distributed feedback laser provide high spectral purity and temperature stability in device applications [@kogelnik]. Translational symmetry determines the lasing modes of this structure without the need for a reflection from external mirrors.
If we consider the interaction of the cavity modes with the gain medium, the semiconductor FP laser geometry is deceptively simple. In a perfectly homogeneously broadened medium, a single lasing mode should always dominate [@haken; @tredicucci]. When driven above threshold, semiconductor FP lasers often oscillate in many modes. This multimode behaviour is characteristic of inhomogeneously broadened gain media, despite the fact that in a semiconductor the carriers are distributed in continuous bands.
A key property of the FP laser in this respect is the fact that all the cavity resonant wavevectors are equally spaced. As a result, four-wave mixing (FWM) interactions, which transfer power among modes, are cavity enhanced. In addition, carrier density pulsations at the intermode frequencies and the finite alpha factor lead to an asymmetric contribution to the nonlinear gain in semiconductor lasers [@yamada; @uskov; @ogita]. Along with the large spontaneous emission rate in semiconductor lasers, this interaction also promotes multimode oscillation and can lead to mode hopping and complex antiphased switching dynamics [@ahmed; @yacomotti].
In this letter we revisit and extend the basic FP laser geometry. We demonstrate that multiwavelength FP lasers can be designed where, apart from the constraint imposed by the half wave resonance requirement, the distribution of lasing modes is chosen *independent* of the cavity length. The basic FP cavity configuration and mode structure are maintained, with the manipulation of the lasing mode spectrum achieved using a non-periodic effective index profile. The precise geometry is determined from the desired lasing spectrum through an inverse scattering approach [@obrien; @obriens_josaB].
Because of its fundamental significance, we present experimental measurements of a two-color semiconductor FP laser with a primary mode spacing in the THz regime. Our measurements demonstrate that the device oscillates simultaneously on two discrete FP modes, and without the requirement for an external cavity arrangement or other external perturbation. In contrast to this ideal behaviour associated with weakly coupled modes, an otherwise identical *plain* FP laser, with a modal spacing determined by the cavity length, displays the mode hopping behaviour and complex dynamics associated with strong mode competition.
Consider the one-dimensional model of the FP cavity geometry represented in Fig. \[capl1\]. The system comprises a FP cavity of length $L_{c}$ with a spatially varying refractive index. The mirror reflectivities are $r_{1}$ and $r_{2}$ (assumed real for simplicity) and there are $N$ additional index steps along the cavity. For each section of the laser cavity (index $i$) we define $\theta_{i} = n_{i}k_{0z}L_{i}$, where $k_{0z}$ is the free space wavenumber along $z$ and $L_{i}$ and $n_{i}$ are the length and the effective refractive index of the $i\mbox{th}$ section respectively. The adjusted complex optical path length across the cavity is then $\sum_{i=1}^{2N+1}\theta_{i}$.
![\[capl1\] One dimensional model of a Fabry-Pérot laser cavity of length $L_{c}$ and including $N$ index steps. The cavity effective index is $n_{1}$ while the additional features providing the index step (shaded regions) have effective index $n_{2}$ as shown. All cavity sections are numbered $1 \leq i \leq 2N+1$ beginning on the left. The matrix $T$ relates the left and right moving fields inside the cavity at the cavity mirrors. The mirror reflectivities are $r_{1}$ and $r_{2}$ as shown.](fig1){height="4.0cm" width="7.5cm"}
We set the background cavity effective index, $n_{1} = n$, and the effective index at the index step features, $n_{2} = n + \Delta n$. Suppose the transfer matrix $T$ relates the right and left moving electric fields, $E^{\pm}(z)$, at the cavity mirrors in Fig. \[capl1\]. Then the lasing modes of the cavity are defined by the relation [@obriens_josaB] $$\label{threshold}
\left(
\begin{array}{c}
1 \\
r_{1}
\end{array}
\right) = T \left(
\begin{array}{c}
r_{2} \\
1
\end{array}
\right).$$ From Eqn. \[threshold\] one can show that the lasing condition at first order in the index step can be written as $$\begin{aligned}
\label{gt_cond}
&1 - r_{1}r_{2}\exp(2i\Sigma \theta_{i}) = \nonumber \\
&i\frac{\Delta n}{n} \sum_{j} \sin \theta_{2j}
\left[r_{1}\exp(2i\phi_{j}^{-}) + r_{2}\exp(2i\phi_{j}^{+})\right], \end{aligned}$$ where the quantities $\phi^{-}_{j} \mbox{ and } \phi^{+}_{j}$ are the optical path lengths from the centre of each additional feature to the left and right facets respectively.
If we neglect a factor describing the background losses associated with each mode, for a vanishing index step, the threshold gain for lasing is determined by the mirror losses. We have $\gamma_{m}^{(0)} = L_{c}^{-1}
\ln 1/r_{1}r_{2}$. In the perturbed case, a set of self-consistent equations for the lasing modes is found by making an expansion in Eqn. \[gt\_cond\] about the cavity resonance condition: $\sum \theta_{i}^{'} = {\phi^{-}_{j}}^{'} + {\phi^{+}_{j}}^{'} = m\pi +
\delta_{m}$ [@obrien], where $\delta_{m} (\ll 1)$ determines the lasing mode frequency shift.
The inverse problem at first order is solved by choosing a particular cavity resonance, $m_{0}$, as an origin in wavenumber space. We assume quarter wave features with $\sin \theta_{2j}^{'} = 1$ in order that the intensity scattered by each feature at the wavelength of mode $m_{0}$ is maximized. Each feature is placed ultimately such that a half wavelength subcavity at the wavelength of mode $m_{0}$ is formed between the feature and one of the external mirrors. The threshold gain can then be expressed at each resonance, $m$, where $m = m_{0} + \Delta m$, as $\gamma_{m} = \gamma_{m}^{(0)} + (\Delta n/n) \gamma_{m}^{(1)}$, where $$\begin{aligned}
\label{gamma_m1}
&\gamma_{m}^{(1)} = \frac{1}{L_{c}\sqrt{r_{1}r_{2}}}\cos(m_{0}\pi) \cos(\Delta m\pi) \nonumber \\
& \times \sum_{j=1}^{N}A(\epsilon_{j}) \sin(2\pi\epsilon_{j}m_{0}) \cos(2\pi\epsilon_{j}\Delta m).\end{aligned}$$ In the above expression, the factor $A(\epsilon_{j}) = r_{1} \exp(\epsilon_{j}L_{c}\gamma_{m}^{(0)})
- r_{2} \exp(-\epsilon_{j}L_{c}\gamma_{m}^{(0)})$ and $\epsilon_{j}$ is the position of the center of each feature measured from the center of the cavity as a fraction of the cavity length.
Using Eqn. \[gamma\_m1\] Fourier analysis can be used in order to build up a particular threshold gain modulation in wavenumber space. Because it represents the simplest system that illustrates the application of classical multimode laser theory [@sargent], the multiwavelength device we describe here is the two-color laser. The appropriate basis for this device is a pair of sinc functions, $\gamma_{m} \sim
\mbox{sinc}(\Delta m + a/2) + \mbox{sinc}(\Delta m - a/2)$. This choice selects two modes, centered at $m_{0}$ and with spacing $a$ modes, while leaving the other FP modes unperturbed. In the inset of Fig. \[ob\_ft\] (a) we have plotted an idealized threshold gain spectrum where the primary mode spacing is $a = 4$ fundamental cavity modes. The Fourier transform of our idealized threshold gain modulation function is $\cos(\pi a \epsilon)$, for $-1/2 \leq \epsilon \leq 1/2$ and is zero otherwise.
The factor $A(\epsilon_{j})$ in Eqn. \[gamma\_m1\] reflects the fact that the change in threshold of a given mode is determined by the difference in the round-trip amplitude gain to the left and to the right of each feature. To determine the appropriate distribution of features, we must therefore correct for the variation of the amplitude of the threshold modulation with position. We take the product of the Fourier transform of our ideal threshold gain modulation with the envelope function, $[A(\epsilon_{j})]^{-1}$. The absolute value of this product determines the feature density function shown in Fig. \[ob\_ft\] (a), which we then sample in order to approximately reproduce the idealized threshold gain spectrum [@obrien].
![\[ob\_ft\] (a) Feature density function (solid line). The dashed lines are the negative of the feature density function in those intervals where the Fourier transform of the function shown in the inset is negative. Inset: Ideal threshold gain of modes in wavenumber space. Lower panel: Laser cavity schematic indicating the locations of the additional features. The device is high-reflection (HR) coated as indicated. (b) Calculation of the threshold gain of modes for the laser cavity schematically pictured in the lower panel of the figure. The horizontal line is at the value of the mirror losses of the plain cavity.](fig2){height="4.5cm" width="7.5cm"}
For the two-color device considered the cavity is asymmetric with one larger facet reflectivity $(r_{1})$. This allows a more uniform density of features along one side of the device center. Once the feature density function is sampled correctly, feature positions are adjusted in order to satisfy the correct phase requirement for resonance. A schematic picture of the device, high-reflection coated as indicated, is shown in the lower panel of Fig. \[ob\_ft\]. With respect to the lasing wavelength of mode $m_{0}$ in the cavity, where $\cos(\pi a \epsilon_{j}) > 0$, the phase requirement corresponds to forming a halfwave resonant subcavity between the corresponding feature and the high-reflection coated mirror. For $\cos(\pi a \epsilon_{j})
< 0$, we form a quarterwave subcavity at the same wavelength. In this way, at each zero of the feature density function a $\pi/2$ phase shift is introduced into the index pattern along the device length. Optical path corrections due to the introduction of the features must also be accounted for when the final feature positions are calculated. The calculated form of the threshold gain spectrum is shown in Fig. \[ob\_ft\] (b) and is an excellent approximation of the ideal form. To appreciate the simplicity of this approach, it is important to note that the device comprises a single, patterned amplifying section with both external mirrors necessary to form the FP mode structure. Thus, unlike distributed feedback approaches, the role of the inhomogeneous cavity effective index is simply to discriminate between the various FP modes.
We now present experimental measurements of a ridge waveguide FP laser fabricated to the design depicted in Fig. \[ob\_ft\]. The device is a multi-quantum well InP/InGaAlAs laser of length 350 $\mu$m with a peak emission near 1.3$\mu$m. The additional features are slotted regions etched into the laser ridge waveguide. This technique is based on standard optical lithography and does not require a regrowth step. The laser was temperature stabilised at 25$^{0}$C to $\pm 0.01^{0}$C and a constant current was applied to the device. Laser emission was spectrally resolved using an optical spectrum analyser with 0.01 nm resolution.
A series of spectra of the device of Fig. \[ob\_ft\] are shown in Fig. \[spectra\]. Fig. \[spectra\] (a) shows the device spectrum below threshold. One can see that the two primary modes are already selected in this regime. Note the good agreement with the calculation shown in Fig. \[ob\_ft\], with the two primary modes separated by four fundamental FP modes. As the current is increased the mode on the short wavelength side reaches threshold first (Fig. \[spectra\] (b)) and as the current is increased further thermal effects lead to the peak power shifting across the primary mode spacing to the long wavelength side (Fig. \[spectra\] (c)). The lasing spectrum at 43.5 mA is shown in Fig. \[spectra\] (d). At this current the time averaged optical power in the primary modes is approximately equal. For comparison, an above threshold spectrum from a plain FP laser fabricated on the same bar is shown in Fig. \[spectra\] (e).
![\[spectra\] (a) Below threshold spectrum of the two-color device of Fig. \[ob\_ft\]. (b) Lasing spectrum at 30 mA. (c) Lasing spectrum at 60 mA. (d) Two-color lasing spectrum at 43.5 mA (e) Above threshold spectrum of a plain Fabry-Pérot device fabricated on the same bar.](fig3){height="5.0cm" width="8.0cm"}
In Fig. \[auto\_corr\] (a) we have plotted a higher resolution picture of part of the spectrum of Fig. \[spectra\] (d). One can see sideband formation due to four wave mixing (FWM) processes in the cavity. FWM is a third order nonlinear process which occurs due to the formation of a dynamic grating in the material complex index [@shen; @park]. The grating is formed through the beating of the primary modes at $\nu_{1}$ and $\nu_{2}$ as shown in Fig. \[spectra\] (d). The sideband shown appears at a frequency of $2\nu_{1} - \nu_{2}$ with the scattering of the primary mode at $\nu_1$ by the grating. Within the bandwidth of the inverted semiconductor, the patterned FP cavity naturally provides gain and resonant feedback for the FWM sideband, which is slightly detuned from the amplified spontaneous emission peak due to the material dispersion and is resonantly enhanced as the two primary modes attain equal time averaged intensity. The presence of a large, narrow linewidth signal, implies that the two modes are oscillating simultaneously with good phase stability.
Phase coherence of the primary modes implies an ultrafast intensity modulation of the laser output. We measured this mode beating in the cw output of the laser at the difference frequency, $\nu_{2}
- \nu_{1} \sim 480$ GHz. The result of the background free intensity autocorrelation measurement is shown in the inset of Fig. \[auto\_corr\] (a) where the contrast ratio observed is close to the theoretical limit of 3:1. In the mode locking regime, the extension of concepts described here to create a comb of discrete wavelengths can lead to compact sources of pulsed radiation with THz repetition rates. In fact, the polarization associated with the mode beating itself is a source of direct THz radiation, generated by the two primary waves in an intracavity process [@hoffmann]. In an intracavity process a single material system simultaneously acts as the pump medium and nonlinear mixing material. Such systems have received considerable interest on account of their potential to enhance nonlinear interactions. Approaches to intracavity wavemixing have tended to focus on engineering of the semiconductor heterostructure to support distinct optical transitions and thus multiple lasing wavelengths [@capasso; @belyanin]. By using the extension of the FP cavity concept described here to suppress unwanted longitudinal cavity mode structure, the efficiency of such approaches can be further improved.
![\[auto\_corr\] (a) Enlarged spectrum showing the presence of a four-wave mixing sideband. Inset: Background free intensity autocorrelation measurement showing mode beating at 480 GHz. (b) Time traces of a primary mode (lower trace) and total device output (upper trace) showing essentially constant total output. Not shown is the second primary mode output which is anticorrelated with the first. (c) Time trace of the output of a single FP laser mode from the device of Fig. \[spectra\] (e). At this current the plain FP laser is exhibiting mode-hopping behaviour.](fig4){height="4.5cm" width="7.5cm"}
We now apply the classical theory of a homogeneously broadened two mode laser to our device near the current of Fig. \[spectra\] (d). In classical laser theory, simultaneous lasing of two modes is possible if the net gain of each mode is positive and the competition due to cross-saturation is sufficiently weakened by a large mode spacing and by the spatial hole burning effect associated with the standing waves of the FP cavity [@sargent].
Measurement of a large intensity modulation at the difference frequency of the two modes indicates that the device is not exhibiting the mode hopping behaviour characteristic of multiwavelength semiconductor lasers with strongly coupled lasing modes [@hioe]. We confirmed this by measurements of time traces of the modal and total intensity output of the two-color device as shown in Fig. \[auto\_corr\] (b). We observe an essentially constant total output and anticorrelated, enhanced intensity noise traces in each of the two primary modes due to mode partition (a single modal intensity time trace is shown for clarity) [@agrawalpra]. The plain FP laser of Fig. \[spectra\] (e) also shows a constant total intensity, but analysis of individual modal intensities reveals complex dynamics including mode hopping behaviour, an example of which is shown in Fig. \[auto\_corr\] (c). This figure shows a sequence of spontaneous switching events where the longitudinal mode in question switches between an “on” state with large intensity to an “off” state with an intensity close to zero.
In the two-color device here, the coupling between the primary modes is determined by various processes that have different strengths depending on the separation between the modes. These include static spectral hole burning, intraband population pulsations and carrier density pulsations. Because the characteristic time associated with interband processes is large, the contribution of asymmetric nonlinear gain due to the interband carrier density pulsations will be much smaller in the two-color device, where the separation between modes is large. If we neglect the asymmetric contribution, weak coupling of modes in the two-color device requires $(4/3)\cdot[1 + (\omega_1 - \omega_2)^{2}
\tau_{\mbox{\tiny in}}^{2}]^{-1} < 1$, where $\omega_{1,2} = 2\pi\nu_{1,2}$ and $\tau_{\mbox{\tiny in}}$ is the intraband relaxation time [@ogita]. An estimated range of values for $\tau_{\mbox{\tiny in}}$ is 100 - 200 fs, which determines a minimum spacing for the two modes of 460-920 GHz. Although this estimate is in agreement with the actual modal separation and stability properties observed, a systematic study of two-color and other multiwavelength FP lasers will be of interest in order to understand separately the roles of the primary mode spacing and the total mode number in determining the stability and dynamical properties of this family of devices.
In conclusion, we have introduced a class of multiwavelength Fabry-Pérot lasers where the number and spacing of the lasing modes is limited only by the bandwidth of the active medium. Measurements of simultaneous lasing in a specially designed two-color Fabry-Pérot cavity geometry with THz mode spacing were presented. The inverse scattering approach to multiwavelength laser design described is likely to open up new avenues for the fundamental studies of semiconductor laser stability and dynamics. In addition, the devices can provide interesting and novel solutions to many applied problems in optoelectronics and nonlinear optics.
*Acknowledgments.* This work was supported by Science Foundation Ireland. The authors thank Guillaume Huyet and John Houlihan for helpful discussions.
[99]{}
G. P. Agrawal and N. K. Dutta, *Long-Wavelength Semiconductor Lasers* (Van Nostrand Reinhold, New York, 1986).
H. Kogelnik and C. V. Shank, Appl. Phys. Lett. **18**, 152 (1971).
H. Haken, *Laser Theory* (Springer-Verlag, Berlin, 1986) (1984).
A. Tredicuicci, C. Gmachl, F. Capasso, D. L. Sivco, A. L. Hutchinson and A. Y. Cho, Nature (London) **396**, 350 (1998).
M. Yamada, J. Appl. Phys. **86**, 81 (1989).
A. Uskov, J. Mork, J. Mark, M. C. Tatham and G. Sherlock, Appl. Phys. Lett. **65**, 944 (1994).
S. Ogita, A. J. Lowery and R. S. Tucker, J. Quantum Electron. **33**, 198 (1997).
M. Ahmed and M. Yamada, J. Quantum Electron. **38**, 682 (2002).
A. M. Yacomotti, L. Furfaro, X. Hachair, F. Pedaci, M. Giudici, J. Tredicce, J. Javaloyes, S. Balle, E. A. Vikorov and P. Mandel, Phys. Rev. A **69**, 053816 (2004).
S. O’Brien and E. P. O’Reilly, Appl. Phys. Lett. **86**, 201101 (2005).
S. O’Brien, A. Amann, R. Fehse, S. Osborne, E. P. O’Reilly and J. M. Rondinelli, J. Opt. Soc. Am. B *in press* (2006).
M. Sargent, M. Scully and W. Lamb, *Laser Physics* (Addison-Wesley, Reading, Massachusetts, 1974).
Y. R. Shen, *The Principles of Nonlinear Optics* (Wiley, New York, 1984).
I. Park, I. Fischer and W. Elsäßer, Appl. Phys. Lett. **84**, 5189 (2004).
S. Hoffmann, M. Hoffmann, E. Brundermann, M. Havenith, M. Matus, J. V. Moloney, A. S. Moskalenko, M. Kira, S. W. Koch, S. Saito and K. Sakai, Appl. Phys. Lett. **84**, 3585 (2004).
N. Owschimikow, C. Gmachl, A. Belyanin, V. Kocharovsky, D. L. Sivco, R. Colombelli, F. Capasso and A. Y. Cho, Phys. Rev. Lett. **90**, 043902 (2003).
A. A. Belyanin, V. Kocharovsky, V. Kocharovsky and M. Scully, Phys. Rev. A **65**, 053824, (2002).
F. T. Hioe and S. Singh, Phys. Rev. A **24**, 2050 (1981).
G. P. Agrawal, Phys. Rev. A **37**, 2488 (1988).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper focuses on the use of unknown input observers for detection and isolation of actuator and effector faults with control reconfiguration in overactuated systems. The control allocation actively uses input redundancy in order to make relevant faults observable. The case study of an overactuated marine vessel supports theoretical developments.'
author:
- 'Andrea Cristofaro and Tor Arne Johansen[^1]'
title: '**Fault-Tolerant Control Allocation: an Unknown Input Observer based approach**'
---
Introduction
============
The main objective of control allocation is to determine how to generate a specified control effect from a redundant set of actuators and effectors. Control effectors are devices or surfaces producing forces and moments, such as thrusters, propellers, fins or rudders, while actuators are electromechanical devices responsible to tune the magnitude, position and orientation of single effectors. Due to input redundancy, several configurations leading to the same generalized force are admissible and for this reason the control allocation scheme commonly incorporates additional secondary objectives [@Bodson:02] [@FoJo:06] [@JoFo:13], such as power or fuel consumption minimization. On the other hand, usually there are also some limitation factors to be accounted for: actuators/effectors dynamics, input saturation and other physical or operational constraints. One further advantage of actuator and effector redundancy is the possibility to reconfigure the control in order to cope with unexpected changes on the system dynamics, such as failures or malfunctions: in particular if the set of actuators and effectors is partially affected by faults, one can modify the control allocation scheme by preventing the use of inefficient/ineffective devices in the generation of control effect or compensating for the loss of efficiency. However, one key point for successfully re-allocating the control is the availability of adequate information about the faults that have occurred; indeed, some accurate fault estimation and/or a correct identification of the faulty actuators or effectors are necessary to address the reconfiguration. Recent results toward fault tolerant control allocation are based on sliding-mode techniques [@AlEd:06] [@Co-etal:05] and adaptive control strategies [@CaGa:10] [@TjJo:08]. Further investigations on this topic, with a more application-oriented character, are proposed for reconfiguration in flight control [@Bu-etal:01], [@Zhang-etal:07] and fault accommodation in automated underwater vehicles [@Sarkar-etal:02]. An interesting bibliographical survey on the general problem of fault-tolerant control reconfiguration is provided in [@ZhJi:08].\
The aim of this paper is to present the use of unknown input observers for fault detection/isolation and control reconfiguration in overactuated systems. Unknown input observers [@CPZ:96] are a very useful tool for generating robust detection filters, as they can be made insensitive to certain input space directions if some structural algebraic conditions on the system are fulfilled. Due to control redundancy, isolating faults affecting single actuators or effectors in overactuated systems can be a difficult task, as the same effects can be produced by faults occurring in different actuators or effectors: the family of filters needed to isolate the faults usually results to be larger compared to a control system framework with full-rank input matrix, and moreover there is an upper bound on the maximum number of simultaneously isolable faults. On the other hand, by constraining the inputs in prescribed configurations without altering system dynamics, control redundancy can be very helpful in separating the effects produced by multiple faults in order to identify which groups of effectors and actuators are losing effectiveness.\
The paper is structured as follows. In Section 2 the basic setup of control allocation is introduced and the general structure of unknown input observers (UIO) is defined; moreover some issues related to control reconfiguration are reported, such as control re-allocation in the presence of cost functions and input constraints. Section 3 is devoted to the presentation of the proposed method for designing families of detection/isolation filters based on UIO, namely [*constrained output fault direction*]{} (COFD) observers; an alternative approach based on a different class of unknown input observers has recently been proposed by the authors [@noiAut]. Finally, in Section 4, the application of the theoretical results is extensively illustrated by the case study of an overactuated ship subject to thruster failures including common mode faults in thruster auxiliary systems.
Control allocation setup
========================
Let us consider the following linear system $$\left\{
\begin{array}{l}
\dot x(t)=A x(t)+B\tau(t)\\ \\
y(t)=C x(t)
\end{array}
\right.$$ with $$\tau(t)=G u(t),$$ where $x\in\mathbb{R}^n$, $\tau\in\mathbb{R}^k$, $y\in\mathbb{R}^p$, $u\in\mathbb{R}^m$, $m>k$ and all matrices except $A$ are assumed to be full-rank. The vector $x$ is the state, which is assumed to be not accessible for direct measurements, while $y$ is the measured output of the system. The vector $u(t)$ represents the redundant control input and $\tau(t)$ is the generalized control effect or virtual input. Without loss of generality, the desired control effect $\tau_c(t)$ is assumed to be given by a suitable known function depending on the system output: \[tau-y\] \_c(t)=f(y(t)). The above condition can also be generalized, assuming that the desired effect $\tau_c$ and the measured output $y(t)$ are related through a suitable dynamic law. A control allocation strategy is defined such that, whenever it is possible, the control $u$ satisfies \[unominal\] G u(t)=\_c(t). Although the above linear equation always admits (uncountable) exact solutions when $rk(G)=k$, there are possible constraints or bounds to be met and this may lead to the existence of approximate solutions only: \[unominalcon\] {
[l]{} u\
Gu(t)=\_c(t)\_c(t)
. We point out that the approximate effect $\hat\tau_c(t)$ may differ from the desired one $\tau_c(t)$ but it is a known quantity, as it can be computed exploiting the input constraints given by $\mathbb{U}$ which is an assigned set.\
In the unconstrained case, a simple solution can be obtained using the right pseudo-inverse matrix [@Golub:83]: \[uncons\] u(t)=G\^[-R]{}\_c(t),G\^[-R]{}:=G\^T(GG\^T)\^[-1]{}. In the case of constrained control $u\in\mathbb{U}$, several methods for control allocation are available in the literature ([@Durham:93], [@Shi:10], [@BuEn:96], [@Ha:04], [@Za:09], [@Bodson:02], [@Frost:10]). In this paper we consider the class of faults acting on effectors and actuators efficiency by changing their effectiveness: these can be modeled by a multiplicative term $\Delta(t)$: $$\tau(t)=G\Delta(t)u(t),\ \Delta(t)=diag[\delta_1(t),...,\delta_m(t)],$$ for some unknown functions $\delta_i(t)$. It follows that, whenever $\delta_i(t)\equiv1\ \forall i=1,...,m$, the controller operates with nominal conditions and hence $$\tau(t)=G u(t)=\hat\tau_c(t)%\ (=\tau_c(t)\ \mathrm{if\ no\ input\ constraint\ is\ considered})$$ On the other hand if one of the actuators is subject to a loss of effectiveness or complete failure, i.e. if $\delta_i(t)\neq1$ for some $i$, the designed control law will no longer be able to ensure the desired effect, this meaning that, in the case of fault presence, one may have $$\tau(t)\neq \hat\tau_c(t)$$ with a consequent deterioration of system performances. Such problems can be avoided by accommodating the fault effects if a suitable control reconfiguration policy is considered. Defining a set of diagnosis signals, usually called residuals, one can detect and isolate the faults; then, performing the correct reconfiguration of the control input, one can track the (approximate) desired effect $\hat\tau_c$ properly again. The approach presented in this paper is based on Unknown Input Observers UIO (see for instance [@CPZ:96]), whose general structure is the following:$$\left\{
\begin{array}{l}
\dot z(t)=Fz(t)+RB v(t) +Ky(t)\\ \\
\hat x(t)=z(t)+Hy(t)
\end{array}
\right.$$ where the matrices $F,R,K$ and $H$ are design parameters and $v(t)$, $y(t)$ are, respectively, a known reference input signal and the measured output of the system to be estimated through the observer; the signal $v(t)$ is usually set equal to the nominal and unperturbed input that is commanded to the system. It is worth to note that, in order to achieve a correct asymptotic state estimation, the matrix $F$ has to be Hurwitz.\
As it will be shown in the following, unknown input observers are useful for the task of isolating faults in overactuated systems. Moreover, thanks to input redundancy, the control can be re-allocated in order to limit or avoid the use of faulty effectors/actuators once these have been isolated. On the other hand, this is not the only advantage of input redundancy in the considered framework: indeed, control allocation can be combined together with the fault isolation scheme in order to enlarge the family of identifiable faulty events. The reconfiguration can be performed by different methods, depending on several factors such as actuator dynamics, bounds on energy consumption, limited control inputs rates or other control constraints. It is worth to note that, due to the negative effects of faults, also the desired control effect $\tau_c(t)$ might be requested to change with respect to the original one in order to recover the deteriorated system performances, this corresponding to update the relation between $\tau_c(t)$ and the output signal $y(t)$ given by (\[tau-y\]).\
In the simplest case of unconstrained inputs, the nominal control allocation law is given by (\[uncons\]) and therefore, if the actuators $i_1,...,i_q$ are faulty and $q\leq m-k$, to get the desired effect $\tau_c(t)$ it is sufficient to re-allocate the control action setting $$u_{i_1}=u_{i_2}=\cdots=u_{i_q}\equiv0$$ and assigning the other components of $u(t)$, which are grouped for convenience in a vector $\tilde u\in\mathbb{R}^{m-q}$, according to \[simple-realloc\] (t)=G\^[-R]{}\_c(t), where the matrix $\tilde G\in\mathbb{R}^{k\times (m-q)}$ is obtained from $G$ by neglecting the columns $i_1,...,i_q$.\
\
On the other hand, since control reconfiguration can be regarded as a reduced-order control allocation problem in which some of the inputs are neglected, the use of the aforementioned techniques for handling input constraints can be straightforward extended. However, by turning off the input signals corresponding to faulty actuators, the redundancy of control inputs is reduced and the error between the desired control effect and the control effect provided by the approximate solution may increase after reconfiguration, as the class of admissible solutions to the allocation problem reduces.\
As already mentioned, control allocation can be used actively also to make faults observable; in particular, by considering additional input constraints (see Section \[club\]) which constrain control devices and surfaces to achieve common modes, one can isolate faults affecting selected groups of effectors. Whenever such constraints are not allowed to be imposed simultaneously in practice due to lack of control design freedom, an iterative control allocation scheme can be defined in order to switch periodically from one common mode to another after a prescribed time interval, until the fault isolation task is accomplished successfully.
Fault detection and isolation
=============================
The estimation error is defined as the difference between the true state $x(t)$ and the estimated state $\hat x(t)$: $$e(t)=x(t)-\hat x(t).$$ Our aim is to design a family of unknown input observers $\{\mathcal{O}_h\}_{h=1}^s$, such that the information provided by the estimation errors allow us to detect and isolate faults. To address such target one can proceed as follows [@CPZ:96].
Unknown Input Observers
-----------------------
The input $v(t)$ in the observer is set equal to the reference control effect $\hat\tau_c(t)$, that is $$v(t)=Gu(t)=\hat\tau_c(t),$$ where $u(t)$ is the nominal (fault free) control (\[unominalcon\]) (or (\[unominal\]) in the unconstrained case). Exploiting the observer structure, the dynamics of the error is ruled by the following equation $$\begin{array}{ll}
\dot e(t)&=\dot x(t)-\dot{\hat x}(t)\\
&=[(I_{n\times n}-HC)A-KC+FHC]x(t)-F\hat x(t)\\
&+(I_{n\times n}-HC)BG\Delta(t)u(t)-RBG u(t).
\end{array}$$ Setting $K=K_1+K_2$, if the following conditions are satisfied $$\begin{aligned}
&R=I_{n\times n}-HC\label{uio1}\\
&F=RA-K_1C,\quad \sigma(F)\in\mathbb{C}^{-}\label{uio2}\\
&K_2=FH\label{uio3}\end{aligned}$$ then the latter equation reduces to $$\dot e(t)=F e(t)+RBG(\Delta(t)-I_{m\times m})u(t),$$ where $\sigma(\cdot)$ stands for the spectrum of a matrix and the set $\mathbb{C}^{-}$ in the left open complex half-plane. Let us denote by $W\in\mathbb{R}^{n\times m}$ the matrix $BG$, whose columns will be indicated with $W_1,...,W_m$, i.e. \[BG=W\] BG=W=\[W\_1W\_m\]. It is worth to note that the matrix $BG(\Delta(t)-I)=W(\Delta(t)-I_{m\times m})$ appearing in the expression of $\dot{e}(t)$ has the following structure: $$W(\Delta(t)-I_{m\times m})=[(\delta_1(t)-1)W_1\ \cdots\ (\delta_m(t)-1)W_m],$$and hence a fault in the $j^{th}$ effector may only affect the $j^{th}$ column of $W$.
Constrained Output Fault Directions (COFD) {#prescrib}
------------------------------------------
This subsection is dedicated to the presentation of the proposed scheme for fault detection and isolation in overactuated systems; such method consists in constraining the residuals in prescribed subspaces of the output space (see for instance [@PaRi:94], [@WhSp:87]).\
Let us consider the canonical basis of $\mathbb{R}^p$, namely $\mathbf{e}_1,...,\mathbf{e}_p$; since by assumption the output matrix $C$ is full-rank, there exists $S\in\mathbb{R}^{n\times p}$ such that $$CS=I_{p\times p}=[\mathbf{e}_1\ \cdots\ \mathbf{e}_p].$$ The general solution of such equation is given by \[SS0\] S=C\^T(CC\^T)\^[-1]{}+\[I\_[nn]{}-C\^T(CC\^T)\^[-1]{}C\]S\_\*, where $S_*\in\mathbb{R}^{n\times p}$ is an arbitrary matrix. Denoting by $S_1,...,S_p$ the columns of the matrix $S$, the basic idea of the method is to design the observer parameters in order to guarantee that, if a fault occurs in the $i^{th}$ actuator, then the estimation error maintains the direction $S_i$ during the system evolution, this corresponding to a fixed direction $\mathbf{e}_i$ for the residual. It is worth to note that a first strong design constraint for the achievability of this condition is that directions $S_1,...,S_p$ need to correspond to eigenvectors of the observer matrix $F$. Moreover, due to the rank deficiency of $BG$, it is not possible in general to address a decoupled distribution of the faults effects over the columns of the matrix $S$ and we are required to deal with linear combinations of such characteristic directions. We recall that the dynamics of the estimation error is given by the equation $$\dot e(t)=F e(t)+RW(\Delta(t)-I_{m\times m})u(t).$$ Since $rank(W)=k<m\leq p$, we can arbitrarily assign only $k$ columns of the matrix $RW$ through the design parameter $R$, as the remaining $m-k$ are consequently constrained; we need therefore to consider several independent observers to achieve a correct fault isolation. One can proceed as follows.
We call a [$multi-index$]{} any vector $J$ of increasing natural numbers, i.e. $J=(j_1,...,j_\ell)$ with $ j_q\in\mathbb{N}\ \forall q=1,...,\ell$ and $1\leq j_1<j_2<\cdots< j_\ell\leq r,\ r\geq\ell$. The positive integers $\ell$ and $r$ are defined as, respectively, the length $L(J)$ and the domain $D(J)$ of multi-index $J$.
It is worth to note that the number $s$ of distinct multi-indices having length $\ell$ and domain $r$ is given by the binomial coefficient \[s-binom\] s==.
\[def2\] Given $W\in\mathbb{R}^{n\times m}$ with $rank(W)=k$, we call uniform sub-rank of $W$ the positive integer $k_0\leq k$ computed as \[kappa0\]
[ll]{} k\_0:=&{k: rank\[W\_[j\_1]{} W\_[j\_]{}\]=,\
& J=(j\_1,...,j\_): D(J)=m }.
[**Notation**]{} [*Given a multi-index $J=(j_1,...,j_\ell)$ with $D(J)=m$, we denote by $W_J\in\mathbb{R}^{n\times\ell}$ the matrix composed by the columns of $W$ corresponding to the indices included in $J$, i.e.*]{} $$W_J=[W_{j_1}\ \cdots\ W_{j_\ell}].$$ Let $k_0\leq k$ the uniform sub-rank of the matrix $W=BG$. Then, if by tuning the matrix $R$ we prescribe the first $k_0$ columns of $RW$, for example imposing that they have to be equal to $S_1,S_2,...,S_{k_0}$, we get $$RW=[S_1\ \cdots\ S_{k_0}\ V_1\ \cdots\ V_{m-k_0}],$$ where $V_j=\sum_{i=1}^{k_0}\alpha_{ij}S_i$ for some coefficients $\alpha_{ij}$. As a consequence we have $$CRW=[\mathbf{e}_1\ \cdots\ \mathbf{e}_{k_0}\ \omega_{1}\ \cdots\ \omega_{m-k_0}],$$ with $\omega_j=\sum_{i=1}^{k_0}\alpha_{ij}\mathbf{e}_i.$ Setting $R^{(1)}=R$, we can iterate this construction by designing $R^{(h)}$ such that \[2met\] R\^[(h)]{}\[W\_[j\_1\^h]{} W\_[j\_[k\_0]{}\^h]{}\]=\[S\_[i\_1\^h]{} S\_[i\_[k\_0]{}\^h]{}\] as multi-indices $J_h=(j_{1}^h,...,j_{k_0}^h)$ and $I_h=(i_{1}^h,...,i_{k_0}^h)$ vary; at the end of the construction we obtain a family of matrices $\{R^{(h)}\}_{h=1}^s$, with $$s=\binom{m}{k_0}\binom{p}{k_0}=\frac{m!p!}{(m-k_0)!(p-k_0)!(k_0!)^2}.$$ We point out that the information provided by the residuals associated to such family of matrices is redundant. In order to reduce the computational burden and avoid overlapping of information, we need to investigate what is the minimum number of matrices $R^{(h)}$ required for a proper fault isolation. We see from the explicit construction of $R^{(1)}$ that, if $k_0>1$, in this case faults affecting the first $k_0$ actuators lead to residual signals directed as $\mathbf{e}_1,...,\mathbf{e}_{k_0}$ respectively, while faults affecting the other actuators lead to residual signal obtained as linear combination of two or more vectors $\mathbf{e}_j$, $j=1,...,k_0$. On the other hand, if more than one fault occurs we get a residual signal defined by a linear combination of vectors $\mathbf{e}_j$ as well, this meaning that with the information provided by this unique observer we are not able to distinguish multiple faults from individual faults affecting one of the last $m-k_0$ actuators. For this reason, the number of observers to be considered has to be sufficient to decouple effects of single and multiple faults. To this purpose we note that, since the observers are designed independently, the particular choice of vectors $S_j$ among $\{S_1\ \cdots\ S_p\}$ in (\[2met\]) does not influence the fault isolation procedure: nevertheless, this freedom of choice may result to be helpful for the design of the Hurwitz matrix $F$. We claim that the maximum number of isolable faults is $k_0-1$ with a required number of observers equal to $$\bar{s}=\binom{m}{k_0}.$$ This can be verified observing that, in order to isolate the faults, at least one residual $r^{(h)}(t)$ needs to have null projection along one of the basis vectors $\mathbf{e}_j$: as a consequence, the number of isolable faults has to be less than the maximum admissible number of independent components of each residual $r^{(h)}(t)$, that is $k_0$. The integer $\bar{s}$ can be obtained simply computing the number of all structurally distinct vectors of $m$ elements with $k_0$ assigned entries.\
For sake of simplicity we fix the vectors $S_j$ in (\[2met\]), assuming that the right-hand side is equal to $\hat{S}:=[S_1\ \cdots\ S_{k_0}]$ for any $h$. Once the properties of the matrices $R^{(h)}$ are defined, one have to deal with the stabilization of the matrix $F^{(h)}$ together with the fulfillment of the rank condition \[rank1\] rank\[S\_j F\^[(h)]{}S\_j (F\^[(h)]{})\^[n-1]{}S\_j\]=1 j=1,...,k\_0, which corresponds to the requirement for $S_j$ to be an eigenvector of the matrix $F^{(h)}$, this last condition being fundamental for ensuring a constant direction of the output. We point out that for any $h=1,...,\bar{s}$, the solution of (\[2met\]) is given by \[acca\] {
[ll]{} R\^[(h)]{}&=(I\_[nn]{}-H\^[(h)]{}C)\
H\^[(h)]{}&=(W\_[J\_h]{}-S)(CW\_[J\_h]{})\^[-L]{}\
&+H\_\*\^[(h)]{}(C-(CW\_[J\_h]{})(CW\_[J\_h]{})\^[-L]{}C)
. where $(\cdot)^{-L}$ stands for the left pseudo-inverse and $H^{(h)}_*\in\mathbb{R}^{n\times p}$ is an arbitrary matrix. Let $M^{(h)}\in\mathbb{R}^{k_0\times k_0}$ be a diagonal and negative definite matrix. Recalling that $F^{(h)}=R^{(h)}A-K_1^{(h)}C$, condition (\[rank1\]) can be achieved by solving \[gain\] K\_1\^[(h)]{}(CS)=R\^[(h)]{}AS-S M\^[(h)]{}, which gives \[gain2\]
[ll]{} K\^[(h)]{}\_1&=(R\^[(h)]{}AS-S M\^[(h)]{})(CS)\^[-L]{}\
&+K\_\*\^[(h)]{}(I\_[pp]{}-(CS)(CS)\^[-L]{}),
where $K_*^{(h)}\in\mathbb{R}^{n\times p}$ is an arbitrary matrix. Denoting by $e^{(h)}(t)$ the estimation error associated to the observer $\mathcal{O}_h$, we define the (vectorial) residual signals: $$r^{(h)}(t)=Ce^{(h)}(t).$$ We are now ready to state the main result of this section.
\[theomain2\] Let $S, R^{(h)}, H^{(h)}, K^{(h)}_1$ be assigned by (\[SS0\]), (\[acca\]) and (\[gain2\]). Let us assume that, for any multi-index $J_h=(j_1^h,...,j^h_{k_0})$, $h=1,...,\bar s$, the following conditions hold true
1. $rank(W_{J_h})=rank(CW_{J_h})=k_0$;
2. the matrices $S_*, H^{(h)}_*$ and $K_*^{(h)}$ can be found such that the matrix $F^{(h)}=R^{(h)}A-K_1^{(h)}C$ is Hurwitz.
Then $(F^{(h)}, H^{(h)}, R^{(h)}, K^{(h)})$ is an unknown input observer and it satisfies condition (\[2met\]) for any $h=1,...,\bar s$, hence the residual signals are able to detect and isolate up to $k_0-1$ faults affecting simultaneously the system actuators.
We can represent residuals as ordered sums of the basis vectors and their combinations; we will employ the $\oplus$ to indicate a logic sum of the vectors $v,w$ depending on the order, i.e. $v\oplus w\neq w\oplus v.$ For example we obtain the following logic representation of the first residual: $$r^{(1)}=\mathbf{e}_1\oplus\mathbf{e}_2\oplus\cdots\oplus\mathbf{e}_{k_0}\oplus\omega_1^1\oplus\cdots\oplus\omega_{m-k_0}^1$$ where $\omega_j^1$ is an arbitrary combination of the vectors $\mathbf{e}_1,...,\mathbf{e}_{k_0}$; in a similar way, the residual associated to the multi-index $J_h=(j_1^h,...,j_{k_0}^h)$ can be represented by $$r^{(h)}=\omega_{1}^h\oplus\cdots\oplus\omega_{j_1-1}^h\oplus\mathbf{e}_1\oplus\omega_{j_1}^h\oplus\cdots\oplus\mathbf{e}_{k_0}\oplus\cdots\oplus\omega_{m-k_0}^h$$
It is worth to note that condition (\[rank1\]) is sufficient for prescribing a fixed direction to the residual signals only if the observer initialization error is zero or if the estimation error $e(t)$ has reached the steady-state at the moment of fault occurrence.
Cluster residuals {#club}
-----------------
We consider here an extended framework, in which the actuators are grouped into $q$ clusters: $$\begin{array}{l}
\mathcal{A}_1=\{u_1,...,u_{i_1}\},\ \mathcal{A}_2=\{u_{i_1+1},...,u_{i_2}\},\\
\ \dots\dots\ \mathcal{A}_q=\{u_{i_{q-1}+1},...,u_m\},
\end{array}$$ where $1\leq i_1<i_2<\cdots i_{q-1}<m$. The introduction of this new model is motivated by the need of isolate common mode faults which may affect simultaneously actuators or effectors sharing the same auxiliary systems. For sake of clarity only the case of non-overlapping clusters is presented; on the other hand, the proposed methods can be readily modified in order to be used also in the case of possibly overlapping clusters.\
Let us denote by $\alpha_h$ the cardinality of the cluster $\mathcal{A}_h$, i.e. $\alpha_h=i_h-i_{h-1}$ and hence $\sum_{i=1}^q\alpha_i=m$. The faults are supposed to act uniformly on the whole cluster $\mathcal{A}_i$, so that they can be modeled as the block-diagonal multiplicative matrix: \[deltaclust\] (t)=diag(d\_1(t)I\_[\_1\_1]{},d\_2(t)I\_[\_2\_2]{},,d\_q(t)I\_[\_q\_q]{}) Let $k_0$ be the uniform sub-rank of $W$ and let us suppose that $k_0\geq3$ and $$\max_{i=1,...,q}\alpha_i\leq k_0-1.$$ As a consequence if a fault is present in a single cluster it can be detected and isolated; in the same way, if there exist two (or more clusters) with $\alpha_i+\alpha_j\leq k_0-1$, a multiple fault can be isolated as well. On the other hand, if for some pair of indices $i,j$, one has $\alpha_i+\alpha_j\geq k_0$, a multiple fault on the associated effectors will lead to completely saturated residuals, this meaning that no fault isolation can be achieved at the present step. However, by introducing additional constraints in the control allocation scheme, one can design a new set of observers to be used to identify faulty clusters of actuators/effectors. For sake of simplicity let us consider first the following case: $$\alpha_1+\alpha_2\geq k_0;$$ in addition, let us assume that \[m-k\] 2+m-k\_1+\_2. We select two finite sequences of real numbers $\{\zeta_1^{(1)},\cdots,\zeta_{\alpha_1-1}^{(1)}\}$ and $\{\zeta_{1}^{(2)},\cdots,\zeta^{(2)}_{\alpha_2-1}\}$; using the control redundancy ensured by (\[m-k\]), we are allowed to impose the constraints \[redund\]
[l]{} =\^[(1)]{}\_[j]{},j=1,...,\_1-1,\
\
=\^[(2)]{}\_[j]{},j=1,...,\_2-1
together with the equality $$\tau_c(t)=Gu(t).$$ Due to (\[redund\]), the overall input signal associated to the first cluster $\mathcal{A}_1$ turns out to be $$\sum_{i=j}^{i_1}W_ju_j(t)=u_{i_1}(t)\left(W_{i_1}+\sum_{j=1}^{\alpha_1-1}{\zeta^{(1)}_j}W_j\right),$$ while the overall signal corresponding to $\mathcal{A}_2$ is $$\sum_{j=i_1+1}^{i_2}W_ju_j(t)=u_{i_2}(t)\left(W_{i_2}+\sum_{j=1}^{\alpha_2-1}{\zeta^{(2)}_j}W_{j+i_1}\right).$$ Let us denote by $W^{\{1\}}$ and $W^{\{2\}}$ the constant vectors $$W_{\{1\}}=W_{i_1}+\sum_{j=1}^{\alpha_1-1}{\zeta^{(1)}_j}W_j,\ W_{\{2\}}=W_{i_2}+\sum_{j=1}^{\alpha_2-1}{\zeta^{(2)}_j}W_{j+i_1};$$ since, without loss of generality, the coefficients $\zeta_j^{(\star)}$, $\star=i_1,i_2$, can be chosen such that the latter vectors are independent, by definition one has \[rango3\] rank\[W\_[{1}]{} W\_[{2}]{} W\_[i\_2+1]{} W\_[m]{}\]3. If the cluster $\mathcal{A}_1$ undergoes a fault, the dynamics of the estimation error turns out to be $$\dot e(t)=F e(t)+ RW_{\{1\}}(\delta_1(t)-1)u_{i_1}(t)$$ and a similar condition holds for faults in $\mathcal{A}_2$. Now, recalling that $rank[W_{\{1\}}\ W_{\{2\}}]= 2$ and designing the observer matrix $R$ such that $$R[ W_{\{1\}}\ W_{\{2\}}\ W_i]=[\mathbf{e_1}\ \mathbf{e}_2\ \mathbf{e}_3],$$ for some $i$ with $u_i\notin \mathcal{A}_1\cup \mathcal{A}_2$, one obtains a residual signal with prescribed output directions associated to the first two clusters of actuators. We will refer to such signal as a [*cluster residual*]{}. The above construction can be readily extended to the case of faults involving more than two effector clusters, if there is enough redundancy to use control allocation. To this purpose, let us set \[wil\] W\_[{h}]{}=W\_[i\_h]{}+\_[j=1]{}\^[\_h-1]{}\^[(i\_h)]{}\_jW\_[i\_[h-1]{}+1]{},h=1,...,q where the coefficients $\zeta^{(i_\ell)}_j\in\mathbb{R}$ have to be defined. Such overall input vectors can be organized into a reduced-order input matrix $$W^\star=[W_{\{1\}}\ W_{\{2\}}\ \cdots\ W_{\{q\}}];$$ applying the observer design scheme proposed in Section III.A with $W$ replaced by $W^*$, one can obtain a family of cluster residuals associated to the effector groups $\mathcal{A}_1,...,\mathcal{A}_q$.
Case study: overactuated marine vessel
======================================
This section is focused on illustrating the application of theoretical results to the case of an overactuated marine vessel. We consider the following ship model $$\left\{
\begin{array}{l}
\dot\eta = P(\eta)\nu\\ \\
M\dot\nu=-V(\eta,\nu)+\tau+P^T(\eta)b(t)
\end{array}
\right.$$ where $M$ is the inertia matrix, $\eta=[x_G,y_G,\psi]^T$ is the ship position coordinates in the earth-fixed reference frame and $\nu=[\nu_x,\nu_y,\dot\psi]$ contains surge, sway and yaw angular velocities with respect to the body-fixed reference frame, which is identified with the vessel center of mass; the vector $V(\eta,\nu)$ includes Coriolis and damping terms, the actual thrust force in surge, sway and the yaw moment are given by $\tau=[\tau_x,\tau_y,m_{\dot\psi}]$ and $P(\eta)=P(\psi)$ is a standard rotation matrix $$P(\psi)=\left[
\begin{array}{rrr}
\cos\psi&-\sin\psi&0\\
\sin\psi&\cos\psi&0\\
0&0&1
\end{array}
\right]$$ The perturbation term $b(t)$, which is assumed to be bounded by a known constant, is used to model disturbances affecting the system, such as slowly-varying forces and moments caused by wave loads, ocean currents or winds (see [@Fossen:94] for further details). Following [@BeFo:97], both $\eta$ and $\nu$ are assumed to be measured (possibly through a state estimation), the constant matrix $M$ is known and $V(\eta,\nu)$ is a known function in the state variables $\eta,\nu$. We consider a ship equipped with 3 azimuth thrusters (rotatable) $T_1, T_2, T_3$ and 2 transverse tunnel thrusters (fixed orientation) $T_4, T_5$. A sketch of the vessel model is depicted in Figure \[fig1\].
\
The actual thrust force is related to the control input through the linear equation $$\tau=Gu(t),$$ with $$G=\left[
\begin{array}{rrrrrrrr}
1&0&1&0&1&0&0&0\\
0&1&0&1&0&1&1&1\\
\gamma_1&\gamma_2&\gamma_3&\gamma_4&\gamma_5&\gamma_6&\chi_7&\chi_8
\end{array}
\right],$$ where the moment arms $\gamma_j$ are associated to azimuth thrusters and the moment arms $\chi_j$ to tunnel thrusters instead. In particular such arms can be computed as $$\begin{array}{rll}
\gamma_{2s-1}&=d_s\sin{\phi_s}& s=1,2,3\\
\gamma_{2s}&=d_s\cos{\phi_s}& s=1,2,3\\
\chi_{s}&=d_s\cos\phi_s& s=4,5
\end{array}$$ where $d_s$ are the distances of thrusters and $\phi_s$ are the angles from the rotation point.\
Assuming that the vessel rotation is negligible with respect to translation motion, setting $X=[\eta,\nu]^T$, for sake of simplicity and without loss of generality, the above nonlinear model can be linearized as follows \[linearmod\] X(t)=AX(t)+B(t)+E(t) with $$A=\left[
\begin{array}{rr}
0_{3\times 3}& P(\bar\psi)\\
0_{3\times 3}&-M^{-1}D
\end{array}
\right],\quad\quad B=\left[
\begin{array}{r}
0_{3\times 3}\\
M^{-1}
\end{array}
\right],$$ where $\bar\psi$ is a constant angle associated to some reference heading direction, $D=D(\bar\nu)$ is a constant damping matrix depending on a nominal reference velocity $\bar\nu$ and $$E(t)=\left[
\begin{array}{r}
0_{3\times 1}\\
M^{-1}P^T(\bar\psi)b(t)
\end{array}
\right].$$ Since by assumption all state variables are measured, without loss of generality the output matrix $C$ is supposed to be equal to the identity matrix $$C=I_{6\times 6}.$$ Assuming that the marine vessel has a mass $\mu=6\cdot10^{6}Kg$, with lenght $L=76 m$ and width $w= 16 m$, the following parameters are obtained [@Fossen:94]: $$\begin{array}{l}
M=10^9\left[
\begin{array}{rrr}
0.0068 & 0. & 0.\\
0. & 0.0113 & -0.0340\\
0. & -0.0340& 4.4524
\end{array}
\right],\\ \\ D=10^8\left[
\begin{array}{rrr}
0.0008 & 0. &0.\\
0 & 0.0025 & -0.0203\\
0 & -0.0340 & 3.8481
\end{array}
\right].
\end{array}$$ Without loss of generality $\bar\psi=0$, that is $P(\bar\psi)=I_{3\times 3}$. Moreover, taking $d_1=d_2=20m$, $d_3=18.5m$, $d_4=30m$, $d_5=35m$ and $\phi_1=\pi+0.3,\ \phi_2=\pi-0.3$, the matrix $G$ is given by $$G\simeq\!\!\left[
\begin{array}{rrrrrrrr}
1&0&1&0&1&0&0&0\\
0&1&0&1&0&1&1&1\\
-5.91 & -19.1 & 5.91 & -19.1 & 0& 18.5 & 30 & 35
\end{array}
\right]\!\!.$$ We suppose the faults to occur in effectors (thrusters) rather than in single actuators: this corresponds to consider a fault matrix $\Delta(t)=diag(\delta_1(t),\delta_1(t),\delta_2(t),\delta_2(t),\delta_3(t),\delta_3(t),\delta_4(t),\delta_5(t))$ with coupled entries representing failures in the thrusters $T_j$, $j=1,...,5$. We set $$W=BG=[W_1\ W_2\ \vline\ W_3\ W_4\ \vline\ W_5\ W_6\ \vline\ W_7\ \vline\ W_8 ],$$ where the vertical rules have been added to easily individuate the actuators corresponding to each thruster.\
We can design a family of $s=4$ COFD unknown input observers $\{\mathcal{O}_h\}$ to isolate faults affecting singularly each thruster. In particular, following the steps of Theorem \[theomain2\], we select the observers matrices $R^{(h)}$ in order to have $$\begin{array}{ll}
R^{(1)}[W_1\ W_2\ W_3]=[\mathbf{e}_1\ \mathbf{e}_2\ \mathbf{e}_3]\\
R^{(2)}[W_3\ W_4\ W_1]=[\mathbf{e}_1\ \mathbf{e}_2\ \mathbf{e}_3]\\
R^{(3)}[W_5\ W_6\ W_1]=[\mathbf{e}_1\ \mathbf{e}_2\ \mathbf{e}_3]\\
R^{(4)}[W_7\ W_8\ W_1]=[\mathbf{e}_1\ \mathbf{e}_2\ \mathbf{e}_3].
\end{array}$$ The observer gains $K_1^{(h)}$ can be chosen in order to assign the eigenvalues of the observer matrices $F^{(h)}$; on the other hand, since the whole state $X$ is measurable, i.e. $C=I_{6\times 6}$, the matrices $F^{(h)}$ turn out to be diagonal. For sake of simplicity we assume $F^{(h)}=F$ for any $h=1,...,5$, with $F=diag(-1,-1,-2,-5,-6,-7)$.\
Different single or multiple thruster fault events have been simulated, assuming the initial conditions $\eta_0=\eta(0)=(1,1,0)$ and $\nu_0=\nu(0)=(2.2,1.9,0)$. The disturbance term is supposed to be given by the sum of two contributions: a constant term with random but fixed input direction representing an irrotational ocean current and an oscillating term with varying input direction representing waves; the overall disturbance effect $b(t)$ is assumed to be bounded by the known constant $\epsilon=5\cdot10^6$. The ship is supposed to be equipped with an $xy$-joystick control device together with a heading autopilot; the nominal operating conditions of the vessel are defined by a constant translational speed regime, this corresponding to the commanded control effect $$\tau_c(t)=D\nu(t)+[0\ 0\ a_\psi]^T,$$ where $a_\psi$ is a PID controller for the yaw angle.\
### Single faults {#single-faults .unnumbered}
We first suppose the azimuth thruster $T_1$ to be affected by a fault that gradually fades and its effect is $\delta_1(t)=e^{-0.03t}$. The behavior of the residual $r^{(1)}(t)$ is depicted in Figure \[fig2\]: while the projections of $r^{(1)}$ along the directions $\mathbf{e}_1$ and $\mathbf{e}_2$ are significant, the projection along the direction $\mathbf{e}_3$ is negligible; on the other hand, as shown in Figure \[fig3\], the projections along the direction $\mathbf{e}_3$ of the other residuals is not negligible. As a consequence the faulty thruster $T_1$ can be identified and the control reconfiguration policy can be applied.\
\
### Multiple faults {#multiple-faults .unnumbered}
The most relevant cases of faults affecting simultaneously two thrusters due to common auxiliaries or power supply failures are reported in the following table:
In particular, it is assumed that $T_1$ and $T_4$ share the same auxiliaries, as well as $T_2$ and $T_5$ do; the thruster $T_3$ is supposed to be equipped with a switching device that enables it to be connected arbitrarily to one sub-group or to the other, depending on the operating conditions of the system.\
Using actively the control allocation, a new family of observers can be designed to obtain cluster residuals (as showed in Section \[club\]); for sake of simplicity we assume the evolution time to be re-initialized at the present step. We choose the coefficients $\zeta^{(1)}=2.27, \zeta^{(2)}=3.41$ and $\zeta^{(3)}=1.38$ according to the initial values of the control inputs, i.e. $\zeta^{(i)}=u_{2i}(0)/u_{2i-1}(0)$; we impose the following conditions on the control input \[glizeta\] \^[(1)]{}u\_1(t)=u\_2(t), \^[(2)]{}u\_3(t)=u\_4(t), \^[(3)]{}u\_5(t)=u\_6(t) and we set \[gliW\]
[l]{} W\_[{1}]{}=\^[(1)]{}W\_1+W\_2, W\_[{2}]{}=\^[(2)]{}W\_3+W\_4,\
W\_[{3}]{}=\^[(3)]{}W\_5+W\_6, W\_[{4}]{}=W\_7, W\_[{5}]{}=W\_8.
By transformation (\[glizeta\])-(\[gliW\]), each thruster $T_i$ is associated to its overall input vector $W_{\{i\}}$, $i=1,2,3,4,5$, in a scheme that corresponds to fixed directions of azimuth thrusters. A family of COFD unknown input observers $\{\mathcal{O}_h\}_{h=1}^{4}$ is then defined such that the corresponding matrices $R^{(h)}$ satisfy: \[rules\]
[l]{} R\^[(1)]{}\[W\_[{1}]{} W\_[{2}]{} W\_[{3}]{}\]=\[\_1 \_2 \_3 \]\
R\^[(2)]{}\[W\_[{1}]{} W\_[{4}]{} W\_[{5}]{}\]=\[\_1 \_2 \_3 \]\
R\^[(3)]{}\[W\_[{2}]{} W\_[{3}]{} W\_[{4}]{}\]=\[\_1 \_2 \_3 \]\
R\^[(4)]{}\[W\_[{2}]{} W\_[{3}]{} W\_[{5}]{}\]=\[\_1 \_2 \_3 \]
Figure \[fig8\] illustrates the case of faults $\delta_2(t)=e^{-0.02t }$ and $\delta_5(t)=e^{-0.01t}$ affecting simultaneously the thrusters $T_2, T_5$. The cluster residual $r^{(4)}$ has a negligible component along the direction $\mathbf{e}_2$ while the other components are significant; moreover it can be verified that all other cluster residuals have significant components along the directions $\mathbf{e}_i$ for $i=1,2,3$. Based on the design rules (\[rules\]), such information is sufficient to allow a correct identification of the pair of faulty thrusters $T_2$ and $T_5$.
### Control reconfiguration {#control-reconfiguration .unnumbered}
The control re-allocation can be performed through the reduced order pseudo-inverse method (\[simple-realloc\]). Figures \[fig5\]-\[fig6\] show the evolution of the ship position in the case of a fault affecting the thruster $T_1$; the reconfiguration procedure is supposed to be activated after $t_0=180s$. The control is successfully re-allocated, and the commanded control effect is modified in order to track the original vessel velocities in surge and sway, while the rotational speed is automatically updated by the heading PID controller given by $a_\psi$.
[99]{} H. Alwi and C. Edwards (2006), Sliding mode FTC with on-line control allocation , [*Proc. 45th IEEE Conf. on Decision and Control*]{}, 5579-5584. S.P. Berge and T.I. Fossen (1997), Robust control allocation of overactuated ships; experiments with a model ship, [*Proc. 4th IFAC Conf. on Manoeuvring and Control of Marine Crafts*]{}, 166-171. M. Bodson (2002), Evaluation of optimization methods for control allocation, [*J. of Guidance, Control and Navigation*]{}, 25, 703-711. J.M. Buffington and D.F. Enns (1996), Lyapunov stability analysis of daisy chain control allocation, [*J. of Guidance, Control and Navigation*]{}, 19, 1226-1230. J.J. Burken, P. Lu, Z.L. Wu and C. Bahm (2001), Two reconfigurable flight control design methods: robust servomechanisms and control allocation, [*J. of Guidance, Control and Navigation*]{}, 24, 482-493. A. Casavola and E. Garone (2010), Fault-tolerant adaptive control allocation schemes for overactuated systems, [*Int. J. Robust and Nonlinear Control*]{}, vol. 20, 1958-1980. J. Chen, R.J. Patton and H.Y. Zhang (1996), Design of unknown input observers and robust detection filters, [*Int. J. of Control*]{}, 63, 85-105. M.L. Corradini, G. Orlando and G. Parlangeli (2005), A fault tolerant sliding mode controller for accommodating actuator failures, [*Proc. 44th IEEE Conf. on Decision and Control*]{}, 3091-3097. A. Cristofaro and T.A. Johansen (2013), Fault-Tolerant Control Allocation using Unknown Input Observers, [*under review*]{}. M.A. Demetriou (2005), Using unknown input observers for robust adaptive fault detection in vector second-order systems, [*Mechanical Systems and Signal Processing*]{}, 19, 291-309. W.C. Durham (1993), Constrained control allocation, [*J. of Guidance, Control and Navigation*]{}, 16, 717-725. T.I. Fossen (1994), Guidance and control of ocean vehicles, Wiley, New York. T.I. Fossen and T.A. Johansen (2006), A survey of control allocation methods for ships and underwater vehicles, [*Proc. 14th Mediterranean Conf. on Control and Automation*]{} S.A. Frost and M. Bodson (2010), Resource balancing control allocation, [*Proc. of 2010 American Control Conference*]{}, 1326-1331. G.H. Golub and C.F. van Loan (1983), Matrix computations, North Oxford academic press. O. Härkegård (2004), Dynamic control allocation using constrained quadratic programming, [*J. of Guidance, Control and Navigation*]{}, 27, 1028-1034. T.A. Johansen and T.I. Fossen (2013), Control allocation: A survey, [*Automatica*]{}, [in press]{}. J.Park and G. Rizzoni (1994), An eigenstructure assignment algorithm for the design of fault detection filters, [*IEEE Trans. on Automatic Control*]{}, 39, 1521-1524. J.A.M. Petersen and M. Bodson (2006), Constrained quadratic programming techniques for control allocation, [*IEEE Trans. on Control Systems Technology*]{}, 14, 91-98. N. Sarkar, T.K. Podder and G. Antonelli (2002), Fault-accommodating thruster force allocation for an AUV considering thruster redundancy and saturation, [*IEEE Trans. on Robotics and Automation*]{}, 18, 223-233. I. Shames, A.M.H. Teixeira, H. Sandberg and K.H. Johansson (2011), Distributed fault detection for interconnected second-order systems, [*Automatica*]{}, 47, 2757-2764. J. Shi, W. Zhang, G. Li and X. Liu (2010), Research on allocation efficiency of the redistributed pseudo inverse algorithm, [*Science China: Information Sciences*]{}, 53, 271-277. J. Tjønnås and T.A. Johansen (2008), Adaptive control allocation, [*Automatica*]{}, 44, 2754-2766. D. Wang and K. Lum (2007), Adaptive unknown input observer approach for aircraft actuator fault detection and isolation, [*Int. J. Adaptive Control and Signal Processing*]{}, 21, 31-48. J.E. White and J.L. Speyer (1987), Detection filter design: spectral theory and algorithms, [*IEEE Trans. on Automatic Control*]{}, 32 (7), 593-603. L. Zaccarian (2009), Dynamic allocation for input redundant control systems, [*Automatica*]{}, 45, 1431-1438. Y. Zhang, S. Suresh, B. Jiang and D. Theilliol (2007), Reconfigurable control allocation against aircraft control effector failures, [*Proc. 16th IEEE Conf. on Control Applications*]{}, 1197-1202. Y. Zhang and J. Jiang (2008) Bibliographical review on reconfigurable fault tolerant control systems, [*Annual Reviews in Control*]{}, 32, 229-252.
[^1]: A. Cristofaro is with School of Science and Technology, University of Camerino, Italy. T.A. Johansen is with Department of Engineering Cybernetics, Center for Autonomous Marine Operations and Systems, Norwegian University of Science and Technology, Norway. email: [[email protected], [email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We review the current state of data mining and machine learning in astronomy. [*Data Mining*]{} can have a somewhat mixed connotation from the point of view of a researcher in this field. If used correctly, it can be a powerful approach, holding the potential to fully exploit the exponentially increasing amount of available data, promising great scientific advance. However, if misused, it can be little more than the black-box application of complex computing algorithms that may give little physical insight, and provide questionable results. Here, we give an overview of the entire data mining process, from data collection through to the interpretation of results. We cover common machine learning algorithms, such as artificial neural networks and support vector machines, applications from a broad range of astronomy, emphasizing those where data mining techniques directly resulted in improved science, and important current and future directions, including probability density functions, parallel algorithms, petascale computing, and the time domain. We conclude that, so long as one carefully selects an appropriate algorithm, and is guided by the astronomical problem at hand, data mining can be very much the powerful tool, and not the questionable black box.'
address:
- |
Herzberg Institute of Astrophysics, National Research Council, 5017 West Saanich Road, Victoria, BC V9E 2E7, Canada\
[email protected]
- |
Department of Astronomy, University of Illinois at Urbana-Champaign, 1002 West Green Street,\
Urbana, IL 61801, USA\
[email protected]
author:
- 'NICHOLAS M. BALL'
- 'ROBERT J. BRUNNER'
title: DATA MINING AND MACHINE LEARNING IN ASTRONOMY
---
Introduction {#Sec: Intro}
============
In its broadest sense, data mining is simply the act of turning raw data from an observation into useful information. This information can be interpreted by hypothesis or theory, and used to make further predictions. This scientific method, where useful statements are made about the world, has been widely employed to great effect in the West since the Renaissance, and even earlier in other parts of the world. What has changed in the past few decades is the exponential rise in available computing power, and, as a related consequence, the enormous quantities of observed data, primarily in digital form. The exponential rise in the amount of available data is now creating, in addition to the natural world, a digital world, in which extracting new and useful information from the data already taken and archived is becoming a major endeavor in itself. This action of [*knowledge discovery in databases*]{} (KDD), is what is most commonly inferred by the phrase data mining, and it forms the basis for our review.
Astronomy has been among the first scientific disciplines to experience this flood of data. The emergence of data mining within this and other subjects has been described[@bell:petascale; @bell:deluge; @hey:4thparadigm] as the [*fourth paradigm*]{}. The first two paradigms are the well-known pair of theory and observation, while the third is another relatively recent addition, computer simulation. The sheer volume of data not only necessitates this new paradigmatic approach, but the approach must be, to a large extent, automated. In more formal terms, we wish to leverage a computational machine to find patterns in digital data, and translate these patterns into useful information, hence [*machine learning*]{}. This learning must be returned in a useful manner to a human investigator, which hopefully results in human learning.
It is perhaps not entirely unfair to say, however, that scientists in general do not yet appreciate the full potential of this fourth paradigm. There are good reasons for this of course: scientists are generally not experts in databases, or cutting-edge branches of statistics, or computer hardware, and so forth. What we hope to do in this review, primarily for the data mining skeptic, is to shed light on why this is a useful approach. To accomplish this goal, we emphasize either algorithms that have or could currently be usefully employed, and the actual scientific results they have enabled. We also hope to give an interesting and fairly comprehensive overview to those who do already appreciate this approach, and perhaps provide inspiration for exciting new ideas and applications. However, despite referring to data mining as a whole new paradigm, we try to emphasize that it is, like theory, observation, and simulation, only a part of the broader scientific process, and should be viewed and utilized as such. The algorithms described are [*tools*]{} that, when applied correctly, have vast potential for the creation of useful scientific results. But, given that it is only part of the process, it is, of course, not the answer to everything, and we therefore enumerate some of the limitations of this new paradigm.
We start in §\[Subsec: Why\] with a summary of some of the advantages of this approach. In §\[Sec: Overview\], we summarize the process from the input of raw data to the visualization of results. This is followed in §\[Sec: Uses\] by the actual application of data mining tools in astronomy. §\[Sec: Overview\] is arranged algorithmically, and §\[Sec: Uses\] is arranged astrophysically. It is likely that the expert in astronomy or data mining, respectively, could infer much of §\[Sec: Uses\] from §\[Sec: Overview\], and vice-versa. But it is unlikely (we hope) that the combination of the two sections does not have new ideas or insights to offer to either audience. Following these two sections, in §\[Sec: Future\], we combine the lessons learned to discuss the future of data mining in astronomy, pointing out likely near-term future directions in both the data mining process and its physical application. We conclude with a summary of the main points in §\[Sec: Conclusions\].
Why Data Mining? {#Subsec: Why}
----------------
Of course, what astronomers care about is not a fashionable new computational method for ever more complex data analysis, but the [*science*]{}. A fancy new data mining system is not worth much if all it tells you is what you could have gained by the judicious application of existing tools and a little physical insight[@hand:illusion]. We therefore summarize some of the advantages of this approach:
[*Getting anything at all*]{}: upcoming datasets will be almost overwhelmingly large. When one is faced with Petabytes of data, a rigorous, automated approach that intelligently extracts pertinent scientific information will be the only one that is tractable.
[*Simplicity*]{}: despite the apparent plethora of methods, straightforward applications of very well-known and well-tested data mining algorithms can quickly produce a useful result. These methods can generate a model appropriate to the complexity of an input dataset, including nonlinearities, implicit prior information, systematic biases, or unexpected patterns. With this approach, [*a priori*]{} data sampling of the type exemplified by elaborate color cuts, is not necessary. For many algorithms, new data can be trivially incorporated as they become available.
[*Prior information*]{}: this can be either fully incorporated, or the data can be allowed to completely ‘speak for themselves’. For example, an unsupervised clustering algorithm can highlight new classes of objects within a dataset that might be missed if a prior set of classifications were imposed.
[*Pattern recognition*]{}: an appropriate algorithm can highlight patterns in a dataset that might not otherwise be noticed by a human investigator, perhaps due to the high dimensionality. Similarly, rare or unusual objects can be highlighted.
[*Complimentary approach*]{}: although there are numerous examples where the data mining approach demonstrably exceeds more traditional methods in terms of scientific return. Even when the approach does not produce a substantial improvement, it still acts as an important complementary method of analyzing data, because different approaches to an overall problem help to mitigate systematic errors in any one approach.
Overview of Data Mining and Machine Learning Methods {#Sec: Overview}
====================================================
In this section, we review the data mining process. Specifically, as described in §\[Sec: Intro\], this data mining review focuses on knowledge discovery in databases (KDD), although our definition of a ‘database’ is somewhat broad, essentially being any machine-readable astronomical data. As a result, this section is arranged algorithmically. To avoid overlap with §\[Sec: Uses\] on the astronomical uses, we defer most of the application examples to that section. Nevertheless, all algorithms we describe have been, or are of sufficient maturity that they could immediately be applied to astronomical data. The reader who is expert in astronomy but not in data mining is advised to read this section to gain the full benefit from §\[Sec: Uses\]. As in any specialized subject, a certain level of jargon is necessary for clarity of expression. Terms likely to be unfamiliar to astronomers not versed in data mining are generally explained as they are introduced, but for additional background we note that there are other useful reviews of the data mining field[@witten:datamining2nd; @bishop:pattern; @hastie:learning2nd]. Another recent overview of data mining in astronomy by Borne has also been published[@borne:datamining].
Data Collection {#Subsec: Data}
---------------
The process of data collection encompasses all of the steps required to obtain the desired data in a digital format. Methods of data collection include acquiring and archiving new observations, querying existing databases according to the science problem at hand, and performing as necessary any cross-matching or data combining, a process generically described as [*data fusion*]{}.
A common motivation for cross-matching is the use of multiwavelength data, i.e., data spanning more than one of the regions of the electromagnetic spectrum (gamma ray, X-ray, ultraviolet, optical, infrared, millimeter, and radio). A common method in the absence of a definitive identification for each object spanning the datasets is to use the object’s position on the sky with some astrometric tolerance, typically a few arcseconds. Cross-matching can introduce many issues including ambiguous matches, variations of the point spread function (resolution of objects) within or between datasets, differing survey footprints, survey masks, and large amounts of processing time and data transfer requirements when cross-matching large datasets.
A major objective of the [*Virtual Observatory*]{} (VO, §\[Subsec: VO\]) is to make the data collection process more simple and tractable. Future VO webservices are planned that will perform several functions in this area, including cross-matches on large, widely distributed, heterogeneous data.
Common astronomical data formats include FITS[@wells:fits], a binary format, and plain ASCII, while an emerging format is VOTable[@ochsenbein:votable]. Commonly used formats from other areas of data mining, such as attribute relation file format (ARFF)[^1], are generally not widely used in astronomy.
Preprocessing of Data {#Subsec: Preprocessing}
---------------------
Some data preprocessing may necessarily be part of the data collection process, for example, sample cuts in database queries. Preprocessing can be divided into steps that make the data to be read meaningful, and those that transform the data in some way as appropriate to a given algorithm. Data preprocessing is often problem-dependent, and should be carefully applied because the results of many data mining algorithms can be significantly affected by the input data. A useful overview of data preprocessing is given by Pyle[@pyle:dataprep].
Algorithms may require the object [*attributes*]{}, i.e., the values in the data fields describing the properties of each object, to be numerical or categorical, the latter being, e.g. ‘star’, or ‘galaxy’. It is possible to transform numerical data to categorical and vice versa.
A common categorical-to-numerical method is scalarization, in which different possible categorical attributes are given different numerical labels, for example, ‘star’, ‘galaxy’, ‘quasar’ labeled as the vectors \[1,0,0\], \[0,1,0\], and \[0,0,1\], respectively. Note that for some algorithms, one should [*not*]{} label categories as, say, 1, 2 and 3, if the output of the algorithm is such that if it has confused an object between 1 and 3 it labels the object as intermediate, in this case, 2. Here, 2 (galaxy) is certainly not an intermediate case between 1 (star) and 3 (quasar). One common algorithm in which such categorical but not ordered outputs could occur is a decision tree with multiple outputs.
Numerical data can be made categorical by transformations such as binning. The bins may be user-specified, or can be generated optimally from the data[@hogg:histogram]. Binning can create numerical issues, including comparing two floating point numbers that should be identical, objects on a bin edge, empty bins, values that cannot be binned such as [*NaN*]{}, or values not within the bin range.
Object attributes may need to be [*transformed*]{}. A common operation is the differencing of magnitudes to create colors. These transformations can introduce their own numerical issues, such as division by zero, or loss of accuracy.
In general, data will contain one or more types of [*bad values*]{}, where the value is not correct. Examples include instances where the value has been set to something such as -9999 or NaN, the value appears correct but has been flagged as bad, or the value is not bad in a formatting sense but is clearly unphysical, perhaps a magnitude of a high value that could not have been detected by the instrument. They may need to be removed either by simply removing the object containing them, ignoring the bad value but using the remaining data, or interpolating a value using other information. Outliers may or may not be excluded, or may be excluded depending on their extremity.
Data may also contain [*missing values*]{}. These values may be genuinely missing, for example in a cross-matched dataset where an object is not detected in a given waveband, or is not in an overlapping region of sky. It is also possible that the data should be present, but are missing for either a known reason, such as a bad camera pixel, a cosmic ray hit, or a reason that is simply not known. Some algorithms cannot be given missing values, which will require either the removal of the object or interpolation of the value from the existing data. The advisability of interpolation is problem-dependent.
As well as bad values, the data may contain values that are correct but are outside the desired range of analysis. The data may therefore need to be [*sampled*]{}. There may simply be a desired range, such as magnitude or position on the sky, or the data may contain values that are correct but are outliers. Outliers may be included, included depending on their extremity (e.g., $n$ standard deviations), downweighted, or excluded. Alternatively, it may be more appropriate to generate a random subsample to produce a smaller dataset.
Outside any normalization of the data prior to its use in the data mining algorithm, for example, calibration using standard sources, input or target attributes of the data will often be further normalized to improve the [*numerical conditioning*]{} of the algorithm. For example, if one axis of the $n$-dimensional space created by $n$ input attributes encompasses a range that, numerically, is much larger than the other axes, it may dominate the results, or create conditions where very large and small numbers interact, causing loss of accuracy. Normalization can reduce this, and examples include linear transformations, like scaling by a given amount, scaling using the minimum and maximum values so that each attribute is in a given range such as 0–1, or scaling each attribute to have a mean of 0 and a standard deviation of 1. The latter example is known as [*standardization*]{}. A more sophisticated transformation with similar advantages is [*whitening*]{}, in which the values are not only scaled to a similar range, but correlations among the attributes are removed via transformation of their covariance matrix to the identity matrix.
Attribute Selection {#Subsec: Attributes}
-------------------
In general, a large number of attributes will be available for each object in a dataset, and not all will be required for the problem. Indeed, use of all attributes may in many cases worsen performance. This is a well-known problem, often called the [*curse of dimensionality*]{}. The large number of attributes results in a high-dimensional space with many low density environments or even empty voids. This makes it difficult to generalize from the data and produce useful new results. One therefore requires some form of [*dimension reduction*]{}, in which one wishes to retain as much of the information as possible, but in fewer attributes. As well as the curse of dimensionality, some algorithms work less well with noisy, irrelevant, or redundant attributes. An example of an irrelevant attribute might be position on the sky for a survey with a uniform mask, because the position would then contain no information, and highly redundant attributes might be a color in the same waveband measured in two apertures.
The most trivial form of dimension reduction is simply to use one’s judgement and select a subset of attributes. Depending on the problem this can work well. Nevertheless, one can usually take a more sophisticated and less subjective approach, such as principal component analysis (PCA)[@karhunen:kl; @loeve:kl; @jolliffe:pca]. This is straightforward to implement, but is limited to linear relations. It gives, as the principal components, the eigenvectors of the input data, i.e., it picks out the directions which contain the greatest amount of information. Another straightforward approach is [*forward selection*]{}, in which one starts with one attribute and selectively adds new attributes to gain the most information. Or, one can perform the equivalent process but starting with all of the attributes and removing them, known as [*backward elimination*]{}.
In many ways, dimension reduction is similar to classification, in the sense that a larger number of input attributes is reduced to a smaller number of outputs. Many classification schemes in fact directly use PCA. Other dimension reduction methods utilize the same or similar algorithms to those used for the actual data mining: an ANN can perform PCA when set up as an autoencoder, and kernel methods can act as generalizations of PCA. A binary genetic algorithm (§\[Subsubsec: Other\]) can be used in which each individual represents a subset of the training attributes to be used, and the algorithm selects the best subset. The epsilon-approximate nearest neighbor search[@arya:epsilon] reduces the dimensionality of nearest neighbor methods. Other methods include information bottleneck[@tishby:ib], which directly uses information theory to optimize the tradeoff between the number of classes and the information contained, Fisher Matrix[@fisher:matrix], Independent Component Analysis[@hyvarinen:ica], and wavelet transforms. The curse of dimensionality is likely to worsen in the future for a similar reason to that of missing values, as more multiwavelength datasets become available to be cross-matched. Classification and dimension reduction are not identical of course: a classification algorithm may build a model to represent the data, which is then applied to further examples to predict their classes.
Selection and Use of Machine Learning Algorithms {#Subsec: Algorithm}
------------------------------------------------
Machine learning algorithms broadly divide into [*supervised*]{} and [*unsupervised*]{} methods, also known as predictive and descriptive, respectively. These can be generalized to form [*semi-supervised*]{} methods. Supervised methods rely on a [*training set*]{}[^2] of objects for which the target property, for example a classification, is known with confidence. The method is trained on this set of objects, and the resulting mapping is applied to further objects for which the target property is not available. These additional objects constitute the [*testing set*]{}. Typically in astronomy, the target property is spectroscopic, and the input attributes are photometric, thus one can predict properties that would normally require a spectrum for the generally much larger sample of photometric objects. The training set must be representative, i.e., the parameter space covered by the input attributes must span that for which the algorithm is to be used. This might initially seem rather restrictive, but in many cases can be handled by combining datasets. For example, the zCOSMOS redshift survey[@lilly:zcosmos], at one square degree, provides spectra to the depth of the photometric portion of the Sloan Digital Sky Survey (SDSS)[@york:sdss], $r \sim 22$ mag, which covers over 8000 square degrees. Since SDSS photometry is available for zCOSMOS objects, one can in principle use the 40,000 zCOSMOS galaxies as a training set to assign photometric redshifts to over 200 million SDSS galaxies.
In contrast to supervised methods, unsupervised methods do not require a training set. This is an advantage in the sense that the data can speak for themselves without preconceptions such as expected classes being imposed. On the other hand, if there is prior information, it is not necessarily incorporated. Unsupervised algorithms usually require some kind of initial input to one or more of the adjustable parameters, and the solution obtained can depend on this input.
Semi-supervised methods attempt to allow the best-of-both-worlds, and both incorporate known priors while allowing objective data interpretation and extrapolation. But given their generality, they can be more complex and difficult to implement. They are of potentially great interest astronomically because they could be used to analyze a full photometric survey beyond the spectroscopic limit, without requiring priors, while at the same time incorporating the prior spectroscopic information where it is available.
### Supervised Methods {#Subsubsec: Supervised}
The most widely used and well-known machine learning algorithm in astronomy to-date, referred to as far back as the mid 1980s,[@jeffrey:annealing] is the [*artificial neural network*]{} (ANN, Fig. \[Fig: ANN\])[@bishop:ann; @ripley:ann; @duda:pattern]. This consists of a series of interconnected nodes with weighted connections. Each node has an activation function, perhaps a simple threshold, or a sigmoid. Although the original motivation was that the nodes would simulate neurons in the brain,[@mcculloch:ann; @hopfield:ann] the ANNs in data mining are of such a size that they are best described as nonlinear extensions of conventional statistical methods.
The supervised ANN takes parameters as input and maps them on to one or more outputs. A set of parameter vectors, each vector representing an object and corresponding to a desired output, or target, is presented. Once the network is trained, it can be used to assign an output to an unseen parameter vector. The training uses an algorithm to minimize a cost function. The cost function, $c$, is commonly of the form of the mean-squared deviation between the actual and desired output: $$c = \frac{1}{N} \sum_{k=1}^N~(o_{k} - t_{k})^{2},$$ where $o_{k}$ and $t_{k}$ are the output and target respectively for the $k$th of $N$ objects.
In general, the neurons could be connected in any topology, but a commonly used form is to have an $a: b_1 : b_2 : \ldots : b_n :c$ arrangement, where $a$ is the number of input parameters, $b_{1,\ldots,n}$ are the number of neurons in each of $n$ one dimensional ‘hidden’ layers, and $c$ is the number of neurons in the final layer, which is equal to the number of outputs. Each neuron is connected to every neuron in adjacent layers, but not to any others. Multiple outputs can each give the Bayesian [*a posteriori*]{} probability that the output is of that specific class, given the values of the input parameters.
The weights are adjusted by the training algorithm. In astronomy this has typically been either the well-known backpropagation algorithm[@werbos:backprop; @parker:backprop; @rumelhart:backprop] or the quasi-Newton algorithm[@bishop:ann], although other algorithms, such as Levenberg-Marquardt[@levenberg:levmarq; @marquardt:levmarq] have also been used.
Another common method used in data mining is the [*decision tree*]{} (DT, Fig. \[Fig: DT\])[@morgan:dt; @breiman:dt; @quinlan:dt; @quinlan:dtbook; @rokach:dt]. Decision trees consist first of a root node which contains all of the parameters describing the objects in the training set population along with their classifications. A node is split into child nodes using the criterion that minimizes the classification error. This splitting subdivides the parent population group into children population groups, which are assigned to the respective child nodes. The classification error quantifies the accuracy of the classification on the test set. The process is repeated iteratively, resulting in layered nodes that form a tree. The iteration stops when specific user-determined criteria are reached. Possibilities include a minimum allowed population of objects in a node (the minimum decomposition population), the maximum number of nodes between the termination node and the root node (the maximum tree depth), or a required minimum decrease resulting from a population split (the minimum error reduction). The terminal nodes are known as the leaf nodes. The split is tested for each input attribute, and can be axis-parallel, or oblique, which allows for hyperplanes at arbitrary angles in the parameter space. The split statistic can be the midpoint, mean, or median of the attribute values, while the cost function used is typically the variance, as with ANN.
In recent years, another algorithm, the [*support vector machine*]{} (SVM, Fig. \[Fig: SVM\])[@cortes:svm; @burges:svm; @vapnik:svm; @cristianini:svm; @kecman:svm; @schlkopf:svm; @abe:svm; @wang:svm; @steinwart:svm], has gained popularity in astronomical data mining. SVM aims to find the hyperplane that best separates two classes of data. The input data are viewed as sets of vectors, and the data points closest to the classification boundary are the support vectors. The algorithm does not create a model of the data, but instead creates the decision boundaries, which are defined in terms of the support vectors. The input attributes are mapped into a higher dimensional space using a kernel so that nonlinear relationships within the data become linear (the ‘kernel trick’)[@aizerman:kerneltrick], and the decision boundaries, which are linear, are determined in this space. Like ANN and DT, the training algorithm minimizes a cost function, which in this case is the number of incorrect classifications. The algorithm has two adjustable hyperparameters: the width of the kernel, and the regularization, or cost, of classification error, which helps to prevent [*overfitting*]{} (§\[Subsec: Improving\]) of the training set. The shape of the kernel is also an adjustable parameter, a common choice being the Gaussian radial basis function. As a result, an SVM has fewer adjustable parameters than an ANN or DT, but because these parameters must be optimized, the training process can still be computationally expensive. SVM is designed to classify objects into two classes. Various refinements exist to support additional classes, and to perform regression, i.e., to supply a continuous output value instead of a classification. Classification probabilities can be output, for example, by using the distance of a data point from the decision boundary.
Another powerful but computationally intensive method is [*k nearest neighbor*]{} ([$k$NN]{})[@fix:knn; @cover:nn; @aha:ib; @dasarathy:knn; @shakhnarovich:knn]. This method is powerful because it can utilize the full information available for each object, with no approximations or interpolations. The training of [$k$NN]{} is in fact trivial: the positions of each of the objects in the input attribute space are simply stored in memory. For each test object, the same attributes are compared to the training set and the output is determined using the properties of the nearest neighbors. The simplest implementation is to output the properties of the single nearest neighbor, but more commonly the weighted sum of $k$ nearest neighbors is used. The weighting is typically the inverse Euclidean distance in the attribute space, but one can also use adaptive distance metrics. The main drawback of this method is that is it computationally intensive, because for each testing object the entire training set must be examined to determine the nearest neighbors. This requires a large number of distance calculations, since the test datasets are often much larger than the training datasets. The workload can be mitigated by storing the training set in an optimized data structure, such as a kd-tree.
However, in the past few years, novel supercomputing hardware (which is discussed in more detail in §\[Subsec: Hardware\]) has become available that is specifically designed to carry out exactly this kind of computationally intensive work, including applications involving a large number of distance calculations. The curve of growth of this technology exceeds that of conventional CPUs, and thus the direct implementation of [$k$NN]{} using this technology has the potential to exceed the performance of conventional CPUs.
### Unsupervised Methods {#Subsubsec: Unsupervised}
[*Kernel density estimation*]{} (KDE)[@parzen:kde; @duda:kde; @silverman:kde; @scott:kde; @taylor:kde; @wasserman:kde; @klemela:kde] is a method of estimating the probability density function of a variable. It is a generalization of a histogram where the kernel function is any shape instead of the top-hat function of a histogram bin. This has the advantages that it avoids the discrete nature of the histogram and does not depend on the position of the bin edges, but the width of the kernel must still be chosen so as not to over- or under-smooth the data. A Gaussian kernel is commonly utilized. In higher numbers of dimensions, common in astronomical datasets, the width of the kernel must be specified in each dimension.
[*K-means clustering*]{}[@steinhaus:kmeans; @macqueen:kmeans] is an unsupervised method that divides data into clusters. The number of clusters must be initially specified, but since the algorithm converges rapidly, many starting points can be tested. The algorithm uses a distance criterion for cluster membership, such as the Euclidean distance, and a stopping criterion for iteration, for example, when the cluster membership ceases to change.
[*Mixture models*]{}[@titterington:mixture; @mclachlan:mixture] decompose a distribution into a sum of components, each of which is a probability density function. Often, the distributions are Gaussians, resulting in Gaussian mixture models. They are often used for clustering, but also for density estimation, and they can be optimized using either expectation maximization or Monte Carlo methods. Many astronomical datasets consist of contributions from different populations of objects, which allows mixture modeling to disentangle these population groups. Mixture models based on the EM algorithm have been used in astronomy for this purpose[@connolly:fast; @dolence:lci].
[*Expectation maximization*]{} (EM)[@dempster:em; @watanabe:em; @mclachlan:em] treats the data as a sum of probability distributions, which each represent one cluster. This method alternates between an expectation stage and a maximization stage. In the expectation stage, the algorithm evaluates the membership probability of each data point given the current distribution parameters. In the maximization stage, these probabilities are used to update the parameters. This method works well with missing data, and can be used as the unsupervised component in semi-supervised learning (§\[Subsubsec: Semisupervised\]) to provide class labels for the supervised learning.
The [*Kohonen self-organizing map*]{} (SOM)[@kohonen:somprev; @kohonen:som] is an unsupervised neural network that forms a general framework for visualizing datasets of more than two dimensions. Unlike many methods which seek to map objects onto a new output space, the SOM is fundamentally topological. This is neatly illustrated by the fact that one astronomical SOM application[@naim:som] is titled ‘Galaxy Morphology Without Classification’. A related earlier method is learning vector quantization[@kohonen:lvq].
[*Independent component analysis*]{} (ICA)[@comon:ica; @lee:ica; @hyvarinen:ica; @roberts:ica; @stone:ica], an example of [*blind source separation*]{}, can separate nonlinear components of a dataset, under the assumption that those components are statistically independent. The components are found by maximizing this independence. Related statistical methods include principal component analysis (§\[Subsec: Attributes\]), singular value decomposition, and non-negative matrix factorization.
### Semi-Supervised {#Subsubsec: Semisupervised}
The semi-supervised approach[@chapelle:semisupervised; @zhu:semisupervised] has been somewhat underused to-date, but holds great potential for the upcoming, large, purely photometric surveys. Supervised methods require a labeled training set, but will not assign new classes. On the other hand, unsupervised methods do not require training, but do not use existing known information. Semi-supervised methods aim to capture the best from both of these methods by retaining the ability to discover new classes within the data, and also incorporating information from a training set when available. An example of a dataset amenable to the approach is shown in Fig. \[Fig: Semi-supervised\].
This is particularly relevant in astronomical applications using large amounts of photometric and a more limited subsample of spectroscopic data, which may not be fully representative of the photometric sample. The semi-supervised approach allows one to use the spectral information to extrapolate into the purely photometric regime, thereby allowing a scientist to utilize all of the vast amount of information present there.
Semi-supervised learning represents an entire subfield of data mining research. Given the nontrivial implementation requirements, this field is a good area for potential fruitful collaborations between astronomers, computer scientists, and statisticians. As one example of a possible issue, a lot of photometric data are likely to be a direct continuation in parameter space of spectroscopic data, with, therefore, a highly overlapping distribution. This means that certain semi-supervised approaches will work better than others, because they contain various assumptions about the nature of the labeled and unlabeled data.
### Other Algorithms {#Subsubsec: Other}
In §§\[Subsubsec: Supervised\]–\[Subsubsec: Unsupervised\] above, we described the main data mining algorithms used to date in astronomy, however, there are numerous additional algorithms available, which have often been utilized to some extent. These algorithms may be employed at more than one stage in the process, such as attribute selection, as well as the classification/regression stage.
While neural networks in some very broad sense mimic the learning mechanism of the brain, [*genetic algorithms*]{}[@holland:genetic; @goldberg:genetic; @coley:ga; @mitchell:ga; @haupt:genetic2nd; @sivanandam:ga] mimic natural selection, as the most successful individuals created are those that are best adapted for the task at hand. The simplest implementation is the binary genetic algorithm, in which each ‘individual’ is a vector of ones and zeros, which represent whether or not a particular attribute, e.g., a training set attribute, is used. From an initial random population, the individuals are combined to create new individuals. The fitness of each individual is the resulting error in the training algorithm run according to the formula encoded by the individual. This process is repeated until convergence if found, producing the best individual.
A typical method of combining two individuals is one-point crossover, in which segments of two individuals are swapped. To more fully explore the parameter space, and to prevent the algorithm from converging too rapidly on a local minimum, a probability of mutation is introduced into the newly created individuals before they are processed. This is simply the probability that a zero becomes a one, or vice-versa. An approximate number of individuals to use is given by $n_{{\mathrm{in}}} \sim 2n_f~{\mathrm{log}}(n_f),$ where $n_f$ is the number of attributes. The algorithm converges in $n_{{\mathrm{it}}} \sim \alpha n_f~{\mathrm{log}}(n_f)$ iterations, where $\alpha$ is a problem-dependent constant; generally $\alpha > 3$.
Numerous refinements to this basic approach exist, including using continuous values instead of binary ones, and more complex methods for combining individuals. Further possibilities for the design of genetic algorithms exist[@goldberg:design], and it is possible in principle to combine the optimization of the learning algorithm and the attribute set.
The [*Information bottleneck*]{} method[@tishby:ib] is based directly on information theory and is designed to achieve the best tradeoff between accuracy and compression for the desired number of classes. The inputs and outputs are probability density functions. [*Association rule*]{} mining[@adamo:associationrules; @zhang:associationrules] is a method of finding qualitative rules within a database such that a rule derived from the occurrence of certain variables together implies something about the occurrence of a variable not used in creating that rule. The [*false discovery rate*]{}[@benjamini:fdr] is a method of establishing a significant discovery from a smaller set of data than the usual $n$ sigma hypothesis test.
This list could continue, broadening into traditional statistical methods such as least squares, and regression, as well as Bayesian methods, which are widely used in astronomy. For brevity we do not consider these additional methods, but we do note that [*graphical models*]{}[@bishop:pattern] are a general way of describing the interrelationships between variables and probabilities, and many of the data mining algorithms described earlier, such as ANNs, are special cases of these models.
### Choice of Algorithm {#Subsubsec: Choice}
Unfortunately, there is no simple method to select the optimal algorithm to use, because the most appropriate algorithm can depend not only on the dataset, but also the application for which it will be employed. There is, therefore, no single best algorithm. Likewise, the choice of software is similarly non-trivial. Many general frameworks exist, for example WEKA[@witten:datamining2nd] or Data to Knowledge[@welge:d2k], but it is unlikely that one framework will be able to perform all steps necessary from raw catalog to desired science result, particularly for large datasets. In Table \[Table: Algorithms\], we summarize some of the advantages and disadvantages of some of the more popular and well-known algorithms used in astronomy. We do not attempt to summarize available software. Various other general comparisons of machine learning algorithms exist[@hastie:learning2nd], as well as numerous studies comparing various algorithms for particular datasets, a field which itself is rather complex[@salzberg:classifiercritique].
Improving Results {#Subsec: Improving}
-----------------
Many of the algorithms previously described involve ‘greedy’ optimization. In these cases, the cost function, which is the measure of how well the algorithm is performing in its classification or prediction task, is minimized in a way that does not allow the value of the function to increase much if at all. As a result, it is possible for the optimization to become trapped in a local minimum, whereby nearby configurations are worse, but better configurations exist in a different region of parameter space. Various approaches exist to overcome local minima. One approach is to simply run the algorithm several times from different starting points. Another approach is [*simulated annealing*]{}[@kirkpatrick:annealing; @cerny:annealing; @vanlaarhoven:annealing; @aarts:annealing], where, in following the metallurgical metaphor, the point in parameter space ‘heats up’, thus perturbing it and allowing it to escape from the local minimum. The point is allowed to ‘cool’, thus having the ability to find a solution closer to the global minimum.
Models produced by data mining algorithms are subject to a fundamental limitation common to many systems in which a predictive model is constructed, the [*bias-variance tradeoff*]{}. The bias is the accuracy of the model in describing the data, for example, a linear model might have a higher bias than a higher order polynomial. The variance is the accuracy of this model in describing new data. The higher order polynomial might have a lower bias than a linear model, but it might be more strongly affected by variations in the data and thus have a higher variance. The polynomial has [*overfit*]{} the data. There is usually an optimal point between minimizing bias and minimizing variance. A typical way to minimize overfitting is to measure the performance of the algorithm on a test set, which is not part of the training set, and adjusting the stopping criterion for training to stop at an appropriate location.
To help prevent overfitting, training can also be [*regularized*]{}, in which an extra term is introduced into the cost function to penalize configurations that add complexity, such as large weights in an ANN. This complexity can cause a function to be less smooth, which increases the likelihood of overfitting. As is the case with supervised learning, unsupervised algorithms can also overfit the data, for example, if some kind of smoothing is employed but its scalelength is too small. In this case, the algorithm will ‘fit the noise’ and not the true underlying distribution.
Another common approach to control overfitting and improve confidence in the accuracy of the results is [*cross-validation*]{}, where subsets of the data are left out of the training and used for testing. The simplest form is the holdout method, where a single subset of the training data is kept out of the training, and the algorithm error is evaluated by running on this subset. However, this can have a high bias (see bias-variance tradeoff, above) if the training set is small, due to a significant portion of the training information being left out. $K$-fold cross-validation improves on this by subdividing the data into $K$ samples and training on $K-1$ samples, or alternatively using $K$ random subsets. Typically, $K=5$ or $K=10$, as small $K$ could still have high bias, as in the holdout method, but large $K$, while being less biased, can have high variance due to the testing set being small. If $K$ is increased to the size of the dataset, so that each subsample is a single point, the method becomes leave-one-out cross-validation. In all instances, the estimated error is the mean error from those produced by each run in the cross-validation.
Another important refinement to running one algorithm is the ability to use a [*committee*]{} of instances of the algorithm, each with different parameters. One can allow these different instantiations to vote on the final prediction, so that the majority or averaged result becomes the final answer. Such an arrangement can often provide a substantial improvement, because it is more likely that the majority will be closer to the correct answer, and that the answer will be less affected by outliers. One such committee arrangement is [*bootstrap aggregating*]{}, or [*bagging*]{}[@breiman:bagging; @witten:datamining2nd], where random subsamples with replacement (bootstrap samples) are taken, and the algorithm trained on each. The created algorithms vote on the testing set. Bagging is often applied to decision trees with considerable success, but it can be applied to other algorithms. The combination of bagging and the random selection of a small subset of features for splitting at each node is known as a Random Forest[@breiman:randomforest].
[*Boosting*]{}[@hastie:learning2nd] uses the fact that several ‘weak’ instances of an algorithm can be combined to produce a stronger instance. The weak learners are iteratively added and misclassified objects in the data gain higher weight. Thus boosting is not the same as bagging because the data themselves are weighted. Boosted decision trees are a popular approach, and many different boosting algorithms are available.
As well as committees of the same algorithm, it is also possible to combine the results of more than one different algorithm on the same dataset. Such a [*mixture of experts*]{} approach often provides an optimal result on real data. The outcome may be decided by voting, or the output of one algorithm can form the input to another, in a chaining approach.
For many astronomical applications, the results are, or would be, significantly improved by utilizing the full probability density function (PDF) for a predicted property, rather than simply its single scalar value. This is because much more information is retained when using the PDF. Potential uses of PDFs are described further in §\[Subsec: PDF\].
Application of Algorithms and Some Limitations {#Subsec: Application}
----------------------------------------------
The purpose of this review is not to uncritically champion certain data mining algorithms, but to instead encourage scientific progress by exploiting the full potential of these algorithms in a considered scientific approach. We therefore end this section by outlining some of the limitations of and issues raised by KDD and the data mining approach to current and future astronomical datasets. Several of these problems might be ameliorated by increased collaboration between astronomers and data mining experts.
[*Extrapolation*]{}: In many astronomical applications, it is common for data with less information content to be available for a greater number of objects over a larger parameter space. The classic example is in surveys where photometric objects are typically observed several magnitudes fainter than spectroscopic objects. For a supervised learning algorithm, it is usually inappropriate to extrapolate beyond the parameter space for which the training set (e.g., the spectroscopic objects) is representative.
[*Non-intuitive results*]{}: It is very easy to run an implementation of a well-known algorithm and output a result that appears reasonable, but is in fact either statistically invalid or completely wrong. For example, randomly subsampled training and testing sets from a dataset will overlap and produce a model that overfits the data.
[*Measurement error*]{}: Most astronomical data measurements have an associated error, but most data mining algorithms do not take this explicitly into account. For many algorithms, the intrinsic spread in the data corresponding to the target property is the measurement of the error.
[*Adjustable parameters*]{}: Several algorithms have a significant number of adjustable parameters, and the optimal configuration of these parameters is not obvious. This can result in large parameter sweeps that further increase the computational requirement.
[*Scalability*]{}: Many data mining algorithms scale, for $n$ objects, as $n^2$, or even worse, making their simple application to large datasets on normal computing hardware intractable. One can often speed up a naïve implementation of an algorithm that must access large numbers of data points and their attributes by storing the data in a hierarchical manner so that not all the data need to be searched. A popular hierarchical structure for accomplishing this task is the kd-tree[@bentley:kdtree]. However, the implementation of such trees for large datasets and on parallel machines remains a difficult problem[@gardner:paralleltree].
[*Learning Curve*]{}: Data mining is an entire field of study in its own right, with strong connections to statistics and computing. The avoidance of some of the issues we present, such as the selection of appropriate algorithms, collaboration where needed, and the full exploitation of their potential for science return, require overcoming a substantial learning curve.
[*Large datasets*]{}: Many astronomical datasets are larger than can be held in machine memory. The exploitation of these datasets thus requires more sophisticated database technology than is currently employed by most astronomical projects.
[*“It’s not science”*]{}: The success of an astronomical project is judged by the science results produced. The time invested by an astronomer in becoming an expert in data mining techniques must be balanced against the expected science gain. It is difficult to justify and obtain funding based purely on a methodological approach such as data mining, even if such an approach will demonstrably improve the expected science return.
[*It does not do the science for you*]{}: The algorithms will output patterns, but will not necessarily establish which patterns or relationships are important scientifically, or even which are causal. The truism ‘correlation is not causation’ is apt here. The successful interpretation of data mining results is up to the scientist.
[*The result can only be as good as the data*]{}: Related to this, the single largest factor in the success of any data mining algorithm is the quality of the input data. If the data are not sufficient for the task, or are poorly collected or incorrectly treated, the result will not be useful.
Uses in Astronomy {#Sec: Uses}
=================
We now turn to the use of data mining algorithms in astronomical applications, and their track record in addressing some common problems. Whereas in §\[Sec: Overview\], we introduced terms for the astronomer unfamiliar with data mining, here for the non-expert in astronomy we briefly put in context the astronomical problems. However, a full description is beyond the scope of this review. Whereas §\[Sec: Overview\] was subdivided according to data mining algorithms and issues, here the subdivision is in terms of the astrophysics. Throughout this section, we abbreviate data mining algorithms that are either frequently mentioned or have longer names according to the abbreviations introduced in §\[Sec: Overview\]: PCA, ANN, DT, SVM, [$k$NN]{}, KDE, EM, SOM, and ICA.
Given that there is no exact definition of what constitutes a data mining tool, it would not be possible to provide a complete overview of their application. This section therefore illustrates the wide variety of actual uses to date, with actual or implied further possibilities. Uses which exist now but will likely gain greater significance in the future, such as the time domain, are largely deferred to §\[Sec: Future\]. Several other overviews of applications of machine learning algorithms in astronomy exist, and contain further examples, including ones for ANN[@miller:annapps; @lahav:annmethods; @bailerjones:ann; @li:annapps; @tagliaferri:nnast], DT[@white:dts], genetic algorithms[@charbonneau:ga], and stellar classification[@bailerjones:stellar].
Most of the applications in this section are made by astronomers utilizing data mining algorithms. However, several projects and studies have also been made by data mining experts utilizing astronomical data, because, along with other fields such as high energy physics and medicine, astronomy has produced many large datasets that are amenable to the approach. Examples of such projects include the Sky Image Cataloging and Analysis System (SKICAT)[@weir:skicat] for catalog production and analysis of catalogs from digitized sky surveys, in particular the scans of the second Palomar Observatory Sky Survey; the Jet Propulsion Laboratory Adaptive Recognition Tool (JARTool)[@burl:jartool], used for recognition of volcanoes in the over 30,000 images of Venus returned by the Magellan mission; the subsequent and more general Diamond Eye[@burl:diamondeye]; and the Lawrence Livermore National Laboratory Sapphire project[@kamath:sapphire]. A recent review of data mining from this perspective is given by Kamath in the book [*Scientific Data Mining*]{}[@kamath:scientific]. In general, the data miner is likely to employ more appropriate, modern, and sophisticated algorithms than the domain scientist, but will require collaboration with the domain scientist to acquire knowledge as to which aspects of the problem are the most important.
Object classification {#Subsec: Classification}
---------------------
Classification is often an important initial step in the scientific process, as it provides a method for organizing information in a way that can be used to make hypotheses and to compare with models. Two useful concepts in object classification are the [*completeness*]{} and the [*efficiency*]{}, also known as recall and precision. They are defined in terms of true and false positives (TP and FP) and true and false negatives (TN and FN). The completeness is the fraction of objects that are truly of a given type that are classified as that type: $${\mathrm{completeness} = \frac{TP}{TP+FN}},$$ and the efficiency is the fraction of objects classified as a given type that are truly of that type $${\mathrm{efficiency} = \frac{TP}{TP+FP}}.$$ These two quantities are astrophysically interesting because, while one obviously wants both higher completeness and efficiency, there is generally a tradeoff involved. The importance of each often depends on the application, for example, an investigation of rare objects generally requires high completeness while allowing some contamination (lower efficiency), but statistical clustering of cosmological objects requires high efficiency, even at the expense of completeness.
### Star-Galaxy Separation {#Subsubsec: Star/gal}
Due to their small physical size compared to their distance from us, almost all stars are unresolved in photometric datasets, and thus appear as point sources. Galaxies, however, despite being further away, generally subtend a larger angle, and thus appear as extended sources. However, other astrophysical objects such as quasars and supernovae, also appear as point sources. Thus, the separation of photometric catalogs into stars and galaxies, or more generally, stars, galaxies, and other objects, is an important problem. The sheer number of galaxies and stars in typical surveys (of order $10^8$ or above) requires that such separation be automated.
This problem is a well studied one and automated approaches were employed even before current data mining algorithms became popular, for example, during digitization by the scanning of photographic plates by machines such as the APM[@maddox:apmstargal] and DPOSS[@djorgovski:dposs]. Several data mining algorithms have been employed, including ANN[@odewahn:autostargal; @odewahn:annstargal; @bazell:preprocessing; @andreon:stargal; @philip:dbnn; @odewahn:dpossstargal; @collister:megazlrg], DT[@weir:stargal; @ball:dtclassification], mixture modeling[@qin:stargal], and SOM[@miller:som], with most algorithms achieving over 95% efficiency. Typically, this is done using a set of measured morphological parameters that are derived from the survey photometry, with perhaps colors or other information, such as the seeing, as a prior. The advantage of this data mining approach is that all such information about each object is easily incorporated. As well as the simple outputs ‘star’ or ‘galaxy’, many of the refinements described in §\[Sec: Overview\] have improved results, including probabilistic outputs and bagging[@ball:dtclassification].
### Galaxy Morphology {#Subsubsec: Morphology}
As shown in Fig. \[Fig: Morph\], galaxies come in a range of different sizes and shapes, or more collectively, morphology. The most well-known system for the morphological classification of galaxies is the Hubble Sequence of elliptical, spiral, barred spiral, and irregular, along with various subclasses[@hubble:extragalnebulae; @hubble:realm; @sandage:hubbatlas; @sandage:carnegieatlas; @vandenbergh:morph; @sandage:classfnhistory]. This system correlates to many physical properties known to be important in the formation and evolution of galaxies[@roberts:hubble; @firmani:hubble]. Other well-known classification systems are the Yerkes system based on concentration index[@morgan:ci; @morgan:ci2; @devaucouleurs:ci], the de Vaucouleurs[@devaucouleurs:devprofile], exponential[@patterson:expprofile; @freeman:expprofile], and Sérsic index[@sersic:australes; @graham:sersic] measures of the galaxy light profile, the David Dunlap Observatory (DDO) system[@vandenbergh:lumclass; @vandenbergh:lumclassb; @vandenbergh:ddo], and the concentration-asymmetry-clumpiness (CAS) system[@conselice:cas].
Because galaxy morphology is a complex phenomenon that correlates to the underlying physics, but is not unique to any one given process, the Hubble sequence has endured, despite it being rather subjective and based on visible-light morphology originally derived from blue-biased photographic plates. The Hubble sequence has been extended in various ways, and for data mining purposes the T system[@devaucouleurs:tsystem; @devaucouleurs:ttype] has been extensively used. This system maps the categorical Hubble types E, S0, Sa, Sb, Sc, Sd, and Irr onto the numerical values -5 to 10.
One can, therefore, train a supervised algorithm to assign T types to images for which measured parameters are available. Such parameters can be purely morphological, or include other information such as color. A series of papers by Lahav and collaborators[@storrielombardi:ann; @lahav:annscience; @naim:eyemorph; @naim:annmorph; @lahav:annmethods; @collister:annz] do exactly this, by applying ANNs to predict the T type of galaxies at low redshift, and finding equal accuracy to human experts. ANNs have also been applied to higher redshift data to distinguish between normal and peculiar galaxies[@naim:peculiarmorph], and the fundamentally topological and unsupervised SOM ANN has been used to classify galaxies from Hubble Space Telescope images[@naim:som], where the initial distribution of classes is not known. Likewise, ANNs have been used to obtain morphological types from galaxy spectra.[@madgwick:morphspec]
Several authors study galaxy morphology at higher redshift by using the Hubble Deep Fields, where the galaxies are generally much more distant, fainter, less evolved, and morphologically peculiar. Three studies[@odewahn:hdfannmorph; @windhorst:hizgals; @cohen:hstmorph] use ANNs trained on surface brightness and light profiles to classify galaxies as E/S0, Sabc and Sd/Irr. Another application[@odewahn:fouriermorph] uses Fourier decomposition on galaxy images followed by ANNs to detect bars and assign T types.
Bazell & Aha[@bazell:ensembles] uses ensembles of classifiers, including ANN and DT, to reduce the classification error, and Bazell[@bazell:features] studies the importance of various measured input attributes, finding that no single measured parameter fully reproduces the classifications. Ball [[*et al.*]{}]{}[@ball:ann] obtain similar results to Naim [[*et al.*]{}]{}[@naim:annmorph], but updated for the SDSS. Ball [[*et al.*]{}]{}[@ball:bivlf] and Ball, Loveday & Brunner[@ball:envt] utilize these classifications in studies of the bivariate luminosity function and the morphology-density relation in the SDSS, the first such studies to utilize both a digital sky survey of this size and detailed Hubble types.
Because of the complex nature of galaxy morphology and the plethora of available approaches, a large number of further studies exist: Kelly & McKay[@kelly:shapelet2] (Fig. \[Fig: Mixture\]) demonstrate improvement over a simple split in $u-r$ using mixture models, within a schema that incorporates morphology. Serra-Ricart [[*et al.*]{}]{}[@serraricart:annastroapps] use an encoder ANN to reduce the dimensionality of various datasets and perform several applications, including morphology. Adams & Woolley[@adams:ann] use a committee of ANNs in a ‘waterfall’ arrangement, in which the output from one ANN formed the input to another which produces more detailed classes, improving their results. Molinari & Smareglia[@molinari:annesolf] use an SOM to identify E/S0 galaxies in clusters and measure their luminosity function. de Theije & Katgert[@detheije:typekinematics] split E/S0 and spiral galaxies using spectral principal components and study their kinematics in clusters. Genetic algorithms have been employed[@cantupaz:evolving; @kamath:bentdouble] for attribute selection and to evolve ANNs to classify ‘bent-double’ galaxies in the FIRST[@becker:first] radio survey data. Radio morphology combines the compact nucleus of the radio galaxy and extremely long jets. Thus, the bent-double morphology indicates the presence of a galaxy cluster. de la Calleja & Fuentes[@delacalleja:ann] combine ensembles of ANN and locally weighted regression. Beyond ANN, Spiekermann[@spiekermann:automorph] uses fuzzy algebra and heuristic methods, anticipating the importance of probabilistic studies (§\[Subsec: PDF\]) that are just now beginning to emerge. Owens, Griffiths & Ratnatunga[@owens:dt] use oblique DTs, obtaining similar results to ANN. Zhang, Li & Zhao[@zhang:morph] distinguish early and late types using k-means clustering. SVMs have recently been employed on the COSMOS survey by Huertas-Company [[*et al.*]{}]{}[@huertascompany:svmmorph; @huertascompany:svmmorph2], enabling early-late separation to $K_{AB} = 22$ mag twice as good as the CAS system. SVMs will also be used on data from the Gaia satellite[@tsalmantza:gaiaclassification].
Recently, the popular [*Galaxy Zoo*]{} project[@lintott:galaxyzoo] has taken an alternative approach to morphological classification, employing [*crowdsourcing*]{}: an application was made available online in which members of the general public were able to view images from the SDSS and assign classifications according to an outlined scheme. The project was very successful, and in a period of six months over 100,000 people provided over 40 million classifications for a sample of 893,212 galaxies, mostly to a limiting depth of $r = 17.77$ mag. The classifications included categories not previously assigned in astronomical data mining studies, such as edge-on or the handedness of spiral arms, and the project has produced multiple scientific results. The approach represents a complementary one to automated algorithms, because, although humans can see things an algorithm will miss and will be subject to different systematic errors, the runtime is hugely longer: a trained ANN will produce the same 40 million classifications in a few minutes, rather than six months.
### Other Galaxy Classifications {#Subsubsec: Other Gal}
Many of the physical properties, and thus classification, of a galaxy are determined by its stellar population. The spectrum of a galaxy is therefore another method for classification[@humason:100redshifts; @morgan:specclass], and can sometimes produce a clearer link to the underlying physics than the morphology. Spectral classification is important because it is possible for a range of morphological types to have the same spectral type, and vice versa, because spectral types are driven by different underlying physical processes.
Numerous studies[@connolly:orthogonal; @connolly:eclass; @madgwick:parametrisation; @yip:spectypes] have used PCA directly for spectral classification. PCA is also often used as a preprocessing step before the classification of spectral types using an ANN[@storrielombardi:annspec]. Folkes, Lahav & Maddox[@folkes:annspec] predict morphological types for the 2dF Galaxy Redshift Survey (2dFGRS)[@colless:2dffinal] using spectra, and Ball [[*et al.*]{}]{}[@ball:ann] directly predict spectral types in the SDSS using an ANN. Slonim [[*et al.*]{}]{}[@slonim:ib] use the information bottleneck approach on the 2dFGRS spectra, which maximally preserves the spectral information for the desired number of classes. Lu [[*et al.*]{}]{}[@lu:ica] use ensemble learning for ICA on components of galaxy spectra. Abdalla [[*et al.*]{}]{}[@abdalla:emission] use ANN and locally weighted regression to directly predict emission line properties from photometry.
Bazell & Miller[@bazell:classdisc] applied a semi-supervised method suitable for class discovery using ANNs to the ESO-LV[@lauberts:esouppsala] and SDSS Early Data Release (EDR) catalogs. They found that a reduction of up to 57% in classification error was possible compared to purely supervised ANNs. The larger of the two catalogs, the SDSS EDR, represents a preliminary dataset about 6% of the final data release of the SDSS, clearly indicating the as-yet untapped potential of this approach. The semi-supervised approach also resembles the hybrid empirical-template approach to photometric redshifts (§\[Subsec: Photo-zs\]), as both seek to utilize an existing training set where available even if it does not span the whole parameter space. However, the approach used by Bazell & Miller is more general, because it allows new classes of objects to be added, whereas the hybrid approach can only iterate existing templates.
### Quasars/AGN {#Subsubsec: QSO Classification}
Most of the emitted electromagnetic radiation in the universe is either from stars, or the accretion disks surrounding supermassive black holes in active galactic nuclei (AGN). The latter phenomenon is particularly dramatic in the case of quasars, where the light from the central region can outshine the rest of the galaxy. Because supermassive black holes are thought to be fairly ubiquitous in large galaxies, and their fueling, and thus their intrinsic brightness, can be influenced by the environment surrounding the host galaxy, quasars and other AGN are important for understanding the formation and evolution of structure in the universe.
The selection of quasars and other AGN from an astronomical survey is a well-known and important problem, and one well suited to a data mining approach. It is well-known that different wavebands (X-ray, optical, radio) will select different AGN, and that no one waveband can select them all. Traditionally, AGN are classified on the Baldwin-Phillips-Terlevich diagram[@baldwin:bpt], in which sources are plotted on the two-dimensional space of the emission line ratios \[O\] $\lambda$ 5007 / H$\beta$ and \[N\] / H$\alpha$, that is separated by a single curved line into star-forming and AGN regions. Data mining not only improves on this by allowing a more refined or higher dimensional separation, but also by including passive objects in the same framework (Fig. \[Fig: BPT ANN\]). This allows for the probability that an object contains an AGN to be calculated, and does not require all (or any) of the emission lines to be detected.
Several groups have used ANNs[@carballo:annqso; @claeskens:gaiaqsos; @carballo:annfirstqso] or DTs[@white:firstqso; @suchkov:dt; @ball:dtclassification; @zhang:qso; @zhang:decisiontable; @zhao:dtactive; @knigge:balqso] to select quasar candidates from surveys. White [[*et al.*]{}]{}[@white:firstqso] show that the DT method improves the reliability of the selection to 85% compared to only 60% for simpler criteria. Other algorithms employed include PCA[@yip:qsoclassification], SVM and learning vector quantization[@zhang:methods], kd-tree[@gao:svmkdqso], clustering in the form of principal surfaces and negative entropy clustering[@abrusco:qsocandidates], and kernel density estimation[@richards:dr6photoqso]. Many of these papers combine multiwavelength data, particularly X-ray, optical, and radio.
Similarly, one can select and classify candidates of all types of AGN[@zhao:agn]. If multiwavelength data are available, the characteristic data mining algorithm ability to form a model of the required complexity to extract the information could enable it to use the full information to extract more complete AGN samples. More generally, one can classify both normal and active galaxies in one system, differentiating between star formation and AGN. As one example, DTs have been used[@ball:dtclassification] to select quasar candidates in the SDSS, providing the probabilities P(star, galaxy, quasar). P(star formation, AGN) could be supplied in a similar framework. Bamford [[*et al.*]{}]{}[@bamford:nonparametric] combine mixture modeling and regression to perform non-parametric mixture regression, and is the first study to obtain such components and then study them versus environment. The components are passive, star-forming, and two types of AGN.
### Other Classifications {#Subsubsec: Other Classification}
Often, the first component of classification is the actual process of object detection, which often is done at some signal-to-noise threshold. Several statistical data mining algorithms have been employed, and software packages written, for this purpose, including the Faint Object Classification and Analysis System (FOCAS)[@jarvis:focas], DAOPHOT[@stetson:daophot], Source Extractor (SExtractor)[@bertin:sextractor], maximum likelihood, wavelets, ICA[@maino:fastica], mixture models[@guglielmetti:background], and ANNs[@andreon:stargal]. Serra-Ricart [[*et al.*]{}]{}[@serraricart:faint] show that ANNs are able to classify faint objects as well as a Bayesian classifier but with considerable computational speedup.
Several studies are more general than star-galaxy separation or galaxy classification, and assign classifications of varying detail to a broad range of astrophysical objects. Goebel [[*et al.*]{}]{}[@goebel:bayesian] apply the AutoClass Bayesian classifier to the IRAS LRS atlas, finding new and scientifically interesting object classes. McGlynn [[*et al.*]{}]{}[@mcglynn:classx] use oblique DTs in a system called ClassX to classify X-ray objects into stars, white dwarfs, X-ray binaries, galaxies, AGN, and clusters of galaxies, concluding that the system has the potential to significantly increase the known populations of some rare object types. Suchkov, Hanisch & Margon[@suchkov:dt] use the same system to classify objects in the SDSS. Bazell, Miller & Subbarao[@bazell:subclasses] apply semi-supervised learning to SDSS spectra, including those classified as ‘unknown’, finding two classes of objects consisting of over 50% unknown.
Stellar classifications are necessarily either spectral or based on color, due to the pointlike nature of the source. This field has a long history and well established results such as the HR diagram and the OBAFGKM spectral sequence. The latter is extended to a two-dimensional system of spectral type and luminosity classes I–V to form the two-dimensional MK classification system of Morgan, Keenan & Kellman[@morgan:mk]. Class I are supergiants, through to class V, dwarfs, or main-sequence stars. The spectral types correspond to the hottest and most massive stars, O, through to the coolest and least massive, M, and each class is subdivided into ten subclasses 0–9. Thus, the MK classification of the sun is G2V.
The use of automated algorithms to assign MK classes is analogous to that for assigning Hubble types to galaxies in several ways: before automated algorithms, stellar spectra were compared by eye to standard examples; the MK system is closely correlated to the underlying physics, but is ultimately based on observable quantities; the system works quite well but has been extended in numerous ways to incorporate objects that do not fit the main classes (e.g., L and T dwarfs, Wolf-Rayet stars, carbon stars, white dwarfs, and so on). Two differences from galaxy classification are the number of input parameters, in this case spectral indices, and the number of classes. In MK classification the numbers are generally higher, of order 50 or more input parameters, compared to of order 10 for galaxies.
Given a large body of work for galaxies that has involved the use of artificial neural networks, and the similarities just outlined, it is not surprising that similar approaches have been employed for stellar classification[@vonhippel:annstellar; @weaver:annstellar; @singh:annstellar; @bailerjones:annstellar; @gulati:annstellar; @bazarghan:stellar], with a typical accuracy of one spectral type and half a luminosity type. The relatively large number of object attributes and output classes compared to the number of objects in each class does not invalidate the approach, because the efforts described generally find that the number of principal components represented by the inputs is typically much lower. A well-known property of neural networks is that they are robust to a large number of redundant attributes (§\[Subsubsec: Choice\]).
Neural networks have been used for other stellar classifications schemes, e.g. Gupta [[*et al.*]{}]{}[@gupta:iras] define 17 classes for IRAS sources, including planetary nebulae and H regions. Other methods have been employed; a recent example is Manteiga [[*et al.*]{}]{}[@manteiga:starmind], who use a fuzzy logic knowledge-based system with a hierarchical tree of decision rules. Beyond the MK and other static classifications, variable stars have been extensively studied for many years, e.g., Wozniak [[*et al.*]{}]{}[@wozniak:mirasvm] use SVM to distinguish Mira variables. The detection and characterization of supernovae is important for both understanding the astrophysics of these events, and their use as standard candles in constraining aspects of cosmology such as the dark energy equation of state. Bailey [[*et al.*]{}]{}[@bailey:sneclassification] use boosted DTs, random forests, and SVMs to classify supernovae in difference images, finding a ten times reduction in the false-positive rate compared to standard techniques involving parameter thresholds (Fig. \[Fig: SNe\]).
Given the general nature of the data mining approach, there are many further classification examples, including cosmic ray hits[@salzberg:dt; @waniak:cosmicray], planetary nebulae[@faundezabans:pne], asteroids[@misra:asteroids], and gamma ray sources[@chattopadhyay:gammaray; @scaringi:isina].
Photometric redshifts {#Subsec: Photo-zs}
---------------------
An area of astrophysics that has greatly increased in popularity in the last few years is the estimation of redshifts from photometric data ([photo-$z$s]{}). This is because, although the distances are less accurate than those obtained with spectra, the sheer number of objects with photometric measurements can often make up for the reduction in individual accuracy by suppressing the statistical noise of an ensemble calculation.
Photo-zs were first demonstrated in the mid 20th century[@stebbins:photoz; @baum:photoz], and later in the 1980s[@koo:photoz; @loh:photoz]. In the 1990s, the advent of the Hubble Space Telescope Deep fields resulted in numerous approaches[@gwyn:photoz; @lanzetta:photoz; @mobasher:photoz; @sawicki:photoz; @connolly:angcf; @wang:photoz; @benitez:photoz], reviewed by Koo[@koo:photozhistory]. In the past decade, the advent of wide-field CCD surveys and multifiber spectroscopy have revolutionized the study of [photo-$z$s]{} to the point where they are indispensable for the upcoming next generation surveys, and a large number of studies have been made.
The two common approaches to [photo-$z$s]{} are the template method and the empirical training set method. The template approach has many complicating issues[@massarotti:template], including calibration, zero-points, priors, multiwavelength performance (e.g., poor in the mid-infrared), and difficulty handling missing or incomplete training data. We focus in this review on the empirical approach, as it is an implementation of supervised learning. In the future, it is likely that a hybrid method incorporating both templates and the empirical approach will be used, and that the use of full probability density functions will become increasingly important. For many applications, knowing the error distribution in the redshifts is at least as important as the accuracy of the redshifts themselves, further motivating the calculation of PDFs.
### Galaxies {#Subsubsec: Galaxy Photo-zs}
At low redshifts, the calculation of photometric redshifts for normal galaxies is quite straightforward due to the break in the typical galaxy spectrum at 4000Å. Thus, as a galaxy is redshifted with increasing distance, the color (measured as a difference in magnitudes) changes relatively smoothly. As a result, both template and empirical [photo-$z$]{} approaches obtain similar results, a root-mean-square deviation of $\sim 0.02$ in redshift, which is close to the best possible result given the intrinsic spread in the properties[@brunner:photoz]. This has been shown with ANNs[@firth:annphotoz; @ball:ann; @collister:annz; @vanzella:hdfannphotoz; @li:photoz; @abrusco:sdssphotoz; @collister:megazlrg; @banerji:desvistaphotoz; @oyaizu:photoz; @niemack:stripe82galexphotoz; @zhang:morph], SVM[@wadadekar:svmphotoz; @wang:novelphotoz], DT[@carliles:photoz], [$k$NN]{}[@ball:pdfphotoz], empirical polynomial relations[@connolly:photoz; @brunner:photoz; @wang:photoz; @sowardsemmerd:photoz; @hsieh:photoz; @lopes:photoz], numerous template-based studies, and several other methods. At higher redshifts, obtaining accurate results becomes more difficult because the 4000Å break is shifted redward of the optical, galaxies are fainter and thus spectral data are sparser, and galaxies intrinsically evolve over time. The first explorations at higher redshift were the Hubble Deep Fields in the 1990s, described above (§\[Subsec: Photo-zs\]), and, more recently, new infrared data have become available, which allow the 4000Å break to be seen to higher redshift, which improves the results. Template-based algorithms work well, provided suitable templates into the infrared are available, and supervised algorithms simply incorporate the new data and work in the same manner as previously described.
While supervised learning has been successfully used, beyond the spectral regime the obvious limitation arises that in order to reach the limiting magnitude of the photometric portions of surveys, extrapolation would be required. In this regime, or where only small training sets are available, template-based results can be used, but without spectral information, the templates themselves are being extrapolated. However, the extrapolation of the templates is being done in a more physically motivated manner. It is likely that the more general hybrid approach of using empirical data to iteratively improve the templates,[@budavari:spectempl; @csabai:photoz; @csabai:edrphotoz; @padmanabhan:photoz; @brodwin:iracphotoz; @budavari:unified] or the semi-supervised method described in §\[Subsubsec: Semisupervised\] will ultimately provide a more elegant solution. Another issue at higher redshift is that the available numbers of objects can become quite small (in the hundreds or fewer), thus reintroducing the curse of dimensionality by a simple lack of objects compared to measured wavebands. The methods of dimension reduction (§\[Subsec: Attributes\]) can help to mitigate this effect.
### Quasars/AGN {#Subsubsec: QSO Photo-zs}
Historically, the calculation of photometric redshifts for quasars and other AGN has been even more difficult than for galaxies, because the spectra are dominated by bright but narrow emission lines, which in broad photometric passbands can dominate the color. The color-redshift relation of quasars is thus subject to several effects, including degeneracy, one emission line appearing like another at a different redshift, an emission line disappearing between survey filters, and reddening. In addition, the filter sets of surveys are generally designed for normal galaxies and not quasars. The assignment of these quasar [photo-$z$s]{} is thus a complex problem that is amenable to data mining in a similar manner to the classification of AGN described in §\[Subsubsec: QSO Classification\].
The calculation of quasar [photo-$z$s]{} has had some success using SDSS data[@budavari:qsophotoz; @richards:qsophotoz; @babbedge:impz; @weinstein:qsophotoz; @wu:qsophotoz; @kitsonias:xrayagn], but they suffer from [*catastrophic failures*]{}, in which, as shown in Fig. \[Fig: Photo-z QSO\], the photometric redshift for a subset of the objects is completely incorrect. However, data mining approaches have resulted in improvements to this situation. Ball [[*et al.*]{}]{}[@ball:ibphotoz] find that a single-neighbor [$k$NN]{} gives a similar result to the templates, but multiple neighbors, or other supervised algorithms such as DT or ANN, pull in the regions of catastrophic failure and significantly decrease the spread in the results. Kumar[@kumar:ml] also shows this effect. Ball [[*et al.*]{}]{}[@ball:pdfphotoz] go further and are able to largely eliminate the catastrophics by selecting the subset of quasars with one peak in their redshift probability density function (§\[Subsec: PDF\]), a result confirmed by Wolf[@wolf:qsopdf]. Wolf [[*et al.*]{}]{}[@wolf:qsophotoz] also show significant improvement using the COMBO-17 survey, which has 17 filters compared to the five of the SDSS, but unfortunately the photometric sample is much smaller.
Beyond the spectral regime, template-based results are sufficient[@salvato:cosmosagnphotoz], but again suffer from catastrophics. Given our physical understanding of the nature of quasars, it is in fact reasonable to extrapolate in magnitude when using colors as a training set, because while one is going to fainter magnitudes, one is not extrapolating in color. One could therefore quite reasonably assign empirical [photo-$z$s]{} for a full photometric sample of quasars.
Other Astrophysical Applications {#Subsec: Other Uses}
--------------------------------
Typically in data mining, information gathered from spectra has formed the training set to apply a predictive technique to objects with photometry. However, it is clear from this process that the spectrum itself contains a large amount of information, and data mining techniques may be used directly on the spectra to extract information that might otherwise remain hidden. Applications to galaxy spectral classification were described in §\[Subsubsec: Other Gal\]. In stellar work, besides the classification of stars into the MK system based on observable parameters, several studies have directly predicted physical parameters of stellar atmospheres using spectral indices. One example is Ramirez, Fuentes & Gulati[@ramirez:ibga], who utilize a genetic algorithm to select the appropriate input attributes, and predict the parameters using [$k$NN]{}. The attribute selection reduces run time and improves predictive accuracy. Solorio [[*et al.*]{}]{}[@solorio:ib] use [$k$NN]{} to study stellar populations and improve the results by using active learning to populate sparse regions of parameter space, an alternative to dimension reduction. Although it has much potential for the future (§\[Subsec: Time Domain\]), the time domain is a field in which a lot of work has already been done. Examples include the classification of variable stars described in §\[Subsubsec: Other Classification\], and, in order of distance, the interaction of the solar wind and the Earth’s atmosphere, transient lunar phenomena, detection and classification of asteroids and other solar system objects by composition and orbit, solar system planetary atmospheres, stellar proper motions, extrasolar planets, novae, stellar orbits around the supermassive black hole at the Galactic center, microlensing from massive compact halo objects, supernovae, gamma ray bursts, and quasar variability. A good overview is provided by Becker[@becker:transientclassification]. The large potential of the time domain for novel discovery lies within the as yet unexplored parameter space defined by depth, sky coverage, and temporal resolution[@djorgovski:rare]. One constraining characteristic of the most variable sources beyond the solar system is that they are generally point sources. As a result, the timescales of interest are constrained by the light crossing time for the source.
The analysis of the cosmic microwave background (CMB) is amenable to several techniques, including Bayesian modeling, wavelets, and ICA. The latter, in particular via the FastICA algorithm[@maino:fastica], has been used in removal of CMB foregrounds[@bottino:foreground], and cluster detection via the Sunyaev-Zeldovich effect[@pires:sz]. Phillips & Kogut[@phillips:nnparamest] use a committee of ANNs for cosmological parameter estimation in CMB datasets, by training them to identify parameter values in Monte Carlo simulations. This gives unbiased parameter estimation in considerably less processing time than maximum likelihood, but with comparable accuracy. One can use the fact that objects cross-matched between surveys will likely have correlated distributions in their measured attributes, for example, similar position on the sky, to improve cross-matching results using pattern classifiers. Rohde [[*et al.*]{}]{}[@rohde:matching] combine distribution estimates and probabilistic classifiers to produce such an improvement, and supply probabilistic outputs. Taylor & Diaz[@taylor:evann] obtain empirical fits for Galactic metallicity using ANNs, whose architectures are evolved using genetic algorithms. This method is able to provide equations for metallicity from line ratios, mitigating the ‘black box’ element common to ANNs, and, in addition, is potentially able to identify new metallicity diagnostics. Bogdanos & Nesseris[@bogdanos:snega] analyze Type Ia supernovae using genetic algorithms to extract constraints on the dark energy equation of state. This method is non-parametric, which minimizes bias from the necessarily a priori assumptions of parametric models. Lunar and planetary science, space science, and solar physics also provide many examples of data mining uses. One example is Li [[*et al.*]{}]{}[@li:svmflare], who demonstrate improvements in solar flare forecasting resulting from the use of a mixture of experts, in this case SVM and [$k$NN]{}. The analysis of the abundance of minerals or constituents in soil samples[@mustard:mixture] using mixture models is another example of direct data mining of spectra. Finally, data mining can be performed on astronomical simulations, as well as real datasets. Modern simulations can rival or even exceed real datasets in size and complexity, and as such the data mining approach can be appropriate. An example is the incorporation of theory[@lemson:votheory] into the Virtual Observatory (§\[Subsec: VO\]). Mining simulation data will present extra challenges compared to observations because in general there are fewer constraints on the type of data presented, e.g., observations are of the same universe, but simulations are not, simulations can probe many astrophysical processes that are not directly observable, such as stellar interiors, and they provide direct physical quantities as well as observational ones. Most of the largest simulations are cosmological, but they span many areas of astrophysics. A prominent cosmological simulation is the Millennium Run[@springel:millennium], and over 200 papers have utilized its data[^3].
The Future {#Sec: Future}
==========
We now turn to the future of data mining in astronomy. Several trends are apparent that indicate likely fruitful directions in the next few years. These trends can be used to make informed decisions about upcoming, very large surveys. This section assumes that the reader is somewhat familiar with the concepts in both §§\[Sec: Overview\] and \[Sec: Uses\], namely, with both data mining and astronomy. We once again arrange the topics by data mining algorithm rather than by astronomical application, but we now interweave the algorithms with examples.
As in the past, it is likely that cross-fertilization with other fields will continue to be beneficial to astronomy, and of particular relevance here, the data mining efforts made by these fields. Examples include high energy physics, whose most obvious spinoff is the World Wide Web from CERN, but the subject has an extensive history of extremely large datasets from experiments such as particle colliders, and has provided well-known and commonly used data analysis software such as ROOT [@brun:root], designed to cope with these data sizes and first developed in 1994. In the fields of biology and the geosciences, the concepts of [*informatics*]{}, the study of computer-based information systems, have been extensively utilized, creating the subfields of bio- and geoinformatics. The official recognition of an analogous subfield within astronomy, [*astroinformatics*]{}, has recently been recommended[@borne:datamining].
Probability Density Functions {#Subsec: PDF}
-----------------------------
A [*probability density function*]{} (PDF, Fig. \[Fig: PDFs\]) is a function such that the probability that the value, $x$, is in the interval $a < x < b$, is the definite integral over the range: $$P(a < x < b) = \int_a^b f(x) dx.$$ Thus the total area under the function is one. PDFs are of great significance for data mining in astronomy because they retain information that is otherwise lost, and because they enable results with improved signal-to-noise from a given dataset. One can think of a PDF as a histogram in the limit of small bins but many objects. Approaches such as supervised learning are in general taking as input the information on objects and providing as output a prediction of properties. The most general way to do this is to work with the full PDFs at each stage. The formalism has recently been demonstrated in an astronomical context by Budav[á]{}ri[@budavari:unified], and it is applicable to the prediction of any astronomical property. For inputs $a,b,c$,..., the output probabilities of a set of properties, $P(x,y,z,...)$ can be predicted. Fully probabilistic cross-matching of surveys has also been implemented by the same author[@budavari:crossid].
Results with PDFs in [photo-$z$s]{} are starting to appear, either with single values and a spread, or the full PDF. Cunha [[*et al.*]{}]{}[@cunha:nzpdfphotoz] show that full PDFs help reduce bias. Margoniner & Wittman[@margoniner:photoz] show that they enable subsamples with improved signal-to-noise, and Wittman[@wittman:pzerror] also demonstrates reduction in error. Ball [[*et al.*]{}]{}[@ball:pdfphotoz] show that generating full [photo-$z$]{} PDFs for quasars allows subsection of a sample virtually free of catastrophic failures, the first time this has been demonstrated, and an important result for their use as tracers of the large scale structure in the universe. Wolf[@wolf:qsopdf] confirms a similar result. Myers, White & Ball[@myers:pdf] show that using the full PDF for clustering measurements will improve the signal-to-noise by four to five times for a given dataset without any alteration of the data (Fig. \[Fig: Clustering\]). This method is applicable to the clustering of any astronomical object. Full PDFs have also been shown to improve performance in the photometric detection of galaxy clusters[@vanbreukelen:cluster], again due to the increased signal-to-noise ratio. Several further efforts use a single [photo-$z$]{} and a spread, but not the full PDF. However, the method of Myers, White & Ball shows that it is the full PDF that will give the most benefit. PDFs will also be important for weak lensing[@margoniner:photoz].
As well as [photo-$z$s]{}, predicting properties naturally incorporates probabilistic classification. Progress has been made, e.g., the SDSS has been classified according to P(galaxy, star, neither)[@ball:dtclassification]. Similar classifications that could be made are P(star formation, AGN) and P(quasar, not quasar). Bailer-Jones [[*et al.*]{}]{}[@bailerjones:gaiaquasars] implement probabilistic classification that emphasizes finding very rare objects, in this case quasars among the stars that will be seen by Gaia.
Ball [[*et al.*]{}]{}[@ball:pdfphotoz] generate a PDF by perturbing inputs for a single-neighbor [$k$NN]{}. The idea of perturbing data has been studied in the field of Privacy Preserving Data Mining[@vaidya:ppdm; @aggarwal:ppdm], but here the aim is to generate a PDF using the errors on the input attributes in a way that is computationally scalable to upcoming datasets. The approach appears to work well despite the fact that at present, survey photometric errors are generally poorly characterized[@scranton:sdsscovariance]. Proper characterization of errors will be of great importance to future surveys as the probabilistic approach becomes more important. Scalability may be best implemented either by using kd-tree like data structures, or by direct implementation on novel supercomputing hardware such as FPGA, GPU, or Cell processors (§\[Subsec: Hardware\]), which can provide enormous performance benefits for applications that require a large number of distance calculations.
Real-Time Processing and the Time Domain {#Subsec: Time Domain}
----------------------------------------
The time domain is already a significant area of study and will become increasingly important over the next decade with the advent of large scale synoptic surveys such as the Large Synoptic Survey Telescope (LSST)[@ivezic:lsst]. A large number of temporal resolved observations over large areas of the sky remains an unexplored area, and the historical precedent suggests that many interesting phenomena remain to be discovered[@djorgovski:rare].
However, as one might expect, this field presents a number of challenges not encountered in the data mining of static objects. These include (i) how to handle multiple observations of objects that can vary in irregular and unpredictable ways, both intrinsic and due to the observational equipment, (ii) objects in difference images (the static background is subtracted, leaving the variation), (iii) the necessarily extremely rapid response to certain events such as gamma ray bursts where physical information can be lost mere seconds after an event becomes detectable, (iv) robust classification of large streams of data in real time, (v) lack of previous information on several phenomena, and (vi) the volume and storage of time domain information in databases. Other challenges are seen in static data, but will assume increased importance as real-time accuracy is needed. For example, the removal of artifacts[@donalek:synoptic] that might otherwise be flagged as unusual objects and incur expensive follow-up telescope time. Variability will be both photometric, a change in brightness, and astrometric, because objects can move. While some astronomical phenomena, such as certain types of variable stars, vary in a regular way, others vary in a nonlinear, irregular, stochastic, or chaotic manner, and the variability itself can change with time (heteroskedasticity)[@studenmund:econometrics]. Time series analysis is a well developed area of statistics, and many of these techniques will be useful.
The combination of available information, but incomplete coverage of the possible phenomena suggests that a probabilistic (§\[Subsec: PDF\]) approach[@mahabal:probabilistic], either involving priors, or semi-supervised (§\[Subsubsec: Semisupervised\]) will in general be the most appropriate. This is because the algorithms can use the existing information, but objectively interpret new phenomena. Supervised learning will perform better for problems where more information and larger datasets are available, and unsupervised or Bayesian priors will perform better when there are fewer observations. Many events will still require followup observations, but since there will be far more events than can ever be followed up in detail, data mining algorithms will help ensure that the observations made are optimal in terms of the targeted scientific results.
As a confederation of data archives and interoperable standards of many of the world’s existing telescopes, the Virtual Observatory (VO, §\[Subsec: VO\]) will be crucial in meeting the challenge of the time domain, and significant infrastructure for the VO already exists. The VOEventNet[@drake:voeventnet] is a system for the rapid handling of real time events, and provides an online federated data stream of events from several telescopes. It can be followed by both human observers and robotic telescopes.
Numerous next-generation wide-field surveys in the planning or construction stages will be synoptic. The largest such survey in the optical is the LSST, which will observe the entire sky, visible from its location, every three nights. These observations will provide a data stream exceeding one petabyte per year, and, as a result, they anticipate many of the challenges described here[@ivezic:lsstclassification]. Like LSST[@borne:lsstmining], the Gaia satellite[@perryman:gaia] has working groups dedicated to data mining. The Classification Working Group has employed several data mining techniques, and developed new approaches[@bailerjones:domain; @bailerjones:gaiaquasars] to be used when the survey comes online. Other ongoing or upcoming synoptic surveys include Palomar-Quest[@djorgovski:pq], the Catalina Real-Time Transient Survey[@drake:catalina], Pan-STARRS[@hodapp:panstarrs], and those at other wavelengths such as instruments leading up to and including the Square Kilometer Array[@johnston:askap]. The time domain will not only provide challenges to existing methods of data mining, but will open up new avenues for the extraction of information, such as using the variability of objects for classification[@eyer:variability] or photometric redshift[@kaczmarczik:astrometricphotoz]. Because they are due to a relatively compact source in the center of galaxies, active galactic nuclei vary on much shorter timescales than normal galaxies. This variability has been proposed as a mechanism to select quasar and other AGN candidates. Other events are suspected theoretically but have not been observed[@mahabal:transients]. But given the dataset sizes, automated detection of such events at some level is clearly required. The computational demands of real time processing of the enormous data streams from these surveys is significant, and will likely be met by the use of newly emerging specialized computing hardware (§\[Subsec: Hardware\]).
Petascale Computing {#Subsec: Petascale}
-------------------
The current state of the art in supercomputing consists of terabyte-sized files and teraflop computing speeds, which is conveniently encapsulated in the term [*terascale computing*]{}. Following Moore’s law[@moore:law], in which computer performance has increased exponentially for the last several decades, the coming decade will feature the similarly-derived [*petascale computing*]{}[@bader:petascale]. Much of the performance increase in the past decade has been driven by increases in processor (CPU) clock frequency, but this rate has now slowed due to physical limitations on the sizes of components, and more importantly power consumption and energy (heat) dissipation. It has therefore become more economical to manufacture chips with multiple processor cores.
The typical supercomputer today is a cluster, which consists of a large number of conventional CPUs connected by a specialized interconnect system, a distributed or shared memory, a shared filesystem, and hosting the Linux operating system. Many systems are heterogeneous because this is scalable and cost-effective, but coordinating and making effective such a system can be challenging. In particular, it will be vital that the system is properly balanced between processing power and disk input/output (I/O) to supply the data. Combined with the increasing number of processor cores, this means that [*parallel and distributed computing*]{} is rapidly increasing in importance.
A useful set of ‘rules of thumb’ for parallel and other aspects of computing were formulated by Amdahl in the 1960s[@amdahl:law], and they remain true today. One of these is that roughly 50,000 CPU cycles are required per byte of data. Most scientific datasets require far fewer cycles than this, and it is thus likely that future performance will be I/O limited, unless sufficient disks are provided in parallel. Bell, Gray & Szalay[@bell:petascale] estimate that a petascale system will require 100,000 one TB disks. The exact details of how to distribute the data for best performance are likely to be system-dependent[@dolence:lci]. The available CPU speed should scale to the data size, although it will not scale to most na[ï]{}vely implemented data mining algorithms (§\[Subsec: Parallel\]).
An example of an upcoming petascale system whose uses will include astronomical data mining is the [*Blue Waters*]{}[^4] system at the National Center for Supercomputing Applications (NCSA), which is due to come online in 2011. Specifications include 200,000 compute cores with 4 GHz 8 core processors, 1 PB of main memory, 10 PB of user disk storage, 500 PB of archival storage, and 400 GB ${\mathrm s}^{-1}$ bandwidth connectivity to provide sustained petascale compute power. It will implement the IBM PERCS (Productive, Easy-to-use, Reliable Computer System)[@ebcioglu:percs], which will integrate their CPU, operating system, parallel programming, and file systems. This provides a method of addressing the issues of running real-world applications at the petascale by balancing the CPU, I/O, networking, and so on. Similarly, a considerable investment of effort is being carried out in the years leading up to deployment in 2011 on the development of applications for the system, in consultation with the scientists who will run them. Several astronomical applications are included, mostly simulations, but also data mining in the form of the analysis of LSST datasets.
Not all petascale computing will be done on systems as large as Blue Waters. In the US, the National Science Foundation Office of Cyberinfrastructure has been advised[@bell:petascale] to implement a power-law type system, with a small number of very large systems, of order ten times more regional centers, and ten times more local facilities (Tiers 1–3). Such local facilities, for example Beowulf clusters, are already common in university departments, and consist of typically a few dozen commodity machines. A recent trend matching the increasing requirements for data-intensive as opposed to CPU-intensive computing is the GrayWulf cluster[@szalay:graywulf], which implements the idea of data ‘storage bricks’: cheap, modular, and portable versions of a balanced system which when added together provide petascale computation.
Parallel and Distributed Data Mining {#Subsec: Parallel}
------------------------------------
As indicated in §\[Subsec: Petascale\] above, because of the slowing increase in raw speed of individual CPUs, processors are becoming increasingly parallelized, both in terms of the number of processor cores on a single chip, and increasing amounts of these chips being deployed in parallel on supercomputing clusters. Providing appropriately scaled systems (CPU, I/O, etc.) is one challenge, but most data mining algorithms not only will be required to run on petascale data, but their na[ï]{}ve implementations scale as $N^2$, or worse. It has been suggested[@szalay:petabyte] that any algorithm that scales beyond $N {\mathrm{log}} N$ will rapidly be rendered infeasible.
McConnell and Skillicorn[@mcconnell:ddm] have promoted parallel and distributed data mining[@freitas:parallel; @kargupta:dpkd; @zaki:parallel; @bhaduri:ddmbib], which is well-known in the data mining field, but virtually unused in astronomy. In this approach, the algorithms explicitly take advantage of available parallelism. The simplest example is task-farming, or the embarrassingly parallel approach, in which a task is divided into many mutually-independent subtasks, each of which is allocated to a single processor. This can be done on an array of ordinary desktop machines as well as a supercomputer. A more complex challenge is when many parts of the data must be accessed, or when an algorithm relies on the outputs from calculations distributed across multiple compute nodes. For a large dataset the hardware required likely includes shared memory (§\[Subsec: Petascale\]), thus shared memory parallelization[@jin:shared] can be important. Many algorithms exist for the implementation of data mining on parallel computer systems beyond simple task farming, but these are not widely used within science, as compared to the commercial sector. The application programming interfaces MPI and OpenMP have been widely used on distributed and shared memory systems, respectively, for simulation and some data analysis, but they do not offer the semantic capabilities[@gray:semantic] needed for data mining, i.e., the metadata describing the meaning of the data being processed and the results produced are not easily incorporated.
Parallel data mining is challenging, as not only must the algorithm be implemented on the hardware, but many algorithms simply cannot be ported as-is to such a system. Instead, parallelization requires that the algorithm itself, as encapsulated in the code, must often be fundamentally altered at the pseudocode level. This can be a time-consuming and counterintuitive process, especially to scientists who are generally not trained or experienced in parallel programming. Progress is slowly being made in astronomy, including a parallel implementation of kd-trees[@gardner:paralleltree], cosmological simulations requiring datasets larger than the node memory size[@norman:enzo], and parallelization of algorithms[@brunner:fpganpcf].
An alternative approach is grid computing, in which the exact resource used is unimportant to the user, although not all data mining algorithms lend themselves to this paradigm. A variant of grid computing is crowdsourcing, in which members of the public volunteer their spare CPU cycles to process data for a project. The most well-known project of this type is SETI@Home, and more recently, the Galaxy Zoo project, which employed large numbers of people to successfully classify galaxies in SDSS images. Such crowdsourcing is likely to become even more important in the future, particularly in combination with greatly improved outreach via astronomical applications on social networking sites such as Facebook[@gomez:socialnetworking].
Scalability is also helped on conventional CPUs by the employment of tree structures, such as the kd-tree, which partition the data. This enables a search to access any data value without searching the whole dataset. Kd-trees have been used for many astronomical applications, including speeding up N-point correlation functions[@moore:npcf]; cross-matching, classification, and photometric redshifts[@gao:kdtree]. They can be extended to more sophisticated structures, for example, the multi-tree[@gray:multitree]. However, implementation of such tree structures on parallel hardware or computational accelerators (§\[Subsec: Hardware\]) remains difficult[@gardner:paralleltree].
The Virtual Observatory {#Subsec: VO}
-----------------------
The Virtual Observatory (VO) is an analogous concept to a physical observatory, but instead of telescopes, various centers house data archives. The VO consists of numerous national-level organizations, and the International Virtual Observatory Alliance. Within the national organizations there are various data centers that house large datasets, computing facilities to process and analyze them, and people with considerable expertise in the datasets stored at that particular center.
Common data standards and web services are necessary for the VO to work. Such standards have emerged, including web services using XML and SOAP, a data format, VOTable[@ochsenbein:votable], a query language based on SQL, the Astronomical Data Query Language[@shirasaki:adql], image access protocols for images (SIAP[@shirasaki:adql]), and spectra (SSAP)[^5], VOEventNet[@drake:voeventnet] for the time domain, plus various standards of interoperability and ways of describing resources such as the Unified Content Descriptor[@derriere:ucd]. Large numbers of high level tools for working with data are also available[^6].
An example of the emerging data standards for archiving is the Common Archive Observation Model[@dowler:caom2008] (CAOM) of the Canadian Astronomical Data Center (CADC). Given that it is likely that the future VO will continue to consist of a number of data centers like the CADC, this model represents a useful and realistic way in which data can be made meaningfully accessible, but not so rigidly presented as to prevent the desired analysis of future researchers with as yet unforeseen science goals. This model consists of the components Artifact, Plane, SimpleObservation, and CompositeObservation, which describe logical parts of the data from individual files to logical sets of observations such as spectra, and forms the basis of all archiving activity at the CADC.
The increasing immobility of large datasets as described in §\[Subsec: Petascale\] will render it uneconomical in terms of time and money to download large datasets to local machines. Rather than bringing the data to the analysis, it will become more sensible to take the analysis to the data[@gray:decade]. To be able to perform complicated data mining analyses, it is necessary that the data be organized well enough to make this tractable, and that the center archiving the data must have sufficient computing power and web services to perform the analyses. The organizational requirement means that the data must be stored as a database with the sophistication found in the commercial sector, where mining of terascale databases is routine. Commercial software and computer science expertise will help, but the task is non-trivial because astronomical data analysis can require particular data types and structures not usually found in commercial software, such as time series observations. An example of such a database already in place is the SDSS, and its underlying schema[@gray:sdss] has been used and copied by other surveys such as GALEX.
Nevertheless, it is likely that considerable analyses will continue to be carried out on smaller subsets of the data, and this data may well continue to be downloaded and analyzed locally, as it has been to date. If one anticipates working exclusively with one survey, it may still be more efficient to implement a GrayWulf-like cluster locally and download the complete dataset.
Another difficult problem faced by the VO is that a significant future scientific benefit from large datasets will be in the cross-matching of multiple datasets, in particular, multiwavelength data. But if such data are distributed among different data centers and are difficult to move, such work may be intractable. What can be done, however, is to make available as part of the VO web services, tools for cross-matching datasets at a given center. A common data format and description, combined with the fact that much of the science is done from small subsets of large datasets, means that this is certainly tractable. As a result, it is not surprising that there is significant demand for such tools[@vignali:vo].
An important consideration for the VO is that many astronomers, indeed many scientists in general, will want to run their own software on the data, and not simply a higher level tool that involves trusting someone else’s code. This will be true even if the source code is available. Or, a scientist might wish to complete an analysis that is not available in a higher level tool. It is thus important that the data are available at a low level of processing so that one can set one’s own requirements as needed. NASA has a categorization of data where 0 is raw, 1 is calibrated, and 2 is a derived product, such as a catalog. An ideal data archive would have available well documented and accessible level 2 catalogs, similarly documented and accessible level 1 data, and perhaps not online but stored level 0 data, to enable, for example, a re-reduction.
Data have been released using the VO publishing interfaces[@gonzalezsolares:iphas], data mining algorithms such as ANNs have been implemented[@brescia:voneural], and applications for analyses with web interfaces are online[@kitching:cloudcosmology]. Multiwavelength analyses are becoming more feasible and useful[@vignali:vo], and it is therefore now possible, but still time-consuming, to perform scientific analyses using VO tools[@chilingarian:vo]. We expect this will be an area where considerable work will still need to be done, however, in order to fully enable the full exploitation of the archives of astronomy data in the future.
Visualization {#Subsec: Future Visualization}
-------------
Visualization of data is an important part of the scientific process, and the combination of terascale computing and data mining poses obvious challenges. Common plotting codes presently in use in astronomy include SuperMongo[^7], PGPlot[^8], Gnuplot[^9], and IDL[^10] [@landsman:idlastro], but these are stand-alone codes that do not easily cope with data that cannot be completely loaded into the available memory space. Newer tools, such as TOPCAT[@taylor:topcat], VisIVO[@comparato:viz], and VOMegaPlot[@urunkar:vomegaplot] support the Virtual Observatory standards such as VOTable and PLASTIC[@taylor:plastic] for interoperability between programs. The full library on which the TOPCAT program is based, STILTS[@taylor:stilts], is able to plot arbitrarily-sized datasets.
As with hardware, software, and data analysis, collaboration with computer scientists and other disciplines has resulted in progress in various areas of scientific visualization. At Harvard, the AstroMed project at the Initiative for Innovative Computing (IIC) has collaborated with medical imaging teams[@borkin:astromed]. The rendering of complex multi-dimensional volumetric and surficial data is a common desire of both fields, and the medical imaging software was considerably more advanced than was typical in astronomy in terms of graphical capability. As with the creation and curation of databases for large datasets, collaboration with the IT sector has enabled significant progress and the use of tools beyond the scope of those that could be created by astronomers alone, such as Google Sky[@scranton:googlesky]. It is likely that such collaboration will continue to increase in importance.
The program S2Plot[@barnes:s2plot], developed at Swinburne, is motivated by the idea of making three-dimensional plots as easy to transfer from one medium to another (interchange) as two-dimensional plots. The existing familiar interface of a plotting code, in this case PGPlot, has been extended[@fluke:interchanging] to enable rendering of multi-dimensional data on several media, including desktop machines, PDF files, Powerpoint-style slides, or web pages. Systems in which the user is able to interact directly with the data are also likely to play a significant role. Partiview[@levy:partiview], developed at NCSA, enables the visualization of particulate data and some isosurfaces either on a desktop or in an immersive CAVE system, and several astronomical datasets have been visualized. Szalay, Springel & Lemson[@szalay:gpuvis] describe using graphical processing units (§\[Subsec: Hardware\]) to aid visualization, in which the data are preprocessed to hierarchical levels of detail, and only rendered to the resolution required to appear to the eye as if the whole dataset is being rendered. Paraview[^11] is a program designed for parallel processing on large datasets using distributed memory systems, or on smaller data on a desktop.
Finally, in recent years, numerous online virtual worlds have become popular, the most well-known of which is Second Life. Hut[@hut:virtual] and Djorgovski[^12] describe their interaction within these worlds, both with other astronomers in the form of avatars in meetings, and with datasets. While it may initially seem to be just a gimmicky way to have a meeting, the interaction with other avatars is described as ‘fundamentally visceral’, much more so than one would expect. This suggests that, along with social networks for outreach, such interaction among astronomers may become more common, as one will be able to attend a meeting without having to travel physically.
Novel Supercomputing Hardware {#Subsec: Hardware}
-----------------------------
For the final part of §\[Sec: Future\], we turn to novel supercomputing hardware. This is a rapidly developing area, but it has enormous potential to speed up existing analyses, and render previously impossible questions tractable. Specialized hardware has been used in astronomy for many years, but until recently only in limited contexts and applications, such as the GRAPE[@ebisuzaki:grape] systems designed specifically for $n$-body calculations, or direct processing of data in instrument-specific hardware. Here, we describe three hardware formats that have emerged in recent years as viable solutions to a more general range of astronomical problems: graphical processing units (GPUs), field-programmable gate arrays (FPGAs), and the Cell processor.
As described in §\[Subsec: Petascale\], the increasing speed of CPU clock cycles has now been largely replaced by increasing parallelism as the main method for continuing improvements in computing power. The methods described there implement [*coarse-grained*]{} parallelism, which is at the level of separate pieces of hardware or application processes. The hardware described here implements [*fine-grained*]{} parallelism, in which, at the instruction level, a calculation that would require multiple operations on a CPU is implemented in one operation. The hardware forms an intermediary between the previously-used application-specific integrated circuits (ASIC), and the general purpose CPU.
Future petascale machines (§\[Subsec: Petascale\]) are likely to include some or all of these three, either as highly integrated components in a cluster-type system, or as part of the heterogeneous hardware making up a distributed grid-like system that has overall petascale performance.
Spurred by the computer gaming industry, the GPUs on graphics cards within desktop-scale computers have increased in performance much more rapidly than conventional processors (CPUs). They are specially designed to be very fast at carrying out huge numbers of operations that are used in the rendering of graphics, by using vector datatypes and streaming the data. Vector processors have been used before in supercomputing, but GPUs have become of great interest to the scientific community due to their commodity-level pricing, which results from their widespread commercial use, and the increasing ease of use for more general operations than certain graphical processes.
At first, GPUs dealt only with fixed-point numbers, but now single-precision floating point and even double-precision are becoming more common. Thus the chips are no longer simply specialized graphics engines, but are becoming much more general-purpose (GPGPUs). Double-precision is required or highly desirable for many scientific applications. The ease of use of GPUs has been increased thanks to NVidia’s Compute Unified Device Architecture development environment (CUDA)[^13] for its cards, and will be further aided by the Open Computing Language (OpenCL)[^14] for heterogeneous environments. These enable the GPU functions to be called in a similar way to a C library, and are becoming a de facto standard. CUDA has also been ported to other higher level languages, including PyCUDA in Python.
GPUs are beginning to be used in astronomy, and several applications have appeared. GPUs can reproduce the functionality of the GRAPE hardware for n-body simulations[@gaburov:sapporo], and CUDA implementations have been shown to outperform GRAPE in some circumstances[@belleman:cuda]. GPUs are beginning to be used for real-time processing of data from next generation instruments[@ord:gpu] as part of the Data Intensive Science Consortium at the Harvard IIC. Significant speedup has been demonstrated of a $k$ nearest neighbor search on a GPU compared to a kd-tree implemented in C on a CPU[@garcia:knngpu].
FPGAs[@brown:fpga; @buell:hprc] are another form of hardware that has become viable for somewhat general-purpose scientific computing. While FPGAs have been widely used as specialized hardware for many years, including in telescopes for data processing or adaptive optics, it is only in the past few years that their speed, cost, capacity, and ease of use have made them viable for more general use by non-specialists. As with GPUs, the ability to work with full double precision floating point numbers is also increasing, and their use is via libraries and development environments that enable the FPGA portion of the code to appear as just another function call in C or a C-like language. These tools implement the hardware description language to program the FPGA, which need not be known by the user.
An FPGA consists of a grid of logic gates which must be programmed via software to implement a specific set of functions before running code (hence field-programmable). If the calculation to be performed can be fully represented in this way on the available gates, this enables a throughput speed of one whole calculation of a function per clock cycle, which given a modern FPGA’s clock speed of 100 MHz or more, is 100 million per second. In practice, however, the actual speed is often limited by the I/O.
One recent example is the direct mapping of an ANN onto an FPGA[@won:annfpga], which can then in principle classify one object per clock cycle, or 100 million objects per second at 100 MHz. FPGAs will continue to be widely used as specialized components for astronomical systems, for example in providing real-time processing of the next generation synoptic surveys. Brunner, Kindratenko & Myers[@brunner:fpganpcf] demonstrated a significant speedup of the N-point correlation function using FPGAs. Freeman, Weeks & Austin[@freeman:similarity] directly implement distance calculations, such as required by the [$k$NN]{} data mining algorithm, on an FPGA.
Finally, the IBM Cell processor[@scarpino:cell] is a chip containing a conventional CPU and and array of eight more powerful coprocessors for hardware acceleration in a similar manner to the GPU and FPGA. Like the NVidia GPU, it has been widely used in mass-production machines such as the Playstation 3, and is or will be incorporated into several ‘hybrid’ petascale machines, including IBM’s Roadrunner, and possibly Blue Waters. Unfortunately, also like the GPU, it is not yet as easy to use as desired for large scale scientific use, but progress in the area is continuing.
Further novel supercomputing hardware such as ClearSpeed may become viable for science and widely used. It is an area of exciting developments and considerable potential. As with many new developments, however, one must be somewhat careful, in this case because the continued development of the hardware is driven by large commercial companies (NVidia, IBM, etc.), and not the scientific community. Nevertheless, the potential scientific gains are so large that it is certainly worth keeping an eye on.
Conclusions {#Sec: Conclusions}
===========
In this review, we have introduced data mining in astronomy, given an overview of its implementation in the form of knowledge discovery in databases, reviewed its application to various science problems, and discussed its future. Throughout, we have tried to emphasize data mining as a tool to enable improved science, not as an end in itself, and to highlight areas where improvements have been made over previous analyses, where they might yet be made, and limitations of this approach.
An astronomer is not a cutting-edge expert in data mining algorithms any more than they are in statistics, databases, hardware, software, etc., but they will need to know enough to usefully apply such approaches to the science problem they wish to address. It is likely that such progress will be made via collaboration with people who are experts in these areas, particularly within large projects, that will employ specialists and have working groups dedicated to data mining. Fully implemented, commercial-level databases will be required since the data will be too big to organize, download, or analyze in any other way.
The available infrastructure should, therefore, be designed so that this data mining approach to research is maximally enabled. The raw or minimally-processed data should be made available in a manner so one can apply user-specific codes either locally or using computational resources local to the data if data size necessitates it. It is unlikely that most researchers will either require or trust the exact resources made available by higher level tools. Instead, they will be useful for exploratory work, but ultimately one must be able to run personal or trusted code on the data, from the level of re-reduction upwards.
A problem arises when one wishes to utilize multiple or distributed datasets, for example in cross-matching data for multi-wavelength studies. Therefore, datasets that can be easily made interoperable via a standard storage schema should be made available. In this manner, a user can bring computing power and algorithms to tackle their particular science question. This problem is particularly acute when large datasets are held at widely separated sites, because transfer of such data across the network is currently impractical. A great deal of science is done on small subsets of the full data, so data will still be frequently downloaded and analyzed locally, but the paradigm of downloading entire datasets is not sustainable.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the referee for a useful and comprehensive report.
The authors acknowledge support from NASA through grants NN6066H156 and NNG06GF89G, from Microsoft Research, and from the University of Illinois.
The authors made extensive use of the storage and computing facilities at the National Center for Supercomputing Applications and thank the technical staff for their assistance in enabling this work.
This research has made use of the SAO/NASA Astrophysics Data System.
[100]{}
G. [Bell]{}, J. [Gray]{} and A. [Szalay]{}, [*IEEE Computer*]{} [**39**]{}, 110 (2006).
G. [Bell]{}, T. [Hey]{} and A. [Szalay]{}, [*Science*]{} [**323**]{}, 1297 (2009).
T. [Hey]{}, S. [Tansley]{} and K. [Talle]{} (eds.), [*[The Fourth Paradigm: Data-Intensive Scientific Discovery]{}*]{} (Microsoft Research, Redmond, WA, 2009).
D. J. [Hand]{}, [*Statistical Science*]{} [**21**]{}, p. 1 (2006).
I. H. [Witten]{} and E. [Frank]{}, [*[Data Mining: Practical Machine Learning Tools and Techniques]{}*]{}, Morgan Kaufmann Series in Data Management Systems, 2nd edn. (Morgan Kaufmann, San Francisco, 2005).
C. M. [Bishop]{}, [*[Pattern Recognition and Machine Learning]{}*]{} (Springer, New York, 2007).
T. [Hastie]{}, R. [Tibshirani]{} and J. [Friedman]{}, [*[The Elements of Statistical Learning: Data Mining, Inference, and Prediction]{}*]{}, Springer Series in Statistics, 2nd edn. (Springer, New York, 2009).
K. [Borne]{}, [*[Scientific Data Mining in Astronomy]{}*]{}, Data Mining and Knowledge Discovery Series Data Mining and Knowledge Discovery Series, (Taylor & Francis: CRC Press, Boca Raton, FL, 2009), pp. 91–114.
D. C. [Wells]{}, E. W. [Greisen]{} and R. H. [Harten]{}, [**]{} [**44**]{}, p. 363 (1981).
F. [Ochsenbein]{} [*et al.*]{}, [VOTable: Tabular Data for the Virtual Observatory]{}, in [*Toward an International Virtual Observatory*]{}, eds. P. J. [Quinn]{} and K. M. [G[ó]{}rski]{} (2004).
D. [Pyle]{}, [*[Data Preparation for Data Mining]{}*]{}, Morgan Kaufmann Series in Data Management Systems (Morgan Kaufmann, San Francisco, 1999).
D. W. [Hogg]{}, preprint, \[arXiv/0807.4820\] (2008).
K. [Karhunen]{}, [*Annales Academiae Scientiarum Fennicae Series A. I. Mathematica-Physica*]{} [**37**]{}, 3 (1947).
M. M. [Lo[è]{}ve]{}, [*[Fonctions Al[é]{}atoires de Second Ordre]{}*]{}, in [ *Processus Stochastiques et Mouvement Brownien*]{}, ed. P. [Levy]{} (Hermann, Paris, 1948), Paris.
I. T. [Jolliffe]{}, [*[Principal Component Analysis]{}*]{}, Springer Series in Statistics, 2nd edn. (Springer, New York, 2002).
S. [Arya]{}, D. M. [Mount]{}, N. S. [Netanyahu]{}, R. [Silverman]{} and A. Y. [Wu]{}, [*Journal of the Association for Computing Machinery*]{} [**45**]{}, 891 (1998).
N. [Tishby]{}, F. C. [Pereira]{} and W. [Bialek]{}, [The Information Bottleneck Method]{}, in [*The 37th annual Allerton Conference on Communication, Control, and Computing*]{}, 1999.
R. A. [Fisher]{}, [*Annals of Eugenics*]{} [**7**]{}, 179 (1936).
A. [Hyvärinen]{}, J. [Karhunen]{} and E. [Oja]{}, [*[Independent Component Analysis]{}*]{} (John Wiley & Sons, New York, 2001).
S. J. [Lilly]{} [*et al.*]{}, [**]{} [**172**]{}, 70 (2007).
D. G. [York]{} [*et al.*]{}, [**]{} [**120**]{}, 1579 (2000).
W. [Jeffrey]{} and R. [Rosner]{}, [**]{} [**310**]{}, 473 (1986).
C. M. [Bishop]{}, [*[Neural Networks for Pattern Recognition]{}*]{} (Oxford University Press, Oxford, 1995).
B. D. [Ripley]{}, [*[Pattern Recognition and Neural Networks]{}*]{} (Cambridge University Press, Cambridge, UK, 2008).
R. O. [Duda]{}, P. E. [Hart]{} and D. G. [Stork]{}, [*[Pattern Classification]{}*]{}, 2nd edn. (Cambridge University Press, Cambridge, UK, 2000).
W. S. [McCulloch]{} and W. H. [Pitts]{}, [*Bulletin of Mathematical Biophysics*]{} [**5**]{}, 115 (1943).
J. J. [Hopfield]{} and D. W. [Tank]{}, [*Science*]{} [**233**]{}, 625 (1986).
P. J. [Werbos]{}, [Beyond regression: new tools for prediction and analysis in the behavioural sciences]{}, PhD thesis, Harvard, (Cambridge, MA, 1974).
D. B. [Parker]{}, [*[Learning Logic]{}*]{}, Tech. Rep. TR-47, Center for Computational Research in Economics and Management Science, MIT (Cambridge, MA, 1985).
D. E. [Rumelhart]{}, G. E. [Hinton]{} and R. J. [Williams]{}, [*Nature*]{} [**323**]{}, 533 (1986).
K. [Levenberg]{}, [*Quarterly of Applied Mathematics*]{} [**2**]{}, p. 164 (1944).
D. W. [Marquardt]{}, [*Journal of the Society of Industrial and Applied Mathematics*]{} [**2**]{}, p. 431 (1963).
A. E. [Firth]{}, O. [Lahav]{} and R. S. [Somerville]{}, [**]{} [**339**]{}, 1195 (2003).
J. N. Morgan and J. A. Sonquist, [*Journal of the American Statistical Association*]{} [**58**]{}, 415 (1963).
L. [Breiman]{}, J. [Friedman]{}, R. [Olshen]{} and C. [Stone]{}, [*[Classification and Regression Trees]{}*]{} (Wadsworth, 1984).
J. R. [Quinlan]{}, [*Machine Learning*]{} [**1**]{}, p. 81 (1986).
J. R. [Quinlan]{}, [*[C4.5: Programs for Machine Learning]{}*]{} (Morgan Kaufmann, San Francisco, 1993).
L. [Rokach]{} and O. [Maimon]{}, [*[Data Mining with Decision Trees: Theory and Applications]{}*]{} (World Scientific, New York, 2008).
S. [Salzberg]{}, R. [Chandar]{}, H. [Ford]{}, S. K. [Murthy]{} and R. [White]{}, [ ** ]{} [**107**]{}, 279 (1995).
C. [Cortes]{} and V. [Vapnik]{}, [*Machine Learning*]{} [**20**]{}, 273 (1995).
C. J. C. [Burges]{}, [*Knowledge Discovery and Data Mining*]{} [**2**]{}, 121 (1998).
V. [Vapnik]{}, [*[The Nature of Statistical Learning Theory]{}*]{}, 2nd edn. (Springer, New York, 1999).
N. [Cristianini]{} and J. [Shawe-Taylor]{}, [*[An Introduction to Support Vector Machines and Other Kernel-based Learning Methods]{}*]{} (Cambridge University Press, 2000).
V. [Kecman]{}, [*[Learning and Soft Computing: Support Vector Machines, Neural Networks, and Fuzzy Logic Models]{}*]{} (MIT Press, Cambridge, MA, 2001).
B. [Schlkopf]{} and A. J. [Smola]{}, [*[Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond]{}*]{} (MIT Press, Cambridge, MA, 2001).
S. [Abe]{}, [*[Support Vector Machines for Pattern Classification]{}*]{} (Springer, New York, 2005).
L. [Wang]{}, [*[Support Vector Machines: Theory and Applications]{}*]{} (Springer, New York, 2005).
I. [Steinwart]{} and A. [Christmann]{}, [*[Support Vector Machines]{}*]{} (Springer, New York, 2008).
M. A. [Aizerman]{}, E. M. [Braverman]{} and L. I. [Rozonoer]{}, [*Automation and Remote Control*]{} [**25**]{}, 1175 (1964).
M. [Huertas-Company]{}, D. [Rouan]{}, L. [Tasca]{}, G. [Soucail]{} and O. [Le F[è]{}vre]{}, [**]{} [**478**]{}, 971 (2008).
E. [Fix]{} and J. [Hodges Jr.]{}, [*[Discriminatory analysis: non-parametric discrimination: Consistency properties.]{}*]{}, Tech. Rep. Report No. 4, USAF School of Aviation Medicine (Randolph Field, TX, 1951).
T. M. [Cover]{} and P. E. [Hart]{}, [*IEEE Transactions on Information Theory*]{} [**13**]{}, 21 (1967).
D. W. [Aha]{}, D. [Kibler]{} and M. K. [Albert]{}, [*Machine Learning*]{} [**6**]{}, 37 (1991).
B. [Dasarathy]{}, [*[Nearest Neighbor Pattern Classification Techniques]{}*]{} (IEEE Computer Society Press, New York, 1991).
G. [Shakhnarovich]{}, T. [Darrell]{} and P. [Indyk]{} (eds.), [*[Nearest-Neighbor Methods in Learning and Vision: Theory and Practice]{}*]{} (MIT Press, Cambridge, MA, 2006).
E. [Parzen]{}, [*Annals of Mathematical Statistics*]{} [**33**]{}, 1065 (1962).
R. O. [Duda]{} and P. E. [Hart]{}, [*[Pattern Classification and Scene Analysis]{}*]{} (John Wiley, New York, 1973).
B. W. [Silverman]{}, [*[Density Estimation for Statistics and Data Analysis]{}*]{}, Monographs on Statistics and Applied Probability (CRC Press, Boca Raton, FL, 1986).
D. W. [Scott]{}, [*[Multivariate Density Estimation: Theory, Practice, and Visualization]{}*]{}, Wiley Series in Probability and Statistics (Wiley-Interscience, New York, 1992).
C. [Taylor]{}, [*Vistas in Astronomy*]{} [**41**]{}, 411 (1997).
L. [Wasserman]{}, [*[All of Statistics: a Concise Course in Statistics]{}*]{} (Springer, New York, 2005).
J. [Klemelä]{}, [*[Smoothing of Multivariate Data: Density Estimation and Visualization]{}*]{}, Wiley Series in Probability and Statistics (John Wiley & Sons, New York, 2009).
H. [Steinhaus]{}, [*Bull. Acad. Polon. Sci.*]{} [**4**]{}, 801 (1956).
J. [MacQueen]{}, [Some Methods for Classification and Analysis of Multivariate Observations]{}, in [*Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability*]{}, eds. L. M. [LeCam]{} and J. Neyman (University of California Press, Berkeley, 1967).
D. M. [Titterington]{}, A. F. M. [Smith]{} and U. E. [Makov]{}, [*[Statistical Analysis of Finite Mixture Distributions]{}*]{} (John Wiley, New York, 1985).
G. J. [McLachlan]{} and D. [Peel]{}, [*[Finite Mixture Models]{}*]{}, Wiley Series in Probability and Statistics (Wiley-Interscience, New York, 2000).
A. J. [Connolly]{}, C. [Genovese]{}, A. W. [Moore]{}, R. C. [Nichol]{}, J. [Schneider]{} and L. [Wasserman]{}, preprint, \[arXiv:astro-ph/0008187\] (2000).
J. [Dolence]{} and R. J. [Brunner]{}, [Fast Two-Point Correlations of Extremely Large Data Sets]{}, [*The 9th LCI International Conference on High-Performance Clustered Computing*]{}, Urbana-Champaign, IL, (2008).
A. [Dempster]{}, N. [Laird]{} and D. [Rubin]{}, [*Journal of the Royal Statistical Society B*]{} [**39**]{}, 1 (1977).
M. [Watanabe]{} and K. [Yamaguchi]{} (eds.), [*[The EM Algorithm and Related Statistical Models]{}*]{}, Statistics: a Series of Textbooks and Monographs (CRC Press, Boca Raton, FL, 2003).
G. J. [McLachlan]{} and T. [Krishnan]{}, [*[The EM Algorithm and Extensions]{}*]{}, Wiley Series in Probability and Statistics (John Wiley & Sons, New York, 2008).
T. [Kohonen]{}, [*Biological Cybernetics*]{} [**43**]{}, 59 (1982).
T. [Kohonen]{}, [*[Self-Organizing Maps, 3rd extended edition]{}*]{}, Springer Series in Information Sciences, Vol. 30, 3rd edn. (Springer, Berlin, 2001).
A. [Naim]{}, K. U. [Ratnatunga]{} and R. E. [Griffiths]{}, [**]{} [**111**]{}, p. 357 (1997).
T. [Kohonen]{}, [*[Self-Organization and Associative Memory]{}*]{}, 3rd edn. (Springer-Verlag, Berlin, 1989).
P. [Comon]{}, [*Signal Processing*]{} [**36**]{}, 287 (1994).
T. [Lee]{}, [*[Independent Component Analysis - Theory and Applications]{}*]{} (Kluwer Academic Publishers, New York, 1998).
S. [Roberts]{} and R. [Everson]{} (eds.), [*[Independent Component Analysis: Principles and Practice]{}*]{} (Cambridge University Press, Cambridge, UK, 2001).
J. V. [Stone]{}, [*[Independent Component Analysis: A Tutorial Introduction]{}*]{} (MIT Press, Cambridge, MA, 2004).
O. [Chapelle]{}, B. [Schölkopf]{} and A. [Zien]{} (eds.), [*[Semi-Supervised Learning]{}*]{} (MIT Press, Cambridge, MA, 2006).
X. [Zhu]{}, A. [Goldberg]{}, R. [Brachman]{} and T. [Dietterich]{}, [*[Introduction to Semi-supervised Learning]{}*]{}, Synthesis Lectures on Artificial Intelligence and Machine Learning (Morgan & Claypool, San Rafael, CA, 2009).
D. [Bazell]{} and D. J. [Miller]{}, [**]{} [**618**]{}, 723 (2005).
J. H. [Holland]{}, [*[Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence]{}*]{} (The University of Michigan Press, Ann Arbor, MI, 1975).
D. E. [Goldberg]{}, [*[Genetic Algorithms in Search, Optimization, and Machine Learning]{}*]{} (Addison-Wesley, Reading, MA, 1989).
D. A. [Coley]{}, [*[An Introduction to Genetic Algorithms for Scientists and Engineers]{}*]{} (World Scientific, New York, 1997).
M. [Mitchell]{}, [*[An Introduction to Genetic Algorithms]{}*]{} (MIT Press, Cambridge, MA, 1998).
R. L. [Haupt]{} and S. E. [Haupt]{}, [*[Practical Genetic Algorithms]{}*]{}, 2nd edn. (Wiley Inter-Science, New York, 2004).
S. N. [Sivanandam]{} and S. N. [Deepa]{}, [*[Introduction to Genetic Algorithms]{}*]{} (Springer, New York, 2007).
, [*Design of innovation: Lessons from and for competent genetic algorithms*]{} (Kluwer Academic Publishers, Boston, MA, 2002).
J. M. [Adamo]{}, [*[Data Mining for Association Rules and Sequential Patterns: Sequential and Parallel Algorithms]{}*]{} (Springer, New York, 2000).
C. [Zhang]{} and S. [Zhang]{}, [*[Association Rule Mining: Models and Algorithms]{}*]{}, Lecture Notes in Computer Science (Springer, New York, 2002).
Y. [Benjamini]{} and Y. [Hochberg]{}, [*Journal of the Royal Statistical Society B*]{} [**57**]{}, 289 (1995).
M. [Welge]{}, W. H. [Hsu]{}, L. S. [Auvil]{}, T. M. [Redman]{} and D. [Tcheng]{}, [High-Performance Knowledge Discovery and Data Mining Systems Using Workstation Clusters]{}, in [*[12th National Conference on High Performance Networking and Computing (SC99)]{}*]{}, 1999.
S. L. [Salzberg]{}, [*Data Mining and Knowledge Discovery*]{} [**1**]{}, 1 (1995).
S. [Kirkpatrick]{}, C. D. [Gelatt]{} and M. P. [Vecchi]{}, [*Science*]{} [**220**]{}, 671 (1983).
V. [[Č]{}ern[ý]{}]{}, [*Journal of Optimization Theory and Applications*]{} [ **45**]{}, 41 (1985).
P. J. [van Laarhiven]{} and E. H. [Aarts]{}, [*[Simulated Annealing: Theory and Applications]{}*]{} (Springer, New York, 1987).
E. [Aarts]{} and J. [Korst]{}, [*[Simulated Annealing and Boltzmann Machines: A Stochastic Approach to Combinatorial Optimization and Neural Computing]{}*]{} (Wiley, New York, 1989).
L. [Breiman]{}, [*Machine Learning*]{} [**26**]{}, 123 (1996).
L. [Breiman]{}, [*Machine Learning*]{} [**45**]{}, 5 (2001).
J. L. [Bentley]{}, [*Communications of the ACM*]{} [**18**]{}, 509 (1975).
J. P. [Gardner]{}, A. [Connolly]{} and C. [McBride]{}, [Enabling rapid development of parallel tree search applications]{}, in [*CLADE ’07: Proceedings of the 5th IEEE workshop on Challenges of large applications in distributed environments*]{}, (ACM, New York, 2007).
A. S. [Miller]{}, [*Vistas in Astronomy*]{} [**36**]{}, 141 (1993).
O. [Lahav]{}, A. [Naim]{}, L. [Sodr[é]{}]{} and M. C. [Storrie-Lombardi]{}, [ ** ]{} [**283**]{}, p. 207 (1996).
C. A. L. [Bailer-Jones]{}, R. [Gupta]{} and H. P. [Singh]{}, [An Introduction to Artificial Neural Networks]{}, in [*Automated Data Analysis in Astronomy*]{}, eds. R. [Gupta]{}, H. P. [Singh]{} and C. A. L. [Bailer-Jones]{} (2002).
L.-L. [Li]{}, Y.-X. [Zhang]{}, Y.-H. [Zhao]{} and D.-W. [Yang]{}, [*Progress in Astronomy*]{} [**24**]{}, 285 (2006).
R. [Tagliaferri]{} [*et al.*]{}, Neural Networks, [**16**]{}, 297 (2003).
R. L. [White]{}, [Astronomical Applications of Oblique Decision Trees]{}, American Institute of Physics Conference Series Vol. 1082 (2008).
P. [Charbonneau]{}, [**]{} [**101**]{}, p. 309 (1995).
C. A. L. [Bailer-Jones]{}, [Automated Stellar Classification for Large Surveys: A Review of Methods and Results]{}, in [*Automated Data Analysis in Astronomy*]{}, eds. R. [Gupta]{}, H. P. [Singh]{} and C. A. L. [Bailer-Jones]{} (2002).
N. [Weir]{}, U. M. [Fayyad]{}, S. G. [Djorgovski]{} and J. [Roden]{}, [**]{} [ **107**]{}, p. 1243 (1995).
M. C. [Burl]{}, L. [Asker]{}, P. [Smyth]{}, U. [Fayyad]{}, P. [Perona]{}, J. [Aubele]{} and L. [Crumpler]{}, [*Machine Learning*]{} [**30**]{}, 165 (1998).
M. C. [Burl]{}, C. [Fowlkes]{}, J. [Roden]{}, A. [Stechert]{} and S. [Mukhtar]{}, [Diamond Eye: a distributed architecture for image data mining]{}, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series Vol. 3695 (1999).
C. [Kamath]{}, [*Journal of Physics Conference Series*]{} [**125**]{}, 012094 (2008).
C. [Kamath]{}, [*[Scientific data mining: a practical perspective]{}*]{} (Society for Industrial and Applied Mathematics, Philadelphia, PA, 2009).
S. J. [Maddox]{}, G. [Efstathiou]{}, W. J. [Sutherland]{} and J. [Loveday]{}, [ ** ]{} [**243**]{}, 692 (1990).
S. G. [Djorgovski]{}, R. R. [Gal]{}, S. C. [Odewahn]{}, R. R. [de Carvalho]{}, R. [Brunner]{}, G. [Longo]{} and R. [Scaramella]{}, [The Palomar Digital Sky Survey (DPOSS)]{}, in [*Wide Field Surveys in Cosmology*]{}, eds. S. [Colombi]{}, Y. [Mellier]{} and B. [Raban]{} (1998).
S. C. [Odewahn]{}, E. B. [Stockwell]{}, R. L. [Pennington]{}, R. M. [Humphreys]{} and W. A. [Zumach]{}, [**]{} [**103**]{}, 318 (1992).
S. C. [Odewahn]{} and M. L. [Nielsen]{}, [*Vistas in Astronomy*]{} [**38**]{}, 281 (1994).
D. [Bazell]{} and Y. [Peng]{}, [**]{} [**116**]{}, p. 47 (1998).
S. [Andreon]{}, G. [Gargiulo]{}, G. [Longo]{}, R. [Tagliaferri]{} and N. [Capuano]{}, [**]{} [**319**]{}, 700 (2000).
N. S. [Philip]{}, Y. [Wadadekar]{}, A. [Kembhavi]{} and K. B. [Joseph]{}, [**]{} [**385**]{}, 1119 (2002).
S. C. [Odewahn]{}, R. R. [de Carvalho]{}, R. R. [Gal]{}, S. G. [Djorgovski]{}, R. [Brunner]{}, A. [Mahabal]{}, P. A. A. [Lopes]{}, J. L. K. [Moreira]{} and B. [Stalder]{}, [**]{} [**128**]{}, 3092 (2004).
A. [Collister]{} [*et al.*]{}, [**]{} [**375**]{}, 68 (2007).
N. [Weir]{}, U. M. [Fayyad]{} and S. [Djorgovski]{}, [**]{} [**109**]{}, p. 2401 (1995).
N. M. [Ball]{}, R. J. [Brunner]{}, A. D. [Myers]{} and D. [Tcheng]{}, [**]{} [ **650**]{}, p. 497 (2006).
D.-M. [Qin]{}, P. [Guo]{}, Z.-Y. [Hu]{} and Y.-H. [Zhao]{}, [*Chinese Journal of Astronomy and Astrophysics*]{} [**3**]{}, 277 (2003).
A. S. [Miller]{} and M. J. [Coe]{}, [**]{} [**279**]{}, 293 (1996).
E. P. [Hubble]{}, [**]{} [**64**]{}, 321 (1926).
E. P. [Hubble]{}, [*[Realm of the Nebulae]{}*]{} (Yale University Press, Newhaven, CT, 1936).
A. [Sandage]{}, [*[The Hubble atlas of galaxies]{}*]{} (Carnegie Institution, Washington, DC, 1961).
A. [Sandage]{} and J. [Bedke]{}, [*[The Carnegie atlas of galaxies]{}*]{} (Carnegie Institution of Washington with The Flintridge Foundation, Washington, DC, 1994).
S. [van den Bergh]{}, [*[Galaxy morphology and classification]{}*]{} (Cambridge University Press, Cambridge, UK, 1998).
A. [Sandage]{}, [**]{} [**43**]{}, 581 (2005).
M. S. [Roberts]{} and M. P. [Haynes]{}, [**]{} [**32**]{}, 115 (1994).
C. [Firmani]{} and V. [Avila-Reese]{}, [Physical processes behind the morphological Hubble sequence]{}, Revista Mexicana de Astronomia y Astrofisica Conference Series Vol. 17 (2003).
W. W. [Morgan]{}, [**]{} [**70**]{}, p. 364 (1958).
W. W. [Morgan]{}, [**]{} [**71**]{}, p. 394 (1959).
G. [de Vaucouleurs]{}, [Qualitative and Quantitative Classifications of Galaxies.]{}, in [*Evolution of Galaxies and Stellar Populations*]{}, eds. B. M. [Tinsley]{} and R. B. [Larson]{} (1977).
G. [de Vaucouleurs]{}, [*Annales d’Astrophysique*]{} [**11**]{}, p. 247 (1948).
F. S. [Patterson]{}, [*Harvard College Observatory Bulletin*]{} [**914**]{}, 9 (1940).
K. C. [Freeman]{}, [**]{} [**160**]{}, p. 811 (1970).
J. L. [Sérsic]{}, [*[Atlas de galaxias australes]{}*]{} (Observatorio Astronomico, Cordoba, Argentina, 1968).
A. W. [Graham]{} and S. P. [Driver]{}, [*[PASA]{}*]{} [**22**]{}, 118 (2005).
S. [van den Bergh]{}, [**]{} [**131**]{}, p. 215 (1960).
S. [van den Bergh]{}, [**]{} [**131**]{}, p. 558 (1960).
S. [van den Bergh]{}, [**]{} [**206**]{}, 883 (1976).
C. J. [Conselice]{}, [**]{} [**147**]{}, 1 (2003).
G. [de Vaucouleurs]{}, [*Memoirs of the Commonwealth Observatory, Mount Stromlo*]{} [**3**]{} (1956).
G. [de Vaucouleurs]{}, [*Handbuch der Physik*]{} [**53**]{}, p. 275 (1959).
M. [Barden]{}, K. [Jahnke]{} and B. [H[ä]{}u[ß]{}ler]{}, [**]{} [**175**]{}, 105 (2008).
M. C. [Storrie-Lombardi]{}, O. [Lahav]{}, L. [Sodré]{} and L. J. [Storrie-Lombardi]{}, [**]{} [**259**]{}, p. 8P (1992).
O. [Lahav]{} [*et al.*]{}, [*Science*]{} [**267**]{}, 859 (1995).
A. [Naim]{} [*et al.*]{}, [**]{} [**274**]{}, 1107 (1995).
A. [Naim]{}, O. [Lahav]{}, L. [Sodré]{} and M. C. [Storrie-Lombardi]{}, [**]{} [**275**]{}, 567 (1995).
A. A. [Collister]{} and O. [Lahav]{}, [**]{} [**116**]{}, 345 (2004).
A. [Naim]{}, K. U. [Ratnatunga]{} and R. E. [Griffiths]{}, [**]{} [**476**]{}, p. 510 (1997).
D. S. [Madgwick]{}, [**]{} [**338**]{}, 197 (2003).
S. C. [Odewahn]{}, R. A. [Windhorst]{}, S. P. [Driver]{} and W. C. [Keel]{}, [ ** ]{} [**472**]{}, p. L13 (1996).
R. [Windhorst]{}, S. [Odewahn]{}, C. [Burg]{}, S. [Cohen]{} and I. [Waddington]{}, [ ** ]{} [**269**]{}, 243 (1999).
S. H. [Cohen]{}, R. A. [Windhorst]{}, S. C. [Odewahn]{}, C. A. [Chiarenza]{} and S. P. [Driver]{}, [**]{} [**125**]{}, 1762 (2003).
S. C. [Odewahn]{}, S. H. [Cohen]{}, R. A. [Windhorst]{} and N. S. [Philip]{}, [ ** ]{} [**568**]{}, 539 (2002).
D. [Bazell]{} and D. W. [Aha]{}, [**]{} [**548**]{}, 219 (2001).
D. [Bazell]{}, [**]{} [**316**]{}, 519 (2000).
N. M. [Ball]{}, J. [Loveday]{}, M. [Fukugita]{}, O. [Nakamura]{}, S. [Okamura]{}, J. [Brinkmann]{} and R. J. [Brunner]{}, [**]{} [**348**]{}, 1038 (2004).
N. M. [Ball]{}, J. [Loveday]{}, R. J. [Brunner]{}, I. K. [Baldry]{} and J. [Brinkmann]{}, [**]{} [**373**]{}, 845 (2006).
N. M. [Ball]{}, J. [Loveday]{} and R. J. [Brunner]{}, [**]{} [**383**]{}, 907 (2008).
B. C. [Kelly]{} and T. A. [McKay]{}, [**]{} [**129**]{}, 1287 (2005).
M. [Serra-Ricart]{}, X. [Calbet]{}, L. [Garrido]{} and V. [Gaitan]{}, [**]{} [ **106**]{}, 1685 (1993).
A. [Adams]{} and A. [Woolley]{}, [*Vistas in Astronomy*]{} [**38**]{}, 273 (1994).
E. [Molinari]{} and R. [Smareglia]{}, [**]{} [**330**]{}, 447 (1998).
P. A. M. [de Theije]{} and P. [Katgert]{}, [**]{} [**341**]{}, 371 (1999).
E. [Cant[ú]{}-Paz]{} and C. [Kamath]{}, [Evolving Neural Networks For The Classification Of Galaxies]{}, in [*GECCO ’02: Proceedings of the Genetic and Evolutionary Computation Conference*]{}, (Morgan Kaufmann Publishers Inc., San Francisco, 2002).
C. [Kamath]{}, E. [Cant[ú]{}-Paz]{}, I. K. [Fodor]{} and N. I. [Tang]{}, [*Computing in Science and Engineering*]{} [**4**]{}, 52 (2002).
R. H. [Becker]{}, R. L. [White]{} and D. J. [Helfand]{}, [**]{} [**450**]{}, p. 559 (1995).
J. [de la Calleja]{} and O. [Fuentes]{}, [**]{} [**349**]{}, 87 (2004).
G. [Spiekermann]{}, [**]{} [**103**]{}, 2102 (1992).
E. A. [Owens]{}, R. E. [Griffiths]{} and K. U. [Ratnatunga]{}, [**]{} [ **281**]{}, 153 (1996).
Y. [Zhang]{}, L. [Li]{} and Y. [Zhao]{}, [**]{} [**392**]{}, 233 (2009).
M. [Huertas-Company]{} [*et al.*]{}, [**]{} [**497**]{}, 743 (2009).
P. [Tsalmantza]{} [*et al.*]{}, [**]{} [**470**]{}, 761 (2007).
C. J. [Lintott]{} [*et al.*]{}, [**]{} [**389**]{}, 1179 (2008).
M. L. [Humason]{}, [**]{} [**83**]{}, p. 10 (1936).
W. W. [Morgan]{} and N. U. [Mayall]{}, [**]{} [**69**]{}, p. 291 (1957).
A. J. [Connolly]{}, A. S. [Szalay]{}, M. A. [Bershady]{}, A. L. [Kinney]{} and D. [Calzetti]{}, [**]{} [**110**]{}, p. 1071 (1995).
A. J. [Connolly]{} and A. S. [Szalay]{}, [**]{} [**117**]{}, 2052 (1999).
D. [Madgwick]{}, O. [Lahav]{}, K. [Taylor]{} and [The 2dFGRS Team]{}, [Parameterisation of Galaxy Spectra in the 2dF Galaxy Redshift Survey]{}, in [*Mining the Sky*]{}, eds. A. J. [Banday]{}, S. [Zaroubi]{} and M. [Bartelmann]{} (2001).
C. W. [Yip]{} [*et al.*]{}, [**]{} [**128**]{}, 585 (2004).
M. C. [Storrie-Lombardi]{}, M. J. [Irwin]{}, T. [von Hippel]{} and L. J. [Storrie-Lombardi]{}, [*Vistas in Astronomy*]{} [**38**]{}, 331 (1994).
S. R. [Folkes]{}, O. [Lahav]{} and S. J. [Maddox]{}, [**]{} [**283**]{}, 651 (1996).
M. [Colless]{} [*et al.*]{}, preprint, \[arXiv:astro-ph/0306581\] (2003).
N. [Slonim]{}, R. [Somerville]{}, N. [Tishby]{} and O. [Lahav]{}, [**]{} [ **323**]{}, 270 (2001).
H. [Lu]{}, H. [Zhou]{}, J. [Wang]{}, T. [Wang]{}, X. [Dong]{}, Z. [Zhuang]{} and C. [Li]{}, [**]{} [**131**]{}, 790 (2006).
F. B. [Abdalla]{}, A. [Mateus]{}, W. A. [Santos]{}, L. [Sodr[è]{}]{}, Jr., I. [Ferreras]{} and O. [Lahav]{}, [**]{} [**387**]{}, 945 (2008).
A. [Lauberts]{} and E. A. [Valentijn]{}, [*[The surface photometry catalogue of the ESO-Uppsala galaxies]{}*]{} (European Southern Observatory, Garching, Germany, 1989).
J. A. [Baldwin]{}, M. M. [Phillips]{} and R. [Terlevich]{}, [**]{} [**93**]{}, 5 (1981).
R. [Carballo]{}, A. S. [Cofi[ñ]{}o]{} and J. I. [Gonz[á]{}lez-Serrano]{}, [ ** ]{} [**353**]{}, 211 (2004).
J.-F. [Claeskens]{}, A. [Smette]{}, L. [Vandenbulcke]{} and J. [Surdej]{}, [**]{} [**367**]{}, 879 (2006).
R. [Carballo]{}, J. I. [Gonz[á]{}lez-Serrano]{}, C. R. [Benn]{} and F. [Jim[é]{}nez-Luj[á]{}n]{}, [**]{} [**391**]{}, 369 (2008).
R. L. [White]{} [*et al.*]{}, [**]{} [**126**]{}, 133 (2000).
A. A. [Suchkov]{}, R. J. [Hanisch]{} and B. [Margon]{}, [**]{} [**130**]{}, 2439 (2005).
Y.-X. [Zhang]{} and Y.-H. [Zhao]{}, [*Chinese Journal of Astronomy and Astrophysics*]{} [**7**]{}, 289 (2007).
Y. [Zhang]{}, Y. [Zhao]{} and D. [Gao]{}, [*Advances in Space Research*]{} [**41**]{}, 1949 (2008).
Y. [Zhao]{} and Y. [Zhang]{}, [*Advances in Space Research*]{} [**41**]{}, 1955 (2008).
C. [Knigge]{}, S. [Scaringi]{}, M. R. [Goad]{} and C. E. [Cottis]{}, [**]{} [ **386**]{}, 1426 (2008).
C. W. [Yip]{} [*et al.*]{}, [**]{} [**128**]{}, 2603 (2004).
Y. [Zhang]{} and Y. [Zhao]{}, [**]{} [**115**]{}, 1006 (2003).
D. [Gao]{}, Y.-X. [Zhang]{} and Y.-H. [Zhao]{}, [**]{} [**386**]{}, 1417 (2008).
R. [D’Abrusco]{}, G. [Longo]{} and N. A. [Walton]{}, [**]{} [**396**]{}, 223 (2009).
G. T. [Richards]{} [*et al.*]{}, [**]{} [**180**]{}, 67 (2009).
M.-F. [Zhao]{}, C. [Wu]{}, A. [Luo]{}, F.-C. [Wu]{} and Y.-H. [Zhao]{}, [*Chinese Astronomy and Astrophysics*]{} [**31**]{}, 352 (2007).
S. P. [Bamford]{}, A. L. [Rojas]{}, R. C. [Nichol]{}, C. J. [Miller]{}, L. [Wasserman]{}, C. R. [Genovese]{} and P. E. [Freeman]{}, [**]{} [**391**]{}, 607 (2008).
J. F. [Jarvis]{} and J. A. [Tyson]{}, [**]{} [**86**]{}, 476 (1981).
P. B. [Stetson]{}, [**]{} [**99**]{}, 191 (1987).
E. [Bertin]{} and S. [Arnouts]{}, [**]{} [**117**]{}, 393 (1996).
D. [Maino]{} [*et al.*]{}, [**]{} [**334**]{}, 53 (2002).
F. [Guglielmetti]{}, R. [Fischer]{} and V. [Dose]{}, [**]{} [**396**]{}, 165 (2009).
M. [Serra-Ricart]{}, V. [Gaitan]{}, L. [Garrido]{} and I. [Perez-Fournon]{}, [ ** ]{} [**115**]{}, p. 195 (1996).
J. [Goebel]{}, J. [Stutz]{}, K. [Volk]{}, H. [Walker]{}, F. [Gerbault]{}, M. [Self]{}, W. [Taylor]{} and P. [Cheeseman]{}, [**]{} [**222**]{}, L5 (1989).
T. A. [McGlynn]{} [*et al.*]{}, [**]{} [**616**]{}, 1284 (2004).
D. [Bazell]{}, D. J. [Miller]{} and M. [SubbaRao]{}, [**]{} [**649**]{}, 678 (2006).
W. W. [Morgan]{}, P. C. [Keenan]{} and E. [Kellman]{}, [*[An atlas of stellar spectra, with an outline of spectral classification]{}*]{} (The University of Chicago press, 1943).
T. [von Hippel]{}, L. J. [Storrie-Lombardi]{}, M. C. [Storrie-Lombardi]{} and M. J. [Irwin]{}, [**]{} [**269**]{}, p. 97 (1994).
W. B. [Weaver]{} and A. V. [Torres-Dodgen]{}, [**]{} [**487**]{}, p. 847 (1997).
H. P. [Singh]{}, R. K. [Gulati]{} and R. [Gupta]{}, [**]{} [**295**]{}, 312 (1998).
C. A. L. [Bailer-Jones]{}, M. [Irwin]{} and T. [von Hippel]{}, [**]{} [ **298**]{}, 361 (1998).
R. K. [Gulati]{} and L. [Altamirano]{}, [**]{} [**273**]{}, 73 (2000).
M. [Bazarghan]{} and R. [Gupta]{}, [**]{} [**315**]{}, 201 (2008).
R. [Gupta]{}, H. P. [Singh]{}, K. [Volk]{} and S. [Kwok]{}, [**]{} [**152**]{}, 201 (2004).
M. [Manteiga]{}, I. [Carricajo]{}, A. [Rodr[í]{}guez]{}, C. [Dafonte]{} and B. [Arcay]{}, [**]{} [**137**]{}, 3245 (2009).
P. R. [Wo[ź]{}niak]{}, S. J. [Williams]{}, W. T. [Vestrand]{} and V. [Gupta]{}, [ ** ]{} [**128**]{}, 2965 (2004).
S. [Bailey]{}, C. [Aragon]{}, R. [Romano]{}, R. C. [Thomas]{}, B. A. [Weaver]{} and D. [Wong]{}, [**]{} [**665**]{}, 1246 (2007).
W. [Waniak]{}, [*Experimental Astronomy*]{} [**21**]{}, 151 (2006).
M. [Faundez-Abans]{}, M. I. [Ormeno]{} and M. [de Oliveira-Abans]{}, [**]{} [ **116**]{}, 395 (1996).
A. [Misra]{} and S. J. [Bus]{}, [Artificial Neural Network Classification of Asteroids in the Sloan Digital Sky Survey]{}, in [*AAS/Division for Planetary Sciences Meeting Abstracts*]{}, 2008.
T. [Chattopadhyay]{}, R. [Misra]{}, A. K. [Chattopadhyay]{} and M. [Naskar]{}, [ ** ]{} [**667**]{}, 1017 (2007).
S. [Scaringi]{}, A. J. [Bird]{}, D. J. [Clark]{}, A. J. [Dean]{}, A. B. [Hill]{}, V. A. [McBride]{} and S. E. [Shaw]{}, [**]{} [**390**]{}, 1339 (2008).
J. [Stebbins]{} and A. E. [Whitford]{}, [**]{} [**108**]{}, p. 413 (1948).
W. A. [Baum]{}, [Photoelectric Magnitudes and Red-Shifts]{}, in [*[IAU Symp. 15: Problems of Extra-Galactic Research]{}*]{}, 1962.
D. C. [Koo]{}, [**]{} [**90**]{}, 418 (1985).
E. D. [Loh]{} and E. J. [Spillar]{}, [**]{} [**303**]{}, 154 (1986).
S. D. J. [Gwyn]{} and F. D. A. [Hartwick]{}, [**]{} [**468**]{}, p. L77 (1996).
K. M. [Lanzetta]{}, A. [Yahil]{} and A. [Fernandez-Soto]{}, [**]{} [**381**]{}, 759 (1996).
B. [Mobasher]{}, M. [Rowan-Robinson]{}, A. [Georgakakis]{} and N. [Eaton]{}, [ ** ]{} [**282**]{}, L7 (1996).
M. J. [Sawicki]{}, H. [Lin]{} and H. K. C. [Yee]{}, [**]{} [**113**]{}, 1 (1997).
A. J. [Connolly]{}, A. S. [Szalay]{} and R. J. [Brunner]{}, [**]{} [**499**]{}, p. L125 (1998).
Y. [Wang]{}, N. [Bahcall]{} and E. L. [Turner]{}, [**]{} [**116**]{}, 2081 (1998).
N. [Ben[í]{}tez]{}, [**]{} [**536**]{}, 571 (2000).
D. C. [Koo]{}, [Overview - Photometric Redshifts: A Perspective from an Old-Timer\[!\] on their Past, Present, and Potential]{}, in [*Photometric Redshifts and the Detection of High Redshift Galaxies*]{}, eds. R. [Weymann]{}, L. [Storrie-Lombardi]{}, M. [Sawicki]{} and R. [Brunner]{}, Astronomical Society of the Pacific Conference Series, Vol. 191 (1999).
M. [Massarotti]{}, A. [Iovino]{} and A. [Buzzoni]{}, [**]{} [**368**]{}, 74 (2001).
R. J. [Brunner]{}, A. J. [Connolly]{}, A. S. [Szalay]{} and M. A. [Bershady]{}, [ ** ]{} [**482**]{}, p. L21 (1997).
E. [Vanzella]{} [*et al.*]{}, [**]{} [**423**]{}, 761 (2004).
L.-L. [Li]{}, Y.-X. [Zhang]{}, Y.-H. [Zhao]{} and D.-W. [Yang]{}, [*Chinese Journal of Astronomy and Astrophysics*]{} [**7**]{}, 448 (2007).
R. [D’Abrusco]{}, A. [Staiano]{}, G. [Longo]{}, M. [Brescia]{}, M. [Paolillo]{}, E. [De Filippis]{} and R. [Tagliaferri]{}, [**]{} [**663**]{}, 752 (2007).
M. [Banerji]{}, F. B. [Abdalla]{}, O. [Lahav]{} and H. [Lin]{}, [**]{} [**386**]{}, 1219 (2008).
H. [Oyaizu]{}, M. [Lima]{}, C. E. [Cunha]{}, H. [Lin]{}, J. [Frieman]{} and E. S. [Sheldon]{}, [**]{} [**674**]{}, 768 (2008).
M. D. [Niemack]{}, R. [Jimenez]{}, L. [Verde]{}, F. [Menanteau]{}, B. [Panter]{} and D. [Spergel]{}, [**]{} [**690**]{}, 89 (2009).
Y. [Wadadekar]{}, [**]{} [**117**]{}, 79 (2005).
D. [Wang]{}, Y.-X. [Zhang]{}, C. [Liu]{} and Y.-H. [Zhao]{}, [*Chinese Journal of Astronomy and Astrophysics*]{} [**8**]{}, 119 (2008).
S. [Carliles]{}, T. [Budav[á]{}ri]{}, S. [Heinis]{}, C. [Priebe]{} and A. [Szalay]{}, [Photometric Redshift Estimation on SDSS Data Using Random Forests]{}, in [ *Astronomical Data Analysis Software and Systems XVII*]{}, eds. R. W. [Argyle]{}, P. S. [Bunclark]{} and J. R. [Lewis]{}, Astronomical Society of the Pacific Conference Series, Vol. 394 (2008).
N. M. [Ball]{}, R. J. [Brunner]{}, A. D. [Myers]{}, N. E. [Strand]{}, S. L. [Alberts]{} and D. [Tcheng]{}, [**]{} [**683**]{}, 12 (2008).
A. J. [Connolly]{}, I. [Csabai]{}, A. S. [Szalay]{}, D. C. [Koo]{}, R. G. [Kron]{} and J. A. [Munn]{}, [**]{} [**110**]{}, p. 2655 (1995).
D. [Sowards-Emmerd]{}, J. A. [Smith]{}, T. A. [McKay]{}, E. [Sheldon]{}, D. L. [Tucker]{} and F. J. [Castander]{}, [**]{} [**119**]{}, 2598 (2000).
B. C. [Hsieh]{}, H. K. C. [Yee]{}, H. [Lin]{} and M. D. [Gladders]{}, [**]{} [ **158**]{}, 161 (2005).
P. A. A. [Lopes]{}, [**]{} [**380**]{}, 1608 (2007).
T. [Budav[á]{}ri]{}, A. S. [Szalay]{}, A. J. [Connolly]{}, I. [Csabai]{} and M. [Dickinson]{}, [**]{} [**120**]{}, 1588 (2000).
I. [Csabai]{}, A. J. [Connolly]{}, A. S. [Szalay]{} and T. [Budav[á]{}ri]{}, [**]{} [**119**]{}, 69 (2000).
I. [Csabai]{} [*et al.*]{}, [**]{} [**125**]{}, 580 (2003).
N. [Padmanabhan]{} [*et al.*]{}, [**]{} [**359**]{}, 237 (2005).
M. [Brodwin]{} [*et al.*]{}, [**]{} [**651**]{}, 791 (2006).
T. [Budav[á]{}ri]{}, [**]{} [**695**]{}, 747 (2009).
T. [Budav[á]{}ri]{} [*et al.*]{}, [**]{} [**122**]{}, 1163 (2001).
G. T. [Richards]{} [*et al.*]{}, [**]{} [**122**]{}, 1151 (2001).
T. S. R. [Babbedge]{} [*et al.*]{}, [**]{} [**353**]{}, 654 (2004).
M. A. [Weinstein]{} [*et al.*]{}, [**]{} [**155**]{}, 243 (2004).
X.-B. [Wu]{}, W. [Zhang]{} and X. [Zhou]{}, [*Chinese Journal of Astronomy and Astrophysics*]{} [**4**]{}, 17 (2004).
S. [Kitsionas]{}, E. [Hatziminaoglou]{}, A. [Georgakakis]{} and I. [Georgantopoulos]{}, [**]{} [**434**]{}, 475 (2005).
N. M. [Ball]{}, R. J. [Brunner]{}, A. D. [Myers]{}, N. E. [Strand]{}, S. [Alberts]{}, D. [Tcheng]{} and X. [Llor[à]{}]{}, [**]{} [**663**]{}, p. 774 (2007).
N. D. [Kumar]{}, [Machine learning techniques for astrophysical modelling and photometric redshift estimation of quasars in optical sky surveys]{}, Master’s thesis, Oxford University (2008).
C. [Wolf]{}, [**]{} [**397**]{}, 520 (2009).
C. [Wolf]{}, L. [Wisotzki]{}, A. [Borch]{}, S. [Dye]{}, M. [Kleinheinrich]{} and K. [Meisenheimer]{}, [**]{} [**408**]{}, 499 (2003).
M. [Salvato]{} [*et al.*]{}, [**]{} [**690**]{}, 1250 (2009).
J. F. [Ram[í]{}rez]{}, O. [Fuentes]{} and R. K. [Gulati]{}, [*Experimental Astronomy*]{} [**12**]{}, 163 (2001).
T. [Solorio]{}, O. [Fuentes]{}, R. [Terlevich]{} and E. [Terlevich]{}, [**]{} [**363**]{}, 543 (2005).
A. C. [Becker]{}, [*Astronomische Nachrichten*]{} [**329**]{}, p. 280 (2008).
S. G. [Djorgovski]{}, A. A. [Mahabal]{}, R. J. [Brunner]{}, R. R. [Gal]{}, S. [Castro]{}, R. R. [de Carvalho]{} and S. C. [Odewahn]{}, [Searches for Rare and New Types of Objects]{}, in [*Virtual Observatories of the Future*]{}, eds. R. J. [Brunner]{}, S. G. [Djorgovski]{} and A. S. [Szalay]{}, Astronomical Society of the Pacific Conference Series, Vol. 225 (2001).
M. [Bottino]{}, A. J. [Banday]{} and D. [Maino]{}, [**]{} [**389**]{}, 1190 (2008).
S. [Pires]{}, J. B. [Juin]{}, D. [Yvon]{}, Y. [Moudden]{}, S. [Anthoine]{} and E. [Pierpaoli]{}, [**]{} [**455**]{}, 741 (2006).
N. G. [Phillips]{} and A. Kogut, preprint, \[arXiv:astro-ph/0108234\] (2001).
D. J. [Rohde]{}, M. R. [Gallagher]{}, M. J. [Drinkwater]{} and K. A. [Pimbblet]{}, [ ** ]{} [**369**]{}, 2 (2006).
M. [Taylor]{} and A. I. [Diaz]{}, [On the Deduction of Galactic Abundances with Evolutionary Neural Networks]{}, in [*From Stars to Galaxies: Building the Pieces to Build Up the Universe*]{}, eds. A. [Vallenari]{}, R. [Tantalo]{}, L. [Portinari]{} and A. [Moretti]{}, Astronomical Society of the Pacific Conference Series, Vol. 374 (2007).
C. [Bogdanos]{} and S. [Nesseris]{}, [*Journal of Cosmology and Astro-Particle Physics*]{} [**5**]{}, p. 6 (2009).
R. [Li]{}, Y. [Cui]{}, H. [He]{} and H. [Wang]{}, [*Advances in Space Research*]{} [ **42**]{}, 1469 (2008).
J. F. [Mustard]{}, L. [Li]{} and G. [He]{}, [**]{} [**103**]{}, 19419 (1998).
G. [Lemson]{} and J. [Zuther]{}, [*Memorie della Societa Astronomica Italiana*]{} [**80**]{}, 342 (2009).
V. [Springel]{} [*et al.*]{}, [**]{} [**435**]{}, 629 (2005).
R. [Brun]{} and F. [Rademakers]{}, [*Nuclear Instruments and Methods in Physics Research A*]{} [**389**]{}, 81 (1997).
T. [Budav[á]{}ri]{} and A. S. [Szalay]{}, [**]{} [**679**]{}, 301 (2008).
C. E. [Cunha]{}, M. [Lima]{}, H. [Oyaizu]{}, J. [Frieman]{} and H. [Lin]{}, [**]{} [**396**]{}, 2379 (2009).
V. E. [Margoniner]{} and D. M. [Wittman]{}, [**]{} [**679**]{}, 31 (2008).
D. [Wittman]{}, [**]{} [**700**]{}, L174 (2009).
A. D. [Myers]{}, M. [White]{} and N. M. [Ball]{}, [**]{} [**399**]{}, 2279 (2009).
C. [van Breukelen]{} and L. [Clewley]{}, [**]{} [**395**]{}, 1845 (2009).
C. A. L. [Bailer-Jones]{}, K. W. [Smith]{}, C. [Tiede]{}, R. [Sordo]{} and A. [Vallenari]{}, [**]{} [**391**]{}, 1838 (2008).
J. [Vaidya]{}, C. [Clifton]{} and M. [Zhu]{}., [*[Privacy Preserving Data Mining]{}*]{} (Springer, New York, 2005).
C. C. [Aggarwal]{} and P. S. [Yu]{} (eds.), [*[Privacy-Preserving Data Mining: Models and Algorithms]{}*]{} (Springer, New York, 2008).
R. [Scranton]{}, A. J. [Connolly]{}, A. S. [Szalay]{}, R. H. [Lupton]{}, D. [Johnston]{}, T. [Budav[á]{}ri]{}, J. [Brinkmann]{} and M. [Fukugita]{}, preprint, \[arXiv:astro-ph/0508564\] (2005).
. [Ivezi[ć]{}]{}, J. A. [Tyson]{}, R. [Allsman]{}, J. [Andrew]{}, R. [Angel]{} and [for the LSST Collaboration]{}, preprint, \[arXiv/0805.2366\] (2008).
C. [Donalek]{}, A. [Mahabal]{}, S. G. [Djorgovski]{}, S. [Marney]{}, A. [Drake]{}, E. [Glikman]{}, M. J. [Graham]{} and R. [Williams]{}, [New Approaches to Object Classification in Synoptic Sky Surveys]{}, American Institute of Physics Conference Series Vol. 1082 (2008).
A. H. [Studenmund]{}, [*[Using Econometrics]{}*]{}, 2nd edn. (Addison-Wesley, New York, 2005).
A. [Mahabal]{}, S. G. [Djorgovski]{}, M. [Turmon]{}, J. [Jewell]{}, R. R. [Williams]{}, A. J. [Drake]{}, M. G. [Graham]{}, C. [Donalek]{}, E. [Glikman]{} and [Palomar-QUEST team]{}, [*Astronomische Nachrichten*]{} [**329**]{}, 288 (2008).
A. J. [Drake]{}, R. [Williams]{}, M. J. [Graham]{}, A. [Mahabal]{}, S. G. [Djorgovski]{}, R. R. [White]{}, W. T. [Vestrand]{} and J. [Bloom]{}, [VOEventNet: An Open Source of Transient Alerts for Astronomers.]{}, Bulletin of the American Astronomical Society Vol. 38 (2007).
. [Ivezi[ć]{}]{} [*et al.*]{}, [Parametrization and Classification of 20 Billion LSST Objects: Lessons from SDSS]{}, American Institute of Physics Conference Series Vol. 1082 (2008).
K. [Borne]{}, J. [Becla]{}, I. [Davidson]{}, A. [Szalay]{} and J. A. [Tyson]{}, [The LSST Data Mining Research Agenda]{}, American Institute of Physics Conference Series Vol. 1082 (2008).
M. A. C. [Perryman]{} [*et al.*]{}, [**]{} [**369**]{}, 339 (2001).
C. A. L. [Bailer-Jones]{}, [A Method for Exploiting Domain Information in Astrophysical Parameter Estimation]{}, in [*Astronomical Data Analysis Software and Systems XVII*]{}, eds. R. W. [Argyle]{}, P. S. [Bunclark]{} and J. R. [Lewis]{}, Astronomical Society of the Pacific Conference Series, Vol. 394 (2008).
S. G. [Djorgovski]{} [*et al.*]{}, [*Astronomische Nachrichten*]{} [**329**]{}, p. 263 (2008).
A. J. [Drake]{} [*et al.*]{}, [**]{} [**696**]{}, 870 (2009).
K. W. [Hodapp]{} [*et al.*]{}, [*Astronomische Nachrichten*]{} [**325**]{}, 636 (2004).
S. [Johnston]{} [*et al.*]{}, [*Publications of the Astronomical Society of Australia*]{} [**24**]{}, 174 (2007).
L. [Eyer]{} [*et al.*]{}, [Variability type classification of multi-epoch surveys]{}, American Institute of Physics Conference Series Vol. 1082 (2008).
M. C. [Kaczmarczik]{}, G. T. [Richards]{}, S. S. [Mehta]{} and D. J. [Schlegel]{}, [ ** ]{} [**138**]{}, 19 (2009).
A. [Mahabal]{} [*et al.*]{}, [Towards Real-time Classification of Astronomical Transients]{}, American Institute of Physics Conference Series Vol. 1082 (2008).
G. E. [Moore]{}, [*Electronics*]{} [**38**]{}, 114 (1965).
D. A. [Bader]{} (ed.), [*[Petascale Computing: Algorithms and Applications]{}*]{}, Computational Science Series (CRC Press, Boca Raton, FL, 2007).
G. [Amdahl]{}, [Validity of the Single Processor Approach to Achieving Large-Scale Computing Capabilities]{}, in [*Spring Joint Computer Conference*]{}, (AFIPS Press, Atlantic City, N.J., 1967).
K. [Ebcioglu]{}, V. [Saraswat]{} and V. [Sarkar]{}, [The IBM PERCS Project and New Opportunities for Compiler-Driven Performance via a New Programming Model]{}, [*Compiler-Driven Performance Workshop (CASCON 2004)*]{}, (2004).
A. S. [Szalay]{} [*et al.*]{}, [GrayWulf: Scalable Clustered Architecture for Data Intensive Computing]{}, [*Hawaii International Conference on System Sciences*]{} (IEEE Computer Society, Los Alamitos, CA, 2009).
A. S. [Szalay]{}, J. [Gray]{} and [Vandenberg, J.]{}, [Petabyte Scale Data Mining: Dream or Reality?]{}, [*SPIE Astronomy Telescopes and Instruments*]{}, Waikoloa, Hawaii, (2002).
S. M. [McConnell]{} and D. B. [Skillicorn]{}, [Distributed Data Mining for Astrophysical Datasets]{}, in [*Astronomical Data Analysis Software and Systems XIV*]{}, eds. P. [Shopbell]{}, M. [Britton]{} and R. [Ebert]{}, Astronomical Society of the Pacific Conference Series, Vol. 347 (2005).
A. A. [Freitas]{} and S. H. [Lavington]{}, [*Mining Very Large Databases with Parallel Processing*]{} (Kluwer Academic Publishers, 1998).
H. [Kargupta]{} and P. [Chan]{}, [*[Advances in Distributed and Parallel Knowledge Discovery]{}*]{} (AAAI/MIT Press, Cambridge, MA, 2000).
M. J. [Zaki]{} and C. [Ho]{} (eds.), [*Large-scale Parallel Data Mining*]{}, Lecture Notes in Artificial Intelligence, State-of-the-Art-Survey, Vol. 1759 (Springer, New York, 2002).
K. [Bhaduri]{}, K. [Liu]{}, H. [Kargupta]{} and J. [Ryan]{}, [Distributed Data Mining Bibliography]{} Online bibliography, (2006).
R. [Jin]{}, G. [Yang]{} and G. [Agrawal]{}, [*IEEE Transactions On Knowledge and Data Engineering*]{} [**17**]{}, 71 (2005).
N. [Gray]{}, [The Fact and Future of Semantic Astronomy]{}, in [*Astronomical Data Analysis Software and Systems XVII*]{}, eds. R. W. [Argyle]{}, P. S. [Bunclark]{} and J. R. [Lewis]{}, Astronomical Society of the Pacific Conference Series, Vol. 394 (2008).
M. L. [Norman]{}, G. L. [Bryan]{}, R. [Harkness]{}, J. [Bordner]{}, D. [Reynolds]{}, B. [O’Shea]{} and R. [Wagner]{}, preprint, \[arXiv/0705.1556\] (2007).
R. J. [Brunner]{}, V. [Kindratenko]{} and A. D. [Myers]{}, [*[Developing and Deploying Advanced Algorithms to Novel Supercomputing Hardware]{}*]{}, tech. rep., NASA (2007).
E. L. [Gomez]{}, H. L. [Gomez]{} and J. [Yardley]{}, preprint, \[arXiv/0903.0266\] (2009).
A. W. [Moore]{} [*et al.*]{}, [Fast Algorithms and Efficient Statistics: N-Point Correlation Functions]{}, in [*Mining the Sky*]{}, eds. A. J. [Banday]{}, S. [Zaroubi]{} and M. [Bartelmann]{} (2001).
D. [Gao]{}, Y. [Zhang]{} and Y. [Zhao]{}, [The Application of kd-tree in Astronomy]{}, in [*Astronomical Data Analysis Software and Systems XVII*]{}, eds. R. W. [Argyle]{}, P. S. [Bunclark]{} and J. R. [Lewis]{}, Astronomical Society of the Pacific Conference Series, Vol. 394 (2008).
A. G. [Gray]{}, A. W. [Moore]{}, R. C. [Nichol]{}, A. J. [Connolly]{}, C. [Genovese]{} and L. [Wasserman]{}, [Multi-Tree Methods for Statistics on Very Large Datasets in Astronomy]{}, in [*Astronomical Data Analysis Software and Systems (ADASS) XIII*]{}, eds. F. [Ochsenbein]{}, M. G. [Allen]{} and D. [Egret]{}, Astronomical Society of the Pacific Conference Series, Vol. 314 (2004).
Y. [Shirasaki]{}, M. [Ohishi]{}, Y. [Mizumoto]{}, M. [Tanaka]{}, S. [Honda]{}, M. [Oe]{}, N. [Yasuda]{} and Y. [Masunaga]{}, [Structured Query Language for Virtual Observatory]{}, in [*Astronomical Data Analysis Software and Systems XIV*]{}, eds. P. [Shopbell]{}, M. [Britton]{} and R. [Ebert]{}, Astronomical Society of the Pacific Conference Series, Vol. 347 (2005).
S. [Derriere]{} [*et al.*]{}, [UCD in the IVOA context]{}, in [*Astronomical Data Analysis Software and Systems (ADASS) XIII*]{}, eds. F. [Ochsenbein]{}, M. G. [Allen]{} and D. [Egret]{}, Astronomical Society of the Pacific Conference Series, Vol. 314 (2004).
P. [Dowler]{}, S. [Gaudet]{}, D. [Durand]{}, R. [Redman]{}, N. [Hill]{} and S. [Goliath]{}, [Common Archive Observation Model]{}, in [*Astronomical Data Analysis Software and Systems XVII*]{}, eds. R. W. [Argyle]{}, P. S. [Bunclark]{} and J. R. [Lewis]{}, Astronomical Society of the Pacific Conference Series, Vol. 394 (2008).
J. [Gray]{}, D. T. [Liu]{}, M. [Nieto-Santisteban]{}, A. S. [Szalay]{}, D. [DeWitt]{} and G. [Heber]{}, [*[Scientific Data Management in the Coming Decade]{}*]{}, Technical Report MSR-TR-2005-10, Microsoft Research (2005).
J. [Gray]{}, A. S. [Szalay]{}, A. R. [Thakar]{}, P. Z. [Kunszt]{}, C. [Stoughton]{}, D. [Slutz]{} and J. [vandenBerg]{}, preprint, \[arXiv:cs/0202014\] (2002).
C. [Vignali]{}, F. [Fiore]{}, A. [Comastri]{}, M. [Brusa]{}, R. [Gilli]{}, N. [Cappelluti]{}, F. [Civano]{} and G. [Zamorani]{}, [Multi-wavelength data handling in current and future surveys: the possible role of Virtual Observatory]{}, in [*Multi-wavelength Astronomy and Virtual Observatory*]{}, ed. [D. Baines & P. Osuna]{} (2009).
E. A. [Gonz[á]{}lez-Solares]{} [*et al.*]{}, [**]{} [**388**]{}, 89 (2008).
M. [Brescia]{} [*et al.*]{}, [*Memorie della Societa Astronomica Italiana*]{} [**80**]{}, p. 565 (2009).
T. [Kitching]{}, A. [Amara]{}, A. [Rassat]{} and A. [Refregier]{}, preprint, \[arXiv/0901.3143\] (2009).
I. V. [Chilingarian]{}, [Virtual Observatory for Astronomers: Where Are We Now?]{}, in [*Multi-wavelength Astronomy and Virtual Observatory*]{}, ed. [D. Baines & P. Osuna]{} (2009).
W. B. [Landsman]{}, [The IDL Astronomy User’s Library]{}, in [*Astronomical Data Analysis Software and Systems II*]{}, eds. R. J. [Hanisch]{}, R. J. V. [Brissenden]{} and J. [Barnes]{}, Astronomical Society of the Pacific Conference Series, Vol. 52 (1993).
M. B. [Taylor]{}, [TOPCAT & STIL: Starlink Table/VOTable Processing Software]{}, in [*Astronomical Data Analysis Software and Systems XIV*]{}, eds. P. [Shopbell]{}, M. [Britton]{} and R. [Ebert]{}, Astronomical Society of the Pacific Conference Series, Vol. 347 (2005).
M. [Comparato]{}, U. [Becciani]{}, A. [Costa]{}, B. [Larsson]{}, B. [Garilli]{}, C. [Gheller]{} and J. [Taylor]{}, [**]{} [**119**]{}, 898 (2007).
N. [Urunkar]{}, A. K. [Kembhavi]{}, A. [Navelkar]{}, J. [Pandya]{}, V. [Moosani]{}, P. [Nair]{} and M. [Shaikh]{}, [*Highlights of Astronomy*]{} [**14**]{}, 620 (2007).
J. D. [Taylor]{}, T. [Boch]{}, M. [Comparato]{}, M. [Taylor]{}, N. [Winstanley]{} and R. G. [Mann]{}, [Binding Applications Together with PLASTIC]{}, in [ *Astronomical Data Analysis Software and Systems XVI*]{}, eds. R. A. [Shaw]{}, F. [Hill]{} and D. J. [Bell]{}, Astronomical Society of the Pacific Conference Series, Vol. 376 (2007).
M. B. [Taylor]{}, [STILTS - A Package for Command-Line Processing of Tabular Data]{}, in [*Astronomical Data Analysis Software and Systems XV*]{}, eds. C. [Gabriel]{}, C. [Arviset]{}, D. [Ponz]{} and S. [Enrique]{}, Astronomical Society of the Pacific Conference Series, Vol. 351 (2006).
M. [Borkin]{}, A. [Goodman]{}, M. [Halle]{} and D. [Alan]{}, [Application of Medical Imaging Software to 3D Visualization of Astronomical Data]{}, in [ *Astronomical Data Analysis Software and Systems XVI*]{}, eds. R. A. [Shaw]{}, F. [Hill]{} and D. J. [Bell]{}, Astronomical Society of the Pacific Conference Series, Vol. 376 (2007).
R. [Scranton]{} [*et al.*]{}, preprint, \[arXiv/0709.0752\] (2007).
D. G. [Barnes]{}, C. J. [Fluke]{}, P. D. [Bourke]{} and O. T. [Parry]{}, [ *Publications of the Astronomical Society of Australia*]{} [**23**]{}, 82 (2006).
C. J. [Fluke]{}, D. G. [Barnes]{} and N. T. [Jones]{}, [*Publications of the Astronomical Society of Australia*]{} [**26**]{}, 37 (2009).
S. [Levy]{}, [Interactive 3-D visualization of particle systems with Partiview]{}, in [*Astrophysical Supercomputing using Particle Simulations*]{}, eds. J. [Makino]{} and P. [Hut]{}, IAU Symposium, Vol. 208 (2003).
T. [Szalay]{}, V. [Springel]{} and G. [Lemson]{}, preprint, \[arXiv/0811.2055\] (2008).
P. [Hut]{}, [Virtual Laboratories and Virtual Worlds]{}, IAU Symposium Vol. 246 (2008).
T. [Ebisuzaki]{}, J. [Makino]{}, T. [Fukushige]{}, M. [Taiji]{}, D. [Sugimoto]{}, T. [Ito]{} and S. K. [Okumura]{}, [**]{} [**45**]{}, 269 (1993).
E. [Gaburov]{}, S. [Harfst]{} and S. [Portegies Zwart]{}, [*New Astronomy*]{} [ **14**]{}, 630 (2009).
R. G. [Belleman]{}, J. [B[é]{}dorf]{} and S. F. [Portegies Zwart]{}, [*New Astronomy*]{} [**13**]{}, 103 (2008).
S. [Ord]{}, L. [Greenhill]{}, R. [Wayth]{}, D. [Mitchell]{}, K. [Dale]{}, H. [Pfister]{} and R. G. [Edgar]{}, preprint, \[arXiv/0902.0915\] (2009).
V. [Garcia]{}, E. [Debreuve]{} and M. [Barlaud]{}, preprint, \[arXiv/0804.1448\] (2008).
S. D. [Brown]{}, R. J. [Francis]{}, J. [Rose]{} and Z. G. [Vranesic]{}, [ *[Field-Programmable Gate Arrays]{}*]{}, The Springer International Series in Engineering and Computer Science (Springer, New York, 1992).
D. [Buell]{}, T. [El-Ghazawi]{}, K. [Gaj]{} and V. [Kindratenko]{}, [*Computer*]{} [ **40**]{}, 23 (2007).
E. [Won]{}, [*Nuclear Instruments and Methods in Physics Research A*]{} [ **581**]{}, 816 (2007).
M. [Freeman]{}, M. [Weeks]{} and J. [Austin]{}, [Hardware implementation of similarity functions]{}, in [*IADIS AC*]{}, 2005.
M. [Scarpino]{}, [*[Programming the Cell Processor: For Games, Graphics, and Computation]{}*]{} (Prentice Hall PTR, New York, 2008).
[^1]: <http://weka.wiki.sourceforge.net/ARFF>
[^2]: For many astronomical applications, one might more properly call it a training [*sample*]{}, but the term training set is in widespread use, so we use that here to avoid confusion.
[^3]: <http://www.mpa-garching.mpg.de/millennium>
[^4]: <http://www.ncsa.uiuc.edu/BlueWaters>
[^5]: <http://www.ivoa.net/Documents>
[^6]: <http://www.us-vo.org>
[^7]: <http://www.astro.princeton.edu/~rhl/sm>
[^8]: <http://www.astro.caltech.edu/~tjp/pgplot>
[^9]: <http://www.gnuplot.info>
[^10]: <http://idlastro.gsfc.nasa.gov>
[^11]: <http://www.paraview.org>
[^12]: <http://blogs.discovermagazine.com/cosmicvariance/2008/11/03/guest-post-george-djorgovski-a-new-world-overture>
[^13]: <http://www.nvidia.com/cuda>
[^14]: <http://www.khronos.org/opencl>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Understanding the magnetic configuration of the source regions of coronal mass ejections (CMEs) is vital in order to determine the trigger and driver of these events. Observations of four CME productive active regions are presented here, which indicate that the pre-eruption magnetic configuration is that of a magnetic flux rope. The flux ropes are formed in the solar atmosphere by the process known as flux cancellation and are stable for several hours before the eruption. The observations also indicate that the magnetic structure that erupts is not the entire flux rope as initially formed, raising the question of whether the flux rope is able to undergo a partial eruption or whether it undergoes a transition in specific flux rope configuration shortly before the CME.'
---
Introduction
============
Twisted bundles of magnetic field lines known as flux ropes are central to all models of coronal mass ejections. They are invoked as being present either before the eruption onset (e.g. [@Toeroekkliem05 Török & Kliem 2005]) or being formed during the eruption itself (e.g. [@Antiochos_etal99 Antiochos et al. 1999]). In light of this, one way to discriminate between these two sets of CME models is to determine the pre-eruption magnetic configuration. However, there is an inherent difficulty in the confident identification of flux ropes as it is not currently possible to directly measure the magnetic field above the photosphere/chromosphere. Instead, proxies for the presence of flux ropes need to be used, either from solar observations or from reconstructions of the coronal magnetic field using the photospheric field as the boundary condition.
One well developed observational approach to investigate the pre-eruption magnetic configuration is to study so-called sigmoidal active regions ([@Rustkumar1996 Rust & Kumar 1996]). Sigmoidal regions contain S shaped EUV and X-ray emission structures and are regions that have a very high likelihood of producing a coronal mass ejection. Sigmoids are seen as predictors of an eruption ([@Canfieldetal99 Canfield 1999]). The sigmoid can be S or reverse S shaped, depending on the chirality of the magnetic field in which it forms ([@Pevtsovetal97 Pevtsov 1997]). [@Rustkumar1996 Rust & Kumar (1996)] suggested a link between sigmoids and kinking flux ropes, which has been a lively area of research ever since. Theoretical expectations from modelled flux ropes show that layers of enhanced current, and presumably heating, are located at the interface between a flux rope and its surrounding magnetic arcade ([@Titovdemoulin99 Titov & D[é]{}moulin 1999]). These layers should appear S shaped when viewed from above, building the case that sigmoids represent flux ropes in the solar atmosphere.
Recent investigations into the magnetic configuration of sigmoids have shown support that, at least a sub-set of sigmoids, do indeed indicate the presence of a magnetic flux rope. [@Greenkliem2009 Green & Kliem (2009)] showed that continuous S shaped threads that make an inverse crossing of the photospheric polarity inversion line (PIL) in their centre strongly suggest field lines that spiral around the flux rope. These threads additionally cross the PIL in the normal direction in the sigmoid’s elbows. They trace field lines at the periphery of the rope and are likely to extend down to the lower atmosphere, where they form bald patches.
As well as forming in active regions, sigmoids have been observed on a smaller scale in X-ray bright points ([@Mandrinietal05 Mandrini 2005]) and on a larger scale in the quiet Sun ([@Jiangetal07 Jiang 2007]). It has been suggested that there are different types of sigmoid relating to their lifetime (transient and long-lived) or detailed observational appearance (continuous S threads or double J’s that overall look S shaped) ([@Pevtsov2002a Pevtsov 2002a]).
In summary, flux ropes can be investigated in the solar atmosphere prior to a coronal mass ejection by using observations of sigmoids. These features provide an opportunity to follow the evolution of the magnetic configuration and investigate how the flux rope forms, the magnetic flux content of the rope and the specific magnetic configuration. In this paper we focus on flux rope formation in a small sample of sigmoidal active regions in the days leading up to a coronal mass ejection.
Observations of flux rope formation
===================================
Here we present the evolution of four active regions; (1) NOAA region 10930 that was observed on the Sun during December 2006, (2) NOAA region 8005 that was seen during December 1996, (3) an un-numbered region that was seen in February 2007 and (4) NOAA active region 10977 that was on the disk during December 2007. This sample includes one region which forms a sigmoid during a flux emergence event and three regions which show sigmoidal structure forming during the decay phase of the region.
NOAA active region 10930
------------------------
Active region 10930 rotated over the eastern limb of the Sun on 5 December 2006. The region contained a large negative polarity sunspot with a corresponding dispersed and spotless positive magnetic field to the east. Immediately to the south of the negative polarity sunspot was a smaller magnetic bipole. On 10 December 2006, new flux began to emerge in the location of the smaller bipole. The positive polarity of the emerging flux underwent an eastward motion and a strong counter clockwise rotation of up to 540 degrees between 10 and 14 December ([@Minchae2009 Min & Chae 2009]). The positive polarity of the emerging flux was directly next to the pre-existing negative spot meaning that flux cancellation was likely to be taking place.
The XRT/Hinode ([@Golubetal07 Golub 2007]) data show that as soon as the flux emergence begins, the magnetic field of the positive sunspot formed connections with the pre-existing negative sunspot producing a magnetic arcade between them. The shear in this arcade field increased rapidly, most likely due to the rapid eastward motion of the emerging flux and its strong counter-clockwise rotation. By 23:15 UT on 11 December the arcade field had taken on the appearance of a ’double J’ configuration and by 19:00 UT on 12 December a continuous S shaped structure was seen. See the top row of Fig. \[fig1\] for the coronal evolution of this region. On 13 December 2006 a GOES X3.4 class flare occurred in the sigmoidal region at around 02:15 UT. A coronal mass ejection was associated with this flare.
NOAA active region 8005
-----------------------
NOAA active region 8005 had a bipolar magnetic configuration. The region emerged on the far side of the Sun and as it rotated over the limb it was already in its decay phase with dispersed magnetic polarities and no sunspots. The photospheric field exhibited dispersal of the magnetic polarities and no episodes of flux emergence. On many occasions opposite polarity fragments approached the active region’s polarity inversion line and cancelled. See [@Greenkliem2009 Green & Kliem (2009)] and [@Mackayetal11 Mackay (2011)] for details of the evolution of the photospheric magnetic field in this region. In the days leading up to the eruption the active region loops as seen in the soft X-ray data had an overall S shape which evolved from a sheared arcade, to a double J shape (by 19 December 07:24 UT) to a sigmoid with continuous S shaped threads being observed by 10:37 UT on 19 December. See the second from top row of Fig. \[fig1\] for the coronal evolution of this region. The active region produced an eruption and a GOES C2.3 class flare on 19 December 1996 at 15:21 UT.
February 2007 region
--------------------
In February 2007 a sigmoid was observed in a very dispersed bipolar active region that had no NOAA number assigned to it. The region produced a coronal mass ejection on 12 February 2007 beginning around 07:00 UT. There was no flare emission associated with this eruption. In the days leading up to the eruption the active region evolved from a double J configuration to having continuous S shaped loops by end of the day on 11 February 2007. See the second row from the bottom of Fig. \[fig1\] for the coronal evolution of this region. The evolution of the photospheric magnetic field showed ongoing dispersal and episodes of flux cancellation at the polarity inversion line. For more details on the evolution of this active region see [@Savchevaetal12 Savcheva (2012)].
NOAA active region 10977
------------------------
NOAA active region 10977 had a bipolar configuration and the whole lifetime of the active region was observed from emergence to dispersal into the surrounding quiet sun. The emergence began on 3 December 2007 and the flux concentrations began to disperse on 4 December 2007. During the decay phase of the active region the photospheric field was dominated by ongoing dispersal, an elongation of the polarities in the north-south direction and cancellation of flux along the polarity inversion line.
Hinode/XRT data show that the active region loops appeared to have relatively little shear during the emergence phase and into 5 December after the emergence had ceased. The shear began to build up during the decay phase when the loops became much more aligned to the polarity inversion line. Early on 6 December the region appeared as a region of double J shaped loops with some remnant arcade field in the south. Then, by 6 December 15:50 UT continuous S shaped sigmoidal threads were seen in the region. See the bottom row of Fig. \[fig1\] for the coronal evolution of this region. On 7 December at around 04:20 UT the region produced a coronal mass ejection. For a more detailed description of the evolution of this active region see [@Greenetal2011 Green (2011)].
![Evolution of sigmoidal active regions showing the three phases of evolution from flux emergence, to sheared arcade, to ’double J’ shaped loops and then finally the continuous S shaped threads of the sigmoid. Top row: NOAA region 10930. Second row from the top: NOAA region 8005. Second row from the bottom: un-numbered region observed on the disk during February 2007. Bottom row: NOAA region 10977.[]{data-label="fig1"}](green_figure1.eps){width="12.5cm"}
Filament formation
==================
All of the above active regions exhibited the formation of a filament along the polarity inversion line where the sigmoid formed. Big Bear Solar Observatory H-alpha data show that a filament was present in NOAA active region 10930 by 11 December 18:00 UT. In NOAA region 8005 the filament had formed by 18 December 17:40 UT. In the un-named region seen on the disk in February 2007, Kanzelh[ö]{}he Observatory data show that there was a filament present by 10 February 2007 12:00 UT. In NOAA region 10977 Hinode/SOT H-alpha data show that the filament was forming by 5 December around 20:00 UT. The filaments in all these regions formed during the phase where the coronal arcade was becoming more sheared. The presence of a filament indicates low lying loops that have dips, or twisted field lines which can support the dense plasma against gravity.
Coronal mass ejections from the sigmoidal regions
=================================================
The observation of the sigmoid allows the magnetic configuration to be probed and a time for the formation of the flux rope to be identified. The S shaped field lines can indicate the presence of magnetic field lines with dips which suggests the presence of a flux rope, even if it is not fully formed. The time between sigmoid (and flux rope) formation and the onset of the coronal mass ejection is given in Fig. \[fig2\]. To increase the study size this figure also includes the sigmoid and coronal mass ejection that occurred in the bright point study of [@Mandrinietal05 Mandrini (2005)] and the sigmoidal region NOAA 11047 that produced a coronal mass ejection and that was studied in [@Savchevaetal12 Savcheva (2012)]. The colours indicate whether the region was observed from the emergence phase, and hence whether all activity is seen, or whether the region emerged on the far-side so that aspects of the photospheric and coronal evolution may have been missed. In all cases the flux rope exhibits a stable phase of several hours between its formation and eruption.
![Bar chart showing the time between first observation of continuous S shaped loops in the regions and the time of the coronal mass ejection. The colours indicate different region characteristics. Red: isolated bipolar regions that have been followed from emergence to eruption and which produce the sigmoid and the eruption in their decay phase. Blue: isolated bipolar regions that emerged on the far side of the Sun and which produced the sigmoid and eruption during their decay phase. Black: a multipolar active region tracked from a flux emergence episode to the sigmoid formation and eruption and which formed the sigmoid during the flux emergence phase.[]{data-label="fig2"}](green_figure2.eps){width="9cm"}
In all four regions of this study the coronal mass ejection does not involve the filament material, as has been previously shown when sigmoidal active regions produce coronal mass ejections ([@Pevtsov2002b Pevtsov 2002b]). The events presented here show that the S shaped loops also do not rise during the eruption. Instead, the erupting structure in three out of the four sigmoidal regions is a faint linear or loop-like feature that rises up as the flare loops brighten underneath. See Fig. 3 in [@Suetal07 Su (2007)] for region 10930, [@McKenziecanfield2008 McKenzie & Canfield (2008)] for the case in February 2007 and [@Greenetal2011 Green (2011)] for region 10977.
Discussion
==========
There is growing observational support that a flux rope is present before the onset of a coronal mass ejection in some cases (see [@Aulanieretal10 Aulanier 2010], [@Greenetal2011 Green 2011] and references therein). Active regions that show a sigmoidal structure have been fruitful in revealing these flux ropes and their formation mechanism. The observations presented here strongly suggest that flux ropes in the solar atmosphere are built by the flux cancellation mechanism proposed by [@vBmartens1989 van Ballegooijen & Martens (1989)]. Flux cancellation involves reconnection low down in the solar atmosphere between converging opposite polarity fragments in a sheared arcade. The evolution of the coronal configuration is driven by the motions of the photospheric plasma and in this study is seen to pass through three stages as arcade field evolves into a flux rope. During the first stage shear in the coronal arcade field increases due to photospheric motions associated with flux emergence or flux dispersal and flux cancellation. Filaments start to form during this stage. During stage two there is an accumulation of a significant amount of axial flux running along the polarity inversion line as flux cancellation, further shearing and/or rotation of the magnetic polarities takes place. Remnant arcade field takes on the appearance of two J’s either side of this axial flux. In stage three, flux cancellation produces field lines that are twisted around the axial flux and which contribute poloidal flux to the rope. This flux cancellation scenario appears to be relevant in dispersed and isolated bipolar active regions in their decay phase, where the flux rope forms along the polarity inversion line of the bipole, or during the flux emergence phase of a multipolar active region where the flux rope forms along the polarity inversion line between neighbouring bipoles which are butted up against each other. The observations of the evolutionary phases suggest that sigmoids do not exhibit different types, rather their appearance evolves as the magnetic configuration changes.
The details of the eruptions from these regions are also important for understanding the pre-eruption magnetic configuration and the aspects of this configuration that are involved in the coronal mass ejection. The eruptions from these regions do not involve either the continuous S shaped threads or the filament. Since both are likely to involve field lines that are located in, or extend down to, the dense plasma of the lower atmosphere, it is not surprising that they are immobile. The flux rope cannot erupt in its entirety in this situation. The observations show that the structure which does erupt is instead a collection of loops that connect between the elbows of the S shaped threads and which can be seen before the onset of the eruption. These loops have been called a linear feature and tentatively been associated with the core of the erupting flux rope in [@Greenetal2011 Green (2011)]. However, in [@Aulanieretal10 Aulanier (2010)] the same feature is interpreted as being a consequence of heating in a current shell above the rope and called an erupting loop-like feature. Such an erupting structure was also seen by [@Mooreetal01 Moore (2001)] but has become more frequently observed with Hinode/XRT due to its large dynamic range and also with SDO due to the increased plasma temperature coverage. See, for example, [@Liuetal10 Liu (2010)] and [@Zharkovetal11 Zharkov (2011)].
The observations presented here suggest that the flux rope is able to partially erupt, allowing the accumulated axial flux to escape and become the erupting linear feature, whilst leaving the flux associated to the S shaped threads and filament behind. In other cases the flux rope may evolve prior to the eruption so that its underside is detached from the lower atmosphere, allowing it to erupt fully, as demonstrated in the simulations of [@Aulanieretal10 Aulanier (2010)].
1999, *ApJ*, 510, 484
2010, *ApJ*, 708, 314
1999, *GeoRL*, 26, 627
2007, *Solar Phys.*, 243, 63
2009, *ApJ*, 700, L83
2011, *A&A*, 526, 2
2007, *Solar Phys.*, 240, 77
2010, *ApJ*, 725, 84
2011, *ApJ*, 729, 97
2005, *A&A*, 434, 725
2008, *A&A*, 481, 65
2009, *Solar Phys.*, 258, 203
2001, *ApJ*, 552, 833
2002a, in P.C.H. Martens and D. Cauffman (eds.), “Multi-Wavelength Observations of Coronal Structure and Dynamics – Yohkoh 10th Anniversary Meeting”, p.125
2002b, *Solar Phys.*, 207, 111
1997, *ApJ*, 481, 973
1996, *ApJ*, 464, 199
2012, *ApJ*, 759, 105
2007, *PASJ*, 59, 785
1999, *A&A*, 351, 707
2005, *ApJ*, 630, L97
1989, *ApJ*, 343, 971
2011, *ApJ*, 741, L35
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Peruru Subrahmanya Swamy, Radha Krishna Ganti, Krishna Jagannathan, [^1][^2]'
bibliography:
- 'myreferences.bib'
title: 'Adaptive CSMA under the SINR Model: Efficient Approximation Algorithms for Throughput and Utility Maximization'
---
CSMA, Gibbs distribution, Bethe approximation, Distributed algorithm, Wireless ad hoc network
Introduction
============
Model and Preliminaries {#sec_model}
=======================
The local Gibbsian problems {#local_gibbs}
===========================
Review of the Bethe approximation {#preliminaries_bethe}
=================================
Equivalence of Local Gibbsian method and the Bethe approximation {#gibbs_bethe}
================================================================
Special case - Conflict graph model {#special_cases}
===================================
Utility Maximization {#util_max}
====================
Numerical results {#simulations}
=================
Conclusions {#conc}
===========
Appendix
========
[^1]: A part of this work [@allerton_bethe] has been presented at IEEE Annual Allerton Conference on Communication, Control, and Computing (Allerton) 2015, held at Monticello, IL, USA
[^2]: P. S. Swamy, R. K. Ganti and K. Jagannathan are with the Department of Electrical Engineering, IIT Madras, Chennai, India 600036. Email:$\{$p.swamy, rganti, krishnaj$\}[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, I introduce weak representations of a Lie groupoid $G$. I also show that there is an equivalence of categories between the categories of 2-term representations up to homotopy and weak representations of $G$. Furthermore, I show that any VB-groupoid is isomorphic to an action groupoid associated to a weak representation on its kernel groupoid; this relationship defines an equivalence of categories between the categories of weak representations of $G$ and the category of VB-groupoids over $G$.'
address: 'Department of Mathematics, University of Illinois, Urbana, IL 61801'
author:
- Seth Wolbert
title: 'Weak representations, representations up to homotopy, and VB-groupoids'
---
Introduction
============
For the duration of this paper, fix a Lie groupoid ${\xymatrix{G=\{G_1 \ar@<3pt>[r] \ar@<-3pt>[r] & G_0\}}}$. Representations up to homotopy of a Lie groupoid were first defined by Arias Abad and Crainic in [@AC]. Their definition, designed as a global model of their representations of Lie algebroids, [@ACAlg], allow one to address the inability to generalize many constructions available for group actions. Indeed, there is no natural Lie groupoid action of a groupoid $G$ on its algebroid $A$ which one may call the adjoint action. However, as noted in [@AC], there is a natural notion of an adjoint representation up to homotopy.
Roughly speaking, one may think of representations up to homotopy as groupoid actions that fail to satisfy the axioms of an action in a controlled manner. For a Lie groupoid $G$, a representation up to homotopy is formally defined as degree one operator $D$ on the bigraded right $C^\bullet(G)$-module $C(G;E)^\bullet$ associated to a graded vector bundle $E^\bullet \to G_0$. This degree one operator must satisfy $D^2=0$ as well as a certain Leibnitz type rule (see Section \[s:Background\] for more details).
In practice, one generally works with the homogeneous pieces of this degree one operator. These may be interpreted as a cochain structure $\delta^i:E^i\to E^{i+1}$ between the pieces of $E^\bullet$, a smooth collection of chain maps $\lambda_g^\bullet:E^\bullet \to E^\bullet$ (called a [*quasi-action*]{} of $G$), a smooth collection of chain homotopies $\Omega_{g,h}:\lambda^\bullet_g\circ \lambda^\bullet_h \Rightarrow \lambda^\bullet_{gh}$, and higher order coherence pieces.
Gracia-Saz and Mehta [@GM] showed that to any 2-term representation up to homotopy corresponds to a [*VB-groupoid over $G$*]{}: a map of Lie groupoids $$\pi:{\xymatrix{V=\{V_1 \ar@<3pt>[r] \ar@<-3pt>[r] & V_0\}}} \to {\xymatrix{G=\{G_1 \ar@<3pt>[r] \ar@<-3pt>[r] & G_0\}}}$$ (i.e., a smooth functor) such that the corresponding maps $\pi_1:V_1\to G_1$ and $\pi_0:V_0\to G_0$ on morphisms and objects are vector bundles and the structure maps are all maps of vector bundles covering the structure maps of $G$. VB-groupoids were first introduced by Pradines [@P]. A simple example is the tangent groupoid associated to any Lie groupoid.
In addition to building a VB-groupoid from any representation up to homotopy, Gracia-Saz and Mehta showed that this establishes a bijection, up to isomorphism between 2-term representations up to homotopy of $G$ and VB-groupoids over $G$. This was result was later extended by del Hoyo and Ortiz [@dHO], who showed that this equivalence in fact corresponds to an equivalence of categories between the categories of 2-term representations up to homotopy of $G$ and VB-groupoids over $G$.
There are some technical issues with representations up to homotopy. Indeed, while the homogeneous pieces may be interpreted as quasi-actions and smooth collections of homotopy operators, there are higher coherence conditions on these objects that defy geometric interpretation. In addition, there are some notable issues with defining right representations up to homotopy.
Also at issue is the lack of greater context. Indeed, the definition of a representation up to homotopy relies heavily on the linear structure of the associated graded vector bundle. It is not clear in general what an “action up to homotopy” on a more general graded manifold should look like.
On the other hand, weak representations of $G$ do not share these issues. For a weak representation, the “space” $G$ acts on is replaced with a [*linear Lie groupoid bundle*]{} $\pi:V\to G_0$; roughly speaking, one should think of $V$ as a smooth collection of linear Lie groupoids. $G$ acts on $V$ via a [*weak action*]{}: this is an action where each groupoid morphism $g\in G_1$ acts via linear equivalences of categories $A_g:\pi^{-1}(s(g))\to \pi^{-1}(t(g))$ while each pair $(g,h)$ of composable arrows corresponds to a linear natural isomorphism between $A_g\circ A_h$ and $A_{gh}$. One may equivalently think of this as a (smooth) pseudofunctor from $G$ to the 2-category of equivalences between the fibers of $V$.
The main purpose of this paper is to show that the category of 2-term representations up to homotopy of $G$ ${{\sf RutH_2}}(G)$ and the category of weak representations of $G$ ${{\sf WRep}}(G)$ are equivalent. Additionally, we show that the category of weak representations of $G$ ${{\sf WRep}}(G)$ and the category of ${{\sf VB}}(G)$ are equivalent. This equivalence is naturally induced by the construction of action groupoids associated to each weak action of $G$ on a groupoid.
This paper is organized as follows: in Section \[s:Background\], we discuss the background associated to representations up to homotopy and VB-groupoids in greater detail. After this has been done, we may give a more detailed description as to the potential issues with the definition of representation up to homotopy. In Section \[s:2-termandLLGB\], we discuss the equivalence between 2-term complexes of vector bundles and linear Lie groupoid bundles. In Section \[s:WReps\], we define weak groupoid actions and weak representations. Additionally, we show that the equivalence between 2-term complexes of vector bundles and Lie groupoid bundles extends to an equivalence between the categories of 2-term representations up to homotopy of $G$ and weak representations of $G$. Finally, in Section \[s:VBsandWR\], we define the action groupoids associated to weak groupoid actions and show this construction defines an equivalence of categories between the category of weak representations up to homotopy of $G$ and VB-groupoids over $G$. 5 pt [**Notation and Conventions:**]{} As stated above, for the duration of this paper, fix a Lie groupoid ${\xymatrix{G=\{G_1 \ar@<3pt>[r] \ar@<-3pt>[r] & G_0\}}}$. Let $s,t:G_1\to G_0$ be the source and target maps of $G$. We assume that the reader is somewhat familiar with the definition of 2-categories. 2-categories and their functors are always assumed to be strict. For all $n>0$, denote by $G_n$ the collection of $n$ composable arrows; i.e., $G_p:=\{(g_1,\ldots, g_n) \in G_1\times\ldots G_1 \; | \; s(g_i)=t(g_{i+1})\}$. 5 pt [**Acknowledgments:**]{} The author would like to acknowledge Eugene Lerman, Rui Loja Fernandes, and Daniel Berwick-Evans for numerous insightful conversations regarding representations up to homotopy and 2-category theory.
Background {#s:Background}
==========
In this section, I will briefly discuss representations up to homotopy and their relationship to VB-groupoids. Essentially, this is a collection of results from [@AC], [@dHO], and [@GM] that will motivate the remainder of the paper.
For each $p$, define the maps $s_p,t_p:G_p\to G_0$ by $t_p(g_1,\ldots, g_p)=t(g_1)$ and $s_p(g_1,\ldots, g_p)=s(g_p)$. In other words, these are the source and targets of the arrow resulting from the product of the string of composable arrows $(g_1,\ldots, g_p)$.
For every $n\geq 0$, [*the space of degree $n$ groupoid cochains*]{} is the space of smooth maps $C^n(G):=C^\infty(G_n)$ with the standard ${\mathbb{R}}$-vector space structure. The coboundary map $\delta:C^\bullet(G)\to C^{\bullet+1}(G)$ for this cochain complex is defined as follows:
- For $n=0$: $\delta f(g):=f(s(g))-f(t(g))$.
- For $n>0$: $$\begin{aligned}
\displaystyle \delta f (g_0,\ldots, g_{n})&:=\\ f(g_1,\ldots, g_n) + \sum_{i=1}^n (-1)^i f(g_0,\ldots, g_{i-1}g_i, \ldots g_n) &+(-1)^{n+1}f(g_0,\ldots, g_{n-1}).\end{aligned}$$
Let $\pi:E\to G_0$. Then [*the space of degree $n$ groupoid cochains valued in $E$*]{} is the vector space of sections $C^n(G;E):=\Gamma(t_n^*E)$ (or, in the case where $n=0$, simply the collection of sections $\Gamma(E)$).
Note that $C^n(G;E)$ is a right $C(G)$-module: for $\omega\in C^p(G;E)$ and $f\in C^q(G)$, $\omega \star f$ is the degree $p+q$ groupoid cochain valued in $E$ defined by $$(\omega \star f)(g_1,\ldots, g_{p+q}):= \omega(g_1,\ldots, g_p)\star f(g_{p+1}, \ldots, g_{p+q})$$
$C^\bullet(G;E)$ contains as a subspace [*the normalized groupoid cochains valued in $E$*]{}. These are cochains $\eta\in C^n(G;E)$ for which $\eta(g_1,\ldots, g_n)=0$ whenever one of the $g_i$ is a unit. More succinctly, for each degeneracy map $s_i:G_{n-1}\to G_n$, we require that $s_i^*\eta=0$. The space of all degree $n$ normalized groupoid cochains is denoted by $\widehat{C}^n(G;E)$.
Let $\pi:E\to G_0$ be a vector bundle. Then a [*quasi-action*]{} of $G$ on $\pi:E\to G_0$ is a smooth map $\lambda:G_1\times_{s,G_0,\pi}G_1\to E$, such that the restriction $\lambda_g:=\lambda(g,\cdot)$ for each $g\in G$ is a linear map $\lambda_g:E_{s(g)}\to E_{t(g)}$.
A quasi-action $\lambda$ is [*unital*]{} if for every $x\in G_0$ with associated unit $u(x)$, $\lambda_{u(x)}=\mathrm{id}_{E_x}$.
\[p:acttoop\] Suppose that $\pi:E\to G_0$ is a vector bundle. Then there is a bijective correspondence $\lambda\mapsto D_\lambda$ between quasi-actions $\lambda$ of $G$ on $E$ and linear degree one operators $D_\lambda$ on $C^\bullet(G;E)$ satisfying the following graded Leibniz identity: $$D_\lambda(f\star\omega)= \omega \star \delta(f)+ (-1)^{|\eta|} D_\lambda\omega \star f.$$ Furthermore:
- The quasi-action $\lambda$ is unital if and only if the degree one operator $D_\lambda$ preserves $\widehat{C}^\bullet(G;E)$
- The quasi-action $\lambda$ is an action if and only if it is unital and $(D_\lambda)^2=0$.
This proposition motivates the definition of a representation up to homotopy.
Suppose that $\displaystyle \pi:E=\bigoplus_{i=0}^nE^i \to G_0$ is a graded vector bundle. For each $n$, define $\displaystyle C(G;E)^n:=\bigoplus_{i=0}^n C^{i}(G;E^{n-i})$. Then [*a representation up to homotopy*]{} is a [*total*]{} linear degree one operator $D$ on $C(G;E)^\bullet$ such that
- $D^2=0$; and
- $D$ satisfies the graded Leibnitz identity: $$D(f\star\omega)= \omega \star \delta(f)+ (-1)^{|\eta|} D\omega \star f.$$ Here, $|\eta|$ refers to the [*total*]{} degree of $\eta$; i.e., for $\eta\in C^k(G;E^l)$, $|\eta|=k+l$.
A [*map of representations up to homotopy*]{} between representations up to homotopy $(C(G;E)^\bullet, D)$ and $(C(G;E')^\bullet, D')$ is a linear map of right $C^\bullet(G)$-modules $\phi:C(G;E)^\bullet \to C(G;E')^\bullet$ such that $\phi\circ D=D'\circ \phi$.
By restricting the above degree one operator to homogeneous elements of $C(G;E)^\bullet$, one may decompose $D$ into a sequence of operators. As we are only interested in the case of 2-term representations up to homotopy for this paper, we will only bother making this explicit in this case. To do this, we must define transformation 2-cochains.
Let $\pi:E\to G_0$ and $\varpi:F\to G_0$ be two vector bundles over $G_0$. Then a [*transformation cochain from $E$ to $F$*]{} is a section $\Omega$ of the vector bundle $\mathrm{Hom}(s_2^*E,t_2^*F)\to G_2$. In other words, $\Omega$ associates to each pair of composable arrows $(g,h)\in G_2$ a linear map $\Omega_{g,h}:=\Omega(g,h):E_{s(h)}\to F_{t(g)}$.
A transformation cochain $\Omega$ from $E$ to $F$ is [*normalized*]{} if $\Omega(g,h)=0$ whenever $g$ or $h$ is a unit.
Now, we may describe 2-term representations up to homotopy in terms of homogeneous pieces:
\[propruth\] A [*representation up to homotopy*]{} of $G$ on a graded vector bundle $E^0\oplus E^1\to G_0$ is a uniquely determined by the tuple $(\delta:E^0\to E^1,\lambda^0,\lambda^1, \Omega)$ for $\delta:E^0\to E^1$ a map of vector bundles, $\lambda^0$ and $\lambda^1$ unital quasi-actions of $G$ on $E^0$ and $E^1$, and a normalized transformation cochain $\Omega$ from $E^1$ to $E^0$ satisfying for every triple of composable arrows $(g_1,g_2,g_3)\in G_3$:
- $\delta\circ \lambda_{g_1}^0 = \lambda_{g_1}^1\circ \delta$
- $\lambda^0_{g_1g_2}-\lambda^0_{g_1} \circ\lambda^0_{g_2} = \Omega_{g_1,g_2}\circ\delta$
- $\lambda^1_{g_1g_2}-\lambda^1_{g_1} \circ\lambda^1_{g_2} = \delta\circ\Omega_{g_1,g_2}$
- $\lambda^0_{g_1}\circ\Omega_{g_2,g_3}-\Omega_{g_1g_2,g_3}+\Omega_{g_1,g_2g_3}-\Omega_{g_1,g_2}\circ\lambda^1_{g_3}=0$
See Theorem 2.1, [@GM].
\[r:ruth\] The first condition of Proposition \[propruth\] says $\delta:E^0\to E^1$ is equivariant with respect to the quasi-actions on $E^0$ and $E^1$. Alternatively, it says that, thinking of $\delta:E^0\to E^1$ as a length 2 cochain complex of vector bundles over $G_0$, $\lambda_g^0$ and $\lambda_g^1$ define an automorphism of cochain complexes $\lambda_g$ for every $g\in G_1$.
The second and third conditions of Proposition \[propruth\] say that $\Omega_{g_1,g_2}$ functions as a chain homotopy relating the automorphisms of cochain complexes $\lambda_{g_1g_2}$ and $\lambda_{g_1}\lambda_{g_2}$ for every pair of composable arrows $(g_1,g_2)$.
Finally, the fourth condition of Proposition \[propruth\] is a cocycle type condition showing up as an extra consequence from the $D^2=0$ condition. This has an interesting interpretation in terms of weak actions of groupoids.
One may also decompose maps of representations up to homotopy in an effective manner.
\[propmapofruth\] Suppose that $(C(G;E)^\bullet,D)$ and $C(G;E')^\bullet,D')$ are representations up to homotopy for graded vector bundles $E=E^0\oplus E^1\to G_0$ and $E'=\oplus E'^0\oplus E'^1\to G_0$. If $(\delta:E^0\to E^1,\lambda^0,\lambda^1, \Omega)$ and $(\delta':E'^0\to E'^1, \lambda'^0,\lambda'^1, \Omega')$ are the unique tuples associated to $(C(G;E)^\bullet,D)$ and $C(G;E')^\bullet,D')$ (see Proposition \[propruth\]), then maps of representations up to homotopy $\phi:(C(G;E)^\bullet,D) \to C(G;E')^\bullet,D')$ are in bijective correspondence with triples $(\phi^0:E^0\to E'^0,\phi^1:E^1\to E'^1,\mu)$ for $\phi^i$ maps of vector bundles and $\mu\in C^1(G;s^*{\mathrm{Hom}}(E^1,E'^0))$ satisfying for all composable $(g,h)\in G_2$:
- $\phi^1\circ \delta = \delta'\circ \phi^0$
- $\phi^0\circ \lambda^0_g-\lambda'^0_g\circ \phi^0 = \mu_g\circ \delta$
- $\phi^1\circ \lambda^1_g-\lambda'^1_g \circ \phi_0 = \delta'\circ \mu_g$
- $\phi^0\circ \Omega_{g,h}+ \mu_g\circ \lambda^1_h + \lambda'^0_g\circ \mu_h = \mu_{gh} + \Omega_{g,h}' \circ \phi^1$
See Proposition 2.6, [@dHO].
\[r:mapofruth\] One may interpret the data of Proposition \[propmapofruth\] as follows. Again, thinking of $\delta:E^0\to E^1$, $\delta':E'^0\to E'^1$ as 2-term complexes of vector bundles, the first item says that $\phi^\bullet$ is a chain map. The second and third items say that $\mu_g$ functions for every $g$ as a homotopy operator between the two composition of chain maps $\phi^\bullet\circ\lambda^\bullet_g$ and $\lambda'^\bullet_g\circ \phi^\bullet$. Like the fourth item of Proposition \[propruth\], the fourth item above will have an interesting interpretation once we have switched to weak actions of groupoids.
From here forward, let ${{\sf RutH_2}}(G)$ denote the category of 2-term representations up to homotopy of $G$. We will primarily think of representations up to homotopy and their morphisms as the data detailed in Proposition \[propruth\] and Proposition \[propmapofruth\].
Now, we turn to discuss VB-groupoids. As noted by Gracia-Saz and Mehta [@GM], VB-groupoids may be thought of either as vector bundles in the category of groupoids or as groupoids in the category of vector bundles.
A [*VB-groupoid over $G$*]{} is a map of Lie groupoids $\pi:V\to G$ such that the maps on objects and morphisms $\pi_0:V_0\to G_0$ and $\pi_1:V_1\to G_1$ are vector bundles and all the structure maps of $V$ are maps of vector bundles.
A [*map of VB-groupoids over $G$*]{} $f: (\pi:V\to G)\to (\varpi:W\to G)$ is a map of Lie groupoids such that $f_0:V_0\to W_0$ and $f_1:V_1\to W_1$ are maps of vector bundles covering ${\mathrm{id}}_{G_0}$ and ${\mathrm{id}}_{G_1}$, respectively.
From here forward, let ${{\sf VB}}(G)$ denote the category of VB-groupoids over $G$.
For any Lie groupoid ${\xymatrix{H=\{H_1 \ar@<3pt>[r] \ar@<-3pt>[r] & H_0\}}}$, the tangent groupoid\
${\xymatrix{TH=\{TH_1 \ar@<3pt>[r] \ar@<-3pt>[r] & TH_0\}}}$ is a VB-groupoid over $H$.
If $\pi:E\to G_0$ is a vector bundle on which $G$ acts, then the action groupoid ${\xymatrix{G\rtimes E=\{G_1\times_{G_0} E \ar@<3pt>[r] \ar@<-3pt>[r] & E\}}}$ is a VB-groupoid over $G$.
As shown by Mehta and Gracia-Saz, to any 2-term representation up to homotopy, one may associate a VB-groupoid called [*the semi-direct product*]{}:
\[c:semidirect\] Let $(\delta:E^0\to E^1,\lambda^0,\lambda^1,\Omega)$ be (the data associated to) a representation up to homotopy (see Proposition \[propruth\]). Construct a VB-groupoid ${\xymatrix{V=\{V_1 \ar@<3pt>[r] \ar@<-3pt>[r] & V_0\}}}$ as follows:
- For objects, take $V_0=E^1$.
- For morphisms, take $V_1=s^*E^1\oplus t^*E^0$.
- For the source and target $\tilde{s},\tilde{t}:V_1\to V_0$ of $V$, define $$\tilde{s}(g,e_0,e_1)=e_1\text{ and }\tilde{t}(g,e_0,e_1)=\delta(e_0)+\lambda^1_g(e_1).$$
- For multiplication, define $$m((g,e_0,e_1),(h,f_0,f_1)):= (gh,e_0+\lambda^0_g(f_0)-\Omega_{g,h}(f_1),f_1)$$
- For units, for each $e\in E^1_x$, $x\in G_0$, take $\tilde{u}(e):=(u(x), 0,e)$.
- For inverses, take $(g,e_0,e_1)^{-1}:=(g^{-1},-\lambda^0_{g^{-1}}(e_0)+\Omega_{g^{-1},g}(e_1),\delta(e_0)+\lambda^1(e_1))$
\[t:ruthsandVB\] Let $\Psi:{{\sf RutH_2}}(G)\to {{\sf VB}}(G)$ be the functor defined as follows:
- On objects: let $\Psi(\delta:E^0\to E^1,\lambda^0,\lambda^1,\Omega)$ be the semi-direct product, as built in Construction \[c:semidirect\].
- On morphisms: let $\Psi(\phi^0,\phi^1,\mu)$ be the morphism of VB-groupoids $(g,e_0,e_1)\mapsto (g,\phi^0(c)+\mu_g(e_1),\phi^1(c))$.
Then $\Psi$ is an equivalence of categories.
In [@GM], Gracia-Saz and Mehta showed that $\Psi$ on objects induces an bijection on isomorphism classes. In Theorem 2.7 of [@dHO], del Hoyo and Ortiz built the functor $\Psi$ and actually showed this bijection came from an equivalence of categories.
One tool required in the proof of Theorem \[t:ruthsandVB\] are connections associated to VB-groupoids. This is a standard tool used by Arias Abad and Crainic [@AC] and Gracia-Saz and Mehta [@GM]. It also has a more general version, known as a cleavage, for general fibrations of groupoids (see for instance Definition 2.1.4, [@dHF]). As we need connections to define the adjoint representation up to homotopy associated to any groupoid and we will also need this tool later, we define it here:
\[d:connection\] Let $\pi:V\to G$ be a VB-groupoid over $M$ with source and target maps $\tilde{s},\tilde{t}:V_1\to V_0$ and unit map $\tilde{u}:V_0\to V_1$. Then a [*connection on $V$*]{} is a map of vector bundles $\sigma:s^*V_0\to V_1$ that is a splitting of the map $\tilde{s}:V_1\to s^*V_0$ (naturally extended to all of $V_1$).
For each $g\in G_1$, denote by $\sigma_g$ the restriction of $\sigma$ to the fiber $s^*V_0|_g$.
A connection is [*unital*]{} if, for every unit $u$ of $G$, $\sigma_u(v)=\tilde{u}(v)$.
As explained by Arias Abad and Crainic [@AC], one may find a unital connection for any VB-groupoid.
For $TG$ the tangent groupoid associated to $G$, the representation up to homotopy associated to $TG$ is the adjoint representation on the algebroid $\rho:A\to TG_0$ of $G$. Explicitly, the adjoint representation was defined by Arias Abad and Crainic as follows (see Section 3.2, [@AC]). First, choose a connection $\sigma$ of $G$. Then the adjoint representation up to homotopy induced by $\sigma$ is that with quasi-actions $$\lambda^1_g(v):=dt(\sigma_g(v))$$ on $TG_0$ and $$\lambda^0_g(X):= \sigma_g(\rho(X))\cdot X\cdot 0_{g^{-1}}$$ on $A$. Here, $\cdot$ denote multiplication in $TG$ and $0_{g^{-1}}$ is the zero element of $T_{g^{-1}}G_1$. As explained by Gracia-Saz and Mehta, $\sigma_g(\rho(X))$ and $0_{g^{-1}}$ are the unique elements of $TG_1$ by which $X$ may be multiplied on the left and right, respectively (see Section 3.3, [@GM]).
As noted by Arias Abad and Crainic (see Proposition 3.16, [@AC]), for any two choices of connection $\sigma$, $\sigma'$ of $G$, the corresponding representations up to homotopy are isomorphic.
2-Term chain complexes and linear Lie groupoid bundles {#s:2-termandLLGB}
======================================================
To establish the equivalence of 2-term representations up to homotopy and linear Lie groupoid bundles, it will be convenient to establish an equivalence of 2-categories between the 2-categories of 2-term chain complexes and linear Lie groupoid bundles. This is just a scaled up version of a result already proven by Baez and Crans (see Theorem 12, [@BC]).
To begin, let’s build the 2-category of linear Lie groupoid bundles:
\[d:VBoverM\] Given a manifold $M$, a [*linear Lie groupoid bundle*]{} over $M$ is a VB-groupoid over the trivial groupoid associated to $M$ (i.e., the groupoid with objects and morphisms $M$ and structure maps ${\mathrm{id}}_M$).
A [*map of linear Lie groupoid bundles*]{} $f:(\pi:V\to M) \to (\varpi:W\to N)$ is a map of Lie groupoids such that $f_0:V_0\to W_0$ and $f_1:V_1\to W_1$ are maps of vector bundles covering the same map on the base ${\bar{f}}:M\to N$.
A [*natural transformation of linear Lie groupoid bundles*]{} is a natural transformation $\alpha:f\Rightarrow g$ of linear Lie groupoid bundle maps covering $\bar{f}:M\to N$ that is also a map of vector bundles covering $\bar{f}$. Explicitly, it is a map of vector bundles $\alpha:V_0\to W_1$ covering $\bar{f}:M\to N$ such that, for every arrow $v:x\to y$ in $V$, the diagram $$\xymatrix{f(x)\ar[r]^{f(v)} \ar[d]_{\alpha(x)} & f(y)\ar[d]^{\alpha(y)} \\ g(x) \ar[r]_{g(v)} & g(y)}$$
As the name suggests, it is useful to think of a linear groupoid bundle $\pi:V\to M$ as smooth family of linear Lie groupoids. Indeed, note that for each $x\in M$, we have a linear groupoid ${\xymatrix{\pi^{-1}(x)=\{\pi_1^{-1}(x) \ar@<3pt>[r] \ar@<-3pt>[r] & \pi_0^{-1}(x)\}}}$ (i.e., a groupoid with objects and morphisms vector spaces and all structure groups linear maps) which is a Lie subgroupoid of $V$.
We may now define the 2-category of linear Lie groupoid bundles over a particular manifold $M$.
Let ${{\sf VB}}$ be the 2-category of linear Lie groupoid bundles defined as follows:
- Objects: linear Lie groupoid bundles $\pi:V\to M$.
- 1-morphisms: maps of linear Lie groupoid bundles $f:V\to W$.
- 2-morphisms: natural transformations of linear Lie groupoid bundles.
1-morphisms compose as expected while 2-morphisms compose vertically and horizontally as natural transformations.
To see this is a well-defined 2-category, we must check vertical and horizontal composition are well-defined.
Natural transformations of linear Lie groupoid bundles are closed under vertical and horizontal composition.
Let $\pi:V\to L$ and $\pi':W\to M$ be linear Lie groupoid bundles and let $f:L\to M$ be a smooth map. Let $F,F', F'':V\to W$ be three maps of linear Lie groupoid bundles and let $\eta:F\Rightarrow F'$ and $\eta:F'\Rightarrow F''$ be two natural transformations of linear Lie groupoid bundles. Recall that the vertical composition of $\eta'$ with $\eta$ takes $v\in V_0$ to the arrow $\eta'(v)\circ \eta(v)$.
Thinking of $\eta$ and $\eta'$ as smooth maps $\eta,\eta':V_0\to W_1$, then, this means that vertical composition $\eta'\circ\eta:V_0\to W_1$ is just the smooth map $\eta'\cdot \eta:V_0\to W_1$, for $\cdot$ multiplication in $W$. Note that, since $\eta$ and $\eta'$ induce a smooth map $(\eta,\eta'):V_0\to W_1\times_{W_0} W_1$ of vector bundles from $V_0$ to the composable arrows of $W$. Thus, $\eta'\circ \eta$ factors as maps of vector bundles $m\circ (\eta,\eta')$.
A very similar argument show that, the horizontal composition $\nu * \eta$ of natural isomorphisms $\eta:F\Rightarrow F'$ and $\nu:G\Rightarrow G'$ which takes $c\in V_0$ to the composition $\nu(F'(v))\circ G(\eta(v))$ must also be a natural transformation of linear Lie groupoid bundles.
Now, we may define the 2-category of 2-term vector bundles over a manifold $M$:
Let ${{\sf 2{\hbox{\sout{ }}}term}}$ be the 2-category of length 2 cochain complexes defined as follows:
- Objects: length two chain complexes of vector bundles $\xymatrix{C_0\ar[r]^\delta & C_1}$ over $M$ (for which $\delta$ covers the identity map on $M$).
- Morphisms: for $C^\bullet$ and $D^\bullet$ chain complexes over manifolds $M$ and $N$, morphisms of ${{\sf 2{\hbox{\sout{ }}}term}}$ are chain maps $f^\bullet: C^\bullet \to D^\bullet$ $$\xymatrix{C^0 \ar[d]_{f^0} \ar[r]^{\delta^C} & C^1 \ar[d]^{f^1} \\ D^0 \ar[r]_{\delta^D} & D^1}$$ such that $f^0$ and $f^1$ both cover a smooth map $f:M\to N$.
- 2-morphisms: for $f^\bullet,f'^\bullet:C^\bullet \to D^\bullet$ two morphisms, a 2-morphism is a chain homotopy $$\Omega:(f^\bullet:C^\bullet \to D^\bullet)\Rightarrow (g^\bullet : C^\bullet \to D^\bullet).$$ Explicitly, this is just a single map $\Omega:C^1\to D^0$ with $\delta^D\circ \Omega = g^1-f^1$ and $\Omega\circ \delta^C = g^0-f^0$.
Composition is defined as follows:
- 1-morphisms are composed as chain maps; i.e., for $h^\bullet:= g^\bullet \circ f^\bullet$, take $h^i:= g^i\circ f^i$.
- 2-morphisms are composed vertically by sum.
- Horizontally, we have the following operation: suppose we have 2-morphisms $\Psi:(f^\bullet:C^\bullet \to D^\bullet)\Rightarrow (g^\bullet : C^\bullet \to D^\bullet)$ and $\Omega:(k^\bullet:D^\bullet \to E^\bullet)\Rightarrow (\ell^\bullet: D^\bullet \to E^\bullet)$. For reference, these fit into the following diagram: $$\xymatrix@C6em@R4em{C^0 \ar[r]^{\delta^C} \ar@<1.5pt>[d]^{f^0} \ar@<-1.5pt>[d]_{g^0} & C^1 \ar@<1.5pt>[d]^{f^1} \ar@<-1.5pt>[d]_{g^1} \ar[dl]_-\Psi \\ D^0 \ar[r]^{\delta^D} \ar@<1.5pt>[d]^{k^0} \ar@<-1.5pt>[d]_{\ell^0} & D^1 \ar@<1.5pt>[d]^{k^1} \ar@<-1.5pt>[d]_{\ell^1} \ar[dl]_-{\Omega} \\ E^0 \ar[r]_{\delta^E} & E^1}$$ Then the horizontal composition is given by $$\Omega'*\Omega:= \left(k^0\circ \Psi + \Omega\circ g^1\right) : (k^\bullet\circ f^\bullet)\Rightarrow (\ell^\bullet \circ g^\bullet)$$
For any two cochain complexes $C^\bullet$ and $D^\bullet$, the zero map from $C^1$ to $D^0$ is the identity 2-morphism.
For completeness, let’s check that this is indeed a 2-category. The only non-trivial fact to check is the interchange law. So consider the diagram of 2-morphisms
$$\xymatrix@C=8em{ C^\bullet \ar@/^2pc/[r]^(.45){f^\bullet} \ar[r]_(.45){g^\bullet} \ar@/_2pc/[r]_(.45){h^\bullet} \ar@/^2pc/[r]^(.5){}="a" \ar@{}[r]^(.5){}="b" \ar@/^-2pc/[r]^(.5){}="c" \ar@{=>}"a";"b"^\Phi \ar@{=>}"b";"c"^X & D^\bullet \ar@/^2pc/[r]^(.45){k^\bullet} \ar[r]_(.45){\ell^\bullet} \ar@/_2pc/[r]_(.45){m^\bullet} \ar@/^2pc/[r]^(.5){}="d" \ar@{}[r]^(.5){}="e" \ar@/^-2pc/[r]^(.5){}="f" \ar@{=>}"d";"e"^\Psi \ar@{=>}"e";"f"^\Omega & E^\bullet}$$ Then the calculation $$\begin{aligned}
\Phi*\Psi+X*\Omega &= (k^0\circ \Phi+\Psi\circ g^1) + (l^0\circ X + \Omega\circ h^1) \\
&= k^0 \circ \Phi + \Psi\circ (g^1-h^1) +\Psi\circ h^1 + (l^0-k^0)\circ X +k^0\circ X +\Omega\circ h^1 \\
&= k^0\circ \Phi + \Psi\circ (-(\delta^D\circ X)) +\Psi\circ h^1 +(\Psi\circ \delta^D)\circ X +k^0\circ X +\Omega\circ h^1 \\&= k^0\circ(\Phi+X)+(\Psi+\Omega)\circ h^1 \\
&= (\Phi+X)*(\Psi+\Omega)\end{aligned}$$ confirms that the vertical and horizontal compositions defined for ${{\sf 2{\hbox{\sout{ }}}term}}$ satisfy the interchange law.
The components of 2-term representations up to homotopy and their morphisms correspond to objects, morphisms, and 2-morphisms in ${{\sf 2{\hbox{\sout{ }}}term}}$:
\[l:ruthtocchain\] Let $(\delta:E^0\to E^1,\lambda^0,\lambda^1,\Omega)$ be (the data corresponding to) a representation up to homotopy. Then:
- $\lambda^0,\lambda^1$ induce a 1-morphism of ${{\sf 2{\hbox{\sout{ }}}term}}$ $\lambda^\bullet: s^*E^\bullet \to E^\bullet$ covering $t:G_1\to G_0$; and
- $\Omega$ induces a 2-morphism of ${{\sf 2{\hbox{\sout{ }}}term}}$ $\Omega:p_1^*\lambda^\bullet \circ p_2^*\lambda^\bullet \Rightarrow m^*\lambda^\bullet$ covering $m:G_2\to G_1$.
Here, $p_i:G_2\to G_1$ are the natural projection maps and $m:G_2\to G_1$ is the multiplication map.
Similarly, if $(\phi^0,\phi^1,\mu)$ is (the data corresponding to) a map of representations up to homotopy from $(\delta:E^0\to E^1,\lambda^0,\lambda^1,\Omega)$ to $(\delta':E'^0\to E'^1,\lambda'^0,\lambda'^1,\Omega')$, then:
- $\phi^0,\phi^1$ induces to a 1-morphism of ${{\sf 2{\hbox{\sout{ }}}term}}$ $\phi^\bullet$ covering ${\mathrm{id}}_M$; and
- $\mu$ corresponds to a 2-morphism of ${{\sf 2{\hbox{\sout{ }}}term}}$ $\mu:\phi^\bullet \circ \lambda^\bullet \Rightarrow \lambda'^\bullet \circ \phi^\bullet$ covering $t:G_1\to G_0$.
This is simple to verify from the conditions established in Propositions \[propruth\] and \[propmapofruth\]; see also Remarks \[r:ruth\] and \[r:mapofruth\].
Now, we may define an equivalence of 2-categories between ${{\sf 2{\hbox{\sout{ }}}term}}$ and ${{\sf VB}}$. On objects, we use the following construction.
\[const:sumgpoid\] Let $\delta:C^0\to C^1$ be a cochain complex over the manifold $M$. Then we may build a VB-groupoid $V$ over $M$ as follows:
- The objects of $V$ are just the vector bundle $C^1$.
- The morphisms of $V$ are the vector bundle $C^0\oplus C^1$.
- The source map is the projection $p:C^0\oplus C^1\to C^1$ while the target map is the sum $\delta+p:C^0\oplus C^1 \to C^1$.
- Multiplication is given as $(c_0,c_1)\cdot (c'_0,c'_1):= (c_0+c_0',c_1')$ (indeed, this choice of multiplication is more or less forced by the choice of source and target maps given above).
- The identity map $u:C^1\to C^0\oplus C^1$ is the inclusion $u(c):=(0,c)$.
- The inversion map $i:C^0\oplus C^1 \to C^0\oplus C^1$ is the map $i(c_0,c_1):=(-c_0,\delta(c_0)+c_1)$.
This construction extends to a strict 2-functor between ${{\sf 2{\hbox{\sout{ }}}term}}$ and ${{\sf VB}}$:
\[p:2-chainsandlLgbs\] Let $\Phi:{{\sf 2{\hbox{\sout{ }}}term}}\to {{\sf VB}}$ be the following (strict) 2-functor:
- Objects: For $C^\bullet:=\left\{ \delta: C^0 \to C^1\right\}$, let $\Phi(C^\bullet)$ be the VB-groupoid\
${\xymatrix{V=\{C^1\oplus C^0 \ar@<3pt>[r] \ar@<-3pt>[r] & C^1\}}}$ built in Construction \[const:sumgpoid\].
- 1-morphisms: given a chain map $f^\bullet:C^\bullet \to D^\bullet$ over a morphism $f:M\to N$, take $\Phi(f^\bullet)$ to be the functor which on objects is just the map $f^1$ and on morphisms is the map $f^0+f^1:C^0\oplus C^1 \to D^0 \oplus D^1$.
- 2-morphisms: given a chain homotopy $\Omega:f^\bullet \Rightarrow g^\bullet$ between maps $f^\bullet, g^\bullet:C^\bullet \to D^\bullet$, let $\Phi(\Omega)$ be the natural transformation given by the map $$\Phi(\Omega):C^1 \longrightarrow D^0\oplus D^1,\;\;\;
\Phi(\Omega)(c):= (f^1(c),\Omega(c))$$
Then $\Phi$ is an equivalence of 2-categories. In fact, for every $C^\bullet$, $D^\bullet\in {{\sf 2{\hbox{\sout{ }}}term}}(M)$, $\Phi$ induces an isomorphism of categories $\Phi:{\mathrm{Hom}}(C^\bullet,D^\bullet)\to {\mathrm{Hom}}(\Phi(C^\bullet),\Phi(D^\bullet))$.
Recall that equivalences of 2-categories is an essentially surjective 2-functor that induces an equivalence of categories on ${\mathrm{Hom}}$ categories.
To begin, let’s show $\Phi$ is essentially surjective. Suppose that $\pi:V\to M$ is a linear Lie groupoid bundle. Let $\tilde{s},\tilde{t}:V_1\to V_0$ be the source and target maps of $V$ and let $\tilde{u}:V_0\to V_1$ be the unit map. Define $C^1:=V_0$ and $C^0:=\ker(\tilde{s})$. Note that $C^0$ is a well-defined subbundle of $V_1$ since $\tilde{s}$ must be full rank everywhere (indeed, it is a submersion and a map of vector bundles).
Next, note that $V_1$ splits as $C^0\oplus \tilde{u}(V_0)\cong C^0\oplus C^1$ and, with respect to this splitting, $\tilde{s},\tilde{t}|_{C^1}\equiv \mathrm{id}_{C^1}$. Then for $\delta:=\tilde{t}|_{C^0}$ and $p:C^0\oplus C^1\to C^1$ the projection, $\tilde{s}=p$ and $\tilde{t}=\delta+p$.
Now, suppose $(c_0,c_1)$ and $(c'_0,c'_1)$ are composable arrows in $V_1$. It follows that $t(c'_0,c'_1)=c'_1+\delta(c'_0)=c_1=s(c_0,c_1)$. The product $(c_0,c_1)\cdot(c'_0,c'_1)$ splits as $$(c_0,c_1)\cdot(c'_0,c'_1)=(c_0, c_1')\cdot (0,c'_1)+(0, \delta(c_0))\cdot (c_0',0)$$ Since $(0,c_1')=u(c_1')$ and $(0,\delta(c_0))=u(\delta(c_0))$, it follows that $$(c_0,c_1)\cdot(c'_0,c'_1) = (c_0,c_1')+(c_0',0)=(c_0+c_0',c_1').$$ Thus, the splitting above induces a smooth, invertible functor from $\Phi(\{\delta: C^0\to C^1\})$ to $V$.
Let $C^\bullet=\{\delta^C: C^0 \to C^1\}$ and $D^\bullet=\{\delta^D:D^0\to D^1\}$ be two objects of ${{\sf 2{\hbox{\sout{ }}}term}}$. Let $F:\Phi(C^\bullet)\to (D^\bullet)$ be any map of Lie groupoids in ${{\sf VB}}$. Then the map on morphisms $F_1:C^0\oplus C^1 \to D^0\oplus D^1$ splits into four maps of vector bundles: $$\begin{aligned}
w:C^0 \to D^0 \;\;&\;\; x: C^0\to D^1 \\
y:C^1\to D^0 \;\;&\;\; z: C^1 \to D^1\end{aligned}$$ that is, $F_1=(w+y)\oplus(x+z)$.
Suppose that $(c_0,c_1)$ and $(c_0',c_1')$ are composable arrows in $\Phi(C^\bullet)$. As $F$ is a functor, it must preserve multiplication. On the one hand, we have $$F((c_0,c_1)\cdot (c_0',c_1')) = F(c_0+c_0',c_1') = (w(c_0)+w(c_0')+y(c_1'),x(c_0)+x(c_0')+z(c_1'))$$ On the other, we have $$\begin{aligned}
F(c_0,c_1)\cdot F(c_0',c_1') &= (w(c_0)+y(c_1),x(c_0)+z(c_1))\cdot (w(c_0')+y(c_1'),x(c_0')+z(c_1')) \\ &= (w(c_0)+y(c_1)+ w(c_0')+y(c_1'), x(c_0')+z(c_1'))\end{aligned}$$ As this must hold for every pair of composable arrows, we must conclude $x=0$ and $y=0$.
It is easy to check that, since $F$ must preserve source maps, $z=F_0$ (i.e., $F$ on objects) and, since $F$ must preserve target maps, $w:C^0\to D^0$ and $z:C^1\to D^1$ yield a chain map between $C^\bullet$ and $D^\bullet$.
Finally, note that it is clear that $\Phi$ induces an isomorphism on 2-morphisms; indeed, $(f^1,\Omega)$ is uniquely determined by $\Omega$. If $\alpha:\Phi(f^\bullet) \Rightarrow \Phi(g^\bullet)$ is a natural transformation, then for any $c\in \Phi(C^\bullet)_0=C^1$, $\alpha(c)$ is an arrow with source $f^1(c)$ and target $g^1(c)$. Thus, the arrow $\alpha(c)=(f^1(c),\bar{\alpha}(c))$ where $f^1(c)+\delta^D(\bar{\alpha}(c))=g^1(c)$. Similarly, one may check that $\bar{\alpha}(\delta^C(c))=g^0(c)-f^0(c)$. Thus, $\bar{\alpha}:f^\bullet\to g^\bullet$ is a chain homotopy.
Weak representations of groupoids {#s:WReps}
=================================
In this section, we will define weak representations of a groupoid $G$. We will also show that the isomorphism $\Phi:{{\sf 2{\hbox{\sout{ }}}term}}\to {{\sf VB}}$ given in Proposition \[p:2-chainsandlLgbs\] induces an isomorphism between the categories of representations up to homotopy and weak representations.
Weak representations are, in particular, weak actions of groupoids; the definition of a weak action is a repurposed version of Burszytyn, Noseda, and Zhu’s definition of the action of a stacky Lie groupoid on a stack (see Definition 3.15, [@BNZ]).
\[d:waction\] Let $H$ be a Lie groupoid and $f:H\to G_0$ be a map of Lie groupoids. Then a [*weak action of $G$ on $H$*]{} is
- a smooth functor $A:G\times_{G_0} H\to H$;
- a smooth natural isomorphism $$\xymatrix@C8em{G_2\times_{G_0} H \ar@/^1pc/[r]^(0.5){}="a" \ar@/^1pc/[r]^{A\circ ({\mathrm{id}}_G\times A)} \ar@/_1pc/[r]_(0.5){}="b" \ar@/_1pc/[r]_{A\circ (m\times {\mathrm{id}}_H)} \ar@{=>}^\alpha"a";"b" & H}$$ i.e., a smooth collection of arrows of $H$ such that, for every morphism $(g,k,h):(g,k,x)\to (g,k,y)\in G_2\times_{G_0}H$, the diagram $$\xymatrix{g\cdot k \cdot x \ar[d]_{\alpha(g,k,x)} \ar[r]^{g\cdot k\cdot h} & g\cdot k \cdot y \ar[d]^{\alpha(g,k,y)} \\ (gk)\cdot x \ar[r]_{(gk)\cdot h} & (gk)\cdot y}$$ commutes (where $g\cdot x:=A(g,x)$); and
- a smooth natural isomorphism $$\xymatrix@C8em{H \ar@/^1pc/[r]^(0.5){}="a" \ar@/^1pc/[r]^{A\circ (u \circ f)} \ar@/_1pc/[r]_(0.5){}="b" \ar@/_1pc/[r]_{{\mathrm{id}}_H} \ar@{=>}^\varepsilon"a";"b" & H}$$ i.e., $\varepsilon$ naturally relates the action of units with the identity functor of $H$.
$A$, $\alpha$, and $\varepsilon$ are subject to the following coherence conditions:
- $f\circ A = t\circ p$, for $p$ the natural projection $p:G\times_{G_0} H\to G$.
- For all appropriate choices of $(g,k,l)\in G_3$ and $x\in H_0$, $\alpha:G_2\times_{G_0}H_0\to H_1$ satisfies $$\alpha(g,kl,x)\cdot g\cdot \alpha(k,l,x) = \alpha(gk,l,x)\alpha(g,k,l \cdot x)$$ More concretely, we require that the diagram $$\label{eq:alphacoh}\xymatrix{& g\cdot k\cdot l\cdot x \ar[dl]_{g\cdot \alpha(k,l,x)} \ar[dr]^{\alpha(g,k,l\cdot x)} & \\g\cdot (kl)\cdot x \ar[dr]_{\alpha(g,kl,x)} & & (gk)\cdot l \cdot x \ar[dl]^{\alpha(gk,l,x)} \\ & (gkl)\cdot x &}$$ commutes.
- For any $(g, x)\in G_1\times_{G_0} H_0$, $\alpha(g,u(f(x)),x)=g\cdot \epsilon(x)$ and $\alpha(u(t(g)),g,x)=\varepsilon(g\cdot x)$.
A weak action of $G$ on $H$ is [*unital*]{} if $\varepsilon={\mathrm{id}}_{{\mathrm{id}}_H}$.
Let’s unpack Definition \[d:waction\] a bit. The fact that $A:G_1\times_{G_0} H \to H$ is a functor subsumes two interesting properties. First note that since $f:H\to G_0$ is a map of Lie groupoids, if $h_1$ and $h_2$ are composable arrows then $f_1(h_1)=f_1(h_2)$. In particular, if $f_1(h_1)=s_G(g)$, then we have that $A_1(g, h_1\cdot h_2)=A(g,h_1)\cdot A(g,h_2)$.
Next, note that any unit $u_H(x)\in H_1$ and $g\in G_1$ satisfying $f_0(x)=s_G(g)$, we have that $A(g,u_H(x))=u_H(A(g,x))$.
If the action is unital, then $\alpha(g,u(f(x)),x)=\alpha(u(t(g)),g,x)={\mathrm{id}}_x$ for all $(g,x)\in G_1\times_{G_0} H_0$.
Finally, suppose additionally that $f_0,f_1$ are submersions. Then for every $a\in G_0$, $f^{-1}(a)$ is a Lie subgroupoid. Furthermore, each $g:a\to b$ in $G$ induces a map of Lie groupoids $A(g,\cdot):f^{-1}(a) \to f^{-1}(b)$. In the case where $A$ is a [*non-weak*]{} action, note this functor is an isomorphism of categories with inverse functor $A(g^{-1},\cdot)$. In the case of a weak action, $A(g,\cdot)$ and $A(g^{-1},\cdot)$ form an equivalence of categories. Indeed, $A(g^{-1},\cdot)\circ A(g,\cdot)$ and ${\mathrm{id}}_{f^{-1}(a)}$ are naturally isomorphic via the smooth natural isomorphism $$\varepsilon\circ \alpha(g^{-1},g,\cdot) : A(g^{-1},\cdot)\circ A(g,\cdot) \Rightarrow {\mathrm{id}}_{f^{-1}(a))}, \;\;\; x\mapsto \varepsilon(x)\cdot \alpha(g^{-1},g,x)$$
To define equivariant maps, we again refer to the work of Bruszytyn, Noseda, and Zhu (see Definition 3.20, [@BNZ]):
\[d:equivmap\] Let $f:H\to G_0$ and $f':H'\to G_0$ be two Lie groupoid maps to $G_0$ and let $(A,\alpha,\varepsilon)$ and $(A',\alpha',\varepsilon')$ be weak actions of $G$ on $H$ and $H'$, respectively. Then a map of Lie groupoids $F:H\to H'$ is [*equivariant*]{} if $f'\circ F=f$ and if there exists there exists a natural isomorphism $\delta:F\circ A\Rightarrow A'\circ ({\mathrm{id}}_{G_1} \times F)$ such that, for all $(g,k,x)\in G_2\times_{G_0}H_0$, the diagrams $$\xymatrix@C=0.5em@R3em{& & F(g\cdot k \cdot x) \ar[dll]_{\delta(g,k\cdot x)} \ar[drr]^{F(\alpha(g,k,x))} & & \\ g\cdot F(k\cdot x)\ar[dr]_{g\cdot \delta(k,x))} & & & & F((gk)\cdot x) \ar[dl]^{\delta(gk,x)} \\ & g\cdot k \cdot F(h) \ar[rr]_{\alpha'(g,k,F(x))} & & (gk)\cdot F(x) }$$ and $$\xymatrix@C7em@R5em{F((f(x))\cdot x) \ar[r]^-{F(\varepsilon(x))} \ar[d]_{\delta(u(f(x)), x)} & F(x) \\ u(f(x))\cdot F(x) \ar[ur]_{\varepsilon'(F(x))} }$$
commute.
The composition $(F'\circ F, ({\mathrm{id}}_{F'}\circ \delta)*(\delta'\circ {\mathrm{id}}_{F\times {\mathrm{id}}}))$ of two equivariant maps $(F,\delta)$ and $(F',\delta')$ is again an equivariant map.
See Lemma 3.21, [@BNZ].
Now, we may define weak representations of a groupoid $G$. As is the case with groupoid actions, weak representations are groupoid actions on smooth linear spaces. In this case, we use linear Lie groupoid bundles:
A [*weak representation of $G$*]{} on a linear groupoid bundle $\pi:V\to G_0$ is a weak unital action $(A,\alpha,{\mathrm{id}}_V)$ such that $A:G_1\times_{G_0}V\to V$ is a map of linear Lie groupoid bundles covering $t:G_1\to G_0$ and $\alpha$ is a natural transformation of linear groupoids covering $t_2:G_2\to G_0$.
A [*map of weak representations of $G$*]{} is any map of linear groupoid bundles which is also equivariant.
From here forward, denote by ${{\sf WRep}}(G)$ the category of weak representations of $G$.
It is a well known fact that a Lie groupoid representation $G$ on a vector bundle $E\to G_0$ corresponds exactly to a smooth functor from $G$ to the frame groupoid of $E$: the Lie groupoid with objects $G_0$ and morphisms linear isomorphisms between the fibers of $E$.
We may make a similar statement for weak representations. Let $\pi:V\to G_0$ be a linear Lie groupoid bundle. Let ${\sf Fr}(V)$ denote the 2-category with objects $G_0$, morphisms linear equivalences of categories $\varphi:\pi^{-1}(x)\to \pi^{-1}(y)$, and 2-morphisms linear natural isomorphisms. It is easy to check then that a weak representation corresponds to a pseudofunctor from $G$ (thought of as a 2-category with 2-morphisms $G_2$, the collection of composable morphisms) to ${\sf Fr}(V)$.
Going from pseudofunctors $F:G\to {\sf Fr}(V)$ to weak representations is more challenging. One should expect that it is smooth pseudofunctors that should correspond to weak representations. However, ${\sf Fr}(V)$ needn’t be a Lie 2-groupoid; indeed, one may construct examples where the source and target maps associated to ${\sf Fr}(V)$ will fail to be submersion for any possible manifold structure on the morphisms of ${\sf Fr}(V)$.
One possible way around this is the use of diffeologies; this approach may be explored in future work.
Now, we may prove the equivalence between the categories of 2-term representations up to homotopy of $G$ and weak representations of $G$.
\[t:ruthsandwr\] The equivalence of 2-categories $\Phi:{{\sf 2{\hbox{\sout{ }}}term}}\to {{\sf VB}}$ induces an equivalence of categories $\Phi(G):{{\sf RutH_2}}(G)\to {{\sf WRep}}(G)$.
Define $\Phi(G)(\delta:E^0\to E^1,\lambda^0,\lambda^1,\Omega):=(\Phi(\delta:E^0\to E^1), \Phi(\lambda^\bullet), \Phi(\Omega))$; here we are using Lemma \[l:ruthtocchain\] to identify the components associated to $(\delta:E^0\to E^1,\lambda^0,\lambda^1,\Omega)$ with elements of ${{\sf 2{\hbox{\sout{ }}}term}}$. Similarly, define $\Phi(G)(\phi^0,\phi^1,\mu):=(\Phi(\phi^\bullet), \Phi(\mu))$. It is easy to verify that the conditions of Propositions \[propruth\] and \[propmapofruth\] (i.e., on the data associated to representations up to homotopy and their maps) correspond exactly to the conditions on weak actions and equivariant maps (see Definitions \[d:waction\] and \[d:equivmap\]). Thus, $\Phi(G)$ is well defined and, since $\Phi$ is essentially surjective and induces an isomorphism on ${\mathrm{Hom}}$ categories, $\Phi(G)$ must be essentially surjective and fully faithful.
VB-groupoids as weak action groupoids {#s:VBsandWR}
=====================================
In this section, we discuss the action groupoids associated to weak groupoid actions. We will also explain how VB-groupoids may be realized as action groupoids associated to weak representations.
Let $f:H\to G_0$ be a map of Lie groupoids $G_0$ and let $(A,\alpha, \varepsilon)$ be a weak action of $G$ on $H$. Write $\tilde{s},\tilde{t}:H_1\to H_0$ and $\tilde{u}:H_0\to H_1$ for the source, target, and unit maps of $H$. Then [*the action groupoid $G\times_A H$*]{} is the groupoid with
- Objects: $H_0$
- Morphisms: For $A_0:G_1\times_{G_0}H_0\to H_0$ the action functor $A$ on objects, the morphisms of $G\times_A H$ are $$\left(G_1\times_{s,{G_0},f_0}H_0\right) \times_{A_0,H_0,\tilde{t}}H_1 =\{(g,x,h)\in G_1\times H_0\times H_1 \;| \; s(g)=\pi_0(x),\; \tilde{t}(h)=g\cdot x\}$$
- Source and target: the source map $ \sigma: (G\times_A H)_1 \to H_0$ is the projection $\sigma(g,x,h):=x$ while the target map $\tau:(G\times_AH)_1\to H_0$ is the map $\tau(g,x,h):=s(h)$.
- Multiplication: multiplication is given by $$(g,x,h)\cdot (g',x',h'):= (gg', x',\alpha(g,g',x')(g\cdot h') h)$$
- Units: the unit map is given by $\mu(x):= (u(f(x)), x, \tilde{u}(x))$.
- Inverses: define $(g,x,h)^{-1}:=(g^{-1}, s(h), (g^{-1}\cdot h)^{-1} \alpha(g^{-1},g,x)^{-1} \varepsilon(x)^{-1})$
For the sake of completeness, let’s verify $G\times_A H$ is indeed a Lie groupoid. First, note that the objects and morphisms indeed must be manifolds; on objects, we are already using $H_0$ and on morphisms, we are pulling back against submersions $s:G_1\to G_0$ and $\tilde{t}:H_1\to H_0$. As projections, $\sigma$ and $\tau$ are certainly submersions. As all structure maps are smooth combinations of the structure maps of $G$ and $H$, it follows that the structure maps of $G \times_A H$ are smooth as well.
To see that multiplication is associative, note that: $$\begin{aligned}
((g,x,h) (g',x',h')) (g'',x'',h'') &= (gg'g'', x'', \alpha(gg',g'',x'')((gg')\cdot h'')\alpha(g,g',x')(g\cdot h')h) \\
&= (gg'g'',x'', \alpha(gg',g'',x'')\alpha(g,g',g''\cdot x'') (g\cdot g'\cdot h'')(g\cdot h')h)\\
&= (gg'g'',x'', \alpha(g,g'g'',x'')(g\cdot \alpha(g',g'',x'')) (g\cdot g'\cdot h'')(g\cdot h')h)\\
&= (gg'g'',x'', \alpha(g,g',g'',x'') (g\cdot(\alpha(g',g'',x'')(g'\cdot h'')h'h)))\\
&= (g,x,h) ((g',x',h') (g'',x'',h''))\end{aligned}$$ Here, we take advantage of the fact that $\alpha$ is a natural transformation, the coherence condition on $\alpha$, and the functoriality of $A$.
Using a similar argument also involving the compatibility conditions on $\alpha$ and between $\alpha$ and $\varepsilon$ (see Definition \[d:waction\]), one may also check that the units and inverses defined above behave as required.
As noted above, weak groupoid actions are inspired by the definition of a stacky Lie groupoid actions of Bursztyn, Noseda, and Zhu [@BNZ]. Similarly, the construction of an action groupoid associated to a weak groupoid action is also inspired by their work. As they note in Corollary 5.3, under certain conditions, the map of stacks associated to a stacky Lie groupoid action is a bibundle inheriting a Lie groupoid structure.
Note that the space of morphisms of $G\times_A H$ in Construction \[c:actiongpoid\] is exactly the (total space of) the bibundle $\langle A \rangle$ associated to the action functor $A$. Multiplication in $G\times_AH$ is also related to the bibundle picture; indeed, one may identify the collection of composable arrows of $G\times_AH$ with the bibundle $\langle A\circ ({\mathrm{id}}_G\times A)\rangle$. Then $\alpha$ induces an isomorphism of bibundles $$\langle \alpha \rangle:\langle A\circ ({\mathrm{id}}_G\times A)\rangle\to \langle A\circ (m\times {\mathrm{id}}_H)\rangle$$ and one may check that, on the arrow component $H_1$ of $G \times_A H$, the multiplication defined above corresponds to the composition of $\langle \alpha \rangle$ with projection to $H_1$.
Action groupoids of weak representations are naturally VB-groupoids.
\[l:WRactiongpoids\] Let $\pi:V\to G_0$ be a linear Lie groupoid bundle and let $(A:G_1\times_{G_0}V\to V, \alpha)$ be a weak representation of $G$ on $V$. Then the action groupoid $G\times_A V$ is a VB-groupoid. Furthermore, for $(B:G_1\times_{G_0}W\to W, \beta)$ another weak representation of $G$ on a linear Lie groupoid bundle $\varpi:W\to G_0$, any map of weak representations $$(F,\delta): (A:G_1\times_{G_0}V\to V, \alpha) \to (A:G_1\times_{G_0}V\to V, \alpha)$$ extends to a map of VB-groupoids ${{\sf Act}}(F,\delta) : G\times_A V \to G\times_BW$.
By assumption, we have that $f_0:V_0\to G_0$ is a vector bundle. Note that the space of arrows $\left(G_1\times_{s,{G_0},\pi_0}V_0\right) \times_{A_0,V_0,\tilde{t}}V_1$ must also naturally inherit the structure of a vector bundle. Indeed, as explained in the proof of Proposition \[p:2-chainsandlLgbs\], for $C:=\ker\left(\tilde{t}\right)$, we may identify $V_1$ with $V_0\oplus C$ and, with respect to this identification, $\tilde{t}$ corresponds to the natural projection $p:V_0\oplus C \to V_0$ while $\tilde{s}$ corresponds to $p+\tilde{s}|_{C}$. It follows that, using this identification, we may identify $\left(G_1\times_{s,{G_0},\pi_0}V_0\right) \times_{A_0,V_0,\tilde{t}}V_1$ with the bundle $s^*V_0\oplus t^*C\to G_1$.
As the structure maps of $G\times_AV$ are all built from maps of vector bundles, it follows that $G\times_A V$ is a VB-groupoid.
For $(F,\delta): (A:G_1\times_{G_0}V\to V, \alpha) \to (A:G_1\times_{G_0}V\to V, \alpha)$ a map of weak representations, define ${{\sf Act}}(F,\delta)$ as follows:
1. On objects: for $x\in V_0$, define ${{\sf Act}}(F,\delta)(x):=F_0$
2. On morphisms: for $(g,x,v)\in (G\times_A V)_1$, define $${{\sf Act}}(F,\delta)(g,x,v):= (g,F(x),\delta(g,x)F(v))$$
As $F$ and $\delta$ correspond to maps of vector bundles, it follows that ${{\sf Act}}(F,\delta)$ corresponds to a pair of maps of vector bundles.
To check that ${{\sf Act}}(F,\delta)$ preserves multiplication, let $(g,x,v)$ and $(g',x',v')$ be a pair of composable arrows. Then: $$\begin{aligned}
{{\sf Act}}(F,\delta)(g,x,v){{\sf Act}}(F,\delta)(g',x',v') &= (gg',F(x'), \beta(g,g',F(x'))(g\cdot (\delta(g',x')F(v')))\delta(g,x)F(v))\\
&= (gg',F(x'), \beta(g,g',F(x'))(g\cdot(g',x'))\delta(g,g'\cdot x')F(g\cdot v')F(v)) \\
&= (gg',F(x'), \delta(gg',x')F(\alpha(g,g',x'))F(g\cdot v')F(v)) \\
&= {{\sf Act}}(F,\delta)((g,x,v)(g',x',v'))\end{aligned}$$ One may use similar arguments to show that ${{\sf Act}}(F,\delta)$ must preserve identities.
${{\sf Act}}:{{\sf WRep}}(G) \to {{\sf VB}}(G)$ given by $$\begin{aligned}
{{\sf Act}}((F,\delta):(V,A,\alpha)\to (W,B,\beta):={{\sf Act}}(F,\delta):G\times_AV\to G\times_BW\end{aligned}$$ (for ${{\sf Act}}(F,\delta)$ defined in Lemma \[l:WRactiongpoids\]) is a well-defined functor.
By Lemma \[l:WRactiongpoids\], we have that $G\times_A V$ is a VB-groupoid for any weak representation $(A,\alpha)$ on $V\to G_0$ and that ${{\sf Act}}(F,\delta)$ is a map of VB-groupoids for any map of weak representations.
So all that remains to be seen is that ${{\sf Act}}$ is a well-defined as a functor, note it is clear that ${{\sf Act}}({\mathrm{id}}_V,{\mathrm{id}}_{{\mathrm{id}}\times F})$ clearly corresponds to the identity functor on $G\times_A V$. For any two maps of weak representations $(F,\delta):(V,A,\alpha)\to (W,B,\beta)$ and $(G,\epsilon):(W,B,\beta)\to (X,C,\gamma)$ and for any $(g,x,v)\in B \times_A V$, $$\begin{aligned}
{{\sf Act}}(G,\epsilon)({{\sf Act}}(F,\delta)(g,x,v))&=(g,G(F(x)), \epsilon(g,F(x))G(\delta(g,x)F(v)))\\
&= (g,G(F(x)), ({\mathrm{id}}_G \circ \delta)*(\epsilon \circ {\mathrm{id}}_{F\times {\mathrm{id}}}))(g,x) G(F(v)))\\
&= {{\sf Act}}((G,\epsilon)\circ(F,\delta))(g,x,v)\end{aligned}$$
We may finally state the main theorem of this paper:
\[t:main\] The functor ${{\sf Act}}:{{\sf WRep}}(G)\to {{\sf VB}}(G)$ is an equivalence of categories.
To prove this, we work in stages. First, let’s prove essential surjectivity. Our technique will be to show that the quasi-action defined on the kernel of a VB-groupoid (as defined in Section 2.1 of [@dHF], for instance) is in fact a weak representation.
\[l:esssurj\] For any VB-groupoid $\pi:V\to G$, there exists a weak representation $(A,\alpha)$ of $G$ on a linear Lie groupoid bundle $\varpi:K\to G_0$ such that $G\times_AK\cong V$.
Let $\tilde{s},\tilde{t}:V_1\to V_0$ be the structure maps of $V$. Let $K$ be the linear Lie groupoid bundle with space of objects $V_0$ and space of morphisms $V_1|_{G_0}$. Note this is the kernel of the Lie groupoid fibration $\pi$ (see Definition 2.1.3 of [@dHF]). Choose a unital connection $\sigma:s^*V_0\to V_1$ of $V$ (see Definition \[d:connection\]). Define $A:G_1\times_{G_0} K\to K$ as follows:
- for $x\in K_0=V_0$, define $A_0(g,x):=\tilde{t}(\sigma_g(v))$; and
- for $k\in K_1$, define $A_1(g,k):= \sigma_g(\tilde{t}(k))k (\sigma_g(\tilde{s}(k)))^{-1}$.
This is a well-defined functor, as, if $\tilde{s}(k)=x$, $\tilde{t}(k)=y$, then: $$\tilde{s}(g\cdot k) = \tilde{s}((\sigma_g(x))^{-1})=\tilde{t}(\sigma_g(x))=g\cdot x \text{ and }\tilde{t}(g\cdot k)=\tilde{t}(\sigma_g(y))=g\cdot y.$$ For composable $k,k'\in K_1$, $$\begin{aligned}
g\cdot (kk') &= \sigma_g(\tilde{t}(kk'))kk' \sigma_g(\tilde{s}(kk'))^{-1} = \sigma_g(\tilde{t}(k))k\sigma_g(\tilde{s}(k))^{-1}\sigma_g(\tilde{s}(k)) k'\sigma_g(\tilde{s}(k'))^{-1} \\ &= \sigma_g(\tilde{t}(k))k\sigma_g(\tilde{s}(k))^{-1}\sigma_g(\tilde{t}(k')) k'\sigma_g(\tilde{s}(k'))^{-1} = (g\cdot k)(g\cdot k') \end{aligned}$$ As the composition of smooth maps of vector bundles, clearly $A_0$ and $A_1$ are both maps of vector bundles.
The fact that $\sigma$ is unital implies that this action must be unital as well.
Next, define for each $(g,g',x)\in G_2\times_{G_0}V_0$ define $$\alpha(g,g',x):=\sigma_{gg'}(x)\sigma_{g'}(x)^{-1}\sigma_g(g'\cdot x)^{-1}.$$ Again, as a combinations of smooth maps of vector bundles, this is again a smooth map of vector bundles. It is easy to check that $\alpha$ gives the natural transformation $A\circ ({\mathrm{id}}_G\times A)\Rightarrow A\circ (m\times {\mathrm{id}}_V)$ and that $\alpha$ satisfies the correct coherence conditions.
Define an isomorphism of VB-groupoids as follows: take $\phi:V \to G\times_AK$ by\
$\phi_0={\mathrm{id}}_{V_0}:V_0\to V_0$ and let $$\phi_1:(G\times_AK)_1,\;(g,x,k)\mapsto k^{-1}\sigma_g(x)$$ Since $(g,x,k)$ is an arrow from $x$ to $s(k)$, it follows that $\phi_1$ preserves source and target maps. It is a straightforward calculation to confirm that it also preserves multiplication and units. As $\phi_1$ has a trivial kernel, it follows that the functor $\phi$ which, on objects is given by $\phi_0$ and on morphisms by $\phi_1$ yields an isomorphism of VB-groupoids over $G$.
Now, we prove that ${{\sf Act}}$ is fully faithful.
\[l:ff\] The functor ${{\sf Act}}:{{\sf WRep}}(G)\to {{\sf VB}}(G)$ is fully faithful.
Fix two weak representations $(A,\alpha)$ and $(B,\beta)$ on linear Lie groupoid bundles $\pi:V\to G_0$ and $\varpi:W\to G_0$, respectively. To see that ${{\sf Act}}(F,\delta)$ uniquely determines both $F$ and $\delta$, first note that we may naturally identify $V$ and $W$ as VB-Lie subgroupoids of the action groupoids $G\times_AV$ and $G\times_BW$. Indeed, for $(G\times_AV)$ the full subgroupoid with morphisms $(G\times_AV)|_{V_0}$, there an isomorphism of of VB-groupoids $\iota_V:V\to (G\times_AV)|_{V_0}$ given on objects by ${\mathrm{id}}_{V_0}$ and on morphisms by $$\iota_V:V_1\to \left((G\times_AV)|_{V_0}\right)_1,\; v\mapsto (u(\pi(v)),\tilde{s}(v),v^{-1})$$ Similarly, let $\iota_W:W\to (G\times_BW)|_{W_0}$ be an isomorphism from $W$ to the kernel subgroupoid of $G\times_BW$.
It follows then that $F$ uniquely determines $\phi|_{V_0}$ for any $\phi={{\sf Act}}(F,\delta)$. Now, note that for any $(g,x,v)\in (G\times_AV)_1$, we have that $$\label{fact}(g,x,v)=(u(\pi(v)),\tilde{t}(v),v)(g,x,\tilde{u}(g\cdot x)).$$ Additionally, recall that ${{\sf Act}}(F,\delta)(g,x,v):=(g,F(x),\delta(g,x)v)$; thus, for $p:(G\times_BW)_1\to W_1$ the natural projection and for $(g,x)\in G_1\times_{G_0}V_0$, take $p(\phi_1(u(\pi(v)),\tilde{t}(v),v))=w$. Then we have that $$\begin{aligned}
(g,F(x),\delta(g,x)v) = \phi_1(g,x,v)&=\phi_1(u(\pi(v)),\tilde{t}(v),v)\phi_1(g,x,\tilde{u}(g\cdot x)) \\ &= (u(\pi(v)),\phi_0(\tilde{t}(v)),w)(g,\phi_0(x),\tilde{u}(\phi_0(g\cdot x))) \\ &= (g,F(x),w) \end{aligned}$$ Thus, $\delta(g,x)$ is also uniquely determined by $\phi$ and so ${{\sf Act}}$ is faithful.
To see it is full, we may reverse this process. Indeed, let $\phi:G\times_AV\to G\times_BW$ be a map of VB-groupoids. Then since any map of VB-groupoids $\phi$ must take $(G\times_AV)|_{V_0}$ to $(G\times_BW)|_{V_0}$, it follows that $\iota_W^{-1}\circ \phi\circ \iota_V$ gives a well-defined functor $F$ from $V$ to $W$. Note this functor is explicitly given by $F_0=\phi_0$ and $F_1=\tilde{i}\circ p\circ \phi_1$ (for $\tilde{i}:W_1\to W_1$ inversion in $W$).
Again using the decomposition above, we have that $$\phi_1(g,x,v)=\phi_1(u(\pi(v)),\tilde{t}(v),v)\phi_1(g,x,\tilde{u}(g\cdot x))$$ So, define $$\delta(g,x)=p(\phi_1(g,x,\tilde{u}(g\cdot x))).$$ It follows then by definition that ${{\sf Act}}(F,\delta)=\phi$. It is not difficult to show that this $\delta$ must satisfy the necessary conditions for $(F,\delta)$ to be a map of weak representations.
By Lemma \[l:esssurj\], ${{\sf Act}}$ is essentially surjective and by Lemma \[l:ff\], ${{\sf Act}}$ is fully faithful. Therefore, it defines an equivalence of categories.
For $\Psi:{{\sf RutH_2}}(G)\to {{\sf VB}}(G)$ the equivalence of categories of Theorem \[t:ruthsandVB\] and $\Phi(G):{{\sf RutH_2}}(G)\to {{\sf WRep}}(G)$ the equivalence of categories of Theorem \[t:ruthsandwr\], one may show that the diagram $$\xymatrix{{{\sf RutH_2}}(G)\ar[rr]^{\Phi(G)} \ar[dr]_{\Psi} & & {{\sf WRep}}(G) \ar[dl]^{{{\sf Act}}} \\ & {{\sf VB}}(G) &}$$ commutes up to a natural isomorphism.
[9]{} C. Arias Abad and M. Crainic; Representations up to homotopy and Bott’s spectral sequence for Lie groupoids. [*Adv. in Math.*]{}, [**248**]{} (2013), 416-452. C. Arias Abad and M. Crainic; Representations up to homotopy of Lie algebroids. [*J. reine agnew. Math.*]{} [**663**]{} (2012), 91-126. J. Baez and A. Crans; Higher-dimensional algebra VI: Lie 2-algebras. [*Theory Appl. Categ.*]{}, [**22**]{} (2009), 542-587. H. Bursztyn, F. Noseda, and C. Zhu; Principal actions of stacky Lie groupoids. [arXiv:1510.09208](https://arxiv.org/abs/1510.09208). M. del Hoyo and R. L. Fernandes; Riemannian metrics on differentiable stacks. [arXiv:1601.05616](http:s//arxiv.org/abs/1601.05616). M. del Hoyo and C. Ortiz; Morita equivalences of vector bundles. [arXiv:1612.09289](https://arxiv.org/abs/1612.09289). A. Gracia-Saz and R. Mehta; VB-groupoids and representation theory of Lie groupoids. [arXiv:1007.3658](https://arxiv.org/abs/1007.3658). J. Pradines; Remarque sur le groupode cotangent de Weinststein-Dazord. [*C. R. Acad. Sci. Paris Sèr. I. Math.,*]{} [*306*]{} 16 (1988), 557-560.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Unmanned aerial vehicles (UAVs), also known as drones, have emerged as a promising mode of fast, energy-efficient, and cost-effective package delivery. A considerable number of works have studied different aspects of drone package delivery service by a supplier, one of which is delivery planning. However, existing works addressing the planning issues consider a simple case of perfect delivery without service interruption, e.g., due to accident which is common and realistic. Therefore, this paper introduces the joint ground and aerial delivery service optimization and planning (GADOP) framework. The framework explicitly incorporates uncertainty of drone package delivery, i.e., takeoff and breakdown conditions. The GADOP framework aims to minimize the total delivery cost given practical constraints, e.g., traveling distance limit. Specifically, we formulate the GADOP framework as a three-stage stochastic integer programming model. To deal with the high complexity issue of the problem, a decomposition method is adopted. Then, the performance of the GADOP framework is evaluated by using two data sets including Solomon benchmark suite and the real data from one of the Singapore logistics companies. The performance evaluation clearly shows that the GADOP framework can achieve significantly lower total payment than that of the baseline methods which do not take uncertainty into account.'
author:
- 'Suttinee Sawadsitang, Dusit Niyato, Puay-Siew Tan, and Ping Wang, [^1][^2][^3] [^4]'
title: 'Joint Ground and Aerial Package Delivery Services: A Stochastic Optimization Approach'
---
Ground-based delivery, drone delivery, UAV, stochastic optimization, uncertainty, breakdown.
Introduction
============
Package delivery becomes an instrumental function of logistics businesses due to the popularity of e-commerce such as online shopping. Traditionally, package delivery has been done through land transportation such as using trucks, cars, motorcycles, and bicycles, which is reasonably affordable, reliable, and accessible. However, due to rising labor costs, an alternative delivery mode by using unmanned aerial vehicles (UAVs) such as drones becomes increasingly attractive. Thanks to the advancement of drone technologies, the cost and reliability of using drones in package delivery services have been improved significantly. Businesses start adopting the drone delivery such as Amazon Prime Air, DHL Parcelcopter, and Google Project Wing, showing the technological and economic feasibility of the services. The benefits of drone package delivery services are palpable as follows. Firstly, drone delivery is faster than land vehicles as drones are not subjected to road traffic jam. Also, it has much less service time as the drones can drop a package directly to the customers with minimal human involved. Secondly, drone delivery reduces resource usage in terms of manpower and energy. Thirdly, drone delivery is applicable to the areas which are difficult to access by land transportation.
Although drone delivery has many advantages over traditional ground-based delivery, using drones still faces many challenges.[^5] For example, the reliability of drone delivery is lower than that of the ground-based delivery. Furthermore, drones have significantly less service coverage and capacity. As a result, using the drones for package delivery is limited to a certain area near the depot. Nevertheless, recently, an integration of ground-based delivery and drone delivery has been introduced [@sidekick]. It has been reported that these two delivery modes can complement each other because of their unique features and advantages. For example, packages can be optimally assigned to be delivered by trucks or drones to maximize service quality and to minimize cost. In this regard, a joint ground and aerial delivery service optimization and planning emerge as an important issue.
In this paper, we introduce the joint ground and aerial delivery service optimization and planning (GADOP) framework for a supplier. In particular, customers’ packages can be delivered by trucks or drones. We consider the case that the supplier needs to make reservations for both trucks and drones in advance. The supplier, after deciding packages to be delivered by trucks, determines route for each truck. Similarly, the supplier determines package serving order of the drones. We model the uncertainty of the drone delivery in terms of takeoff and breakdown conditions. Specifically, the drone may not be able to take off, e.g., due to bad weather, and if it can take off, the drone may experience accident, e.g., due to technical issues. Therefore, we adopt the stochastic programming technique to formulate the joint ground and aerial delivery service optimization and planning in which the decisions are made in different stages. The contributions of this paper are summarized as follows.
- The proposed joint optimization is able to utilize and trade off benefits and limitations of truck and drone delivery. In particular, the proposed optimization can yield the minimum total payment[^6] for the supplier by considering the routing path of trucks and the serving order of drones.
- The formulated stochastic programming model can find and achieve an optimal solution taking the uncertainty of the drone delivery into account. Consequently, the supplier can obtain the better planning for the truck and drone usage given certain constraints such as capacity limit and traveling distance limit.
- To reduce the computational time of obtaining the solution of the stochastic programming model, we introduce the decomposed GADOP by adopting the L-shape method. The decomposition allows the model to be split into multiple smaller sub-problems, which are more computationally tractable.
This paper is structured as follows. Section \[sec\_related\] presents the literature review of vehicle routing problem, ground and aerial delivery. The system model and the formulations of the proposed joint optimization are presented in Section \[sec\_system\] and Section \[sec\_formulation\], respectively. Section \[sec\_decom\] presents the decomposition of the proposed joint optimization. Section \[sec\_eva\] presents the experiment setting and the numerical results of the proposed joint optimization. At the end, the conclusion and future works are discussed in Section \[sec\_con\].
Related Work {#sec_related}
============
Vehicle Routing Problem (VRP) has played a significant role in goods distribution for half century. Hundreds of different models and solution algorithms have been proposed since then. Many surveys on a variety of issues in VRP exist in the literature, e.g., [@ref_survey_2007-general], [@ref_greensurvey]. Some of them consider specific aspects of the problem, e.g., VRP with time windows [@ref_survey_1988], pick-up and delivery VRP [@ref_survey_2007-pickup], dynamic VRP [@ref_survey_2013-dynamic] and stochastic VRP [@ref_survey_1996], and solution algorithms [@ref_survey_1992], [@ref_survey_2000]. However, almost all of the existing VRP works consider only cars or trucks because they are the most popular mode of transportation. Recently, drones has been introduced which has a great potential to be used in businesses. A number of researchers and engineers studied and improved the technical aspects of drones, i.e., enhance endurance, safety, and features. Recently, planning and scheduling problems of drones have been studied in [@UAV-communication], [@UAV-refuel]. The authors in [@UAV-communication] foresaw that drones could be used as mobile wireless communication nodes. In particular, the drones fly to collect data and return to transmit the data to the base station. The authors formulated the problem based on the pick-up and delivery VRP and added the communication constraints, which were not considered in the previous works. The authors in [@UAV-refuel] addressed the drone path planning problem, which is also one type of VRP, by minimizing the total drone resource requirement, i.e., fuel or battery. In the problem, all the customers need to be visited once, and multiple depots are available for the drones to refuel or change their battery. The authors adopted resource constraint and developed a heuristic algorithm for the problem.
However, the authors in [@UAV-communication] and [@UAV-refuel] did not consider uncertainty, which is unavoidable in a real environment. A stochastic drone mission planning is presented in [@robust] and [@online]. The authors in [@robust] proposed a drone planning model with the stochastic resource consumption, i.e., fuel or energy consumption, traveling from one location to another location. The authors presented the comparison between the robust solutions and the solution with average resource consumption. Instead of considering stochastic resource consumption, the authors in [@online] considered the uncertainty in recording time and traveling time from one location to another location, as well as a visiting time interval of each location, which is known as time-windows. The authors combined two objectives together by adding and balancing two weight parameters. The objectives are to maximize (i) the expected profits of visiting the locations which are already known before taking off and (ii) the percentage of reaching new locations, which are not known before taking off, on time. The new locations are emergency and unplanned cases. The authors proposed a re-planning approach as integer programming and developed a heuristic algorithm to solve it. However, in [@robust] and [@online], the authors formulated the problems as orienteering problems. In particular, a solution of the orienteering problem is to visit only some locations while the VRP requires all locations to be visited.
For drone-related research, most of the existing works consider surveillance applications. Only few works are dedicated to parcel delivery [@drone_delivery]. The drone parcel delivery can be found in [@drone_delivery] and [@sidekick]. The authors in [@drone_delivery] proposed the drone delivery planning considering drones’ capacity, battery weights, and changing payload weights. The authors modeled multiple drones, and each drone is able to serve multiple customers before returning to the depot. Drones are also allowed to come back to refuel or charge a battery at the depot for multiple times. The authors in [@sidekick] considered the heterogeneous VRP for package delivery using both trucks and drones. The authors proposed the Flying Sidekick Traveling Salesman Problem (FSTSP) and the Parallel Drone Scheduling Traveling Salesman Problem (PDSTSP). In the FSTSP, the truck driver can set up a drone and let it visit different customers. After visiting and dropping packages to the assigned customers, the drone returns to the truck. Hence, the truck is referred to as a mobile depot for the drone separation and re-connection. On the other hand, in the PDSTSP, the authors formulated the problem with only one depot, but with multiple drones. Both the problems were formulated as a mix-integer programming model, where the objectives of the FSTSP and the PDSTSP are to minimize the serving time for both vehicles, i.e., one truck and one drone, and to minimize the traveling time of the truck, respectively.
However, all the above works ignored the random breakdown of the drone delivery which can happen more frequently than land transportation, e.g., truck. Therefore, we propose the joint ground and aerial delivery service optimization and planning (GADOP) framework considering realistic drone operations. For example, (i) drones cannot take off due to weather conditions such as raining and strong wind, (ii) drones may have technical problems during flying, and (iii) drones can have an accident such as a bird strike. Therefore, we formulate an optimization of the GADOP framework as a stochastic integer programming problem. We model the uncertainty of drones including takeoff and breakdown conditions. Furthermore, to accelerate the computational time of solving the problem, we adopt L-shape [@ref_L-shaped] to decompose the GADOP optimization.
System Model and Assumptions {#sec_system}
============================
![The time-line of the scenarios.[]{data-label="f_timeline"}](timeline.pdf){width="50.00000%"}
![An example of scenario tree with one drone. []{data-label="f_scenariotree"}](scenariotree.pdf){width="50.00000%"}
We consider package delivery service of a supplier. In this paper, we consider two delivery modes, i.e., truck delivery mode and drone delivery mode (Figure \[f\_system\]). The truck delivery mode is for the supplier to rent vehicles, e.g., lorry trucks, and use the vehicles to deliver packages to customers. The drone delivery mode is for the supplier to rent drones and use the drones to deliver packages to customers. While trucks have a higher initial cost, e.g., rental fee and driver wage, and a higher traveling cost, i.e., fuel, than those of the drones, the trucks have larger capacity and are more reliable.
Drone delivery is less reliable than the truck delivery as the drones can be affected by weather and other technical problems. Consequently, the drone delivery has a high breakdown probability. In this paper, we specifically consider two operability conditions of the drone, i.e., a takeoff condition and a breakdown condition. The takeoff condition indicates whether the drone can take off from the depot or not, which is observable before the drone departs. For example, the drone cannot takeoff from the depot if there is rain or strong wind. The breakdown condition happens after the drone already takes off from the depot, e.g., due to an accident. Therefore, the mulfunctioned drone needs to be collected from the site and repaired, which also incurs a certain cost. Furthermore, the supplier will pay a penalty composed of the fee for outsourcing the delivery to a carrier and the compensation to the customer due to the delayed delivery.
The supplier faces a problem of utilizing trucks and drones for package delivery for their customers. Trucks and drones need to be reserved in advance. Furthermore, the supplier has to take the uncertainties including drone takeoff and breakdown conditions into account. Therefore, we formulate the GADOP framework as a three-stage stochastic integer programming model.
- **First Stage:** the supplier decides (i) whether customers will be served by either trucks or drones, (ii) how many trucks and drones need to be reserved, (iii) the routing path of trucks, and (iv) the serving order of drones’ delivery. The decisions are made based on the probability of the uncertainties.
- **Second Stage:** the supplier calculates the penalty payment for customers which cannot be served due to the drone takeoff condition.
- **Third Stage:** the supplier calculates the penalty payment for customers which cannot be served due to the drone breakdown condition. The breakdown condition is associated with each drone and each customer. Additionally, the supplier takes the repair cost into account when the breakdown happens.
Figure \[f\_timeline\] shows the stages of the GADOP framework. In the first stage, the decisions are made before the actual drone takeoff condition is observed. In the second stage, the decisions are made after the takeoff condition is observed and before the delivery begins. The third stage decisions are made when an actual breakdown of the drone is observed. Figure \[f\_scenariotree\] shows an example of a scenario tree with one drone for the supplier. The formal definition of the scenarios is presented in Section \[sec\_uncertainty\].
The supplier has a set of customers to be served denoted by $\mathcal{C}=\{C_1,C_2,\dots,C_{c'}\}$, where $c'$ denotes the total number of customers. Without loss of generality, each customer in $\mathcal{C}$ has one package to be delivered, and $g_i$ denotes the package weight of customer $i$. The supplier will assign customers’ packages to either trucks or drones in the first stage. We assume that all the customers have a dropoff station for the drone delivery. The supplier can choose to reserve and use trucks from the set $\mathcal{T} = \{T_1,T_2, \dots,T_{t'}\}$ and drones from the set $\mathcal{D} = \{D_1,D_2, \dots,D_{d'}\}$. $t'$ and $d'$ denote the total numbers of available trucks and drones, respectively. Every truck in the set $\mathcal{T}$ has a capacity limit ($\bar{f}_t$), daily distance traveling limit ($\bar{l_t}$), and daily traveling time limit ($\bar{h_t}$). The package dropoff time of a truck is denoted by $\bar{r}$. Similarly, every drone in the set $\mathcal{D}$ has a capacity limit ($\widehat{f}_d$), daily flying distance limit ($\widehat{l}_d$), and flying distance limit per trip ($\widehat{e}_d$). The package dropoff time of a drone is assumed to be negligible.
The supplier has one depot. The traveling distance from location $i'$ to location $j'$ is represented as $k_{i',j'}$, where $i'$ and $j'$ are the locations being members of the set $\mathcal{C}\cup\{0\}$, where $0$ represents the depot. The average truck driving speed and the drone flying speed are denoted by $\bar{q}_{i',j'}$ and $\widehat{q}_{i',j'}$, respectively. We assume that a drone can carry only one package per trip, and therefore the drone needs to return to the depot after delivering one package. A similar assumption is made in the literature, e.g., [@sidekick], [@ref_roundtrip].
Uncertainty {#sec_uncertainty}
-----------
The drone delivery is much less reliable than the truck delivery [@ref_reliable]. Therefore, we consider the uncertainty of the drones including takeoff and breakdown conditions. Again, the takeoff condition and breakdown condition are observable in the second and third stages, i.e., before and after taking off, respectively. Let $\Omega=\{\omega_1,\omega_2, \dots,\omega_{\omega'}\}$ be the set of the takeoff scenarios, where $\omega'$ represents the total number of the scenarios. The takeoff scenario is defined as $\omega = (\mathbb{R}_1, \mathbb{R}_2, \dots, \mathbb{R}_{d'})$, the maximum number of takeoff scenarios is based on the total number of drones, i.e., $d'$. $\mathbb{R}_d =1$ when drone $D_d$ cannot take off from the depot, and $\mathbb{R}_d =0$ otherwise. For example, $\omega = (0,0,1)$ means that the supplier has three drones, and drones $D_1$ and $D_2$ can take off while drone $D_3$ cannot take off from the depot. Let $\Lambda = \{\lambda_1,\lambda_2,\dots,\lambda_{\lambda'}\}$ denote the set of breakdown scenarios, where $\lambda_{\lambda'}$ represents the total number of the scenario. The breakdown scenario $\lambda$ contains the breakdown parameters indicating which drones are broken during the delivery. The scenario is defined as follows:
=
\_[1,1]{} & \_[1,2]{} &…& \_[1,d’]{}\
\_[2,1]{} & \_[2,2]{} &…& \_[2,d’]{}\
& & …&\
\_[c’,1]{} &\_[c’,2]{} & …&\_[c’,d’]{}
, \[eq\_matrix\]
where $\mathbb{B}_{i,d} =1$ when drone $D_d$ breaks down while serving customer $C_i$, and $\mathbb{B}_{i,d} = 0$ otherwise. Additionally, $\mathbb{P}(\omega)$ and $\mathbb{P}(\lambda)$ denote the probabilities that scenario $\omega$ and scenario $\lambda$ will happen, respectively. In reality, the probabilities can be obtained from the history records.
Payments {#sec_sys_payment}
--------
The supplier involves six types of payments, i.e., costs, including (i) the initial costs of trucks, (ii) the initial costs of drones, (iii) the traveling costs of trucks, (iv) the traveling costs of drones, (v) penalty of unsuccessful drone delivery, and (vi) drone repair cost. The initial costs are the fixed costs such as rental fee, driver or staff wages. $\bar{c}_t^{(i)}$ and $\widehat{c}_d^{(i)}$ represent the initial costs of trucks and the initial costs of drones, respectively. The traveling costs are functions of traveling distance, fuel cost, or battery recharge cost. $\bar{c}_t^{(r)}$ and $\widehat{c}_d^{(r)}$ represent the traveling costs of trucks and the traveling costs of drones, respectively. The penalty and the repair cost, denoted as $p$ and $m$, respectively, are incurred when a drone does not successfully deliver a package.
Problem Formulation {#sec_formulation}
===================
$$\begin{aligned}
& \mbox{Minimize:} \nonumber \\
& \sum_{t \in \mathcal{T}}\bar{c}^{(i)}_t \bar{W}_{t} +
\sum_{d \in \mathcal{D}}\widehat{c}^{(i)}_d \widehat{W}_d + \sum_{\substack{i', j'\in \\ \mathcal{C}\cup\{0\}}}\sum_{t\in \mathcal{T}} \bar{c}^{(r)}_{i',j'}V_{i',j',t} + \mathbb{E}(\mathscr{L}(\widehat{X}_{i,d})),
\label{eq_obj1}\end{aligned}$$
$$\begin{aligned}
&\text{where}\nonumber \\
&\mathscr{L}(\widehat{X}_{i,d}) &=& \sum_{\omega \in \Omega} \mathbb{P}(\omega)
\left(
\sum_{i \in \mathcal{C}}\sum_{d \in \mathcal{D}}\widehat{c}^{(r)}_{i}\widehat{X}_{i,d}(1-\mathbb{R}_d(\omega) )+\right.\nonumber \\
&&&\left.\sum_{d \in \mathcal{D}} pZ^{(b)}_{i,d}(\omega) + \mathbb{E}(\mathscr{M}(\widehat{X}_{i,d}, \omega)) \right), \label{eq_obj2}\\
&\mathscr{M}(\widehat{X}_{i,d},\omega) &=& \sum_{\lambda \in \Lambda}\mathbb{P}(\lambda) \left(\sum_{i \in \mathcal{C} }\sum_{d \in \mathcal{D} }pZ^{(a)}_{i,d}(\omega,\lambda) \right.
\nonumber \\
&&&\left. + \sum_{d \in \mathcal{D}} mZ^{(m)}_{d}(\omega,\lambda)\right),
\label{eq_obj3}\end{aligned}$$ subject to: (\[con\_ini\_ftl\]) to (\[eq\_con\_lastbound\]).
The GADOP framework is a three-stage stochastic integer programming as modeled in (\[eq\_obj1\]) for the first stage, (\[eq\_obj2\]) for the second stage, and (\[eq\_obj3\]) for the third stage. $\mathbb{E}(\mathscr{L}(\widehat{X}_{i,d}))$ denotes the expectation of the traveling cost and the penalty of drones over random takeoff conditions occurred in the second stage. $\mathbb{E}(\mathscr{M}(\widehat{X}_{i,d},\omega))$ denotes the expectation of the penalty cost and the repair cost of drones over random takeoff conditions and random breakdown conditions occurred in the third stage. The objective is to minimize the total delivery payments as discussed in Section \[sec\_sys\_payment\]. The decision variables in the framework are listed below.
- $\bar{W}_{t}$ is an indicator whether truck $t$ will be used or not. When $\bar{W}_{t} =1$, truck $t$ will be used, and $\bar{W}_{t}=0$ otherwise.
- $\widehat{W}_d$ is an indicator whether drone $d$ will be used or not. When $\widehat{W}_d =1$, drone $d$ will be used, and $\widehat{W}_d =0$ otherwise.
- $V_{i',j',t}$ is a routing variable in which $V_{i',j',t} =1$ if the route from location $i'$ to location $j'$ will be used by truck $t$, and $V_{i',j',t}=0$ otherwise.
- $\widehat{X}_{i,d}$ is an allocation variable of drone $d$. $\widehat{X}_{i,d}=1$ if drone $d$ will serve customer $i$, and $\widehat{X}_{i,d}=0$ otherwise.
- $\bar{X}_{i,t}$ is an allocation variable of truck $t$. $\bar{X}_{i,t} =1$ if truck $t$ will serve customer $i$, and $\bar{X}_{i,t}=0$ otherwise.
- $Z^{(b)}_{i,d}(\omega)$ is a penalty variable. $Z^{(b)}_{i,d}(\omega)=1$ if the penalty associated to customer $i$, which will be served by drone $d$, but the drone cannot take off from the depot, is paid in scenario $\omega$, and $Z^{(b)}_{i,d}(\omega)=0$ otherwise.
- $Z^{(a)}_{i,d}(\omega,\lambda)$ is another penalty variable. $Z^{(a)}_{i,d}(\omega,\lambda)=1$ if the penalty associated to customer $i$, which will be served by drone $d$, but the drone breaks down, is paid in scenarios $\omega$ and $\lambda$, $Z^{(a)}_{i,d}(\omega,\lambda)=0$ otherwise.
- $Z^{(m)}_{d}(\omega,\lambda)$ is a repair variable. $Z^{(m)}_{d}(\omega,\lambda)=1$ if drone $d$ in scenario $\omega$ and $\lambda$ breaks down during delivery, and $Z^{(m)}_{d}(\omega,\lambda)=0$ otherwise.
- $S_{i,t}$ is an auxiliary variable for sub-tour elimination in truck delivery.
- $U_{i,d}$ is a serving order of the drones. $U_{i,d} < U_{j,d}$ means that drone $d$ will serve customer $i$ before customer $j$.
- $M_{i,j,d}$ is an auxiliary binary variable for ensuring the nonequivalent of $U_{i,d}$.
The types and bounds of all decision variables are indicated in constraints (\[eq\_con\_firstbound\])-(\[eq\_con\_lastbound\]).
The constraints in (\[con\_ini\_ftl\]) and (\[con\_ini\_uav\]) ensure that the initial cost of trucks and the initial cost of drones will be paid if they are used, where $\Delta$ is any large number $\Delta \geq c'$. The constraints in (\[con\_capacity\_ftl\]) and (\[con\_capacity\_uav\]) ensure that the total weight of packages does not exceed the capacity limit of a truck and a drone, respectively. The constraint in (\[con\_fly\_trip\]) ensures that the flying distance between the depot to a customer and vice verse does not exceed the flying limit. The daily traveling distance limits are controlled by the constraints in (\[con\_fly\_day\]) and (\[con\_drive\_day\]) for drone delivery and truck delivery, respectively. The total time of the truck delivery, i.e., traveling time and serving time, must not exceed the working hours as imposed in the constraint in (\[con\_serving\]).
&\_[i ]{} |[X]{}\_[i,t]{} |[W]{}\_[t]{}, & t \[con\_ini\_ftl\]\
&\_[i ]{} \_[i,d]{} \_d, & d \[con\_ini\_uav\]\
&\_[i]{}g\_i|[X]{}\_[i,t]{} |[f]{}\_t, &t \[con\_capacity\_ftl\]\
&g\_i\_[i,d]{} \_d, & i , d \[con\_capacity\_uav\]\
&(k\_[0,i]{}+k\_[i,0]{})\_[i,d]{} \_d, & i , d \[con\_fly\_trip\]\
&\_[i C]{}(k\_[0,i]{}+k\_[i,0]{})\_[i,d]{} \_d, & d \[con\_fly\_day\]\
&\_k\_[i’,j’]{}V\_[i’,j’,t]{} |[l]{}\_[t]{}, & t \[con\_drive\_day\]
The allocation of trucks and drones is imposed by the constraint in (\[con\_allocation\]), where one customer can be served by either a truck or a drone. The constraints in (\[con\_routing\_1\])-(\[con\_subtour\]) are to find the route for truck delivery. Only the allocated truck can take the arrival and departing route to serve the customer. The constraint in (\[con\_subtour\]) ensures that no subtour exists in the solution.
If drones cannot take off from the depot, the penalty must be paid for all associated customers, as indicated in the constraint in (\[con\_2\_stage\]). In the case that the drones can take off, but they are broken after departing, the constraint in (\[con\_3\_stage\]) ensures that the penalty must be paid. When breakdown occurs after departing from the depot, the drone will be no longer able to serve all the next customers. Consequently, the penalty associated to these customers must be paid as indicated in the constraint in (\[con\_order\_main\]). Note that the next customer $j$ after customer $i$ has a higher serving order, i.e., $U_{j,d} > U_{i,d}$. Moreover, the constraint in (\[con\_repair\]) ensures that the repair cost must be paid if the drone is broken after departing from the depot.
The constraints in (\[eq\_con\_order3\])-(\[eq\_con\_order2\]) decide serving orders of drones. The constraint in (\[eq\_con\_order3\]) ensures that $U_{i,d}$ must not be zero when drone $d$ is allocated to serve customer $i$, i.e., $\ddot{X}_{i,d}=1$. $U_{i,d} = 0$ which represents the drone serving order is thus not needed. The constraints in (\[eq\_con\_order1\])-(\[eq\_con\_order2\]) ensure that the serving orders must not be the same when $\ddot{X}_{i,d}=1$. Again, $M_{i,j,d}$ is an auxiliary binary variable for ensuring that $U_{i,d} \neq U_{j,d}$ when $\widehat{X}_{i,d} =1$.
&\_ ( + |[r]{} ) V\_[i’,j’,t]{} |[h]{}\_t, & t \[con\_serving\]\
&\_[t]{}|[X]{}\_[i,t]{} + \_[d ]{} \_[i,d]{} = 1, & i \[con\_allocation\]
&\_[i ]{} V\_[0,i,t]{}1, & t \[con\_routing\_1\]\
&\_[i ]{} V\_[i,0,t]{} 1, & t\
&\_[i’ {0}]{} V\_[i’,i,t]{} = |[X]{}\_[i,t]{}, & t , i\
&\_[i’ {0}]{} V\_[i,i’,t]{} = |[X]{}\_[i,t]{}, & t , i\
& V\_[i’,i’,t]{} = 0, & t , i’ {0}
&S\_[i,t]{} - S\_[j,t]{} + c’V\_[i,j,t]{} c’-1, i , j , t \[con\_subtour\]\
&\_[i,d]{}\_d() = Z\^[(b)]{}\_[i,d]{}(), i , d , \[con\_2\_stage\]\
&\_[i,d]{}(1-\_d())\_[i,d]{}() Z\^[(a)]{}\_[i,d]{}(,),\
& i , d , , \[con\_3\_stage\]\
&( U\_[i,d]{}- U\_[j,d]{}) ( 1 - Z\^[(a)]{}\_[j,d]{}(,) + Z\^[(a)]{}\_[i,d]{}(,) ),\
& i, j , ij, d , , \[con\_order\_main\]\
&Z\^[(m)]{}\_[d]{}(,) Z\^[(a)]{}\_[i,d]{}(,),\
& i , d , , \[con\_repair\]
&\_[i,d]{} U\_[i,d]{}, & d , i \[eq\_con\_order3\]\
&0 U\_[j,d]{} \_[i ]{}\_[i,d]{}, & d , j \[eq\_con\_order4\]
&U\_[i,d]{} - U\_[j,d]{} M\_[i,j,d]{}-\_[i,d]{},\
& i ,j , i j,d \[eq\_con\_order1\]\
&U\_[i,d]{} - U\_[j,d]{} \_[i,d]{} - ( 1- M\_[i,j,d]{}),\
& i ,j , i j,d \[eq\_con\_order2\]
In order to enhance customers’ satisfaction, suppliers allow their customers to specific a delivery time slot, which can be referred to as time windows. Time windows constraint can be hard time windows or soft time windows. In the hard time window, all customers must be served within time windows. On the other hand, the soft time window is more relaxed, and thus the supplier can choose to pay for the penalty of not delivery on time. We consider the hard time window in this paper. The constraints in (\[eq\_con\_tw\_1\])-(\[eq\_con\_tw\_2\]) ensures that the delivery is done within the assigned time windows. $\mathcal{L}^{(m)}$ and $\mathcal{L}^{(a)}$ denote a set of customers who require to be served in the morning and afternoon, respectively, and we have $\mathcal{L}^{(m)} \subseteq \mathcal{C}$, $\mathcal{L}^{(a)} \subseteq \mathcal{C}$ and $\mathcal{L}^{(m)} \cap \mathcal{L}^{(a)} = \emptyset$. The constraints in (\[eq\_con\_tw\_1\]) and (\[eq\_con\_tw\_2\]) ensure that all morning customers must be served before the afternoon customers. The constraints in (\[eq\_con\_tw\_d\_1\])-(\[eq\_con\_tw\_t\_2\]) ensure that traveling time of trucks and drones does not exceed the morning and afternoon traveling time limits, i.e., time windows. The morning and afternoon traveling time limits are defined as $l^{(m)}$ and $l^{(a)}$, respectively. More specific time slots and explaination of hard time window can be found in [@ref_VTC1].
& U\_[i,d]{} U\_[j,d]{} + (1- \_[i,t]{}),\
& i \^[(m)]{}, j \^[(a)]{}, d , \[eq\_con\_tw\_1\]\
& S\_[i,t]{} S\_[j,t]{} + (1- |[X]{}\_[i,t]{}),\
&i \^[(m)]{}, j \^[(a)]{}, t , \[eq\_con\_tw\_2\]
& \_[m \^[(m)]{}]{}( + )\_[m,d]{} l\^[(m)]{},& d \[eq\_con\_tw\_d\_1\]\
& \_[a \^[(a)]{}]{}( + )\_[a,d]{} l\^[(a)]{},& d \[eq\_con\_tw\_d\_2\]\
&\_\_ ( + |[r]{} ) V\_[i’,m,t]{} l\^[(m)]{}, & t \[eq\_con\_tw\_t\_1\]\
&\_\_ ( + |[r]{} ) V\_[i,a’,t]{} l\^[(a)]{}, & t \[eq\_con\_tw\_t\_2\]
& \_[i,d]{}, |[X]{}\_[i,t]{}, \_d, |[W]{}\_t {0,1}, d , t , i \[eq\_con\_firstbound\]\
& V\_[i’,j’,t]{} {0,1}, i ’,j’ {0}, t\
& U\_[i,d]{}, S\_[i,t]{} {0,1,…, c’}, d , t , i \[eq\_con\_secondlast\]\
& Z\^[(b)]{}\_[i,d]{}(), Z\^[(a)]{}\_[i,d]{}(,), Z\^[(m)]{}\_[d]{}(,), M\_[i,j,d]{} {0,1},\
& d , i,j , , \[eq\_con\_lastbound\]
The above optimization problem can be complex to solve when the numbers of customers, trucks, and drones grow. Therefore, in the next section, we present the decomposition method to obtain an optimal solution of the problem more efficiently.
Decomposition {#sec_decom}
=============
The GADOP optimization, which is presented in the last section, is an NP-hard problem. The computational time for solving the problem depends on the numbers of variables, parameters, and constraints. To address the computational issue, we decompose the GADOP optimization into multiple sub-problems according to the decision stages, i.e., one first-stage problem, $\omega'$ second-stage sub-problems, and $\omega'$ third-stage sub-problems. Each second-stage sub-problem and each third-stage sub-problem are associated with a certain takeoff scenario, i.e., $\omega \in \Omega$. The decomposition is executed iteratively as described in the following steps:
(i) Solving the first-stage problem, which is the master problem,
(ii) Solving all the second-stage sub-problems and the third-stage sub-problems using the results from the master problem as the parameters,
(iii) Calculating feedback parameters, i.e., $E^{(p)}_{i,d}$, $E^{(r)}_{i,d}$, and $E^{(m)}_{i,d}$ representing the penalty, the traveling cost of drones, and drone repair cost from the second-stage sub-problems and third-stage sub-problem, and
(iv) Checking the convergence condition, if the convergence condition is met, then the solution is found, and otherwise the feedback parameters are sent to the master problem, and Step (i) repeats with the additional constraints applying the feedback parameters.
This decomposition is based on the L-Shape method [@ref_L-shaped]. The algorithm to execute the decomposition is shown in Algorithm \[al\_decom\].
![Decomposition of scenarios and stages where a circle represents one optimization master problem or sub-problem.[]{data-label="f_parallel"}](parallel.pdf){width="50.00000%"}
Figure \[f\_parallel\] presents an example of the GADOP optimization with five takeoff scenarios and ten breakdown scenarios. After decomposed, the example has one master problem, five second-stage sub-problems, and five third-stage sub-problems to be solved. Note that the numbers of the second-stage sub-problems and the third-stage sub-problems depend on the number of the takeoff scenarios. In fact, after applying the decomposition, we have found that the solution of the second-stage problem can be obtained by a simple calculation due to special structure of the GADOP problem. As a result, for the same takeoff scenario, the second-stage sub-problem and the third-stage sub-problem do not have any interdependency with each other, and thus we can solve the second-stage sub-problems and the third-stage sub-problems simultaneously. Additionally, a takeoff scenario does not have any interdependency with other takeoff scenarios. However, dependencies of breakdown scenarios exist in serving orders ($U_{i,d}$). Since the serving orders need to be decided before the delivery begins, the same serving order ($U_{i,d}$) is used for all breakdown scenarios. As a result, the third-stage optimization cannot be further decomposed according to breakdown scenarios.
Master Problem or First-Stage Problem
-------------------------------------
Minimizing:
\_[t ]{}|[c]{}\^[(i)]{}\_t |[W]{}\_[t]{} + \_[d ]{}\^[(i)]{}\_d \_d + \_\_[t]{} |[c]{}\^[(r)]{}\_[i’,j’]{}V\_[i’,j’,t]{} + \_1 + \_2, \[eq\_obj\_dec\_master\]
subject to: (\[con\_ini\_ftl\]) to (\[con\_subtour\]), (\[eq\_con\_tw\_2\]) to (\[eq\_con\_tw\_t\_2\]), (\[eq\_cutting\_theta1\]), and (\[eq\_cutting\_theta2\]).
&\_[i ]{}\_[d ]{}(E\^[(p)]{}\_[i,d]{} + E\^[(r)]{}\_[i,d]{} )\_[i,d]{} + \_1 0 & \[eq\_cutting\_theta1\]\
&\_[d ]{}E\^[(m)]{}\_[i,d]{}\_[d]{} + \_2 0, & i \[eq\_cutting\_theta2\]
The objective function of the master problem is presented in (\[eq\_obj\_dec\_master\]). The constraints in (\[eq\_cutting\_theta1\]) and (\[eq\_cutting\_theta2\]) are the feedback cutting constraints [@ref_L-shaped]. Again, $E^{(p)}_{i,d}$, $E^{(r)}_{i,d}$, and $E^{(m)}_{i,d}$ are the feedback parameters, which are dedicated to the penalty, the drone traveling cost, and the drone repair cost, respectively. $\theta_1$ and $\theta_2$ are initially set at small values for the first iteration. After the first iteration, $\theta_1$ and $\theta_2$ become positive decision variables. The results of $\theta_1$ and $\theta_2$ are denoted by $\theta^*_1$ and $\theta^*_2$, respectively, which are used in checking the convergence condition. The result of $\widehat{X}_{i,d}$ is denoted by $\widehat{X}^{*}_{i,d}$, which is used as a parameter in the second-stage sub-problems and the third-stage sub-problems. After decomposition, the first-stage problem is still an NP-hard problem. However, the first-stage problem becomes less complex than the original GADOP because the number of decision variables in the first-stage problem is fewer than that in the original problem.
Second-Stage Sub-Problem
-------------------------
Minimize:
& \_[i ]{}\_[d ]{}\^[(r)]{}\_[i]{}\^[\*]{}\_[i,d]{}(1-\_d )+ \_[d ]{} p\^[(b)]{}\_[i,d]{} , \[eq\_obj\_dec\_2\] d
subject to: (\[eq\_dec\_2\]).
&\^[\*]{}\_[i,d]{}\_d = \^[(b)]{}\_[i,d]{}, & i , d \[eq\_dec\_2\]
We decompose the second-stage sub-problem given a certain takeoff scenario, and thus the variable $\omega$ is removed from the objective function in (\[eq\_obj\_dec\_2\]) and the constraint in (\[eq\_dec\_2\]). To minimize the expected cost in the second stage, i.e., $\mathbb{E}(\mathscr{L}(\widehat{X}_{i,u})) $, we need to solve all the second-stage sub-problems associated to a takeoff scenario ($\omega$). Note that random parameters $\widetilde{\mathbb{R}}_d$ and $\widetilde{\mathbb{B}}_{i,d}$, i.e., for scenario variables ${\mathbb{R}}_d$ and ${\mathbb{B}}_{i,d}$ defined in Section \[sec\_uncertainty\], respectively, as well as the decision variable $\widetilde{Z}^{(b)}_{i,d}$ is associated with the takeoff scenario, and therefore the $\omega$ is omitted.
In general, the second-stage problem is an optimization, and thus we present the second-stage problem in the optimization form. Nevertheless, in our exact problem, there is a special structure that makes the second-stage problem is able to be solved directly through the assignment of $\widehat{X}_{i,d}$ and $\widetilde{\mathbb{R}}_d$.
After applying the decomposition, there is only one decision variable in the second-stage problem that is $\widetilde{Z}^{(b)}_{i,d}$. Furthermore, the decision variable $\widehat{X}_{i,d}$ is solved in the first-stage problem, and thus it becomes a parameter, i.e., $\widehat{X}^*_{i,d}$, in the second-stage problem. According to the constraint in (\[eq\_dec\_2\]), $\widetilde{Z}^{(b)}_{i,d}$ depends on parameters only, and consequently the second-stage problem is no longer an optimization problem.
Third-Stage Sub-Problem
-----------------------
Minimize:
& \_()(\_[i ]{}\_[d ]{}p\^[(a)]{}\_[i,d]{}() + \_[d ]{}m\^[(m)]{}\_[d]{}()), \[eq\_obj\_dec\_3\]
subject to: (\[eq\_con\_order3\]), (\[eq\_dec\_3\_1\]) to (\[eq\_dec\_order2\]).
&\^[\*]{}\_[i,d]{}(1-\_d)\_[i,d]{} \^[(a)]{}\_[i,d]{}(), i , d \[eq\_dec\_3\_1\]\
& U\_[i,d]{}- U\_[j,d]{} ( 1 - \^[(a)]{}\_[j,d]{}() + \^[(a)]{}\_[i,d]{}() ),\
& i, j , ij, d \[eq\_dec\_3\_2\]\
&\^[(m)]{}\_[d]{}() \^[(a)]{}\_[i,d]{}(), i , d \[eq\_dec\_3\_3\]\
&\^[\*]{}\_[i,d]{} U\_[i,d]{}, d , i \[eq\_dec\_order1\]\
&0 U\_[j,d]{} \_[i ]{}\^[\*]{}\_[i,d]{}, d , j \[eq\_dec\_order2\]\
&U\_[i,d]{} - U\_[j,d]{} M\_[i,j,d]{}-\^\*\_[i,d]{},\
& i ,j , i j,d \[eq\_dec\_order3\]\
&U\_[i,d]{} - U\_[j,d]{} \^\*\_[i,d]{} - ( 1- M\_[i,j,d]{}),\
& i ,j , i j,d \[eq\_dec\_order4\]
Similar to the second-stage sub-problem, each third-stage sub-problem is formulated given a certain takeoff scenario. The objective function of the third-stage sub-problem is expressed in (\[eq\_obj\_dec\_3\]) which is the expected cost of the penalty and repair cost of drones. The constraints in (\[eq\_dec\_3\_1\]), (\[eq\_dec\_3\_2\]), and (\[eq\_dec\_3\_3\]) contain the third-stage decision variables, which are $\widetilde{Z}^{(a)}_{i,d}(\lambda)$ and $\widetilde{Z}^{(m)}_{d}(\lambda)$. The decision variables $\widetilde{Z}^{(a)}_{i,d}(\lambda)$ and $\widetilde{Z}^{(m)}_{d}(\lambda)$ are associated with the takeoff scenario, and therefore $\omega$ is omitted. Since there is the interdependency among $U_{i,d}$, i.e., drone serving order, the constraints in (\[eq\_dec\_order1\])-(\[eq\_dec\_order4\]) are directly adopted in this sub-problem. The reason is that $U_{i,d}$ of each takeoff scenario, which depends on $\widehat{X}^{*}_{i,d}$, has already obtained in the master problem. Again, $\widehat{X}^{*}_{i,d}$ is the assignment of a drone to a customer.
Calculating Feedback Parameters and Convergence
-----------------------------------------------
This section explains the feedback parameter calculation and the convergence condition checking. $e_{i,d}^{(p)}(\omega)$ and $e_{i,d}^{(m)}(\omega)$ are the auxiliary parameters in the third-stage sub-problems, which are associated with the drone penalty and the drone repair cost, respectively. The feedback parameters, i.e., $E_{i,d}^{(r)}$, $E_{i,d}^{(p)}$, and $E_{i,d}^{(m)}$, as expressed in (\[eq\_feedback\_routing\]), (\[eq\_feedback\_penalty\]), and (\[eq\_feedback\_maintenance\]), respectively, are returned to the master problem. $E_{i,d}^{(r)}$, $E_{i,d}^{(p)}$, and $E_{i,d}^{(m)}$ are the expected values associated with the drone traveling cost, the drone penalty, and the drone repair cost, respectively. The constraints in (\[eq\_cutting\_theta1\]) and (\[eq\_cutting\_theta2\]) ensure that the supplier pays the routing cost, the penalty, and the repair cost from the second-stage and third-stage when the associated $\widehat{X}_{i,d}$ is selected. Therefore, $E_{i,d}^{(r)}$, $E_{i,d}^{(p)}$, and $E_{i,d}^{(m)}$ must be negative in order to match with $\theta_1$ and $\theta_2$. The formulations of the feedback parameters are given as follow.
E\_[i,d]{}\^[(r)]{} &&=& -\^[(r)]{}\_[i,d]{}\_()( 1- \_d() ), \[eq\_feedback\_routing\]\
E\_[i,d]{}\^[(p)]{} &&=& -\_()( p\_d() + e\_[i,d]{}\^[(p)]{}() ), \[eq\_feedback\_penalty\]\
e\_[i,d]{}\^[(p)]{}() &&=& p\_()(1-\_d())\_[i,d]{}(), \[eq\_feedback\_penalty\_third\]\
E\_[i,d]{}\^[(m)]{} &&=& -\_() e\_[i,d]{}\^[(m)]{}(), \[eq\_feedback\_maintenance\]\
e\_[i,d]{}\^[(m)]{}() &&=& m\_()(1-\_d())\_[i,d]{}(). \[eq\_feedback\_maintenance\_third\]
We next define a convergence parameter denoted by $B$ which is used for checking the convergence condition. If $B$ is less than or equal to $\theta_1^* + \theta_2^*$, the convergence condition is met, and then Algorithm \[al\_decom\] is terminated and the solution is taken from the latest iteration. The convergence parameter is calculated from (\[eq\_convergence\_1\]) and (\[eq\_convergence\_2\]).
The proof of convergence of the L-shape method can be found in [@ref_L-shaped]. In the GADOP optimization problem, we observe that the feedback constraints are based on the probabilities of scenarios, the penalty, the routing cost, and the repair cost. As we observe from the GADOP problem, these feedback parameters can be calculated directly from parameters, i.e., the probabilities of scenarios ($\mathbb{P}(\omega),\mathbb{P}(\lambda)$), the penalty ($p$), the routing cost ($\widehat{C}_{i,d}^{(r)}$), and the repair cost ($m$). Since the feedback parameters do not depend on the solution at an iteration $k$, the feedback terms are constants. According to Algorithm 1, we observe that the first-stage solutions at $k=1$ and $k=2$ are identical because the feedback terms are the same. Therefore, the convergence of the decomposed GADOP is guaranteed at $k=2$.
&B = \_[i ]{}\_[d ]{}\^[\*]{}\_[i,d]{}(E\_[i,d]{}\^[(p)]{}+E\_[i,d]{}\^[(r)]{}) + M\_d\^[\*]{}\_[d]{},\[eq\_convergence\_1\]\
&M\_d = (E\_[i,d]{}\^[(m)]{}), i \[eq\_convergence\_2\]
Evaluation {#sec_eva}
==========
Parameter setting
-----------------
We implement the GADOP framework as a three-stage integer programming by using GAMS [@ref_gams]. The experiments presented in this section are solved by the CPLEX solver. We use CPLEX default setting which comes with the heuristic option. Unless otherwise stated, we consider three drones and one truck in the GADOP framework. All the three drones are of the same type, the initial costs of which are $\widehat{c}^{(i)}_1,\widehat{c}^{(i)}_2, \widehat{c}^{(i)}_3 = S\$100$ and the capacity limit of which are $\widehat{f}_1,\widehat{f}_2,\widehat{f}_3= 2$ kilograms. The daily flying distance limit and the flying distance limit per trip are $\widehat{l}_d = 150$ kilometers and $\widehat{e}_d = 15$ kilometers for each drone, respectively. The truck is a panel van, the initial cost of which is $\bar{c}^{(i)}_1 = S\$280$ and the capacity limit of which is $\bar{f}_1 = 1,060$ kilograms. The daily distance traveling limit and daily traveling time limit are $\bar{l_t} = 200$ kilometers and $\bar{h_t}= 8$ hours, respectively. The driving speed is set at $q_{i',j'} = 50$ kilometers per hour. The initial costs of the drones and the truck are synthesized based on Singapore shipping services [@ref_vehicle]. The traveling cost of the truck is approximated to $\bar{c}^{(r)}_{i',j'} = k_{i',j'} \times 1.05 \times 0.1$ based on Singapore government statistics [@ref_gov], where $1.05$ is the approximate fuel price factor and $0.1$ is the average fuel consumption factor. The cost of a drone traveling round-trip between location $i$ and the depot is $\bar{c}^{(r)}_i = 0.005 \times (k_{i,0} + k_{0,i}) $, where $0$ is the index of the depot.
Furthermore, all customers have the same package weight $g_i = 1$ kilogram. The serving time of the truck is $\bar{r} = 15$ minutes while the serving time of the drone is negligible because the drone just only drops a package off. The penalty and repair cost are $p = S\$20$ and $m = S\$50$, respectively. The penalty is set based on the cost of outsourcing package delivery to the Speedpost service [@ref_singpost] by Singpost carrier plus additional cost for collecting the package or giving the compensation to a customer. Note that all customers have no specific time windows due to the space limit. Nevetheless, the experiment associated with time windows can be found in [@ref_VTC1].
We evaluate the GADOP framework using two geographically data sets, i.e., Solomon Benchmark suite C101 [@ref_solomon] and a real trace from one of the Singapore logistics companies. Since Singapore has many high buildings, i.e., residential buildings, we use Singapore road network as the drone flying paths. The distances in kilometers between one location to another are obtained from the Google Map [@ref_googleAPI]. Solomon Benchmark suite is used in all experiments except the experiment in Section \[sec\_real\].
$\begin{array}{ccc}
\includegraphics[trim={5em 0 0 0},clip,width=0.35\textwidth]{nob-input.eps} & \hspace{-2.2em}
\includegraphics[trim={5em 0 0 0},clip,width=0.35\textwidth]{nob-ftl.eps} & \hspace{-2.2em}
\includegraphics[trim={5em 0 0 0},clip,width=0.35\textwidth]{nob-uav.eps} \vspace{-1.5em}\\
\vspace{-0.5em}
(a)& \hspace{-2.2em}(b) &\hspace{-2.2em} (c)
\end{array}$
Tradeoff between taking and not-taking uncertainty into account {#sec_tradeoff}
----------------------------------------------------------------
We evaluate the GADOP framework with 40 customers. The Solomon benchmark suite is tested under (i) the deterministic case where the drones are always able to takeoff and no breakdown occurs and (ii) the random case where the drones may not be able to take off and breakdown may occur.
In the deterministic case, the scenarios are fixed as $\Omega =\{\omega_1\}$, $\Lambda = \{\lambda_1\}$, and all $\mathbb{R}_d(\omega_1) =0, \mathbb{B}_{i,d}(\lambda_1) = 0$. The customer locations and the drone flying distance limit are presented in Figure \[f\_solomon\_ex\_nob\](a). The solution of the optimization in the deterministic case is to serve 17 and 23 customers by one drone and one truck, respectively. The routing path of the truck is presented in Figure \[f\_solomon\_ex\_nob\](b), and the customers that are served by the drone are presented in Figure \[f\_solomon\_ex\_nob\](c). The total payment is $S\$386.29$, which includes the initial cost of the truck $S\$280$, the initial cost of the drone $S\$100$, the routing cost of the truck $S\$5.59$, and the routing cost of the drone $S\$0.70$. The total traveling distances of the drone and truck are 140.27 and 53.26 kilometers, respectively.
Furthermore, in the random case, there are two scenarios for each of takeoff condition and breakdown condition ($\Omega =\{\omega_1,\omega_2\}$, $\Lambda = \{\lambda_1,\lambda_2\}$). The probabilities of two takeoff scenarios are $\mathbb{P}(\omega_1)=0.9$ and $\mathbb{P}(\omega_2)=0.1$, where all drones can takeoff in scenario $\omega_1$ and cannot takeoff in scenario $\omega_2$. The probabilities of two random breakdown scenarios are also set to $\mathbb{P}(\lambda_1)=0.9$ and $\mathbb{P}(\lambda_2) =0.1$. The solution of the optimization in the random case is to serve 14 and 26 customers by one drone and one truck, respectively. The solution is presented in Figure \[f\_solomon\_ex\_b\]. The total payment is $S\$414.25$ including the initial cost of the truck $S\$280$, the initial cost of the drone $S\$100$, the routing cost of the truck $S\$5.736$, the routing cost of the drone $S\$0.51$, and the penalty incurred from the takeoff scenarios $S\$28$. The total traveling distances of the drone and the truck are 114.44 and 54.63 kilometers, respectively. Obviously, only drone delivery cannot serve all customers because some customers are outside the flying area. We observe that the solution assigns the drone to serve as many customers as possible in the deterministic case because the traveling cost of the drone is cheaper. Note that daily traveling distance limit of the drone is 150 kilometer. On the contrary, in the random case, since the drone may not be able to take off or the breakdown can occur, the solution avoids the penalty and repair cost by assigning more number of customers to the truck.
$\begin{array}{cc}
\includegraphics[trim={5em 0 0 0},clip,width=0.35\textwidth]{map-ftl-2.eps} & \hspace{-2.2em}
\includegraphics[trim={5em 0 0 0},clip,width=0.35\textwidth]{map-uav.eps} \vspace{-1.5em}
\\(a) & \hspace{-2.2em} (b)
\end{array}$
\[f\_solomon\_ex\_b\]
Routing, payment, and explanation of the real dataset {#sec_real}
-----------------------------------------------------
We next test the GADOP framework with the real data set from one of the Singapore logistics companies. In this experiment, we use the same takeoff and breakdown scenarios as the random case in Section \[sec\_tradeoff\], i.e., the two takeoff scenarios and two breakdown scenarios.
The solution of the optimization is presented in Figure \[f\_routeMap\], which is to use one drone to serve 13 customers and one truck to serve 27 customers. The total payment is $S\$411.76$ including the initial cost of the truck $S\$280$, the initial of the drone $S\$100$, the routing cost of the truck $S\$5.17$, the routing cost of the drone $S\$0.58$, and the penalty incurred from the takeoff scenarios $S\$26$. The total traveling distances of the drone and the truck are 129.20 and 49.31 kilometers, respectively. We observe that the truck is used for the customers which are clustered in the same area or located along the route. Furthermore, the drone tends to be used for the customer far away from each other such as customer 1 at the bottom right corner of the map (Figure \[f\_routeMap\]).
$\begin{array}{cc}
\includegraphics[trim={0 0 15em 5em},clip,width=0.49\textwidth]{UAV-route-red-new.pdf} &
\includegraphics[trim={0 0 15em 5em},clip,width=0.49\textwidth]{FTL-route-new.pdf} \\
(a)\text{ Drone delivery} & (b)\text{ Truck delivery}
\end{array}$
{width="\textwidth"}\
$(a)$
\
$(b)$
{width="\textwidth"}\
\[f\_increase\_customers\]
\[f\_compare\]
(0, 0) node\[inner sep=0\] [{width="\textwidth"} ]{}; (0.3, -2.8) node [($\mathbb{P}(\omega_2)$)]{};
\
(0, 0) node\[inner sep=0\] [{width="\textwidth"} ]{}; (-2, -2.1) node [($\lambda_2$)]{}; (1.2, -1.9) node [($\mathbb{P}(\lambda_2)$)]{};
{width="\textwidth"}
\[f\_breakdown\_before\]
\[f\_penalty\]
$\begin{array}{ccc}
\hspace{-3.1em}
\includegraphics[width=0.39\textwidth]{maintenance-nopenalty.eps} &
\hspace{-2.8em}
\includegraphics[width=0.39\textwidth]{maintenance-penalty20.eps} &
\hspace{-2.8em}
\includegraphics[width=0.39\textwidth]{sim3.eps} \\
\hspace{-2.8em} (a) &\hspace{-2.8em} (b) &
\end{array}$
\[f\_repair\]
\[f\_simulation\]
Impacts of the number of customers and drone flying limit
---------------------------------------------------------
We vary the number of customers and the drone flying limit per trip. In this experiment, we test the GADOP framework with one scenario which all drones can take off and no breakdown occurs, i.e., the deterministic case. The results and the locations of the customers are presented in Figure \[f\_increase\_customers\]. If the drone flying limit per trip is large, i.e., 12.5 kilometers, all the customers are assigned to a single or multiple drones without trucks. However, when the number of the customers is large, i.e., 40, the truck and drone are used together. The reason is that three drones cannot serve all 40 customers in one day. If the drone flying limit per trip is small, i.e., 5 kilometers, the truck is selected to serve all the customers when the number of the customers is less than 25. The reason is that the drone alone cannot serve the customers farther than its flying distance limit while the truck is able to serve all the customers. When the number of the customers is larger than or equal to 30, the truck alone cannot serve all the customers because of its daily traveling time limit, and thus both the truck and drone are used together when the drone flying limit per trip is less than or equal to 10 kilometers.
We conduct the performance comparison between (i) stand-alone ground-based delivery in which only ground-based trucks are used to serve customers, and (ii) joint truck and drone delivery services. As presented in Figure \[f\_compare\], the stand-alone ground-based delivery is always the most expensive delivery mode. This result clearly shows the effectiveness of the proposed joint truck and drone delivery.
Impact of random takeoff scenarios
----------------------------------
We evaluate the GADOP framework with 25 customers, two takeoff scenarios, and one breakdown scenario, i.e., $\Lambda = \{\lambda_1\}$, and all the drones do not break down, i.e., $\mathbb{B}_{i,d}(\lambda_1) = 0$. The two takeoff scenarios are (i) all the drones can take off ($\omega_1$) and (ii) all the drones cannot take off ($\omega_2$). We vary the probability of that all the drones cannot take off, i.e., $\mathbb{P}(\omega_2)$, and $\mathbb{P}(\omega_1) = 1- \mathbb{P}(\omega_2)$. We ignore the flying distance limit in this case for the ease of presenting the impact of the takeoff scenarios. The results are presented in Figure \[f\_breakdown\_before\]. The drone is used to serve all the customers when the probability is less than or equal to 0.3. Otherwise, the truck is used to serve all the customers.
Impact of random breakdown scenarios {#sec_afterdepart}
------------------------------------
Next we consider 25 customers, one takeoff scenario $\Omega=\{\omega_1\}$, i.e., all the drones can take off $\mathbb{R}_d(\omega_1) = 0$, and two breakdown scenarios, i.e., $\Lambda=\{\lambda_1,\lambda_2\}$. For the two breakdown scenarios, (i) there is no breakdown, i.e., all $\mathbb{B}_{i,d}(\lambda_1) = 0$, and there is breakdown for some customers. For example, breakdown percentage equal to $20\%$ means 5 customers have breakdown and 20 customers have no breakdown, which can be referred to $\mathbb{B}_{i,d}(\lambda_2) = 1$ for $i=1,\dots,5$, and $\mathbb{B}_{i,d}(\lambda_2) = 0$ for $i=6,\dots,25$. We vary the probability of the breakdown scenario $\mathbb{P}(\lambda_2)$ and the percentage of the breakdown customers. Note that we again ignore the flying distance limit. The results are presented in Figure \[f\_breakdown\_after\_3d\]. From the results, the GADOP framework always selects the drone to serve the customers when the percentage of the breakdown customers is less than or equal to $20\%$ or the probability of the breakdown scenario $\mathbb{P}(\lambda_2)\leq 0.2$.
Impact of penalty {#sec_penalty}
-----------------
We use the same setting and scenarios as in Section \[sec\_afterdepart\], but the probabilities of the breakdown scenarios are equal, i.e., $\mathbb{P}(\lambda_1) = \mathbb{P}(\lambda_2) = 0.5$. The results are presented in Figure \[f\_penalty\]. When the breakdown penalty cost is high, i.e., $p \geq 15$, the truck is used to serve all the customers. On the contrary, the drone is used when the penalty cost is $p < 15$. The same result is also observed when the percentage of breakdown customers is $80\%$, $60\%$, and $40\%$ when $p \geq 20, 25$, and $35$, respectively. Apparently, the number of breakdown customers has a significant impact on the solution.
Impact of repair cost
---------------------
Next, we consider two specific scenarios that (i) there is no breakdown and (ii) there is breakdown to all customers. We vary the drone repair cost and the probability of these two scenarios. Figure \[f\_repair\] shows the total payments, where (a) the penalty is ignored $p=0$ and (b) the penalty is $p=20$. When the probability that the breakdown occurs to all customers, i.e., $\mathbb{P}(\lambda_2) = 0$, the repair cost is not taken into account. As shown in Figure \[f\_repair\](a), the drone is used to serve all the customers. On the other hand, Figure \[f\_repair\](b) shows that the drone is selected when the probability that the breakdown occurs to all customers is $\mathbb{P}(\lambda_2) \leq 0.25$. Otherwise, the truck is used. We observe that the impact of the drone repair cost is significantly less than the penalty $p$ and the probability of that the breakdown occurs to all customers.
Simulation
----------
We conduct a simulation and compare the performance in terms of the total payments obtained from the solutions of the GADOP optimization, expected value formulation (EVF), the parallel drone scheduling traveling salesman problem (PDSTSP) [@sidekick]. We adapt the PDSTSP from minimizing the delivery time to minimizing total payments. The EVF and adapted PSDTSP formulations can be found in the Appendix.
Here, we set the initial cost of a truck to $S\$400$ and the daily traveling time limit is 10 hours. For the EVF, all the decisions are optimized only for the first stage. For the PDSTSP, the uncertainties are not considered. Clearly, the total payment obtained from solving the GADOP optimization is significantly lower than that of the EVF and the PSDTSP. The reason is that the solution of the GADOP optimization takes into account the takeoff and breakdown conditions explicitly which lower the chance of incurring penalty, i.e., by using trucks for more customers. On the contrary, the solution of the EVF considers only the probability of unsuccessful delivery and implicitly ignores the effect of the penalty, even though the penalty is considered. Note that when the penalty is large, e.g., $S\$14$ and $S\$16$ for the cases with 30 and 25 customers, respectively, the trucks are used for all customers which is the same for both GADOP optimization and EVF. However, the PSDTSP decisions are not influenced by the penalty. Therefore, the total payments of GADOP and the EVF are identical, and the total payment of PSDTSP is the highest.
Decomposition {#decomposition}
-------------
$\begin{array}{c}
\hspace{0em}
\includegraphics[width=0.5\textwidth]{time+cost.eps}
\end{array}$ \[f\_decom\]
We study the decomposition of the GADOP optimization. The penalty, repair cost, and the drone initial cost are $p=5$, $m=5$, and $\widehat{c}^{(i)}_d =10$ respectively. In this experiment, we consider one truck and 10 breakdown scenarios ($\Lambda = \{\lambda_1,\lambda_2,\dots, \lambda_{10}\}$). The number of drones and the number of takeoff scenarios are varied. The takeoff scenarios include all possibilities of takeoff events that can occur to all the drones. To illustrate, five drones are experimented with $32$ takeoff scenarios ($\Omega = \{\omega_1,\omega_2,\dots, \omega_{32}\}$), six drones are experimented with $64$ takeoff scenarios, and so on.
The comparison is shown in Figure \[f\_decom\]. When the number of takeoff scenarios grows, the difference between the computational time of the original and decomposed problems becomes larger. As shown in Figure \[f\_decom\], the decomposed GADOP achieves $407\%$, $648\%$, $854\%$, and $1372\%$ speedup with 32, 64, 128, and 256 scenarios, respectively. Meanwhile, the decomposed GADOP always achieves lower total payment. Note that, in some cases, the CPLEX solver cannot give an optimal solution for the original problem when the problem is too large. Thus, the CPLEX solver may return a suboptimal solution instead of the optimal solution. In this regard, the suboptimal solution achieves much higher payment than that of the decomposition while the computational time of the decomposition may not significantly improve [@ref_VTC2]. In summary, the decomposition enables the GADOP to handle larger number of scenarios.
Conclusion and Future works {#sec_con}
===========================
We have proposed the joint ground and aerial delivery service optimization and planning (GADOP) framework. The uncertainty of drone delivery has been taken into account by considering takeoff and breakdown conditions. The aim of the GADOP framework is to help a supplier to make the decision either to use ground-based trucks, drones, or the combination to serve customers. We have formulated the GADOP optimization as a three-stage stochastic integer programming model. The performance evaluation of both Solomon benchmark suite and the Singapore real road network has been conducted. The results clearly show that using drones in the package delivery can reduce the total payment of the supplier significantly. Moreover, simulation results have shown that the GADOP framework always achieves the lower total payment than that of the expected value formulation (EVF) and the Parallel Drone Scheduling Traveling Salesman Problem (PDSTSP). We have adopted the L-shape decomposition method to enhance the solvability of the optimization problem.
For future work, the uncertainty in customers’ demand and traveling time will be considered and, we will also incorporate multiple-stage scenarios of the breakdown condition, which depend on each customer. Moreover, different types of subtour elimination, which aim to speedup computational time, will be studied.
Appendix
========
Expected Value Formulation (EVF)
--------------------------------
The EVF objective function is similar to the original objective function in (\[eq\_obj1\]), but only the formulation of the $\mathbb{E}(\mathscr{L}(\widehat{X}_{i,d}))$ is different. The equation is indicated in (\[eq\_eva\]), where $\mathbb{P}(d)$, $\mathbb{P}(i)$, and $\mathbb{M}(d)$ denote the probability that drone $d$ cannot take off, the probability that breakdown occurs while serving customer $i$, and the probability that drone $d$ needs to be repaired, respectively. The EVF subjects to the constraints in (\[con\_ini\_ftl\])-(\[con\_subtour\]) and (\[eq\_con\_order3\])-(\[eq\_con\_lastbound\]).
$$\begin{aligned}
\mathbb{E}(\mathscr{L}(\widehat{X}_{i,d})) = &
\sum_{i \in \mathcal{C}}\sum_{d \in \mathcal{D}}\widehat{c}^{(r)}_{i}\widehat{X}_{i,d}\mathbb{P}(d)\nonumber +\sum_{d \in \mathcal{D}} p\widehat{X}_{i,d}(1-\mathbb{P}(d)) \nonumber\\
&+ \sum_{i \in \mathcal{C} }\sum_{d \in \mathcal{D} }p\widehat{X}_{i,d}(1-\mathbb{P}(d))\mathbb{P}(i)\nonumber\\
&+ \sum_{d \in \mathcal{D}} m(1-\mathbb{P}(d))\mathbb{M}(i)
\label{eq_eva}\end{aligned}$$
Parallel Drone Scheduling Traveling Salesman Problem (PDSTSP)
-------------------------------------------------------------
The PDSTSP can handle multiple drones with only one truck. It does not take the serving time constraint and the constraints of traveling limits ,i.e., time and distance, into account. Furthermore, the the uncertainties i.e., takeoff and breakdown conditions, are ignored in the PDSTSP. We adapt the PDSTSP objective function, which is proposed in [@sidekick], from minimizing the delivery time to minimizing the total payment. The new objective function is presented in (\[eq\_psdtsp\]). The two delivery time constraints in [@sidekick] are replaced by the constraints in (\[con\_ini\_psd\_ftl\]) and (\[con\_ini\_psd\_uav\]). The constraint in (\[con\_ini\_psd\_ftl\]) ensures that the initial cost of the truck is paid when the truck is used. Similarly, the constraint in (\[con\_ini\_psd\_uav\]) ensures that the initial cost of drones are paid when the drones are used. The rest of the constraints in [@sidekick] are kept as the original.
$$\begin{aligned}
& \mbox{Minimize:} \nonumber \\
& \bar{c}^{(i)} \bar{W} +
\sum_{d \in \mathcal{D}}\widehat{c}^{(i)}_d \widehat{W}_d + \sum_{\substack{i', j'\in \\ \mathcal{C}\cup\{0\}}}\bar{c}^{(r)}_{i',j'}V_{i',j'} +
\sum_{i \in \mathcal{C}}\sum_{d \in \mathcal{D}}\widehat{c}^{(r)}_{i}\widehat{X}_{i,d}
\label{eq_psdtsp}\end{aligned}$$
$$\begin{aligned}
&\sum_{i \in \mathcal{C}} V_{i} \leq \Delta \bar{W}, & \forall i \in \mathcal{C} \label{con_ini_psd_ftl}\\
&\sum_{i \in \mathcal{C}} \widehat{X}_{i,d} \leq \Delta \widehat{W}_d, & \forall d \in \mathcal{D}\label{con_ini_psd_uav}\end{aligned}$$
Acknowledgment
==============
This work was partially supported by Singapore Institute of Manufacturing Technology-Nanyang Technological University (SIMTech-NTU) Joint Laboratory and Collaborative research Programme on Complex Systems.
[1]{} C. C. Murray and A. G. Chu, “The flying sidekick traveling salesman problem: Optimization of drone-assisted parcel delivery," [*Transportation Research Part C: Emerging Technologies*]{}, vol. 54, pp. 86–109, May 2015.
Y. Marinakis, and A. Migdalas. “Annotated Bibliography in Vehicle Routing," [*International Journal in Operational Research*]{}, vol. 7, pp. 27-46, 2007.
C. Lin, K. L. Choy, G. T. S. Ho, S. H. Chung and H. Y. Lam, “Survey of Green Vehicle Routing Problem: Past and future trends," [*Expert Systems with Applications*]{}, vol. 41, pp. 1118-1138, 2014.
M. M. Solomon, “ Time Window Constrained Routing and Scheduling Problem," [*Transportation Science*]{}, vol. 22, pp. 1-13, 1988.
G. Berbeglia, J.-F. Cordeau, I. Gribkovskaia and G. Laporte, “Static Pickup and Delivery Problems: A Classi¯cation Scheme and Survey," [*TOP*]{}, vol. 15, pp. 1-31, Jul 2007.
V. Pillac , M. Gendreau, C. Guéret and A. L. Medaglia. “A review of dynamic vehicle routing problems," [*European Journal of Operational Research*]{}, vol. 225, pp. 1-11, Feb 2013.
M. Gendreau, G. Laporte and R. Seguin, “ Stochastic vehicle routing," [*European Journal of Operational Research*]{}, vol. 88, pp. 3-12, 1996.
G. Laporte, “The Vehicle Routing Problem: An overview of exact and approximate algorithms," [*European Journal of Operational Research*]{}, vol. 59, pp. 345-358, 1992.
G. Laporte, M. Gendreau, J. Y. potvin, and F. Semet, “Classical and modern heuristics for the vehicle routing problem," [*International transactions in Operational Research*]{}, vol. 7, pp. 285-300, 2000.
C . Sabo, D. Kingston and K. Cohen, “A Formulation and Heuristic Approach to Task Allocation and Routing of UAVs under Limited Communication," [*Unmanned Systems*]{}, vol. 2, pp. 1-17, Dec 2013.
K. Sundar and S. Rathinam, “Algorithms for Routing an Unmanned Aerial Vehicle in the Presence of Refueling Depots," in [*IEEE Transactions on Automation Science and Engineering*]{}, vol. 11, no. 1, pp. 287-294, Jan 2014.
L. Evers, T. Dollevoet and A. I. Barros, “Robust UAV mission planning," [*Annals of Operations Research*]{}, vol. 222, pp. 293-315, Nov 2014.
L. Evers, A. I. Barros, H. Monsuur and A. Wagelmans, “Online stochastic UAV mission planning with time windows and time-sensitive targets," [ *European Journal of Operational Research*]{}, vol. 238, pp. 348-362 , Oct 2014.
K. Dorling, J. Heinrichs, G. G. Messier and S. Magierowski, “Vehicle Routing Problems for Drone Delivery," in [*IEEE Transactions on Systems*]{}, Man, and Cybernetics: Systems, vol. 47, no. 1, pp. 70-85, Jan 2017.
J. R. Birge and F. Louveaux, “Introduction to Stochastic Programming". New York: Springer.
L. A. Haidari, S. T. Brown, M. Ferguson, E. Bancroft, M. Spiker, A. Wilcox, R. Ambikapathi, V. Sampath, D. L. Connor, B. Y. Lee, “The economic and operational value of using drones to transport vaccines," In [*Vaccine*]{}, vol. 34, pp. 4062-4067, Jul 2016.
S. Sawadsitang, R. Kaewpuang, S. Jiang, D. Niyato, P. Wang, “Optimal Stochastic Delivery Planning in Full-Truckload and Less-Than-Truckload Delivery," [*2017 IEEE 85th Vehicular Technology Conference (VTC Spring)*]{}, Sydney, Australia, Jun 2017.
F. Schenkelberg, “How reliable does a delivery drone have to be?," [*2016 Annual Reliability and Maintainability Symposium (RAMS)*]{}, Tucson, Arizona, USA, 2016, pp. 1-5.
D. Chattopadhyay, “Application of General Algebraic Modeling System to Power System Optimization," [*IEEE Transactions on Power Systems*]{}, vol. 14, no. 1, pp. 15-22, Feb 1999.
Rentalorrysingapore.com \[Online\]. Available: [https://rentalorrysingapore .com/lorry-dimensions/](https://rentalorrysingapore .com/lorry-dimensions/)
Land Transportation Authority of Singapore, “Singapore Land Transportation Statistic in brief," 2009. \[Online\]. Available: <https://www.lta.gov.sg/content/dam/ltaweb/corp/PublicationsResearch/files/FactsandFigures/Statistics_in_Brief_2009.pdf>
Singpost. \[Online\]. Available: <http://www.singpost.com/>
M. M. Solomon,“Algorithms for the vehicle routing and scheduling problems with time window constraints," [*Operations Research*]{}, vol. 35, no. 2, pp. 254-265, 1987.
Google. \[Online\]. Available: <https://developers.google.com/maps/documentation/distance-matrix/>
S. Sawadsitang, S. Jiang, D. Niyato, P. Wang, “Optimal Stochastic Package Delivery Planning with Deadline: A Cardinality Minimization in Routing," [*2017 IEEE 86th Vehicular Technology Conference (VTC fall)*]{}, Toronto, Canada, Sept 2017.
\[[{width="1in" height="1.25in"}]{}\][Suttinee Sawadsitang]{} received her B.Eng in Computer Engineering from King Mongkuts University of Technology Thonburi (KMUTT), Thailand in 2012 and M.Eng from Shanghai Jiao Tong University, China in 2015. She is currently pursuing a Ph.D. degree at SIMTech-NTU Joint Lab on Complex Systems, Nanyang Technological University, Singapore. Her research interests are in the area of transportation systems, operations research, optimization, and high performance computing.
\[[{width="1in" height="1.25in"}]{}\][Dusit Niyato]{} (M’09-SM’15-F’17) is currently a professor in the School of Computer Science and Engineering, at Nanyang Technological University, Singapore. He received B.Eng. from King Mongkuts Institute of Technology Ladkrabang (KMITL), Thailand in 1999 and Ph.D. in Electrical and Computer Engineering from the University of Manitoba, Canada in 2008. His research interests are in the area of energy harvesting for wireless communication, Internet of Things (IoT) and sensor networks.
\[[{width="1in" height="1.25in"}]{}\] [Puay Siew Tan]{} received the Ph.D. degree in computer science from the School of Computer Engineering, Nanyang Technological University, Singapore. She is presently an Adjunct Associate Professor with the School of Computer Science and Engineering, Nanyang Technological University and also the Co-Director of the SIMTECH-NTU Joint Laboratory on Complex Systems. In her full-time job at Singapore Institute of Manufacturing Technology (SIMTech), she leads the Manufacturing Control Tower^TM^ (MCT^TM^) as the Programme Manager. She is also the Deputy Division Director of the Manufacturing System Division. Her research interests are in the cross-field disciplines of Computer Science and Operations Research for virtual enterprise collaboration, in particular sustainable complex manufacturing and supply chain operations in the era of Industry 4.0.
\[[{width="1in" height="1.25in"}]{}\] [Ping Wang]{} (M’08, SM’15) received the PhD degree in electrical engineering from University of Waterloo, Canada, in 2008. She was with Nanyang Technological University, Singapore. Currently she is an associate professor at the department of Electrical Engineering and Computer Science, York University, Canada. Her current research interests include resource allocation in multimedia wireless networks, cloud computing, and smart grid. She was a corecipient of the Best Paper Award from IEEE Wireless Communications and Networking Conference (WCNC) 2012 and IEEE International Conference on Communications (ICC) 2007. She served as an Editor of IEEE Transactions on Wireless Communications, EURASIP Journal on Wireless Communications and Networking, and International Journal of Ultra Wideband Communications and Systems.
[^1]: S. Sawadsitang and D. Niyato are with the School of Computer Science and Engineering, Nanyang Technological University, Singapore, 639798, e-mail: (e-mail: [email protected]; [email protected]).
[^2]: P.-S. Tan is with the Singapore Institute of Manufacturing Technology, Singapore 638075 (e-mail: [email protected]).
[^3]: P. Wang is with the Department of Electrical Engineering and Computer Science, York University, Canada (e-mail: [email protected]).
[^4]: Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org.
[^5]: In this paper, ground-based delivery refers to the delivery by trucks.
[^6]: In this paper, total cost and total payment refer to the same.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The velocity dependence of energy and momentum is studied. It is shown that in the case of STR in the space-time of only one spatial dimension the standard energy and momentum definition can be naturally modified without lost of local Lorenz invariance, conservation rules and additivity for multiparticle system. One parameter family of energies and momenta is constructed and it is shown that within natural conditions there is no further freedom. Choosing proper family parameter one can obtain energy and momentum increasing with velocity faster or slower in comparison with the standard case, but almost coinciding with them in the wide velocity region.'
author:
- |
Marek Pawlowski\
Soltan Institute for Nuclear Studies\
Hoza 69, 00-681 Warsaw, Poland\
e-mail: [email protected]
title: 'One-parameter family of additive energies and momenta in 1+1 dimensional STR'
---
It is generally believed that the structure of space-time at Planck scale may differ from our everyday experience. Quantum phenomena, including quantum gravity, are expected to modify local space-time structure. Such modification can break local Lorenz invariance, can modify it or leave it untouched. A class of models exploring the last possibility is known as Double Special Relativity (DSR) [@dsr]. Their authors, being motivated by theoretical arguments and some experimental question marks [@cosmic], introduce additional scale (connected with length or energy) and assume that the local Lorenz invariance is preserved. They pay the price of loose of standard definition of momentum and energy. In fact the price is even higher: momentum is no longer an additive quantity (momentum of a system is not a sum of momenta of its components) in DSR models; the same concerns energy [@kosinski]. The reason is clear. Simple textbook arguments [@Feynman] lead to the conclusion that the standard STR definition of momentum: \[p-einstein\] [**p**]{}(m, [**v**]{}) = is the unique Lorenz covariant function of $\bf v$ being additive and satisfying reasonable set of conditions. Similarly the energy function is also unique: \[e-einstein\] E(m, [**v**]{}) = .
We show below that the standard arguments are not sufficient to select (\[p-einstein\],\[e-einstein\]) in 1+1-dimensional Minkowski space. The admissible form of momentum and energy is more general and depends on a new arbitrary dimensionless parameter.
It is well known that the one dimensional relativistic velocity group $(V^1,\oplus)$ is commutative and associative (in contrast to the case of dimension 3). In fact the group is isomorphic with ordinary $(R,+)$ group. The group isomorphism can be written in example in the following form:
\[isomorphism\] V\^1 v y(v)= R
The one dimensional velocity addition is given by \[v-composition\] v u = , (we have assumed $c=1$ for simplicity)
Its counterpart takes the simple “linear” form \[y-composition\] y(v u) = y(v) + y(u).
Observe the following properties: \[zero\] y(0) = 0, \[minus\] y(-v) = -y(v).
These observations will be very useful for our subsequent considerations.
Our aim is to define additive relativistic momentum.
Assume that there exists the momentum which is a real function $p(m,v)$ that can be ascribed to each separate body (of low energy defined mass $m$) and that fulfils the following conditions:
1. $p(m,-v)=-p(m,v)$ (antisymmetry)
2. For a two body elastically scattering system we define the center of mass reference frame $U_{CM}$ (it is the frame in which the asymptotic relations hold: $v^{in}_I =v^{CM}_I =
-v^{out}_I$, $I=1,2$). We’ll demand that the sum of momenta vanishes in $U_{CM}$: \[CM\] p(m\_1,v\^[CM]{}\_1)+p(m\_2,v\^[CM]{}\_2)=0.
3. There holds the relativistic invariant momentum conservation rule: $$\begin{aligned}
\nonumber p(m_1,v^{CM}_1 \oplus u) +
p(m_2,v^{CM}_2 \oplus u)=\\ \label{momentumconservation}
p(m_1,-v^{CM}_1 \oplus u)+p(m_2,-v^{CM}_2 \oplus u)
\end{aligned}$$ for arbitrary $u$ and arbitrary set of $m_I, v_I^{CM}$ satisfying (\[CM\]).
4. $p(m,v)\rightarrow mv$ for $v\ll 1$ (correspondence principle).
We will show that the most general function satisfying conditions 1.-4. is \[momentum\] p(m,v)=( ) where $a$ is an arbitrary real or pure imaginary parameter.
Observe that by means of the isomorphism (\[isomorphism\]) the momentum (\[momentum\]) can be concisely rewritten: \[y-momentum\] p(m,y)=().
The proof of (\[momentum\]) goes as follows:
Using antisymmetry of $p$ and $y$ one can rewrite (\[momentumconservation\]) $$\begin{aligned}
\nonumber p(m_1,y_1 + w) - p(m_1,y_1 - w) +
\\ \label{proof1} p(m_2,y_2 + w) + p(m_2,y_2 - w) =0.
\end{aligned}$$ where $w=y(u)$. Differentiating (\[proof1\]) $2n$ times with respect to the second argument and putting $w=0$ we get the set of relations \[proof2\] p\^[(2n)]{}(m\_1,y\_1)+p\^[(2n)]{}(m\_2,y\_2)=0. Let us treat the first of them (for $n=0$) as an involved relation between $y_1$ and $y_2$. Now, differentiating it once and twice with respect to $y_1$ and combining the results with (\[proof2\]) for $n=1$ (in order to eliminate the second body momentum) we get \[proof3\] ()=. The similar equation but with the third derivative under the logarithm, can be obtained on the similar way starting from $n=1$. As the RHS of the both formulas are identical, we get the differential equation \[proof4\] ()= () where the subscripts have been omitted.
Applying again the conditions 1. and 4. in order to fix some integration constants we get the general solution of (\[proof4\]) in the form given by (\[y-momentum\]).
The obtained momentum (\[momentum\]) comes to the standard expression (\[p-einstein\]) for $a=1$. The plots of velocity dependencies of (\[momentum\]) for several values of $a$ are compared on the Figure 1. If the value of $a$ is close to unity, the dependencies are almost undistinguishable in the wide range of $v$.
The solution of (\[proof4\]) admits also pure imaginary parameters $a$. Some relevant plots are given on the Figure 2. The momenta with imaginary $a$ are undistinguishable from the others for small velocities. However their asymptotic behavior is bizarre.
Despite the fact, that the momentum formula (\[momentum\]) was derived from considerations based on the elastic scattering, its applicability is universal. Consider for example a two body decay. Let $M$ be the decaying mass and $m_1, m_2$ be the masses of the decay products. Then assuming momentum conservation principle and its relativistic invariance we come to the conclusion that \[decay\] M=m\_1 (ay\_1)+m\_2 (ay\_2). We’ll make use of this relation deriving generalized energy formula.
The generalized energy can be derived from the set of assumptions similar to 1.-4. Energy is expected to be a symmetric function of velocity (i), to be additive and conserved in every inertial reference frame (ii) and to have a proper low energy limit (iii). In addition it is demanded that energy is conserved in two-body decays (iv).
A similar consideration as in the case of momentum leads to the generic formula carrying out all the conditions: \[energy\] E(m,v)=m((1-)+()) or in an equivalent notation \[y-energy\] E(m,y)=m((1-)+()). It follows from the proof of (\[energy\]) that the constant $a$ is the same as previously. Again, for $a=1$ the energy (\[energy\]) takes its standard STR form (\[e-einstein\]).
The momentum and energy of a body given by (\[momentum\]) and (\[energy\]) are connected by energy/momentum dispersion relation \[dispersion\] a\^2(E-m(1-))\^2-p\^2= which is a simple consequence of the hypergeometric unity relation.
The new dimensionless constant $a$ that appears naturally in energy and momentum definitions in 1+1 dimensional space-time, allows us to modify slightly the relation between energy/momentum and velocity. For $a<1$ both the energy and momentum are reduced in high energy limit in comparison with the standard definition. For $a>1$ they increase. This is just the phenomenon that makes attractive DSR models. In example it helps to explain eventual ultra high energy events in cosmic radiation.
The presented analysis was done in a space-time of reduced dimension. The analysis was strictly dependent on this dimension and it is obvious that there is no straight way to generalize it. If we want to obtain similar result in the physical space-time we have to attenuate some natural conditions that usually are placed on momentum and energy. Presented analysis shows a new type of dispersion relation - in a sense alternative to the relations studied in the literature in the context of DSR.
[99]{} G. Amelino-Camelia, Phys. Lett. B [**510**]{}, 255 (2001), Int. J. Mod. Phys. D [**11**]{}, 35 (2002) K. Greisen, Phys. Rev. Lett. [**16**]{}, 748 (1966); G.T. Zatsepin and V.A. Kuzmin, JETP Lett. [**4**]{}, 78 (1966). P. Kosinski, P. Maslanka, Phys. Rev. [**D68**]{} 067702, 2003 see e.g. R.P. Feynman,[ *Feynman lectures on Physics*]{}, Pearson P T R (1970).
Figure Captions
Fig. 1 $p(1,v)$ plots for $a=0.00001, 0.5, 0.7, 1, 1.2, 1.5$
Fig. 2 $p(1,v)$ plots for imaginary parameter $a=0.5 i, 0.8 i, i,
1.2 i, 1.5 i$ compared with $a=1$ and $a=0.00001$. (Plot range is not symmetric.)
(40,55)
(0,-174)
[**Fig. 1**]{}
(40,55)
(0,-174)
[**Fig. 2**]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider the problem of finding an $\epsilon$-approximate stationary point of a smooth function on a compact domain of $\R^d$. In contrast with dimension-free approaches such as gradient descent, we focus here on the case where $d$ is finite, and potentially small. This viewpoint was explored in 1993 by Vavasis, who proposed an algorithm which, for [*any fixed finite dimension $d$*]{}, improves upon the $O(1/\epsilon^2)$ oracle complexity of gradient descent. For example for $d=2$, Vavasis’ approach obtains the complexity $O(1/\epsilon)$. Moreover for $d=2$ he also proved a lower bound of $\Omega(1/\sqrt{\epsilon})$ for deterministic algorithms (we extend this result to randomized algorithms).
Our main contribution is an algorithm, which we call [*gradient flow trapping*]{} (GFT), and the analysis of its oracle complexity. In dimension $d=2$, GFT closes the gap with Vavasis’ lower bound (up to a logarithmic factor), as we show that it has complexity $O\left(\sqrt{\frac{\log(1/\epsilon)}{\epsilon}}\right)$. In dimension $d=3$, we show a complexity of $O\left(\frac{\log(1/\epsilon)}{\epsilon}\right)$, improving upon Vavasis’ $O\left(1 / \epsilon^{1.2} \right)$. In higher dimensions, GFT has the remarkable property of being a [*logarithmic parallel depth*]{} strategy, in stark contrast with the polynomial depth of gradient descent or Vavasis’ algorithm. In this higher dimensional regime, the total work of GFT improves quadratically upon the only other known polylogarithmic depth strategy for this problem, namely naive grid search.
author:
- |
Sébastien Bubeck\
Microsoft Research
- |
Dan Mikulincer [^1]\
Weizmann Institute
bibliography:
- 'bib.bib'
title: How to trap a gradient flow
---
Introduction
============
Let $f : \R^d \rightarrow \R$ be a smooth function (i.e., the map $x \mapsto \nabla f(x)$ is $1$-Lipschitz, and $f$ is possibly non-convex). We aim to find an $\epsilon$-approximate stationary point, i.e., a point $x \in \R^d$ such that $\|\nabla f(x)\|_2 \leq \epsilon$. It is an elementary exercise to verify that for smooth and bounded functions, gradient descent finds such a point in $O(1/\epsilon^2)$ steps, see e.g., [@Nes04]. Moreover, it was recently shown in [@CDHS19] that this result is [*optimal*]{}, in the sense that any procedure with only black-box access to $f$ (e.g., to its value and gradient) must, [*in the worst case*]{}, make $\Omega(1/\epsilon^2)$ queries before finding an $\epsilon$-approximate stationary point. This situation is akin to the non-smooth convex case, where the same result (optimality of gradient descent at complexity $1/\epsilon^2$) holds true for finding an $\epsilon$-approximate optimal point (i.e., such that $f(x) - \min_{y \in \R^d} f(y) \leq \epsilon$), [@NY83; @Nes04].
There is an important footnote to both of these results (convex and non-convex), namely that optimality only holds in [*arbitrarily high dimension*]{} (specifically the hard instance in both cases require $d = \Omega(1/\epsilon^2)$). It is well-known that in the convex case this large dimension requirement is actually necessary, for the cutting plane type strategies (e.g., center of gravity) can find $\epsilon$-approximate optimal points on compact domains in $O(d \log(1/\epsilon))$ queries. It is natural to ask: **Is there some analogue to cutting planes for non-convex optimization?**[^2] In dimension $1$ it is easy to see that one can indeed do a binary search to find an approximate stationary point of a smooth non-convex function on an interval. The first non-trivial case is thus dimension $2$, which is the focus of this paper (although we also obtain new results in high dimensions, and in particular our approach does achieve $O(\mathrm{poly}(d) \log(1/\epsilon))$ [*parallel depth*]{}, see below for details).
This problem, of finding an approximate stationary point of a smooth function on a compact domain of $\R^2$, was studied in 1993 by Stephen A. Vavasis in [@Vav93]. From an algorithmic perspective, his main observation is that in finite dimensional spaces one can speed up gradient descent by using a [*warm start*]{}. Specifically, observe that gradient descent only needs $O(\Delta/\epsilon^2)$ queries when starting from a $\Delta$-approximate optimal point. Leveraging smoothness (see e.g., Lemma \[lem:discretizationerror\] below), observe that the best point on a $\sqrt{\Delta}$-net of the domain will be $\Delta$-approximate optimal. Thus starting gradient descent from the best point on $\sqrt{\Delta}$-net one obtains the complexity $O_d\left( \frac{\Delta}{\epsilon^2} + \frac{1}{\Delta^{d/2}} \right)$ in $\R^d$. Optimizing over $\Delta$, one obtains a $O_d\left(\left(\frac{1}{\epsilon}\right)^{\frac{2d}{d+2}}\right)$ complexity. In particular for $d=2$ this yields a $O(1/\epsilon)$ query strategy. In addition to this algorithmic advance, Vavasis also proved a lower bound of $\Omega(1/\sqrt{\epsilon})$ for deterministic algorithms. In this paper we close the gap up to a logarithmic term. Our main contribution is a new strategy loosely inspired by cutting planes, which we call [*gradient flow trapping*]{} (GFT), with complexity $O\left(\sqrt{\frac{\log(1/\epsilon)}{\epsilon}} \right)$. We also extend Vavasis lower bound to randomized algorithms, by connecting the problem with unpredictable walks in probability theory [@BPP98].
Although we focus on $d=2$ for the description and analysis of GFT in this paper, one can in fact easily generalize to higher dimensions. Before stating our results there, we first make precise the notion of approximate stationary points, and we also introduce the [*parallel query*]{} model.
Approximate stationary point
----------------------------
We focus on the constraint set $[0,1]^d$, although this is not necessary and we make this choice mainly for ease of exposition. Let us fix a differentiable function $f : [0,1]^d \rightarrow \R$ such that $\forall x,y \in [0,1]^d$, $\|\nabla f(x) - \nabla f(y)\|_2 \leq \|x-y\|_2$. Our goal is to find a point $x \in [0,1]^d$ such that for any $y \in [0,1]^d$, $$f(x) \leq f(y) + \epsilon \cdot \|x-y\|_2 \,.$$ We say that such an $x$ is an $\epsilon$-stationary point (its existence is guaranteed by the extreme value theorem). In particular if $x \in (0,1)^d$ this means that $\|\nabla f(x)\|_2 \leq \epsilon$. More generally, for $x = (x^1, \hdots, x^d) \in [0,1]^d$ (possibly on the boundary), let us define the [*projected gradient*]{} at $x$, $g(x) = (g_1(x), \hdots, g_d(x))$ by: $$g_i(x) = \begin{cases}
\max \left(0, \frac{df}{dx^i}(x) \right)& \text{ if } x^i = 0 \,,\\
\frac{df}{dx^i}(x) & \text{ if } x^i \in (0,1) \,,\\
\min\left(0, \frac{df}{dx^i}(x)\right )& \text{ if } x^i = 1 \,.
\end{cases}$$ It is standard to show (see also [@Vav93]) that $x$ is an $\epsilon$-stationary point of $f$ if and only if $\|g(x)\|_2 \leq \epsilon$.
Parallel query model
--------------------
In the classical black-box model, the algorithm can sequentially query an oracle at points $x \in [0,1]^d$ and obtain the value[^3] of the function $f(x)$. An extension of this model, first considered in [@Nem94], is as follows: instead of submitting queries one by one sequentially, the algorithm can submit any number of queries in parallel. One can then count the [*depth*]{}, defined as the number of rounds of interaction with the oracle, and the [*total work*]{}, defined as the total number of queries.
It seems that the parallel complexity of finding stationary points has not been studied before. As far as we know, the only low-depth algorithm (say depth polylogarithmic in $1/\epsilon$) is the naive grid search: simply query all the points on an $\epsilon$-net of $[0,1]^d$ (it is guaranteed that one point in such a net is an $\epsilon$-stationary point). This strategy has depth $1$, and total work $O(1/\epsilon^d)$. As we explain next, the high-dimensional version of GFT has depth $O(\mathrm{poly}(d) \log(1/\epsilon))$, and its total work improves at least quadratically upon grid search.
Complexity bounds for GFT
-------------------------
In this paper we give a complete proof of the following near-optimal result in dimension $2$:
\[thm:main\] Let $d=2$. The gradient flow trapping algorithm (see Section \[sec:formal\]) finds a $4 \epsilon$-stationary point with less than $10^5 \sqrt{\frac{\log(1/\epsilon)}{\epsilon}}$ queries to the value of $f$.
It turns out that there is nothing inherently two-dimensional about GFT. At a very high level, one can think of GFT as making hyperplane cuts, just like standard cutting planes methods in convex optimization. While in the convex case those hyperplane cuts are simply obtained by gradients, here we obtain them by querying a $\tilde{O}(\sqrt{\epsilon})$-net on a carefully selected small set of hyperplanes. Note also that the meaning of a “cut" is much more delicate than for traditional cutting planes methods (here we use those cuts to “trap" gradient flows). All of these ideas are more easily expressed in dimension $2$, but generalizing them to higher dimensions presents no new difficulties (besides heavier notation). In Section \[sec:highdim\] we prove the following result:
\[thm:highdim\] The high-dimensional version of GFT finds an $\epsilon$-stationary point in depth $O(d^2 \log(d/\epsilon))$ and in total work $d^{O(d)} \cdot \left(\frac{\log(1/\epsilon)}{\epsilon}\right)^{\frac{d-1}{2}} $.
In particular we see that the three-dimensional version of GFT has complexity $O\left(\frac{\log(1/\epsilon)}{\epsilon} \right)$. This improves upon the previous state of the art complexity $O(1/\epsilon^{1.2})$ [@Vav93]. However, on the contrary to the two-dimensional case, we believe that here GFT is suboptimal. As we discuss in Section \[sec:heuristic\], in dimension $3$ we conjecture the lower bound $\Omega(1/\epsilon^{0.6})$.
In dimensions $d \geq 4$, the total work given by Theorem \[thm:highdim\] is worse than the total work $O\left(\left(\frac{1}{\epsilon}\right)^{\frac{2d}{d+2}}\right)$ of Vavasis’ algorithm. On the other hand, the depth of Vavasis’ algorithm is of the same order as its total work, in stark contrast with GFT which maintains a logarithmic depth even in higher dimensions. Among algorithms with polylogarithmic depth, the total work given in Theorem \[thm:highdim\] is more than a quadratic improvement (in fixed dimension) over the previous state of the art (namely naive grid search).
Paper organization
------------------
The rest of the paper (besides Section \[sec:LB\] and Section \[sec:open\]) is dedicated to motivating, describing and analyzing our gradient flow trapping strategy in dimension $2$ (from now on we fix $d=2$, unless specified otherwise). In Section \[sec:basic\] we make a basic “local to global" observation about gradient flow which forms the basis of our “trapping" strategy. Section \[sec:informal\] is an informal section on how one could potentially use this local to global phenomenon to design an algorithm, and we outline some of the difficulties one has to overcome. In Section \[sec:formal\] we formally describe our new strategy and analyze its complexity. In Section \[sec:LB\] we extend Vavasis’ $\Omega(1/\sqrt{\epsilon})$ lower bound to randomized algorithms. Finally we conclude the paper in Section \[sec:open\] by introducing several open problems related to higher dimensions.
A local to global phenomenon for gradient flow {#sec:basic}
==============================================
We begin with some definitions. For an axis-aligned rectangle $R =[a,b] \times [c,d]$ we denote its volume and diameter by $$\mathrm{diam}(R) : = \sqrt{(b-a)^2 + (d-c)^2} \text{ and } \mathrm{vol}(R) : = (b-a)(d-c).$$ We further define the aspect ratio of $R$ as $\frac{\max\left(b-a,d-c\right)}{\min\left(b-a,d-c\right)}$. The $4$ edges of $R$ are the subsets $$\{a\}\times [c,d], \ \{b\}\times [c,d],\ [a,b]\times \{c\} \text{ and } [a,b] \times \{d\},$$ and the boundary of $R$, which we denote $\partial R$ is the union of all edges.\
If $E \subset [0,1]^2$ is a segment and $\delta > 0$, we say that $N \subset E$ is a $\delta$-net of $E$, if for any $x \in E$, there exists some $y \in N$ such that $\|x-y\|_2 \leq \delta$. We will always assume implicitly that if $N \subset E$ is a $\delta$-net, then the endpoints of $E$ are elements of $N$.\
We denote $f^*_{\delta}(E)$ for the largest value one can obtain by minimizing $f$ on a $\delta$-net of $E$. Formally, $$f^*_\delta(E) = \sup\limits_N\inf\limits_{x\in N}f(x),$$ where the supremum is taken over all $\delta$-nets of $E$. We say that a pair $(E,x)$ of segment/point in $[0,1]^2$ (where $E$ is [*not*]{} a subset of an edge of $[0,1]^2$) satisfies the property $P_c$ for some $c\geq0$ if there exists $\delta>0$ such that $$f(x) < f^*_{\delta}(E) - \frac{\delta^2}{8} + c \cdot \mathrm{dist}(x,E) \,,$$ where $$\mathrm{dist}(x,E) := \inf\limits_{y \in E} \|x-y\|_2.$$ When $E$ is a subset of $\partial[0,1]^2$ we [*always*]{} say that $(E,x)$ satisfies $P_c$ (for any $c \geq 0$ and any $x \in [0,1]^2$).\
For an axis-aligned rectangle $R$ and $x \in R$, we say that $(R,x)$ satisfies $P_c$ if, for any of the four edges $E$ of $R$, one has that $(E,x)$ satisfies $P_c$. We refer to $x$ as the [*pivot*]{} for $R$.
Our main observation is as follows:
\[lem:main\] Let $R$ be a rectangle such that $(R,x)$ satisfies $P_c$ for some $x \in R$ and $c \geq 0$. Then $R$ must contain a $c$-stationary point (in fact the gradient flow emanating from $x$ must visit a $c$-stationary point before exiting $R$).
This lemma will be our basic tool to develop cutting plane-like strategies for non-convex optimization. From “local" information (values on a net of the boundary of $R$) one deduces a “global" property (existence of approximate stationary point in $R$).
Let us assume by contradiction that $R$ does not contain a $c$-stationary point, and consider the unit-speed gradient flow $(x(t))_{t \geq 0}$ constrained to stay in $[0,1]^2$. That is, $x(t)$ is the piecewise differentiable function defined by $x(0)=x$ and $\frac{d}{dt} x(t) = - \frac{g(x(t))}{\|g(x(t))\|_2}$, where $g$ is the projected gradient defined in the previous section. Since there is no stationary point in $R$, it must be that the gradient flow exits $R$. Let us denote $T = \inf \{ t \geq 0 : x(t) \not\in R\}$, and $E$ an edge of $R$ such that $x(T) \in E$. Remark that $E$ cannot be part of an edge of $[0,1]^2$. Furthermore, for any $0 \leq t \leq T$, one has $$f(x(t)) - f(x(0)) = \int_{0}^t g(x(s)) \cdot \frac{d}{ds} x(s) ds \leq - c \cdot t \leq - c \cdot \|x(t) - x(0)\|_2 \,.$$ In particular $f(x(T)) - f(x) \leq - c \cdot \mathrm{dist}(x,E)$, so that $$\min_{y \in E} f(y) \leq f(x) - c \cdot \mathrm{dist}(x, E) \,.$$ Lemma \[lem:discretizationerror\] below shows that for any $\delta >0$ one has $f^*_{\delta}(E) \leq \min_{y \in E} f(y) + \frac{\delta^2}{8}$, and thus together with the above display it shows that $(E,x)$ does [*not*]{} satisfy $P_c$, which is a contradiction.\
\[lem:discretizationerror\] For any segment $E \subset [0,1]^2$ and $\delta > 0$ one has: $$f^*_{\delta}(E) \leq \min_{y \in E} f(y) + \frac{\delta^2}{8} \,.$$
Let $x \in E$ be such that $f(x) = \min_{z \in E} f(z)$. If $x$ is an endpoint of $E$, then we are done since we require the endpoints of $E$ to be in the $\delta$-nets. Otherwise $x$ is in the relative interior of $E$, and thus one has $\nabla f(x) \cdot (y - x) = 0$ for any $y \in E$. In particular by smoothness one has: $$\begin{aligned}
f(y) &= f(x) + \int_{0}^1 \nabla f(x + t (y-x)) \cdot (y-x) dt\\
&\leq f(x) + \int_{0}^1 t\cdot \|y-x\|_2^2 dt = f(x) + \frac{1}{2} \|y-x\|_2^2 \,.\end{aligned}$$ Moreover for any $\delta$-net of $E$ there exists $y$ such that $\|y-x\|_2 \leq \frac{\delta}{2}$, and thus $f(y) \leq f(x) + \delta^2 / 8$, which concludes the proof.
Our algorithmic approach to finding stationary points will be to somehow shrink the domain of consideration over time. At first it can be slightly unclear how the newly created boundaries interact with the definition of stationary points. To dispell any mystery, it might be useful to keep in mind the following observation, which states that if $(R,x)$ satisfies $P_c$, then $x$ cannot be on a boundary of $R$ which was not part of the original boundary of $[0,1]^2$.
\[lem:minor\] Let $R$ be a rectangle such that $(R,x)$ satisfies $P_c$ for some $x \in R$ and $c \geq 0$. Then $x\notin \partial R \setminus \partial [0,1]^2$.
Let $E$ be an edge or $R$ which is not a subset of $\partial [0,1]^2$. Then by definition of $P_c$, and by invoking Lemma \[lem:discretizationerror\], one has: $$f(x) < f_\delta^*(E) -\frac{\delta^2}{8} + c \cdot \mathrm{dist}(x,E) \leq \min_{y \in E} f(y) + c \cdot \mathrm{dist}(x,E) \,.$$ In particular if $x \in E$ then $\mathrm{dist}(x,E) = 0$, and thus $f(x) < \min_{y \in E} f(y)$ which is a contradiction.
From Lemma \[lem:main\] to an algorithm {#sec:informal}
=======================================
Lemma \[lem:main\] naturally leads to the following algorithmic idea (for sake of simplicity in this discussion we replace squares by circles): given some current candidate point $x$ in some well-conditioned domain (e.g., such that the domain contains and is contained in balls centered at $x$ and of comparable sizes), query a $\sqrt{\epsilon}$-net on the circle $C = \{ y : \|y-x\|_{2} = 1\}$, and denote $y$ for the best point found on this net. If one finds a significant enough improvement, say $f(y) < f(x) - \frac{3}{4} \epsilon$, then this is great news, as it means that one obtained a **per query** improvement of $\Theta(\epsilon^{-3/2})$ (to be compared with gradient descent which only yields an improvement of $\Theta(\epsilon^{-2})$). On the other hand if no such improvement is found, then the gradient flow from $x$ must visit an $\epsilon$-stationary point inside $C$.[^4] In other words one can now hope to restrict the domain of consideration to a region inside $C$, which is a constant fraction smaller than the original domain.
Optimistically this strategy would give a $\tilde{O}(B/\epsilon^{3/2})$ rate for $B$-bounded smooth functions (since at any given scale one could make at most $O(B/\epsilon^{3/2})$ improvement steps). In particular together with the warm start this would tentatively yield a $\tilde{O}(1/\epsilon^{3/4})$ rate, thus already improving the state-of-the-art $O(1/\epsilon)$ by Vavasis.
There is however a difficulty in the induction part of the argument. Indeed, what we know after a shrinking step is that the current point $x$ satisfies $f(x) \leq f(y) + \epsilon$ for any $y \in C$. Now we would like to query a net on $\{y : \|y-x\|_{2} = 1/2\}$. Say that after such querying we find that we can’t shrink, namely we found some point $z$ with $f(z) < f(x) - \frac{\epsilon}{2} + \frac{\delta^2}{8}$, and in particular $f(z) < f(y) + \frac{1}{2} \epsilon + \frac{\delta^2}{8}$ for any $y \in C$. Could the gradient flow from $z$ escape the original circle $C$ without visiting an $\epsilon$-stationary point? Unfortunately the answer is yes. Indeed (because of the discretization error $\delta^2/8$) one cannot rule out that there would be a point $y \in C$ with $f(y) < f(z) - \frac{\epsilon}{2}$, and since $C$ is only at distance $1/2$ from $z$, such a point could be attained from $z$ with a gradient flow without $\epsilon$-stationary points. Of course one could say that instead of satisfying $P_{\epsilon}$ we now only satisfy $P_{\epsilon + \delta^2 / 4}$, and try to control the increase of the approximation guarantee, but such an approach would not improve upon the $1/\epsilon^2$ of gradient descent (simply because we could additively worsen the approximation guarantee too many times).
The core part of the above argument will remain in our full algorithm (querying a $\sqrt{\epsilon}$-net to shrink the domain). However it is made more difficult by the discretization error as we just saw. We also note that this discretization issue does not appear in discrete spaces, which is one reason why discrete spaces are much easier than continuous spaces for local optimization problems.
Technically we observe that the whole issue of discretization comes from the fact that when we update the center, we move closer to the boundary, which we “pay" in the term $\mathrm{dist}(x,E)$ in $P_c$, and we cannot “afford" it because of the discretization error term that we suffer when we update. Thus this issue would disappear if in our induction hypothesis we had $P_0$ for the boundary. Our strategy will work in two steps: first we give a querying strategy for a domain with $P_0$ that ensures that one can **always** shrink with $P_{\epsilon}$ guaranteed for the boundary, and secondly we give a method to essentially turn a $P_{\epsilon}$ boundary into $P_0$.
Gradient flow trapping {#sec:formal}
======================
We say that a pair $(R,x)$ is a [*domain*]{} if $R$ is an axis-aligned rectangle with aspect ratio bounded by $3$, and $x \in R$. The gradient flow trapping (GFT) algorithm is decomposed into two subroutines:
1. The first algorithm, which we call the [*parallel trap*]{}, takes as input a domain $(R,x)$ satisfying $P_0$. It returns a domain $(\tilde{R}, \tilde{x})$ satisfying $P_{\epsilon}$ and such that $\mathrm{vol}(\tilde{R}) \leq 0.95 \ \mathrm{vol}(R)$. The cost of this step is at most $2 \sqrt{\frac{\mathrm{diam}(R)}{\epsilon}}$ queries.
2. The second algorithm, which we call [*edge fixing*]{}, takes as input a domain $(R,x)$ satisfying $P_{\epsilon'}$ (for some $\epsilon' \in [\epsilon, 2 \epsilon]$) and such that for $k \in \{0,1,2,3\}$ edges $E$ of $R$ one also has $P_0$ for $(E,x)$. It returns a domain $(\tilde{R}, \tilde{x})$ such that either (i) it satisfies $P_{\epsilon'}$ and for $k+1$ edges it also satisfies $P_0$, or (ii) it satisfies $P_{\left(1+ \frac{1}{500 \log(1/\epsilon)} \right) \epsilon'}$ and furthermore $\mathrm{vol}(\tilde{R}) \leq 0.95 \ \mathrm{vol}(R)$. The cost of this step is at most $90 \sqrt{\frac{\mathrm{diam}(R) \log(1/\epsilon)}{\epsilon}}$ queries.
Equipped with these subroutines, GFT proceeds as follows. Initialize $(R_0, x_0) = ([0,1]^2, (0.5,0.5))$, $\epsilon_0 = \epsilon$, and $k_0 = 4$. For $t \geq 0$:
- if $k_t =4$, call [*parallel trap*]{} on $(R_t, x_t)$, and update $k_{t+1}=0$, $(R_{t+1},x_{t+1})=(\tilde{R_t}, \tilde{x_t})$, and $\epsilon_{t+1} = \epsilon$.
- Otherwise call [*edge fixing*]{}, and update $(R_{t+1},x_{t+1})=(\tilde{R_t}, \tilde{x_t})$. If $R_{t+1} = R_t$ then set $k_{t+1}=k_t+1$ and $\epsilon_{t+1} = \epsilon_t$, and otherwise set $k_{t+1} = 0$ and $\epsilon_{t+1} = \left(1+ \frac{1}{500 \log(1/\epsilon)} \right) \epsilon_t$.
We terminate once the diameter of $R_t$ is smaller than $2 \epsilon$.
Next we give the complexity analysis of GFT assuming the claimed properties of the subroutines [*parallel trap*]{} and [*edge fixing*]{} in 1. and 2. above. We then proceed to describe in details the subroutines, and prove that they satisfy the claimed properties.
Complexity analysis of GFT {#sec:complexity}
--------------------------
The following three lemmas give a proof of Theorem \[thm:main\].
\[lem:proof1\] GFT stops after at most $200 \log(1/\epsilon)$ steps.
First note that at least one out of five steps of GFT reduces the volume of the domain by $0.95$ (since one can do at most four steps in a row of edge fixing without volume decrease). Thus on average the volume decrease per step is at least $0.99$, i.e., $\mathrm{vol}(R_T) \leq 0.99^T$. In particular since $R_T$ has aspect ratio smaller than $3$, it is easy to verify $\mathrm{diam}(R_T) \leq 2\sqrt{\mathrm{vol}(R_T)} \leq 2 \times 0.99^{T/2}$. Thus for any $T \geq \log_{100/99}(1/\epsilon^{2})$, one must have $\mathrm{diam}(R_T) \leq 2 \epsilon$. Thus we see that GFT performs at most $\log_{100/99}(1/\epsilon^{2}) \leq 200 \log(1/\epsilon)$ steps.
When GFT stops, its pivot is a $4 \epsilon$-stationary point.
First note that $\epsilon_T \leq \left(1+\frac{1}{500 \log(1/\epsilon)}\right)^T \epsilon$, thus after $T \leq 200 \log(1/\epsilon)$ steps we know that $(R_T,x_T)$ satisfies at least $P_{2 \epsilon}$. In particular by Lemma \[lem:main\], $R_T$ must contain a $2 \epsilon$-stationary point, and since the diameter is less than $2 \epsilon$, it must be (by smoothness) that $x_T$ is a $4 \epsilon$-stationary point.
GFT makes at most $10^5 \sqrt{\frac{\log(1/\epsilon)}{\epsilon}}$ queries before it stops.
As we saw in the proof of Lemma \[lem:proof1\], one has $\mathrm{diam}(R_t) \leq 2 \times 0.99^{t/2}$. Furthermore the $t^{th}$ step requires at most $90 \sqrt{\frac{\mathrm{diam}(R_t) \log(1/\epsilon)}{\epsilon}}$ queries. Thus the total number of queries is bounded by: $$90 \sqrt{\frac{2 \log(1/\epsilon)}{\epsilon}} \sum_{t=0}^{\infty} 0.99^{t/4} \leq 10^5 \sqrt{\frac{\log(1/\epsilon)}{\epsilon}} \,.$$
A parallel trap {#sec:paralleltrap}
---------------
Let $(R,x)$ be a domain. We define two segments $E$ and $F$ in $R$ as follows. Assume that $R$ is a translation of $[0,s] \times [0,r]$. For sake of notation assume that in fact $R=[0,s] \times [0,r]$ with $s \in [r, 3r]$ and $x^1 \geq r/2$, where $x = (x^1,x^2)$ (in practice one always ensures this situation with a simple change of variables). Now we define $E = \{r/6\} \times [0,r]$ and $F=\{r/3\} \times [0,r]$ (See Figure \[fig:figure1\]).
![The *parallel trap*[]{data-label="fig:figure1"}](Figure1){width="10cm"}
The parallel trap algorithm queries a $\sqrt{r \epsilon}$-net on both $E$ and $F$ (which cost at most $2 \frac{r}{\sqrt{r \epsilon}} = 2 \sqrt{\frac{r}{\epsilon}}$). Denote $\bar{x}$ to be the best point (in terms of $f$ value) found on the union of those nets. That is, denoting $N \subset F \cup E$ for the queried $\sqrt{r\epsilon}$-net, then $$\bar{x} = \arg\min\limits_{y\in N} f(y).$$ One has the following possibilities (see Figure \[fig:figure2\] for an illustration):
- If $f(x) \leq f(\bar{x})$ then we set $\tilde{x}= x$ and $\tilde{R} = [r/3,s] \times [0,r]$.
- Otherwise we set $\tilde{x} = \bar{x}$. If $\bar{x} \in F$ we set $\tilde{R} = [r/6,s] \times [0,r]$, and if $\bar{x} \in E$ we set $\tilde{R} = [0,r/3] \times [0,r]$.
The above construction is justified by the following lemma (a trivial consequence of the definitions), and it proves in particular the properties of [*parallel trap*]{} described in 1. at the beginning of Section \[sec:formal\].
The rectangle $\tilde{R}$ has aspect ratio smaller than $3$, and it satisfies $\mathrm{vol}(\tilde{R}) \leq 0.95 \ \mathrm{vol}(R)$. Furthermore if $(R,x)$ satisfies $P_0$, then $(\tilde{R}, \tilde{x})$ satisfies $P_{\epsilon}$.
The first sentence is trivial to verify. For the second sentence, first note that for any edge $E$ of $R$ one has $P_0$ for $(E,\tilde{x})$ since by assumption one has $P_0$ for $(E,x)$ and furthermore $f(\tilde{x}) \leq f(x)$. Next observe that $\tilde{R}$ has at most one new edge $\tilde{E}$ with respect to $R$, and this edge is at distance at least $r/6$ from $\tilde{x}$, thus in particular one has $\epsilon \cdot \mathrm{dist}(\tilde{x}, \tilde{E}) - \delta^2 / 8 > 0$ for $\delta=\sqrt{r \epsilon}$. Furthermore by definition $f(\tilde{x}) \leq f^*_{\delta}(\tilde{E})$, and thus $f(\tilde{x}) < f^*_{\delta}(\tilde{E}) - \frac{\delta^2}{8} + \epsilon \cdot \mathrm{dist}(\tilde{x}, \tilde{E})$, or in other words $(\tilde{E},\tilde{x})$ satisfies $P_{\epsilon}$.
![The three possible cases for $(\tilde{R},\tilde{x})$. $\tilde{R}$ is marked in red.[]{data-label="fig:figure2"}](Figure2){width="\columnwidth"}
Edge fixing {#sec:edgefixing}
-----------
Let $(R,x)$ be a domain satisfying $P_{\epsilon'}$ for some $\epsilon' \in [\epsilon, 2 \epsilon]$, and with some edges possibly also satisfying $P_0$. Denote $\mathcal{E}$ for the closest edge to $x$ that does not satisfy $P_0$, and let $r=\mathrm{dist}(x,\mathcal{E})$. We will consider three[^5] candidate smaller rectangles, $R_1$, $R_2$ and $R_3$, as well as three candidate pivots (in addition to $x$) $x_1 \in \partial R_1$, $x_2 \in \partial R_2$ and $x_3 \in \partial R_3$. The rectangles are defined by $R_i = R \cap \{y : \|x_{i-1}-y\|_{\infty} \leq \frac{r}{3} \}$, where we set $x_0 = x$. The possible output $(\tilde{R}, \tilde{x})$ of [*edge fixing*]{} will be either $(R_i, x_{i-1})$ for some $i\in\{1,2,3\}$, or $(R, x_3)$ (see Figure \[fig:figure3\] for a demonstration of how to construct $R_1$).
To guarantee the properties described in 2. at the beginning of Section \[sec:formal\] we will prove the following: if the output is $(R_i, x_{i-1})$ for some $i$ then all edges will satisfy $P_{\left( 1 + \frac{1}{500 \log(1/\epsilon)} \right) \epsilon'}$ (Lemma \[lem:Ri\] below) and the domain has shrunk (Lemma \[lem:shrunkdomain\] below), and if the output is $(R,x_3)$ then one more edge satisfies $P_0$ compared to $(R,x)$ while all edges still satisfy at least $P_{\epsilon'}$ (Lemma \[lem:notRi\] below).
\[lem:shrunkdomain\] For any $i \in \{1,2,3\}$ one has $\mathrm{vol}(R_i) \leq \frac{2}{3} \mathrm{vol}(R)$. Furthermore if the aspect ratio of $R$ is smaller than $3$, then so is the aspect ratio of $R_i$.
Let us denote $\ell_1(R)$ for the length of $R$ in the axis of $\mathcal{E}$ (the edge whose distance to $x$ defines $r$), and $\ell_2(R)$ for the length in the orthogonal direction (and similarly define $\ell_1(R_i)$ and $\ell_2(R_i)$).
Since $R_i \subset R$ one has $\ell_1(R_i) \leq \ell_1(R)$. Furthermore $\ell_2(R_i) \leq \frac{2}{3} r$ and $\ell_2(R) \geq r$, so that $\ell_2(R_i) \leq \frac{2}{3} \ell_2(R)$. This implies that $\mathrm{vol}(R_i) \leq \frac{2}{3} \mathrm{vol}(R)$.
For the second statement observe that $\ell_1(R) \geq \frac{\ell_2(R)}{3} \geq \frac{r}{3}$ (the first inequality is by assumption on the aspect ratio of $R$, the second inequality is by definition of $r$). Given this estimate, the construction of $R_i$ implies that $\frac{1}{3}r\leq \ell_2(R_i),\ell_1(R_i) \leq \frac{2}{3}r$, which concludes the fact that $R_i$ has aspect ratio smaller than $2$.
#### Queries and choice of output.
The edge fixing algorithm queries a $\sqrt{\frac{\epsilon' r}{500\log(1/\epsilon)}}$-net on $\partial R_i$ for all $i \in \{1,2,3\}$ (thus a total of $4 \sqrt{\frac{500 r \log(1/\epsilon)}{\epsilon'}} \leq 90 \sqrt{\frac{r \log(1/\epsilon)}{\epsilon}}$ queries), and we define $x_i$ to be the best point found on each respective net.
If for all $i \in \{1,2,3\}$ one has $$\label{eq:noimprov}
f(x_i) \leq f(x_{i-1}) - \frac{\epsilon' r}{3} \,,$$ then we set $(\tilde{R},\tilde{x})=(R,x_3)$. Otherwise denote $i^* \in \{1,2,3\}$ for the smallest number which violates , and set $(\tilde{R},\tilde{x})=(R_{i^*},x_{i^*-1})$.
![*Edge fixing*: the rectangle $R_1$ is marked in red.[]{data-label="fig:figure3"}](Figure3)
\[lem:notRi\] If $(\tilde{R},\tilde{x})=(R,x_3)$ then $(\mathcal{E},x_3)$ satisfies $P_0$. Furthermore for any edge $E$ of $R$, if $(E,x)$ satisfies $P_0$ (respectively $P_{\epsilon'}$) then so does $(E,x_3)$.
Since $(\tilde{R},\tilde{x})=(R,x_3)$ it means that $f(x_3) \leq f(x_0) - \epsilon' r$. In particular since $(\mathcal{E},x_0)$ satisfies $P_{\epsilon'}$ one has $f(x_0) < f^*_{\delta}(\mathcal{E}) - \frac{\delta^2}{8} + \epsilon' r$, and thus now one has $f(x_3) < f^*_{\delta}(\mathcal{E}) - \frac{\delta^2}{8}$ which means that $(\mathcal{E},x_3)$ satisfies $P_0$.
Let us now turn to some other edge $E$ of $R$. Certainly if $(E,x_0)$ satisfies $P_0$ then so does $(E,x_3)$ since $f(x_3) \leq f(x_0)$. But, in fact, even $P_{\epsilon'}$ is preserved since by the triangle inequality (and $\|x_3 - x_0\|_2 \leq r$) one has $$f(x_3) - \epsilon' \cdot \mathrm{dist}(x_3,E) \leq f(x_3) + \epsilon' r - \epsilon' \cdot \mathrm{dist}(x_0,E) \leq f(x_0) - \epsilon' \cdot \mathrm{dist}(x_0,E).$$
\[lem:Ri\] If $(\tilde{R},\tilde{x})=(R_{i},x_{i-1})$ for some $i \in \{1,2,3\}$, then $(\tilde{R},\tilde{x})$ satisfy $P_{\left( 1 + \frac{1}{500 \log(1/\epsilon)} \right) \epsilon'}$.
By construction, if $(\tilde{R},\tilde{x})=(R_{i},x_{i-1})$, then for any edge $E$ of $R_i$ one has $f(x_{i-1}) < f^*_{\delta}(E) + \frac{\epsilon' r}{3}$. Furthermore one has $\frac{\epsilon' r}{3} = - \frac{\epsilon' r}{8 \times 500 \log(1/\epsilon)} + \left(1+ \frac{3}{8 \times 500 \log(1/\epsilon)} \right) \frac{\epsilon' r}{3}$, and thus by definition one then has $P_{\left(1+ \frac{3}{8 \times 500 \log(1/\epsilon)} \right) \epsilon}$ for $(E, x_{i-1})$ whenever $\mathrm{dist}(x_{i-1}, E) = \frac{r}{3}$. If $\mathrm{dist}(x_{i-1}, E) < \frac{r}{3}$ then by the triangle inequality, $\mathrm{dist}(x_0,E) < r$, and moreover $E$ is also an edge with respect to $R$. Thus from the definition of $r$, $(E,x_0)$ satisfies $P_0$. Also by our choice of $x_{i-1}$, we know that $f(x_{i-1}) \leq f(x_0)$. Hence $(E,x_{i-1})$ satisfies $P_0$ as well.
Generalization to higher dimensions {#sec:highdim}
-----------------------------------
As explained in the introduction, there is no reason to restrict GFT to $[0,1]^2$ and, in fact, the algorithm may be readily adapted to higher-dimensional spaces, such as $[0,1]^d$, for some $d > 2$. We now detail the necessary changes and derive the complexity announced in Theorem \[thm:highdim\].
First, if $F$ is an affine hyperplane, and $x \in [0,1]^d$, we define $P_c$ for $(F,x)$ in the obvious way (i.e., same definition except that we consider a $\delta$-net of $F$). Similarly for $(R,x)$, when $R$ is an axis-aligned hyperrectangle.
Gradient flow trapping in higher dimensions replaces every line by a hyperplane, and every rectangle by a hyperrectangle. In particular at each step GFT maintains a domain $(R,x)$, where $R$ is a hyperrectangle with aspect ratio bounded by 3, and $x \in R$. The two subroutines are adapted as follows:
1. [*Parallel trap*]{} works exactly in the same way, except that the two lines $E$ and $F$ are replaced by two corresponding affine hyperplanes. In particular the query cost of this step is now $O\left(\left(\frac{\mathrm{diam}(R)}{\epsilon}\right)^\frac{d-1}{2}\right)$, and the volume shrinks by at least $0.95$.
2. In [*edge fixing*]{}, we now have three hyperrectangles $R_i$, and we need to query nets on their $2d$ faces. Thus the total cost of this step is $O \left( d\left(\frac{\mathrm{diam(R)}\log(1/\epsilon)}{\epsilon}\right)^\frac{d-1}{2} \right)$. Moreover, suppose that domain does not shrink at the end of this step and the output is a domain $(R,\tilde{x})$ for some other $\tilde{x} \in R$. In this case we know that $R$ has some face $F$, such that $(F,x)$ did not satisfy $P_0$, but $(F,\tilde{x})$ does satisfy $P_0$. It follows that we can run [*edge fixing*]{}, at most $2d$ times before the domain shrinks.
We can now analyze the complexity of the high-dimensional version of GFT:
First observe that, if $R$ is a hyperrectangle in $[0,1]^d$ with aspect ratio bounded by $3$, then we have the following inequality, $$\mathrm{diam}(R) \leq 3\sqrt{d}\cdot\mathrm{vol}(R)^{\frac{1}{d}}.$$ By repeating the same calculations done in Lemma \[lem:proof1\] and the observation about *parallel trap* and *edge fixing* made above, we see that the domain shrinks at least once in every $2d + 1$ steps, so that at step $T$, $$\mathrm{vol}(R_T) \leq 0.95^{\frac{T}{2d+1}},$$ and $$\mathrm{diam}(R_T) \leq 3\sqrt{d}\cdot 0.95^{\frac{T}{(2d+1)d}}.$$ Since the algorithm stops when $\mathrm{diam}(R_T) \leq 2\epsilon$, we get $$T = O\left(d^2\log\left(\frac{d}{\epsilon}\right)\right).$$ The total work done by the algorithm is evident now by considering the number of queries at each step.
Lower bound for randomized algorithms {#sec:LB}
=====================================
In this section, we show that any randomized algorithm must make at least $\tilde{\Omega}\left(\frac{1}{\sqrt{\epsilon}}\right)$ queries in order to find an $\epsilon$-stationary point. This extends the lower bound in [@Vav93], which applied only to deterministic algorithms. In particular, it shows that, up to logarithmic factors, adding randomness cannot improve the algorithm described in the previous section.
For an algorithm $\mathcal{A}$, a function $f: [0,1]^2\to \R$ and $\epsilon >0$ we denote by $\mathcal{Q}\left(\mathcal{A},f,\epsilon\right)$ the number of queries made by $\mathcal{A}$, in order to find an $\epsilon$-stationary point of $f$. Our goal is to bound from below $$\mathcal{Q}_{\mathrm{rand}}(\epsilon):=\inf\limits_{\mathcal{A} \text{ random} }\sup\limits_f\E_\mathcal{A}\left[\mathcal{Q}\left(\mathcal{A},f,\epsilon\right)\right],$$ where the infimum is taken over all random algorithms and the supremum is taken over all smooth functions, $f$. The expectation is with respect to the randomness of $\mathcal{A}$. By Yao’s minimax principle we have the equality $$\mathcal{Q}_{\mathrm{rand}}(\epsilon) = \sup\limits_{\mathcal{D}}\inf\limits_{\mathcal{A} \text{ determinstic}}\E_{f\sim\mathcal{ D}}\left[\mathcal{Q}\left(\mathcal{A},f,\epsilon\right)\right].$$ Here, $\mathcal{A}$ is a deterministic algorithm and $\mathcal{D}$ is a distribution over smooth functions. The rest of this section is devoted to proving the following theorem:
\[thm: lower bound\] Let $h:\N \to \R$ be a decreasing function such that $$\sum\limits_{k=1}^\infty \frac{h(k)}{k}< \infty,$$ and set $$\label{eq: def sum}
S_h(n):= \sum\limits_{k=1}^n \frac{1}{k\cdot h(k)}.$$ Then, $$\mathcal{Q}_\mathrm{rand}(\epsilon)=\Omega\left(\frac{1}{\sqrt{\epsilon}\cdot S_h\left(\left\lceil\frac{1}{\sqrt{\epsilon}}\right\rceil\right)}\right).$$
Remark that one may take $h(k):= \frac{1}{\log(k)^2+1}$ in the theorem. In this case $S_h(k) = O(\log^3(k))$, and $\mathcal{Q}_\mathrm{rand}(\epsilon)=\Omega\left(\frac{1}{\log^3(1/\epsilon) \sqrt{\epsilon} }\right)$, which is the announced lower bound.\
One of the main tools utilized in our proof is the construction introduced in [@Vav93]. We now present the relevant details.
A reduction to monotone path functions
--------------------------------------
Let $G_n = (V_n,E_n)$ stand for the $n+1 \times n+1$ grid graph. That is, $$V_n = \{0,\hdots, n\} \times \{0,\hdots, n\} \text{ and } E_n =\{(v,u)\in V_n\times V_n : \|v-u\|_1 = 1\}.$$ We say that a sequence of vertices, $(v_0,...,v_n)$ is a *monotone path* in $G_n$ if $v_0 = (0,0)$ and for every $0<i\leq n$, $v_i - v_{i-1}$ either equals $(0,1)$ or $(1,0)$. In other words, the path starts at the origin and continues each step by either going right or up. If $(v_0,...,v_n)$ is a monotone path, we associate to it a *monotone path function* $P:V_n \to \R$ by $$P(v) = \begin{cases}
-\|v\|_1& \text{if } v\in \{v_0,...,v_n\}\\
\hfill \|v\|_1& \text{otherwise}
\end{cases}.$$ By a slight abuse of notation, we will sometimes refer to the path function and the path itself as the same entity. If $i = 0,...,n$ we write $P_i$ for $P^{-1}(-i)$ and $P[i]$ for the prefix $(P_0,P_1,...,P_{i})$. If $v \in V_n$ is such that $P(v) > 0$, we say that $v$ does not lie on the path.\
We denote the set of all monotone path functions on $G_n$ by $\mathrm{F}_n.$ It is clear that if $P \in \mathrm{F}_n$ then $P_n$ is the only local minimum of $P$ and hence the global minimum.\
Informally, the main construction in [@Vav93] shows that for every $P \in \mathrm{F}_n$ there is a corresponding smooth function $\hat{P}:[0,1]^2\to \R$, which ’traces’ the path in $P$ and preserves its structure. In particular, finding an $\epsilon$-stationary point of $\hat{P}$ is not easier than finding the minimum of $P$.\
To formally state the result we fix $\epsilon >0$ and assume for simplicity that $\frac{1}{\sqrt{\epsilon}}$ is an integer. We henceforth denote $n(\epsilon) := \frac{1}{\sqrt{\epsilon}}$ and identify $V_{n(\epsilon)}$ with $[0,1]^2$ in the following way: if $(i,j) = v\in V_{n(\epsilon)}$ we write $\mathrm{square}(v)$ for the square: $$\mathrm{square}(v) = \left[\frac{i}{n(\epsilon) + 1}, \frac{i+1}{n(\epsilon) + 1}\right] \times \left[\frac{j}{n(\epsilon) + 1}, \frac{j+1}{n(\epsilon) + 1}\right].$$ If $\varphi: [0,1]^2 \to \R$, then $\mathrm{supp}(\varphi)$ denotes the closure of the set $\{x \in[0,1]^2 : \varphi(x) \neq 0\}$.
\[lem: vavsis\] Let $P \in \mathrm{F}_{n(\epsilon)}$. Then there exists a function $\hat{P}:[0,1]^2\to\R$ with the following properties:
1. $\hat{P}$ is smooth.
2. $\hat{P} = f_P + \ell$, where $\ell$ is a linear function, which does not depend on $P$, and $$\mathrm{supp}(f_P)\subset \bigcup\limits_{i=0}^n\mathrm{square}\left(P_i\right).$$
3. If $x \in [0,1]^2$ is an $\epsilon$-stationary point of $\hat{P}$ then $x \in \mathrm{square}\left(P_n\right)$.
4. if $P' \in \mathrm{F}_{n(\epsilon)}$ is another function and for some $i = 0,...,n$, $(P'_{i-1},P'_i,P'_{i+1}) = (P_{i-1},P_i,P_{i+1})$. Then $$\hat{P}'\vert_{\mathrm{square}\left(P_i\right)} = \hat{P}\vert_{\mathrm{square}\left(P_i\right)}$$
We now make precise of the fact that finding the minimum of $P$ is as hard as finding an $\epsilon$-stationary point of $\hat{P}$. For this we define $\mathcal{G}(\mathcal{A},P)$, the number of queries made by algorithm $\mathcal{A}$, in order to find the minimal value of the function $P$.
\[lem: reduction\] For any algorithm $\mathcal{A}$, which finds an $\epsilon$-stationary point of smooth functions on $[0,1]^2$, there exists an algorithm $\tilde{\mathcal{A}}$ such that $$\mathcal{Q}(\mathcal{A}, \hat{P},\epsilon) \geq \frac{1}{5} \mathcal{G}(\tilde{\mathcal{A}},P),$$ for any $P \in \mathrm{F}_{n(\epsilon)}.$
Given an algorithm $\mathcal{A}$ we explain how to construct $\tilde{\mathcal{A}}$. Fix $P \in \mathrm{F}_{n(\epsilon)}$. If $\mathcal{A}$ queries a point $x \in \mathrm{square}\left(v\right)\subset [0,1]^2$. Then $\tilde{\mathcal{A}}$ queries $v$ and all of its neighbors. When $\mathcal{A}$ terminates it has found an $\epsilon$-stationary point. By Lemma \[lem: vavsis\], this point must lie in $\mathrm{square}\left(P_n\right)$. By querying $P_n$ and its neighbors, $\tilde{\mathcal{A}}$ will determine that $P_n$ is a local minimum and hence the minimum of $P$.\
Since each vertex has at most $4$ neighbors, it will now suffice to show that $\tilde{\mathcal{A}}$ can remain consistent with $\mathcal{A}$. We thus need to show that after querying the neighbors of $v$, $\tilde{\mathcal{A}}$ may deduce the value of $\hat{P}(x)$.\
As we are only interested in the number of queries made by $\tilde{\mathcal{A}}$, it is fine to assume that $\tilde{\mathcal{A}}$ has access to the construction used in Lemma \[lem: vavsis\]. Now, suppose that $P(v) > 0$ and $v$ does not lie on the path. In this case, by Lemma \[lem: vavsis\], $\hat{P}(x) = \ell(x)$, which does not depend on $P$ itself and $\ell(x)$ is known. Otherwise $v = P_i$ for some $i =0,...,n$. So, after querying the neighbors of $v$, $\tilde{\mathcal{A}}$ also knows $P_{i-1}$ and $P_{i+1}$. The lemma then tells us that $\hat{P}\vert_{\mathrm{square}\left(v\right)}$ is uniquely determined and, in particular, the value of $\hat{P}(x)$ is known.
A lower bound for monotone path functions
-----------------------------------------
Denote $\mathcal{D}_{\mathrm{p}}(n)$ to be the set of all distributions supported on $\mathrm{F}_n$. By Lemma \[lem: reduction\], $$\mathcal{Q}_{\mathrm{rand}}(\epsilon) \geq \sup\limits_{\mathcal{D}\in\mathcal{D}_{\mathrm{p}}\left(n(\epsilon)\right) }\inf\limits_{\mathcal{A} \text{ determinstic}}\E_{P\sim\mathcal{ D}}\left[\mathcal{Q}\left(\mathcal{A},\hat{P},\epsilon\right)\right] \geq \frac{1}{5} \sup\limits_{\mathcal{D}\in\mathcal{D}_{\mathrm{p}}\left(n(\epsilon)\right) }\inf\limits_{\mathcal{A} \text{ determinstic}}\E_{P\sim\mathcal{ D}}\left[\mathcal{G}\left(\tilde{\mathcal{A}},P\right)\right].$$ In [@sun2009quantum], the authors present a family of random paths $(X_\delta)_{\delta>0} \subset \mathcal{D}_{\mathrm{p}}(n)$. Using these random paths it is shown that for every $\delta > 0$, $$\mathcal{G}_{\mathrm{rand}}(n):=\sup\limits_{\mathcal{D}\in\mathcal{D}_{\mathrm{p}}\left(n\right) }\inf\limits_{\mathcal{A} \text{ determinstic}}\E_{P\sim\mathcal{ D}}\left[\mathcal{G}\left(\mathcal{A},P\right)\right] = \Omega(n^{1-\delta}).$$ This immediately implies, $$\mathcal{Q}_{\mathrm{rand}}(\epsilon) = \Omega\left(\left(\frac{1}{\sqrt{\epsilon}}\right)^{1-\delta}\right).$$ Their proof uses results from combinatorial number theory in order to construct a random path which, roughly speaking, has unpredictable increments. This distribution is then used in conjunction with a method developed by Aaronson ([@aaronson2006lower]) in order to produce a lower bound.\
We now present a simplified proof of the result, which also slightly improves the bound. We simply observe that known results concerning unpredictable random walks, can be combined with Aaronson’s method. Theorem \[thm: lower bound\] will then be a consequence of the following theorem:
\[thm: discrete lower bound\] Let the notations of Theorem \[thm: lower bound\] prevail. Then $$\mathcal{G}_\mathrm{rand}(n) = \Omega\left(\frac{n}{S_h(n)}\right).$$
The theorem of Aaronson, reformulated using our notations (see also [@sun2009quantum Lemma 2]), is given below.
\[thm: aaronson\] Let $w :\mathrm{F}_n \times \mathrm{F}_n \to \R^+$ be a weight function with the following properties:
- $w(P,P') = w(P',P)$.
- $w(P,P') = 0$, whenever $P_n = P'_n$.
Define $$T(w,P) := \sum\limits_{Q \in \mathrm{F}_n} w(P,Q),$$ and for $v \in V_n$ $$T(w,P,v):= \
\sum\limits_{Q \in \mathrm{F}_n: Q(v) \neq P(v)}w(P,Q).$$ Then $$\mathcal{G}_{\mathrm{rand}}(n) = \Omega\left( \min\limits_{\substack{P,P',v\\P(v)\neq P'(v), w(P,P') > 0}}\max\left(\frac{T(w,P)}{T(w,P,v)},\frac{T(w,P')}{T(w,P',v)}\right)\right).$$
For $P \in \mathrm{F}_n$, one should think about $w$ as inducing a probability measure according to $w(P,\cdot)$. If $Q$ is sampled according to this measure, then the quantity $\frac{T(w,P,v)}{T(w,P)}$ is the probability that $P(v) \neq Q(v)$. That is, either $v \in P$ or $v \in Q$, but not both. The theorem then says that if this probability is small, for at least one path in each pair $(P, P')$ such that $P_n \neq P'_n$, then any randomized algorithm must make as many queries as the reciprocal of the probability.\
We now formalize this notion; For a random path $X \in \mathcal{D}_\mathrm{p}(n)$, define the following weight function: $$w_X(P,P') = \begin{cases}
0& \text { if } P_n = P'_n\\
\P(X = P)\cdot\sum\limits_{i=0}^{n-1}\P(X = P'| X[i] = P[i])& \text{ otherwise }
\end{cases}.$$ Here $w_X(P,P')$ is proportional to the probability that $X = P'$, conditional on agreeing with $P$ on the first $i$ steps, where $i$ is uniformly chosen between $0$ and $n-1$. Note that, for any $i$, $$\begin{aligned}
\P\left(X = P\right)\cdot&\P\left(X = P'| X[i] = P[i]\right) \\
&=\P(X[i] = P[i])\cdot\P(X = P|X[i] = P[i])\cdot\P(X = P'| X[i] = P[i])\\
&=\P\left(X = P'\right)\cdot\P\left(X = P| X[i] = P'[i]\right).\end{aligned}$$ Hence, $w_X(P, P') = w_X(P', P)$. We will use the following theorem from [@haggstrom1998nearest], which generalizes the main result of [@BPP98].
Let $h$ be as in Theorem \[thm: lower bound\], Then there exists a random path $X^h \in \mathcal{D}_\mathrm{p}(n)$ and a constant $c_h>0$, such that for all $m\geq k$, and for every $(v_0,v_1...,v_{m-k})$, sequence of vertices, $$\label{eq: unpredictablity}
\sup\limits_{\|u\|_1 = m}\P\left(X^h_m = u|X^h_0 = v_0,...,X^h_{m-k} = v_{m-k}\right) \leq \frac{c_h}{kf(k)}.$$
For $X^h$ as in the theorem abbreviate $w_h := w_{X^h}$ and recall $S_h(n) : = \sum\limits_{k=1}^{n}\frac{1}{k\cdot h(k)}$. We now prove the main quantitative estimates which apply to $w_h$.
For any $P \in \mathrm{F}_n$, $$\sum\limits_{Q \in \mathrm{F}_n}w_h(P,Q) \geq \P(X^h = P)\cdot\left(n - c_hS_h(n)\right).$$
We write $$\begin{aligned}
\sum\limits_{Q \in \mathrm{F}_n}w_h(P,Q) &= \P(X^h = P)\sum\limits_{Q: Q_n \neq P_n}\sum\limits_{i=0}^{n-1}\P\left(X^h = Q| X^h[i] = P[i]\right)\\
&=\P(X^h = P)\sum\limits_{i=0}^{n-1}\left(1 - \sum\limits_{Q: Q_n = P_n}\P\left(X^h = Q| X^h[i] = P[i]\right)\right)\\
&=\P(X^h = P)\left(n - \sum\limits_{i=0}^{n-1}\sum\limits_{Q: Q_n = P_n}\P\left(X^h = Q| X^h[i] = P[i]\right)\right).
\end{aligned}$$ Using , we get $$\sum\limits_{Q: Q_n = P_n}\P\left(X^h = Q| X^h[i] = P[i]\right) \leq \P(X^h_n = P_n|X^h[i] = P[i])\leq \frac{c_h}{(n-i)\cdot h(n-i)},$$ and $$\sum\limits_{i=0}^{n-1}\sum\limits_{Q: Q_n = P_n}\P\left(X^h = Q| X^h[i] = P[i]\right) \leq \sum\limits_{k=1}^{n}\frac{c_h}{k\cdot h(k)} = c_hS_h(n).$$
Let $P \in \mathrm{F}_n$ and $v \in V_n$ such that $\|v\|_1 = \ell$ and $P_{\ell} \neq v$. Then, $$\sum\limits_{\substack{Q\in \mathrm{F}_n\\ Q_\ell =v}}w_h(P,Q) \leq 2\P(X^h = P)c_hS_h(n).$$
$$\begin{aligned}
\sum\limits_{\substack{Q\in \mathrm{F}_n\\ Q_\ell =v}}w_h(P,Q) &=\P(X^h = P)\sum\limits_{i=0}^{\ell - 1}\sum\limits_{\substack{Q: Q_n \neq P_n\\Q_\ell = v}}\P\left(X^h = Q| X^h[i] = P[i]\right)\\
&\leq\P(X^h = P)\sum\limits_{i=0}^{\ell - 1}\sum\limits_{\substack{Q: Q_\ell = v
}}\P\left(X^h = Q| X^h[i] = P[i]\right).
\end{aligned}$$
Observe that if $Q_\ell = v$, then $Q_{\ell+1}$ must equal $v + (0,1)$ or $v + (1,0)$. In particular, for $i < \ell$, shows $$\begin{aligned}
\sum\limits_{\substack{Q: Q_{\ell} = v
}}\P&\left(X^h =Q| X^h[i] = P[i]\right) \\
&\leq \P\left(X^h_{\ell+1} = v + (0,1) \text{ or }X^h_{\ell+1} = v + (1,0)|X^h[i] = P[i]\right)\\
&\leq \P\left(X^h_{\ell+1} = v + (0,1)|X^h[i] = P[i]\right) + \P\left(X^h_{\ell+1} = v + (1,0)|X^h[i] = P[i]\right)\\
&\leq \frac{2c_h}{(\ell+1 - i)\cdot h(\ell+1 - i)}.
\end{aligned}$$ So, $$\begin{aligned}
\sum\limits_{i=0}^{\ell - 1}\sum\limits_{\substack{Q: Q_\ell = v
}}\P\left(X^h =Q| X^h[i] = P[i]\right) &\leq \sum\limits_{i=0}^{\ell - 1}\frac{2c_h}{(\ell+1 - i)\cdot h(\ell+1 - i)}\\
\leq 2c_hS_h(n).
\end{aligned}$$
We are now in a position to prove Theorem \[thm: discrete lower bound\].
Let $P \in \mathrm{F}_n$ and let $v \in V_n$, with $$\|v\|_1 = \ell \text{ and } P_\ell \neq v.$$ Note that $P(v) = \ell$. So, if $Q \in \mathrm{F}_n$ is such that $Q(v) \neq P(v)$, then necessarily $Q_{\ell} = v$. We now set $P' \in \mathrm{F}_n$, with $P(v) \neq P'(v)$. In this case, the previous two lemmas show $$\begin{aligned}
\max&\left(\frac{T(w_h,P)}{T(w_h,P,v)},\frac{T(w_h,P')}{T(w_h,P',v)}\right) \\
&\geq \frac{T(w_h,P)}{T(w_h,P,v)}
= \frac{\sum\limits_{Q\in \mathrm{F}_n}w_h(P,Q)}{\sum\limits_{\substack{Q\in \mathrm{F}_n\\Q(v) \neq P(v)}}w_h(P,Q)} = \frac{\sum\limits_{Q\in \mathrm{F}_n}w_h(P,Q)}{\sum\limits_{\substack{Q\in \mathrm{F}_n\\Q_{\ell}=v}}w_h(P,Q)}\geq \frac{n - c_hS_h(n)}{2c_hS_h(n)}.
\end{aligned}$$ Since we are trying to establish a lower bound, we might as well assume that $S_h(n) = o(n)$. So, for $n$ large enough $$\frac{n - c_hS_h(n)}{2c_hS_h(n)} \geq\frac{n}{4c_hS_h(n)}.$$ Plugging this estimate into Theorem \[thm: aaronson\] yields the desired result
Heuristic extension to higher dimensions {#sec:heuristic}
----------------------------------------
In this section we propose a heuristic approach to extend the lower bound to higher dimensions. In the $2$ dimensional case, the proof method of Section \[sec:LB\] consisted of two steps: first reduces the problem to the discrete setting of monotone paths in $[n]^2$, and then analyze the query complexity of finding the minimal point for such path functions. Thus, to extend the result we should consider path functions on the $d$-dimensional grid, as well as a way to build smooth functions on $[0,1]^d$ from those paths.
The lower bound for finding minimal points of path functions in high-dimensional grids was obtained in [@Zha06], where it was shown that, in the worst case, any randomized algorithm must make $\Omega\left(n^{\frac{d}{2}}\right)$ queries in order to find the end point of a path defined over $[n]^{d}$. Thus, if we can find a discretization scheme, analogous to Lemma \[lem: vavsis\], in higher dimensions, we could obtain a lower bound for finding $\epsilon$-stationary points. What are the constraints on such a discretization?
First note that necessarily the construction of [@Zha06] must be based on paths of lengths $\Omega\left(n^{\frac{d}{2}}\right)$, for otherwise one could simply trace the path to find its endpoint. In particular, an analogous construction to Lemma \[lem: vavsis\] will reach value smaller than $- \epsilon \cdot n^{\frac{d}{2}}$ at the stationary point (i.e., the endpoint of the path). On the other hand, in at least one of the neighboring cube (which are at distance less than $1/n$ from the stationary point), the background linear function should prevail, meaning that the function should reach a positive value. Since around the stationary point the function is quadratic, we get the constraint: $$- \epsilon \cdot n^{\frac{d}{2}} + \left(\frac{1}{n}\right)^2 > 0 \Leftrightarrow n < \left( \frac{1}{\epsilon} \right)^{\frac{2}{d+2}} \,.$$ In particular the lower bound $\Omega\left(n^{\frac{d}{2}}\right)$ now suggests that for finding stationary point one has the complexity lower bound $\left( \frac{1}{\epsilon} \right)^{\frac{d}{d+2}}$.
Discussion {#sec:open}
==========
In this paper we introduced a near-optimal algorithm for finding $\epsilon$-stationary points in dimension $2$. Finding a near-optimal algorithm in dimensions $d \geq 3$ remains open. Specific challenges include:
1. Finding a strategy in dimension $3$ which improves upon GFT’s $\tilde{O}(1/\epsilon)$ complexity.
2. The heuristic extension of the lower bound in Section \[sec:heuristic\] suggests $\Omega\left(\frac{1}{\epsilon^{\frac{d}{d+2}}}\right)$ as a complexity lower bound for any dimension $d$ (note in particular that the exponent tends to $1$ as $d$ tends to infinity). On the other hand, [@CDHS19] proved that for $d = \Omega(1/\epsilon^2)$, one has the complexity lower bound $\Omega(1/\epsilon^2)$. How do we reconcile these two results? Specifically we raise the following question: Is there an algorithm with complexity $C_d / \epsilon$ for some constant $C_d$ which depends only on $d$? (Note that $C_d$ as small as $O(\sqrt{d})$ would remain consistent with [@CDHS19].) Alternatively we might ask whether the [@CDHS19] lower bound holds for much smaller dimensions, e.g. when $d = \Theta(\log(1/\epsilon))$, are we in the $1/\epsilon$ regime as suggested by the heuristic, or are we already in the high-dimensional $1/\epsilon^2$ of [@CDHS19]?
3. Especially intriguing is the limit of low-depth algorithms, say as defined by having depth smaller than $\mathrm{poly}(d \log(1/\epsilon))$. Currently this class of algorithms suffers from the curse of dimensionality, as GFT’s total work degrades significantly when the dimension increases (recall from Theorem \[thm:highdim\] that it is $\tilde{O}\left(\frac{1}{\epsilon^{\frac{d-1}{2}}}\right)$). Is this necessary? A much weaker question is to simply show a separation between low-depth and high-depth algorithms. Namely can one show a lower bound $\Omega(1/\epsilon^c)$ with $c>2$ for low-depth algorithms? We note that lower bounds on depth have been investigated in the convex setting, see [@Nem94], [@BJLLS19].
4. A technically challenging problem is to adapt the construction in \[Section 3, [@Vav93]\] to non-monotone paths in higher dimensions. In particular, to formalize the heuristic argument from Section \[sec:heuristic\], such construction should presumably avoid creating saddle points.
Many more questions remain open on how to exploit the low-dimensional geometry of smooth gradient fields, and the above four questions are only a subset of the fundamental questions that we would like to answer. Other interesting questions include closing the logarithmic gap in dimension $2$, or understanding better the role of randomness for this problem (note that GFT is deterministic, but other type of strategies include randomness, such as Hinder’s non-convex cutting plane [@Hin18]).
### Acknowledgment {#acknowledgment .unnumbered}
We thank Ronen Eldan, Yin Tat Lee, Yuanzhi Li and Mark Selke for many helpful discussions on this problem.
[^1]: This work was done while D. Mikulincer was an intern at Microsoft Research. Supported by an Azrieli foundation fellowship.
[^2]: We note that a different perspective on this question from the one developed in this paper was investigated in [@Hin18], where the author asks whether one can adapt [*actual cutting planes*]{} to non-convex settings. In particular [@Hin18] shows that one can improve upon gradient descent and obtain a complexity $O(\mathrm{poly}(d)/\epsilon^{4/3})$ with a cutting plane method, under a higher order smoothness assumption (namely third order instead of first order here).
[^3]: Technically we consider here the zeroth order oracle model. It is clear that one can obtain a first order oracle model from it, at the expense of a multiplicative dimension blow-up in the complexity. In the context of this paper an extra factor $d$ is small, and thus we do not dwell on the distinction between zeroth order and first order.
[^4]: In “essence" $(C,x)$ satisfies $P_{\epsilon}$, this is only slightly informal since we defined $P_{\epsilon}$ for rectangles and $C$ is a circle. In particular we chose the improvement $\frac{3}{4} \epsilon$ instead of the larger $\frac{7}{8} \epsilon$ (which is enough to obtain $P_c$) to account for an extra term due to polygonal approximation of the circle. We encourage the reader to ignore this irrelevant technicality.
[^5]: We need three candidates to ensure that the domain will shrink.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
It is well known that magnification bias, the modulation of galaxy or quasar source counts by gravitational lensing, can change the observed angular correlation function. We investigate magnification-induced changes to the shape of the observed correlation function $w(\theta)$, and the angular power spectrum $C_{\ell}$, paying special attention to the matter-radiation equality peak and the baryon wiggles. Lensing effectively mixes the correlation function of the source galaxies with that of the matter correlation at the lower redshifts of the lenses distorting the observed correlation function. We quantify how the lensing corrections depend on the width of the selection function, the galaxy bias $b$, and the number count slope $s$. The lensing correction increases with redshift and larger corrections are present for sources with steep number count slopes and/or broad redshift distributions. The most drastic changes to $C_{\ell}$ occur for measurements at high redshifts ($z{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}1.5$) and low multipole moment ($\ell {\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$<$}}}100$). For the source distributions we consider, magnification bias can shift the location of the matter-radiation equality scale by $1$–$6\%$ at $z\sim1.5$ and by $z\sim3.5$ the shift can be as large as $30\%$. The baryon bump in $\theta^2w(\theta)$ is shifted by ${\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$<$}}}1\%$ and the width is typically increased by $\sim10\%$. Shifts of ${\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}0.5\%$ and broadening ${\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}20\%$ occur only for very broad selection functions and/or galaxies with $(5s-2)/b{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}2$. However, near the baryon bump the magnification correction is not constant but is a gently varying function which depends on the source population. Depending on how the $w(\theta)$ data is fitted, this correction may need to be accounted for when using the baryon acoustic scale for precision cosmology.
author:
- 'Marilena LoVerde$^{1,2}$, Lam Hui$^{1,2,3}$, Enrique Gaztañaga$^{4}$'
title: 'Lensing corrections to features in the angular two-point correlation function and power spectrum'
---
Introduction
============
The galaxy two-point angular correlation function and its spherical harmonic transform, the angular power spectrum, provide information about dark matter clustering. These statistics have scale dependent features which may be used as “standard rulers” to measure cosmological distances [@Peebles1973; @Cooray06]. Features in the three-dimensional power spectrum which appear at a comoving wave number $k$ will appear in the angular power spectrum at redshift $z_0$ at multipole $\ell \sim k\chi(z_0)$, where $\chi(z_0)$ is the comoving distance to $z_0$. One such feature is the peak in the power spectrum that separates the modes which entered the horizon during radiation domination from those modes that entered during matter domination. This scale is determined by the horizon size at matter-radiation equality. Also present in the power spectrum are the so-called baryon acoustic oscillations (BAO). This series of wiggles in Fourier space is a signature of acoustic oscillations in the photon-baryon fluid that was present in the early universe. The location of the peaks of the BAO depends on the horizon size at the time of recombination [@BAO1; @BAO2; @BAO3; @BAO4; @BAO5; @EH98; @BAO6]. This scale is robustly measured from the cosmic microwave background [@WMAPI; @WMAPIII], and thus can serve as a “standard ruler”; given the scale of the BAO a measurement of the angular size of the BAO at some redshift will yield the angular diameter distance to that redshift . In fact the first measurements of this peak have occurred recently [@BAOmeas1; @BAOmeas2; @BAOmeas3; @BAOmeas4; @BAOmeas5]. In recent years, much effort has gone towards using these features in the correlation function for precision measurements [@BAOth1].
Gravitational lensing changes the observed number density of galaxy or quasar sources - an effect called magnification bias [@Gunn67; @Narayan1989; @BTP1996]. (Hereafter, the terms ‘galaxy’ and ‘quasar’ can be considered synonymous.) It is well known that magnification bias modifies the galaxy angular correlation function [@Villumsen1995; @VFC97; @MJV98; @MJ98; @EGmag03; @ScrantonSDSS05; @menard; @JSS03]. In this paper we extend the previous analyses by investigating and quantifying how magnification bias changes the shape of the angular two point correlation function, and its spherical harmonic transform the angular power spectrum, paying special attention to important features such as the turnover in the power spectrum and the BAO.
Corrections from gravitational lensing enter as follows. The two-point function is measured from the galaxy number density fluctuation. \_n= The effect of gravitational lensing is to alter the area of the patch of sky being observed and to change the observed flux of the source. Both effects change the measured galaxy number density, the first by changing the area, the second by changing the number of sources observed in a flux limited survey. Together these effects are called magnification bias [@Gunn67; @Narayan1989; @Villumsen1995; @BTP1996]. To first order (i.e. the weak lensing limit), they lead to a correction term $\delta_\mu$ being added to the intrinsic galaxy fluctuation $\delta_g$ \_n=\_g+\_\[deltan\] With this term the observed autocorrelation function becomes, \_n\_n=\_g\_g+\_g\_+\_\_g+\_\_. The galaxy-galaxy term $\langle\delta_g\delta_g\rangle$ depends on the matter distribution at the source galaxies. The magnification terms, especially $\langle\delta_\mu\delta_\mu\rangle$, depend on the matter distribution spanning the range between the sources and the observer.
The lensing of high redshift quasars by low redshift galaxies have been detected confirming the presence of magnification bias for these systems. The most recent measurements of this effect are discussed in [@EGmag03; @JSS03; @ScrantonSDSS05] (see also [@Myers2003]). Discussions of earlier measurements can be found in the references therein. These measurements work by cross-correlating the angular positions of galaxies/quasars at widely separated redshifts. Here, we will focus on the angular correlation of galaxies at similar redshifts. As was first pointed out by [@Matsubara], magnification bias alters observations of the 3D clustering of galaxies. In two separate papers [@3Dpaper1; @3Dpaper2], we further consider the effect of magnification bias on the 3D correlation function and power spectrum, which turn out to have qualitatively new and interesting features.
While this paper was in preparation, a paper addressing weak gravitational lensing corrections to baryon acoustic oscillations was posted [@otherpaper]. In that paper they consider the effect of magnification bias (a first order correction) and stochastic deflection (a second order correction) on BAO in the real space correlation function for a delta function source distribution. In contrast, we consider the effect of magnification bias on the angular correlation function and the angular power spectrum for an extended source distribution. As we will show the width of the source distribution strongly affects the magnitude of the magnification bias correction. While some conclusions we reach are addressed in [@otherpaper], the observables considered and the analyses here are different.
This paper is organized as follows. In section \[corrfunctions\] we present expressions for the angular auto-correlation function and the angular power spectrum when magnification bias is included. In section \[ampsection\] we discuss the factors affecting the relative magnitude of the magnification bias terms compared with the galaxy term. In section \[shapesection\] we examine the effect of magnification on the shape of the angular power spectrum for a few different source distributions. In section \[BAOsec\] we consider the effect of magnification on the baryon bump in the angular correlation function. In \[discussion\] we conclude and discuss the implications of our work for future surveys.
Anisotropies and Correlation functions {#corrfunctions}
======================================
We consider the two-dimensional galaxy fluctuation in direction $\hat{\thetaB}$ integrated over a selection function with mean redshift $z_0$ \[deltandef\] \_n(,z\_0)= where $\bar{n}$ is the mean number of galaxies in the redshift bin. We include this label, $z_0$, to remind the reader that the angular fluctuation depends on the source selection function.
When the effect of magnification bias is considered, the expression for the net measured galaxy overdensity becomes a sum of two terms (Equation \[deltan\]). In the expressions that follow and throughout this paper we assume a flat universe, generalizing to an open or closed universe is fairly straightforward. The first term in Equation \[deltan\] is the intrinsic galaxy fluctuation integrated over a normalized selection function given by $W(z,z_0)$, $$\label{deltag}
\delta_g(\hat{\thetaB},z_0)= \int_{0}^\infty\!\! dz \, b(z) W(z,z_0) \delta(\chi(z)\hat{\thetaB},z).$$ Where $\chi(z)$ is the comoving distance to redshift $z$, $\delta=(\rho-\bar{\rho})/\bar{\rho}$ is the matter overdensity and $b(z)$ is a bias factor relating the galaxy number density fluctuation to the matter density fluctuation $\delta_g=b(z)\delta$. For simplicity we have assumed that the bias is scale independent, this should be accurate for $k{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$<$}}}0.05 h\textrm{Mpc}^{-1}$, however at smaller scales it may be important (see for example [@RS3]). If we further assume that $b(z)$ is slowly varying across $W(z,z_0)$ then we can set $b(z)=b(z_0)$ and pull it out of the integral in Equation \[deltag\]. The second term in Equation \[deltan\] is the correction due to magnification bias, $$\begin{aligned}
\label{mutermfull}
\delta_{\mu}(\hat{\thetaB},z_0)=\int_0^{\infty}\!\!dz\frac{c}{H(z)}\nabla_{\perp}^2\phi(\chi(z)\hat{\thetaB},z_0)\qquad\qquad\qquad\\ \nonumber
\times\,\,\chi(z)\int_z^{\infty} dz'\left(5s(z')-2\right)\frac{\chi(z')-\chi(z)}{\chi(z')}W(z',z_0). \end{aligned}$$ $H(z)$ is the Hubble parameter and $c$ is the speed of light [@Narayan1989]. The magnification bias term depends on the Laplacian of the gravitational potential $\phi$ (with respect to the comoving coordinates in the direction perpendicular to $\hat{\thetaB}$) and the slope of the number count function. For a survey with limiting magnitude $m$ this is $$s=\frac{d\textrm{log}_{10}N(<m)}{dm}.$$ If $s(z)$ is slowly varying across $W(z,z_0)$, then we can set $s(z)=s(z_0)$ and the expression for $\delta_\mu$ becomes $$\begin{aligned}
\label{muterm}
\delta_{\mu}(\hat{\thetaB},z_0)&=&\left(5s(z_0)-2\right)\nonumber\\
&\times&\int_0^{\infty}\!\!dz\frac{c}{H(z)}g(z,z_0)\nabla_{\perp}^2\phi(\chi(z)\hat{\thetaB},z_0).\end{aligned}$$ Where we have introduced the lensing weight function $$\label{lenswt}
g(z,z_0)=\chi(z)\int_z^{\infty} dz'\frac{\chi(z')-\chi(z)}{\chi(z')}W(z',z_0).$$ The lensing weight function can be thought of as roughly proportional to the probability for sources in $W(z,z_0)$ to be lensed by density perturbations at $z$. The lensing weight function increases in magnitude as $z_0$ increases and is peaked at a redshift $z$ corresponding to about half the comoving distance to $z_0$.
On scales $k\gg aH$ Poisson’s equation can be used to relate the gravitational potential to the matter fluctuation -k\^2 ([**k**]{},z)= \_m (1+z)([**k**]{},z) where ${\bf k}$ is the comoving wave vector, $k$ is its magnitude and $\Omega_m$ is the matter density today [@dodelson]. .
We consider the two-point correlation function of the galaxy overdensity $$\begin{aligned}
\label{wtheta1}
w_{nn}(\theta,z_0)&=&\langle\delta_g(\hat{\thetaB},z_0)\delta_g(\hat{\thetaB}',z_0)\rangle+2\langle\delta_g(\hat{\thetaB},z_0)
\delta_{\mu}(\hat{\thetaB}',z_0)\rangle\nonumber\\
&+&\langle\delta_{\mu}(\hat{\thetaB},z_0)\delta_{\mu}(\hat{\thetaB}',z_0)\rangle\nonumber\\
&\equiv& w_{gg}(\theta,z_0)+2w_{g\mu}(\theta,z_0)+w_{\mu\mu}(\theta,z_0)\end{aligned}$$ where $\cos\theta=\hat{\thetaB}\cdot\hat{\thetaB}'$. In what follows we will consider the Legendre coefficients of the correlation functions. These are defined as \[wthetaeq\] w\_[gg]{}(,z\_0)=\_ C\^[gg]{}\_(z\_0)P\_() where $P_\ell(\cos\theta)$ are the Legendre polynomials. The Legendre components of the observed correlation function, $w_{nn}(\theta,z_0)$ will of course be a sum of all the terms, C\^[nn]{}\_(z\_0)=C\^[gg]{}\_(z\_0)+2C\^[g]{}\_(z\_0)+C\^\_(z\_0) We calculate the angular power spectra, $\Cgg$, $\Cgm$ and $\Cmm$, using the Limber approximation [@Limber] which is accurate for $\ell{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}10$. To simplify the expressions and provide some insight into the effects of the magnification terms we introduce the following functions $$\begin{aligned}
\label{phiggeq}
\phi^{gg}(z,z_0)=b(z_0)^2W(z,z_0)^2\,\qquad\qquad\qquad\qquad\qquad\quad\,\,\\
\label{phigmeq}
\label{phimmeq}
\phi^{\mu\mu}(z,z_0)=\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\,\\
\frac{9}{4}(5s(z_0)-2)^2\Omega^2_m \frac{H_0^4}{c^2H(z)^2}g^2(z,z_0)(1+z)^2\nn
\phi^{g\mu}(z,z_0)=\sqrt{\phi^{gg}(z,z_0)\phi^{\mu\mu}(z,z_0)}\qquad\qquad\qquad\qquad\quad\end{aligned}$$
With this notation the angular correlation functions can all be expressed as \[cxxleq\] C\^[xx]{}\_(z\_0)=\_0\^\^[xx]{}(z,z\_0)P(,z) where $xx$ symbolizes $gg$, $g\mu$ or $\mu\mu$ and $P(k=\ell/\chi(z),z)$ is the matter power spectrum.
For illustration we use Gaussian selection functions \[sel\] W(z,z\_0)=of various widths to demonstrate how the shape of the selection function alters the effect of magnification.
We assume a flat $\Lambda$CDM cosmology with $\Omega_m=0.27$, $\Omega_{\Lambda}=0.73$ and $\Omega_b=0.0224/h^2$ as the fractional energy densities in matter, vacuum and baryons today. The Hubble constant $H_0=100h$ km/s/Mpc is set to $h=0.7$, the fluctuation amplitude to $\sigma_8=0.8$, and scalar spectral index $n_s=0.95$. For the linear power spectrum we use the Eisenstein and Hu transfer function [@EH98]. In a few places (e.g. Figure \[CnnCnoBfig\]) we discuss the no-BAO spectrum, this is calculated using the BBKS transfer function [@BBKS] with a modified shape parameter $\Gamma=\Omega_mh\,\textrm{exp}[-\Omega_b(1+\sqrt{2h}/\Omega_m)]$ [@GammaSugiyama]. The non-linear evolution of both power spectra is calculated using the prescription of Smith *et. al.* [@Smithetal]. The nonlinear power spectrum is used in all plots and discussions.
Unless otherwise stated we set the galaxy sample dependent ratio $(5s-2)/b=1$ (see [@mepaper; @3Dpaper1] on how this varies with galaxy sample and redshift). The first correction term, $\Cgm/b^2$ is linear in this quantity while the second term, $\Cmm/b^2$ is quadratic. Thus the magnitude of the net correction, calculated here cannot in general be scaled by $(5s-2)/b$. However, at low redshifts $2\Cgm/b^2$ dominates over $\Cmm/b^2$ and at higher redshifts $\Cmm > 2\Cgm$. The redshift of the transition between the two regimes increases with the width of the selection function. Specifically when $\sigma=0.07$, $2\Cgm <\Cmm$ for $z_0 \ge 1.5$, when $\sigma=0.15$, $2\Cgm <\Cmm$ for $z_0\ge 2.0$ and when $\sigma=0.30$, $2\Cgm <\Cmm$ for $z_0\ge 2.5$.
amplitude of the lensing corrections {#ampsection}
====================================
The relative magnitude of the lensing magnification terms, $\Cgm$ and $\Cmm$, to the intrinsic galaxy term, $\Cgg$, depends on several things: first the galaxy-sample dependent quantities $b$ and $s$, and second the selection function and cosmological quantities in Equations \[phiggeq\]–\[phimmeq\]. Here we discuss how these quantities affect the relative magnitudes of $\Cgg$, $\Cgm$ and $\Cmm$.
The magnification terms are scaled by the galaxy-sample dependent factors $(5s-2)/b$ and $(5s-2)^2/b^2$. Thus for a given galaxy bias, galaxies residing on the steep end of the luminosity function will have larger magnification corrections. However, if $s<2/5$, the galaxy-magnification term $\Cgm$ is negative, while the magnification-magnification term, $\Cmm$ is always positive. Unless otherwise stated we fix $(5s-2)/b=1$ so both terms are positive.
The difference between the galaxy-galaxy, galaxy-magnification and magnification-magnification terms can be seen more clearly by considering equations \[phiggeq\]–\[cxxleq\]. From Equation \[cxxleq\], we see that all the angular power spectra are simply integrals of the power spectrum over some redshift distribution $\phi^{xx}(z,z_0)$. Examples of these distributions for a selection function with $\sigma=0.15$ are shown in Figure \[SelRad\]. In the case of $\phi^{gg}(z,z_0)$ this distribution is determined by the selection function. This suggests that one can think of the magnification terms as adding to $\Cgg$ measurements of the correlation function with “selection functions” determined by $\phi^{g\mu}(z,z_0)$ and $\phi^{\mu\mu}(z,z_0)$.
Figure \[SelRad\] shows that $\phi^{g\mu}(z,z_0)$ is peaked not too far from where the selection function is peaked, and is similar in shape to the selection function. Thus we expect $\Cgm$ to be similar in shape to $\Cgg$. On the other hand, the magnification-magnification term $\phi^{\mu\mu}$ is peaked at a much lower redshift than the selection function and is also much more broadly distributed in redshift. Thus we expect $\Cmm$ to be peaked at lower $\ell$ (larger angular scales) than $\Cgg$ because it is probing structure at lower redshifts which is nearer to the observer and occupies a larger angle in the sky. Additionally, since $\phi^{\mu\mu}(z,z_0)$ is rather broadly distributed, we expect sharp features such as baryon oscillations to be smoothed out in $\Cmm$.
One might infer from looking at Figure \[SelRad\] that the magnification bias terms $\Cgm$ and $\Cmm$ will be quite small compared to $\Cgg$. However, the other quantities in the integrand of Equation \[cxxleq\] decrease with redshift so the relative magnitudes of $\phi^{gg}$, $\phi^{g\mu}$ and $\phi^{\mu\mu}$ do not completely determine the relative magnitudes of $C^{gg}_{\ell}$,$C^{g\mu}_{\ell}$ and $C^{\mu\mu}_{\ell}$.
the shape of the Angular power spectrum {#shapesection}
=======================================
Using equations \[phiggeq\]–\[phimmeq\] and \[cxxleq\] we calculate the angular power spectra using Gaussian selection functions centered at a variety of redshifts. Because the galaxy-magnification cross term $\Cgm$ depends strongly on the width of the selection function, we consider several different widths given by $\sigma=0.07,0.15,0.30$. The magnification-magnification term $\Cmm$ is largely independent of the width of the selection function, but the galaxy-magnification term $\Cgm$ increases with increasing width of the selection function. The effect of increasing $\sigma$ is therefore to increase the overall contribution of magnification to $C^{nn}_\ell$.
In Figure \[CnnCnoBfig\] we show the angular power spectrum with and without magnification bias for a selection function of width $\sigma=0.15$. In this figure the power spectra are normalized by the no-BAO power spectrum, which is calculated neglecting magnification bias and the effects of baryons. We see immediately several features: the effect of magnification is largest at the low $\ell$ values (small $\ell$ correspond to large angular scales), the magnitude of the magnification correction increases with redshift, and the effect at high $\ell$ near the baryon wiggles is mostly (though not completely) to boost the amplitude rather than change the shape of the spectrum.
Magnification clearly changes the shape of the angular power spectrum. The most significant changes are at $\ell{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$<$}}}100$ where the peak in the power spectrum resides. The peak in the power spectrum is an important feature because its location is related to the Hubble scale when the universe transitions from radiation domination to matter domination. This feature in the power spectrum can be used as a standard ruler [@Cooray06].
We identify the location $\ell_{peak}$ and amplitude of the peak in $\Cgg$ and $\Cnn$. The shift in $\ell_{peak}$, and the fractional change in amplitude, $(\Cnn-\Cgg)/\Cgg$ (where $\Cnn$ is evaluated at the $\ell_{peak}$ for $\Cnn$ and $\Cgg$ is evaluated at the $\ell_{peak}$ found for $\Cgg$) when magnification is included are shown in Table \[lpeaktable\]. The change in $\ell_{peak}$ and the change in amplitude increase both with redshift and with increasing width of the selection function $\sigma$. The value of $\ell_{peak}$ is also dependent on the width of the selection function, with $\ell_{peak}$ taking slightly larger values for narrower selection functions. One additional consequence of magnification bias is to change the redshift and scale at which non-linear evolution of the power spectrum becomes important. This is because even for sources at high redshift where non-linearity is small, $\Cmm$ will depend on structure at low redshift where non-linearity is important. Indeed, even at $z_0=3.5$ the difference between $\Cnn$ calculated with the non-linear power spectrum and $\Cnn$ calculated with the linear power spectrum is ${\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}10\%$ by $\ell \sim 1800, 1400,1000$ for $\sigma =0.07$, $0.15$ and $0.30$ respectively. This is to be compared with $\Cgg$, the no magnification case, for which the difference between $\Cgg(z_0=3.5)$ calculated with and without non-linear evolution does not approach $10\%$ until $\ell \sim 2600$ for each value of $\sigma$.
------- --------------- --------------------- ----------------------------------------------- --------------- ---------------------- ----------------------------------------------- --------------- ---------------------- -----------------------------------------------
$z_0$ $\ell_{peak}$ $\Delta\ell_{peak}$ $\Delta C_{\ell_{peak}}/C^{gg}_{\ell_{peak}}$ $\ell_{peak}$ $\Delta \ell_{peak}$ $\Delta C_{\ell_{peak}}/C^{gg}_{\ell_{peak}}$ $\ell_{peak}$ $\Delta \ell_{peak}$ $\Delta C_{\ell_{peak}}/C^{gg}_{\ell_{peak}}$
$0.5$ $19$ $0$ $0.02$ $16$ $0$ $0.04$ $-$ $-$ $-$
$1.0$ $34$ $0$ $0.01$ $33$ $0$ $0.04$ $30$ $-2$ $0.14$
$1.5$ $46$ $-1$ $0.02$ $45$ $-1$ $0.05$ $44$ $-3$ $0.14$
$2.0$ $55$ $-1$ $0.03$ $54$ $-3$ $0.07$ $54$ $-5$ $0.17$
$2.5$ $62$ $-2 $ $0.04$ $61$ $-5$ $0.10$ $61$ $-9$ $0.22$
$3.0$ $67$ $-3$ $0.06$ $67$ $-8$ $0.13$ $67$ $-15$ $0.29$
$3.5$ $72$ $-5$ $0.08$ $72$ $-11$ $0.17$ $72$ $-21$ $0.38$
------- --------------- --------------------- ----------------------------------------------- --------------- ---------------------- ----------------------------------------------- --------------- ---------------------- -----------------------------------------------
the angular correlation function and the baryon bump {#BAOsec}
====================================================
We now turn our attention to the real space angular correlation function $w_{nn}(\theta,z_0)$ (Equation \[wtheta1\]). In multipole space, the baryon oscillations appear as a series of peaks in the power spectrum (e.g. Figure \[CnnCnoBfig\]). In real space the signature of baryon oscillations is a single bump in the correlation function. The location of the bump in $w_{nn}(\theta,z_0)$ is determined by the comoving sound horizon at recombination and the distance to $z_0$. The sound horizon at recombination is measured quite precisely from the cosmic microwave background anisotropy, thus a measurement of $\theta_{BAO}$ at $z_0$ can give a measure of the comoving distance to $z_0$.
Before we address the lensing corrections to the baryon bump, we will first discuss a few issues that arise (whether or not lensing is included) when using the angular correlation function to measure the baryon oscillation scale. The angular correlation function is in some sense averaging the matter power spectrum across different redshifts. Thus, if too broad of a selection function is used the baryon feature will be washed out. For a fixed width in redshift, the washing out is more severe at low redshifts because $\Delta \chi \sim \Delta z /H(z)$ and $H(z)$ decreases with decreasing $z$. Additionally, the baryon bump is a subtle feature in the angular correlation function. It is customary to multiply the angular correlation function by $\theta^2$ to make the baryon bump easier to identify and characterize. Indeed for the cases we have considered this is necessary in order for there to be a local maximum at the acoustic scale at all. Figure \[wggNear\] illustrates this point: while there is a baryon feature, there is no local maximum in $w(\theta)$. In Figure \[th2zoomfig\] we show $\theta^2w(\theta)$ in the region near the baryon bump, in this case the baryon feature is quite visible for $\sigma=0.07$ but is very hard to identify for $\sigma=0.30$, fortunately the baryon bump becomes more prominent with increasing redshift.
Since the baryon bump is broadened at low redshifts we limit our analysis to $z_0 \ge 1.0$ for $\sigma=0.07$ and $0.15$ and $z_0\ge 2.0$ for $\sigma=0.30$. We have experimented with multiplying by $\theta^3$: in this case even for the broadest selection function $\sigma=0.30$ the baryon bump is visible at all redshifts we consider. Interestingly, the magnitude of the shift in the baryon oscillation scale due to magnification bias is sensitive to whether one looks at $\theta^2 w(\theta)$ or $\theta^3 w(\theta)$. The shift is much larger in the latter case, suggesting that the importance of magnification bias for baryon oscillation measurements depends on precisely how the baryon bump is fitted. In this paper, we focus on the effects on $\theta^2 w(\theta)$.
We calculate the two-point function $w_{nn}(\theta,z_0)$ by performing the sum in Equation \[wthetaeq\]. In Figure \[wggNear\] we plot $w_{nn}(\theta)$ and $w_{gg}(\theta)$ with redshift bins centered at $z_0=1.5$ and $z_0=3.0$ for a small angular range near the baryon bump. One can see from Figure \[wggNear\] that, as expected, the angular correlation function is changed by magnification bias. The lensing correction to the angular correlation function is scale dependent. To address how the baryon bump is changed by lensing magnification we consider the location of the baryon peak $\theta_{BAO}$, the height at the peak $\theta^2w(\theta_{BAO})$, and the peak width. The peak width is defined to be $\left[\left(\frac{\partial^2}{d\theta^2} \theta^2w \right)/(\theta^2w)\right]^{-1/2}$ evaluated at $\theta_{BAO}$.
Figure \[th2DiffAll\] summarizes the fractional changes in the peak location, height and width as a function of $(5s-2)/b$ for selection functions with $\sigma=0.07$, $0.15$ and $0.30$, and for several redshifts. The magnitude of the changes to the peak increase with redshift for all $\sigma$. The changes in peak location and width are largest at the highest redshifts ($z_0=3-3.5$) but the exact redshift dependence varies depending on the selection function. This is in part because broader selection functions grant a larger $w_{g\mu}(\theta)$ term which dominates at low redshifts. In the region near the baryon bump, this term is more strongly scale dependent than $w_{\mu\mu}(\theta)$. Additionally, the baryon peak is sharper and the amplitude of $w_{gg}(\theta)$ is larger for narrower selection functions (e.g. Figure \[th2zoomfig\]), both of these factors make it more difficult for magnification to effect the peak location. Magnification shifts $\theta_{BAO}$ by $<0.5\%$ for all $\sigma$ and $z_0$ values we considered. The changes to the width and amplitude are much larger. We should point out that as the width of the selection function is increased the baryon bump is also broadened making the peak harder to identify – this is true regardless of whether magnification bias is present. This is illustrated in Figure \[th2zoomfig\], where we show $\theta^2w(\theta)$ for a narrow angular range for each $\sigma$. One can see that for $\sigma=0.30$ there is no maximum at this redshift.
We have shown that the change in location of the baryon bump found in our calculated $w_{nn}(\theta)$ and $w_{gg}(\theta)$ is likely to be small. However, an actual measurement of $\theta_{BAO}$ will involve fitting a curve to data points with error bars. From this perspective, magnification bias – which adds a scale and source population dependent correction to $w_{gg}(\theta)$ – may be of more concern. This fact is illustrated in the lower panel of Figure \[wggNear\]. Here we show the difference between the two-point functions with and without magnification ($w_{nn}-w_{gg}$). Although this difference is slowly varying in the region around the baryon peak, the value of the difference is not constant. In the regime very near the peak the offset between $w_{nn}(\theta)$ and $w_{gg}(\theta)$ can be approximated as a straight line. The precise slope of this line will depend on the population of galaxies via $b$ and $s$. Interestingly, the magnification correction to $\theta^2w(\theta)$ is closer to a constant than it is for $w(\theta)$. When the baryon bump in $\theta^3w(\theta)$ is considered we find the magnification correction is more strongly scale dependent than for $\theta^2w(\theta)$. Thus the changes to the baryon peak location in $\theta^3w(\theta)$ are significantly larger (by a factor of $5-6$). We emphasize then, that while Figure \[th2DiffAll\] gives an indication of the effect of magnification bias on the baryon bump, the actual effect of magnification on measurements of the acoustic scale will depend on the method by which the peak location is fitted from the data. When using the BAO peak as a standard ruler this scale and population dependent bias should be taken into account.
discussion
==========
The angular two-point correlation function and the angular power spectrum are important cosmological statistics, but a complete understanding of systematics is necessary to use these to derive precise constraints on cosmological parameters. Magnification bias, a gravitational lensing correction to the observed number density of sources, has been known to change the correlation function at high redshifts [@Villumsen1995; @VFC97; @MJV98; @MJ98]. We have examined and quantified how magnification bias changes the shape of the angular correlation function and power spectrum. The effect of magnification is to adjust both the scale-dependence and amplitude of the correlation function. We have shown that the scale-dependence changes the location of the matter-radiation equality peak in the angular power spectrum, and can potentially lead to a bias in determining the location of the baryon bump in the angular correlation function (see Figures \[CnnCnoBfig\], \[wggNear\], \[th2DiffAll\] and Table \[lpeaktable\]). Magnification bias becomes important at high redshifts ($z{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}1.5$) with the most drastic scale dependent changes occurring at low multipoles $\ell {\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$<$}}}100$. Precisely how, if ignored, magnification would bias measurements of the matter-radiation equality scale or the BAO size will likely depend on how these quantities are extracted from the correlation function as well as the population of galaxies or quasars used for the measurement. Finally, the magnification terms are sensitive to non-linear evolution of structure at low redshifts even for sources at high redshift. Consequently, the presence of magnification bias changes the redshift and scale for which non-linear evolution of the power spectrum is visible.
Magnification bias is a source-population dependent effect, that is the magnitude (and even the sign) of the correction depend on the population of galaxies or quasars being observed. On the one hand, one may be able to select a population of sources with number count slope $s=2/5$, in which case magnification bias vanishes. However, such a selection may reduce the number of sources available for analysis so this is not necessarily the best option. Current measurements of the galaxy angular correlation are at redshifts $z<1$, so the correction from magnification to these observations is expected to be negligible. Projections for measurements of galaxy clustering from future high redshift galaxy surveys will need to address the effect of magnification bias.
Finally, let us reiterate that this paper focuses exclusively on the angular correlation. The effect of magnification bias on the 3D correlation has some surprising new features. For instance, in certain situations, magnification bias can be important even for low redshift measurements. This was originally noted by [@Matsubara] and is further analyzed in two separate papers [@3Dpaper1; @3Dpaper2].
LH thanks Ming-Chung Chu and the Institute of Theoretical Physics at the Chinese University of Hong Kong for hospitality where part of this work was done. Research for this work is supported by the DOE, grant DE-FG02-92-ER40699, and the Initiatives in Science and Engineering Program at Columbia University. EG acknowledges support from Spanish Ministerio de Ciencia y Tecnologia (MEC), project AYA2006-06341 with EC-FEDER funding, and research project 2005SGR00728 from Generalitat de Catalunya.
[99]{} P. J. E. Peebles, 185, 413 (1973); M. G. Hauser, P. J. E. Peebles 185, 757 (1973); E. J. Groth, P. J. E. Peebles 217, 385 (1977); M. Tegmark 79, 3806 (1997); M. Tegmark, A. J. S. Hamilton, M. A. Strauss, M. S. Vogeley, A. S. Szalay 499, 555 (1998); Y. Wang, D. N. Spergel, M. A. Strauss 510, 20 (1999); W. Hu, D. J. Eisenstein, M. Tegmark, M. White 59, 023512 (1998); D. J. Eisenstein, W. Hu, M. Tegmark, 518, 2 (1999) A. Cooray, 651, L77 (2006)
J. Silk, 151, 459 (1968) P. J. E. Peebles, J. T. Yu, 162, 815 (1970) R. A. Sunyeav, Ya. B. Zel’dovich, Astrophysics and Space Science, 7, 3S (1970) J. R. Bond, G. Efstathiou, 285, 45 (1984) J. A. Holtzman, 71, 1 (1989) Eisenstein, D. J., Hu, W., 496, 605 (1998) A. Meiksin, M. White, J. A. Peacock [Mon. Not. R. Astron. Soc. ]{}304, 851 (1999)
D. Spergel [[*et al.*]{}]{}[Astrophys. J. Suppl.]{}148, 175 (2003) D. Spergel [[*et al.*]{}]{}, , in press
D. J. Eisenstein, W. Hu, M. Tegmark 504, 57 (1998); H.- J. Seo, D. J. Eisenstein 598, 720 (2003); E. V. Linder, 68, 083504 (2003); T. Matsubara, A. S. Szalay, 90, 021302 (2003); C. Blake, K. Glazebrook 594, 655 (2003); W. Hu, Z. Haiman, 68, 063004 (2003); T. Matsubara 615, 573 (2004); H.- J. Seo, D. J. Eisenstein 633, 575 (2005); K. Glazebrook, C. Blake 631, 1 (2005); M. White, Astropart. Phys., 24, 334 (2005); C. Blake, S. Bridle, [Mon. Not. R. Astron. Soc. ]{}363, 1329 (2005); R. Angulo, C.M. Baugh, C.S. Frenk, R. G. Bower, A. Jenkins, S. L. Morris [Mon. Not. R. Astron. Soc. ]{}262, L25 (2005); C. Blake, D. Parkinson, B. Basset, K. Glazebrook, M. Kunz, R. C. Nichol [Mon. Not. R. Astron. Soc. ]{}365, 255 (2006); D. Dolney, B. Jain, M. Takada [Mon. Not. R. Astron. Soc. ]{}366, 884 (2006); D. Dolney, B. Jain, M. Takada [Mon. Not. R. Astron. Soc. ]{}366, 884 (2006); D. Jeong, E. Komatsu, 651, 619 (2006); E. Huff, A. E. Schulz, M. White, D. Schlegel, M. S. Warren, Astropart. Phys. 26, 351 (2007); R. Angulo, C. M. Baugh, C. S. Frenk, C. G. Lacey [Mon. Not. R. Astron. Soc. ]{}in press (2007) S. Cole, [[*et al.*]{}]{}, [Mon. Not. R. Astron. Soc. ]{}362, 505 (2005) D. J. Eisenstein, [[*et al.*]{}]{}633,560 (2005) G. Huetsi, [Astron. & Astrophys. ]{}submitted (2005) astro-ph$/0507678$ N. Padmanabhan, [[*et al.*]{}]{}[Mon. Not. R. Astron. Soc. ]{}submitted (2006) astro-ph$/0605302$ W. Percival, [[*et al.*]{}]{}657, 51 (2007) J. E. Gunn, 147 61 (1967) E. L. Turner, J. P. Ostriker, J. R. Gott, 284, 1 (1984); R. L. Webster, P. C. Hewett, M. E. Harding, G. A. Wegner, Nature 336, 358 (1988); W. Fugmann, [Astron. & Astrophys. ]{}204, 73 (1988); R. Narayan, [Astrophys. J. Lett. ]{}339, 53 (1989); P. Schneider, [Astron. & Astrophys. ]{}221, 221 (1989). T. J. Broadhurst, A. N. Taylor, J. A. Peacock, 438, 49 (1996)
J. V. Villumsen, unpublished preprint (1995) astro-ph$/9512001$ J. Villumsen, W. Freudling, L. N. da Costa, 481, 578 (1997) R. Moessner, B. Jain, J. Villumsen, [Mon. Not. R. Astron. Soc. ]{}294, 291 (1998) R. Moessner, B. Jain,[Mon. Not. R. Astron. Soc. ]{}294, L18 (1998) E. Gaztañaga, 589, 82 (2003) R. Scranton, [[*et al.*]{}]{}, 633, 589 (2005) A. D. Myers, [[*et al.*]{}]{}[Mon. Not. R. Astron. Soc. ]{}342, 467 (2003) B. Menard, M. Bartelmann, [Astron. & Astrophys. ]{}386, 784 (2002). B. Jain, R. Scranton, R. V. Sheth, [Mon. Not. R. Astron. Soc. ]{}345, 62 (2003) T. Matsubara, 537, L77 (2000) L. Hui, E. Gaztanaga, M. LoVerde, submitted to (2007) arxiv$/0706.1071$ L. Hui, E. Gaztanaga, M. LoVerde, submitted to (2007b) arxiv$/0710.419$ A. Vallinotto, S. Dodelson, C. Schimd, J. P. Uzan (2007) astro-ph$/0702606$ For a review, see S. Dodelson, Modern Cosmology, Academic Press (2003). R. E. Smith, R. Scoccimarro, R. K. Sheth, 75, 063512 (2007) D. N. Limber, 119, 655 (1954) J. M. Bardeen, J. R. Bond, N. Kaiser, A. S. Szalay, 304, 15B (1986) N. Sugiyama, 100, 281 (1995) R. E. Smith, [[*et al.*]{}]{}[Mon. Not. R. Astron. Soc. ]{}341 1311 (2003) M. LoVerde, L. Hui, E. Gaztañaga, 75, 043519 (2007)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using a suite of cosmology simulations of a sample of $\textgreater 120$ galaxy clusters with $\log(M_{DM, vir}) \le 14.5$. We compare clusters that form in purely dark matter run and their counterparts in hydro runs and investigate 4 independent parameters, that are normally used to classify dynamical state. We find that the virial ratio $\eta$ in hydro-dynamical runs is $\sim 10$ per cent lower than in the DM run, and there is no clear separation between the relaxed and unrelaxed clusters for any parameter. Further, using the velocity dispersion deviation parameter $\zeta$, which is defined as the ratio between cluster velocity dispersion $\sigma$ and the theoretical prediction $\sigma_t = \sqrt{G M_{total}/R}$, we find that there is a linear correlation between the virial ratio $\eta$ and this $\zeta$ parameter. We propose to use this $\zeta$ parameter, which can be easily derived from observed galaxy clusters, as a substitute of the $\eta$ parameter to quantify the cluster dynamical state.'
author:
- '\'
bibliography:
- 'bibliography.bib'
date: 'Accepted XXX. Received YYY; in original form ZZZ'
title: ' On the dynamical state of galaxy clusters: insights from cosmological simulations II.'
---
\[firstpage\]
galaxies: clusters: general – galaxies: kinematics and dynamics – galaxies: halos – galaxies: evolution – cosmology: theory
Introduction {#i}
============
Currently favored models of cosmological structure formation are hierarchical – lower mass systems merge progressively to form more massive structures, with galaxy clusters representing the final state of this process. The dynamical process, driven by gravity, determines the final properties of the dark matter halo, as well as the baryonic contents in it – galaxies, intra cluster medium (ICM), etc. However, even at the final state of hierarchical structure formation, the galaxy clusters are not always in dynamic equilibrium. In observations, galaxy cluster systems can be roughly separated into relaxed and unrelaxed; the ICM in relaxed clusters is normally in hydrostatic equilibrium, while dynamically unrelaxed clusters are undergoing, or have undergone, a merger, which leaves the ICM turbulent [see @Wen2013 and references therein]. In simulations, there are a vague of ways dynamical state can be evaluated.
Using dark-matter-only simulations, [@Jing2000] found that about $30$ per cent of the simulated dark matter halos can not be fitted by the NFW profile [@NFW], and these halos that showed larger deviations from the NFW profile exhibited significant internal substructures. Using the integral virial ratio parameter $2T/|W|+1$, here T is the kinetic energy, W is the potential energy, [@Bett2007] suggested $2T/W +1 \textless 1.5$ to select halos in quasi-equilibrium states [see also @Klypin2016]. [@Neto2007] expanded the criteria by including substructure mass fraction and centre-of-mass offset. However, they adopted a narrower limit for their virial ratio $2T/|W| < 1.35$ [see also @Ludlow2012]. [@Shaw2006; @Poole2006; @Davis2011] modified the virial ratio by taking the surface pressure energy $E_s$ into account. This is because halos are not isolated in cosmology simulations, and infalling materials alter $2T/W$. Besides the surface pressure energy, [@Davis2011] also considered the potential energy from particles outside of halos – $W_{ext}$ for the virial ratio. However, they found that $W_{ext}$ is negligible. Nevertheless, different limits are used to calculate the virial ratio: $(2T-E_s)/W +1 > -0.2$ for [@Shaw2006]; $|1+2T/(E_s + W)| < 0.02$ for [@Poole2006]; While [@Knebe2008] suggested $-0.15 \leq (2T-E_s)/W +1
\leq 0.15$ (with a mass dependence at $z$ = 1) to select out relaxed halos. [@Power2012] studied the relation between centre-of-mass offset and equilibrium state. Instead of using virial ratio, they suggested a centre of mass offset value of $0.04$ to select relaxed halos.
All of these studies were based on dark matter only simulations. However, as numerical simulations with sophisticated sub-grid baryon models have become more mature and successful in producing observed-like galaxies, there has been great interest in studying the baryonic effects on galaxy cluster properties [e.g. @Schaller2015b; @Cui2016]; on power spectrum [e.g. @Daalen2011]; on halo mass as well as halo mass function [e.g. @Cui2012a; @Cui2014b; @Velliscig2014]; and on substructure shapes and alignments [e.g. @Knebe2010; @Velliscig2015]. It is timely and interesting to study and how baryons affect the dynamical state of galaxy clusters. Baryons, especially gas, are subject to other forces in addition to gravity to dark matter, which will lead changes on T and W.
In this paper, we study the dynamical state of galaxy clusters with a volume- and mass-complete sample from a series of cosmological simulations with three different baryon models, which we have presented in @Cui2015 [hereafter Paper I]. We investigate how different measures of dynamical state change between dark-matter-only and hydro-dynamical runs. In the following sections, we briefly describe these hydro-simulations with different baryon models [see also @Cui2012a; @Cui2014b] and the statistical sample of clusters (§ \[simulation\]), and present our dynamical state classification methods (§ \[method\]). In section \[results\] we present our results. Finally, we summarise our conclusions in § \[concl\], and comment on the implications for interpretation of observations of galaxy clusters.
Simulated Galaxy Cluster Catalogue {#simulation}
==================================
These simulations use a flat $\Lambda$CDM cosmology, with cosmological parameters of $\Omega_{\rm m} = 0.24$ for the matter density parameter, $\Omega_{\rm b} = 0.0413$ for the baryon contribution, $\sigma_8=0.8$ for the power spectrum normalization, $n_{\rm s} = 0.96$ for the primordial spectral index, and $h =0.73$ for the Hubble parameter in units of $100 {~h ~{\rm km}~s^{-1}~{\rm Mpc}^{-1}}$. They used the same realization of the initial matter power spectrum, and were run with the TreePM-SPH code [GADGET-3]{}, an improved version of the public [GADGET-2]{} code [@Gadget2]. Three simulations were run, we refer to the dark-matter-only simulation as the DM run; the hydrodynamical simulations including radiative cooling, star formation and kinetic feedback from supernovae: in one case we ignore feedback from AGN (which is referred as the CSF run), while in the other we include it (which is referred as the AGN run). The DM run has two families of dark matter particles: the one with larger particle mass shares the same ID as the dark matter particles in the CSF and AGN runs; while the one with smaller particle mass has equal mass as the gas particles in the CSF and AGN runs at the initial condition of $z$ = 49. With this particular setup, we can make an explicated investigation on the baryon effect.
Halos are identified using the Python spherIcAl Overdensity (SO) algorithm [<span style="font-variant:small-caps;">P</span>IAO]{} [^1] [@Cui2014b], and are selected from the DM run with a mass cut. We reselect 123 halos, which have the virial mass of $\log_{10} (M_{vir}) > 14.5
{~h^{-1}\>{\rm M_\odot}}$. We use [@Bryan1998] t oestimate $\Delta_{vir}$ and compute $M_{vir}$. Counter parts SO halos in AGN and CSF runs are identified by cross-matching dark matter components using their unique particle IDs [also see @Cui2014b more details].
Methods {#method}
=======
#### Virial Ratio {#virial-ratio .unnumbered}
The exact virial theorem for a self-gravitating system is $$\frac{1}{2} \frac{d^2 I}{dt^2} = 2T + W - E_s,
\label{eq:vt}$$ where I is the moment of inertia. The proper way of calculating the equation \[eq:vt\] is by using the time averaged values of these quantities [see discussion in @Poole2006]. However, due to the limited outputs of the simulation, we only calculate these quantities at $z$ = 0.
Total kinetic energy T is calculated differently for collisionless (dark matter and star) particles and collisional (gas) particles. After removing the halo motion, which is given by the mass-weighted mean velocity from particles within $30 {\rm kpc}$, and the Hubble flow, T is simply $\frac{1}{2}m_i v_i^2$, where $i$ is for all collisionless particles; We use the gas thermal energy U for its kinetic energy. Total potential energy W is directly calculated by using all particles inside halos without any approximation. $E_s$ is the energy from surface pressure P at the halo boundary. As described in [@Chandrasekhar1961], $E_s$ is $$E_s = \int P(r){\mathbfit{r}} \cdot {\mathbfit{dS}}.$$ Assuming the ideal gas law, P for collisionless particles [see @Shaw2006 for more details] can be written as $$P_c = \frac{\sum_i m_i v^2_i}{3 V},$$ this summation is over all particles with mass $m_i$, velocity $v_i$ inside volume V, while P for gas particles [see @Poole2006 for more details] is $$P_g = \frac{\sum_i N_i k_B T_i}{V},$$ here $N_i$, $T_i$ are the gas number and temperature respectively, $k_B$ is the Boltzmann constant.
We follow @Shaw2006 to calculate P: first, we rank order all particles by their radius and select the outermost 20 per cent; then we label the radius of the innermost particle in this shell as $R_{0.8}$, the outermost as $R_{vir}$, and the median as $R_{0.9}$. V is the volume occupied by the outermost 20 per cent particles, $V = \frac{4 \upi}{3} (R^3_{vir} - R^3_{0.8}).$ The surface pressure energy from collisionless component can be approximated by $$E_{s,c} \approx 4 \upi R^3_{0.9} P_c = \frac{R^3_{0.9}}{R^3_{vir} -
R^3_{0.8}} \sum_i m_i v^2_i.$$
For gaseous particles, the gas number density n can be expressed in terms of the mean molecular weight: $\mu = {\rho}/(n m_p)$, where $m_p$ is the mass of a proton, $\rho$ is gas density. Following @Mo2010, we assume the elements heavier than helium have a mass number $M_i \approx 2(Q_i +1)$, here $Q_i +1$ is the charge number of a fully ionized atom. If we define the total mass as $X_i = 1$, where $X_i$ is the mass abundance of element i, then we have $\mu = {4}/(6 X_H + X_{He} + 2)$. Normally we assume the metallicity $Z = 1 - X_H - X_{He}$ is very small, and the mass fraction for hydrogen is around 0.76. Thus, we can have $\mu \approx 0.588$ and the gas number $$N = nV = \frac{\rho V}{\mu m_p} = \frac{m}{\mu m_p}.$$ Finally, we can calculate the surface pressure energy from the gas component as, $$E_{s, g} \approx 4 \upi R^3_{0.9} P_g = \frac{R^3_{0.9}}{R^3_{vir} - R^3_{0.8}} \frac{3
k_B}{\mu m_p} \sum_i m_i T_i,$$ where summation is over all the gas particles lying between $R_{0.8}$ and $R_{vir}$. $E_s$ is contributed by both collisionless and gas particles.
If the system is in a steady state and dynamical equilibrium, equation (\[eq:vt\]) will reduce to $2T + W - E_s = 0$, which can be rewritten as $(2T - E_s)/|W| = 1$. Therefore, we define $\eta = (2T - E_s)/|W|$, and expect $\eta \rightarrow 1$ for dynamically relaxed galaxy clusters.
#### Total subhalo mass fraction {#total-subhalo-mass-fraction .unnumbered}
Subhalos are identified by <span style="font-variant:small-caps;">SubFind</span> [@Springel2001; @Dolag2009; @Cui2014a]. For all the galaxy clusters identified by <span style="font-variant:small-caps;">PIAO</span>, we run <span style="font-variant:small-caps;">SubFind</span> on them one by one. The smallest subhalo has at least 32 particles. Subhalos with only gas particles are not taken into account [@Dolag2009]. The subhalo mass fraction $f_s$ does not include the most massive substructure as this is simply the bound component of the main halo.
#### Centre-of-mass offset {#centre-of-mass-offset .unnumbered}
For all the particles within the virial radius $R_{vir}$, we compute the centre of mass as $${\mathbfit{R}}_{cm} = \frac{1}{M} \sum_{i=1}^n m_i r_i,$$ here $m_i$ is the $i^{th}$ particle mass, $r_i$ is its position, M is the halo virial mass, and n is the total particle number within $R_{vir}$. The centre offset is defined as $\Delta_r = |{\mathbfit{R}}_{cm} - {\mathbfit{R}}_c|/{R_{vir}}$. We note here that the density peak position is used as cluster centre ${\mathbfit{R}}_{c}$ .
#### Velocity dispersion deviation {#velocity-dispersion-deviation .unnumbered}
The velocity dispersion $\sigma$ is always an important quantity for cluster dynamics. It is often used to predict the cluster’s dynamical mass through the virial theorem: $$\frac{1}{2} M_{total} \sigma^2 \propto \frac{G M_{total}^2}{R},$$ where G is the universal gravitational constant, and $M_{total}$ and R are the cluster mass and radius. Thus, one can easily get the predicted dynamical mass through $M_{total} \propto (R \sigma^2)/{G}$. However, this is based on the assumption that the cluster is in dynamical equilibrium, which is normally not true. Therefore, we define a parameter to quantify the deviation to the dynamical equilibrium: $\zeta = \frac{\sigma}{\sigma_t}$, here $\sigma_t = {\sqrt{(G M_{total})/{R}}}$. Note that the velocity dispersion deviation $\zeta$ can be different from unity even for perfectly relaxed clusters, because its exact value also depends on the density profile.
Results
=======
Radial profiles {#4.1}
---------------
{width="\textwidth"}
We first show the radial profiles of these four parameters: $\eta$, $\zeta$, $\Delta_r$ and $f_s$, in Fig. \[fig:rp\] from upper to lower panels respectively. We select three clusters with different $\eta$ values order as in plot (>1, $\approx$1 and <1 from the DM run at $R_{vir}$). The results from all particles are shown with solid lines and filled symbols, while the symbols with dotted lines from only dark matter particles (DMPs). Different colours and symbols styles represent different versions of simulations, which are indicated on the legend of the top middle panel.
It is worth to note at here again that the DM run has two family dark matter particles: the first (more massive) one shares the same mass and ID to the dark matter particle in the two hydro-dynamical runs; while the second family only has its mass the same as the initial gas particles in the two hydro-dynamical runs. We have verified that this separation in our DM run does not show signs of mass segregation. This particular set in the DM run allows us to make equal comparisons to the two hydro-dynamical runs. If it is not particularly noted, the dark matter particle (DMP) from the DM run refers to the first family (heavier) particle in the lower part.
To calculate the values of these four parameters at each radius $R_i$, we simply use the corresponding particles within that radius. However, only particles inside the spherical shell $R_{i, 0.8} \le R \le R_i$ are used to calculate the surface pressure energy $E_{s, i}$.
1. At inner region, the values of $\eta$ are all larger than 1 for all three galaxy clusters, which means that the values of $2T - E_s$ are always larger than their potential energy $|W|$. At outer radius (mostly $R \gtrsim 0.6 R_{vir}$), $\eta$ becomes more flat for all three clusters and three runs. $\eta$ from both the CSF run and the AGN run is normally smaller than from the DM run over all radii. However, there is a better agreement between these three runs, when only DMP is taken into account, especially at outer regions. It means that DMPs are less affected by baryons.
2. $\zeta$ normally has a value smaller than 1, and shows a declining trend from inner to outer radii, which is basically the same saw in $\eta$. In agreement with $\eta$, galaxy clusters simulated with baryon models also have smaller $\zeta$ values than the DM run. It not surprising that the result from DMP is also similar to $\eta$. However, there is a slightly larger disagreement between the three runs, especially for the two with $\eta \leq 1$.
3. The radial profile for the centre-of-mass offset $\Delta_r$ shows large difference between the three galaxy clusters and the three simulation versions. As DMPs contribute the largest mass for galaxy clusters, it is not surprising to see that the dotted lines basically follow the solid lines. It seems to have less correlation between $\Delta$ and $\eta$, seeing from these radial profiles.
4. It is not surprising that $f_s$ from the CSF run normally has a larger value than the other two runs. This is caused by the over-cooling problem, which affects not only central galaxies, but also satellite galaxies. Similar to $\Delta_r$, there is very little differences between the total (solid lines) and DMP (dotted lines) for the $f_s$ profile. $f_s$ for all three clusters show a clear increasing trend from inner to outer regions. This is simply because the closer to the centre, the higher possibility that substructures are destroyed. This trend is anti-correlated with the radial profile from $\eta$.
Fro these three example clusters, $\eta$ shows a decreasing trend from inner to outer radii, which means that galaxy clusters can be highly un-virialized at their centres than the outer region. In agreement with [@Shaw2006], $\eta$ at outer radius ($R \gtrsim 0.6 R_{vir}$) becomes more flat, which means that $\eta$ is primely determined by materials inside $0.6 R_{vir}$. The in-falling materials at outer region has less effect on $\eta$. It is interesting to see that baryons give a systematic decreasing effects on $\eta$ over the whole radii. However, the $\eta$ from DMP seems to be less affected. Because gravity is the only interaction between dark matter and gas, and gas only occupy a small mass fraction of clusters with a smoother distribution, it not surprising to see this results. Because larger $\sigma$ at fixed radius corresponds to larger $T$, it is also not surprising to see that $\zeta$ basically follows the trend of $\eta$.
There is no clear trend for the profile of $\Delta_r$. This is because the centre of mass is largely relying on the mass distributions, especially the substructure position. However, $f_s$ shows an increasing trend as radius increases.
The baryonic effects
--------------------
{width="\textwidth"}
We further investigate baryon effects on the four parameters in Fig. \[fig:be\]. To explicitly show and understand this effect, changes of these four parameters from the DM run to the two hydro-dynamical runs are separated into two rows: the upper row shows the results from all types of particles; while the lower row is from DMPs. These results are shown as a function of their halo masses $M_{vir}$. As shown in the legends of right panels, the different color and style symbols indicate different simulations; while the different color and style lines are the mean of data points. The upper row shows the results from all particles; while the lower row is from DMPs.
Through these comparisons, we find:
1. The upper panel from the first column shows the ratio of $\eta$, which is calculated with all particles. It is clear that $\eta$ from both the CSF run and the AGN run is about 10 per cent lower than the one from the DM run. Nevertheless, there is very small difference between the two hydro-dynamical runs evident from their mean values. The ratio of $\eta$ shows almost no dependence on cluster masses.
The lower panel shows the results from DMPs. There is almost no difference between the two hydro-dynamical runs and the DM run, which is consistent with the finding from Fig. \[fig:rp\]. Although the red dashed line (the CSF run) is on top of the green dotted line (the AGN run), there is very little difference between the CSF run and the AGN run without any dependence on cluster masses.
2. We show the ratio of $\zeta$ in the second column of Fig. \[fig:be\]. Again, the mean of $\zeta$ from both the CSF run and the AGN run is slightly lower ($\sim 0 - 10$ per cent) than from the DM run. However, $\zeta$ from the AGN run is closer to the DM results than from the CSF run. Again, this ratio shows almost no dependence on cluster masses.
Similar to the $\eta$ results from DMPs, the mean ratio of $\zeta$ from both the CSF run and the AGN run to the DM run is around 1. The difference between the CSF run and the AGN run is in consistent the result from the upper panel: red dashed line (the CSF run) is always on top of green dotted line (the AGN run).
3. The changes of $\Delta_r$ are shown in log space in the third column. Due to its sensitivity to the position of substructures, which seems to be easily affected by baryons, there is a large scatter for these data points. However, the mean ratio of $\Delta_r$ is around 1 for the CSF run; while this is also true for the AGN run at smaller mass, but a slightly smaller $\Delta_r$ than the DM run is shown at larger mass.
Since dark matter normally occupies more than 80 per cent of total cluster mass, it is not surprising that the lower panel, which shows the result from DMPs, gives very similar results as the upper panel.
4. We show how $f_s$ changes in the last column. $f_s$ is clear larger from the CSF run than the DM run: it increases about 40 per cent at smaller cluster mass; it is still about 20 per cent higher at larger cluster mass. $f_s$ from the AGN run is 20 per cent lower than from the CSF run; 20 per cent higher than the DM run at smaller cluster masses, and almost no difference between the two runs at $M \gtrsim 10^{14.8} {~h^{-1}\>{\rm M_\odot}}$.
For the changes of $f_s$ from DMPs in the lower panel, we see very similar result for the AGN run as in the upper panel. Nevertheless, $f_s$ from the CSF run is around 10 per cent closer to the DM run than its result in the upper panel over all cluster mass range.
The bottom-left panel in Fig. \[fig:be\] shows that $\eta$ is consistent for DMPs between the three runs, which implies that baryons have little effect on both the kinetic and potential energy of dark matter if $E_s$ is ignored. We find a fixed value of $\eta_{DMP}/\eta_{All}$ for all clusters. This value from the two hydro-dynamical runs is $\sim 10$ per cent higher than from the DM run. This means that baryons have a systematic change on $\eta$. This is consistent with the finding from the top-left panel of Fig. \[fig:be\]. We study this below.
Similar to the findings from radial profiles in Fig. \[fig:rp\], $\zeta$ also shows the closest correlation with $\eta$ for the change caused by baryons. Although the ratio between $\zeta_{AGN}$ and $\zeta_{DM}$ is very similar to the ratio of $\eta$, $\zeta_{CSF}$ is much closer to $\zeta_{DM}$ than $\eta_{CSF}$ to $\eta_{DM}$. This shows that the over-cooling problem in the CSF run has more effect on $\zeta$ than $\eta$.
In agreement with Fig. \[fig:rp\], baryons have a large influence on $\Delta_r$. It is not surprising that $\Delta_{r, DMP}$ follows $\Delta_{r, All}$, and both have a large scatter. However, the mean changes of $\Delta_r$ seem to rest on 1, except the drop at high mass end from the AGN run. This large scatter can be caused by the sensitivity of the mass distribution to baryons: 1) galaxy cluster centers can be changed from the DM run to the two hydro-dynamical runs; 2) the positions and masses of substructures can be altered by baryons.
$f_s$ from the AGN run seems to suffer a weak baryon effect, except at smaller mass clusters, which tend to have higher ($\sim 20$ per cent) substructure mass fraction than the DM run, while the over-cooling problem is more obvious for $f_s$: substructures from the CSF run are more massive than from the DM run.
{width="\textwidth"}
$\eta$ is calculated from kinetic energy T, potential energy W, and surface pressure energy $E_s$. In \[fig:beta\], we study how $\eta$ is derivative to T, W, $E_s$. From left to right columns, we show the baryon effects on T, W, $E_s$ and $M_{vir}$ respectively. The upper row shows the results from all particles; while the lower panel results are coming from DMPs. The key findings of Fig. \[fig:beta\] are summarized as:
1. The ratio of kinetic energy T is shown in the first column. Again, the upper panel shows the results from all particles. $T_{CSF}/T_{DM}$ is about 0.95. However, the mean of this ratio drops to $\sim 0.9$ at both larger and smaller mass end; while it reaches $\sim 1.0$ at $M \approx 10^{14.8} {~h^{-1}\>{\rm M_\odot}}$. $T_{AGN}/T_{DM}$ is about 0.85 - 0.9. For the result coming from DMPs on the lower panel, both ratios have a constant shift up of $\sim 10$ per cent.
2. The second column shows the ratio of potential energy W. For the results from all particles in the upper panel, $W_{CSF}/W_{DM}$ is about 1.05, which is gradually reaching $\sim 1.0$ at the massive mass end. On the contrary, $W_{AGN}/W_{DM}$ is about 0.95, increasing to $\sim 1.0$ for the most massive clusters. For the results from DMPs on the lower panel, this ratio for the CSF run is almost the same; while the AGN run slightly ($\sim 3$ per cent) shift up.
Although both the CSF and AGN runs tend to have similar virial and dark matter masses as the DM run (actually the total mass from the AGN run is a little lower than the DM run at smaller halo masses, see the fourth column of this figure for more detail), the over-cooling problem in the CSF run tends to result a much higher concentration [see more discussion in @Cui2016], and so a higher potential energy than the AGN run.
3. We show ratios of $E_s$ in the third column. We have verified that $E_s$ only occupies $\sim 20$ per cent of the total kinetic energy T. It means that $E_s$ has a minor contribution to $\eta$. As shown on the upper panel, the baryon effect on the total $E_s$ is very similar ($\sim 5$ per cent lower than the DM run) between the AGN run and the CSF run. It is not surprising that DMPs contribute similar to $E_s$ between these three runs, that is shown on the lower panel.
From this we conclude that $E_s$ is irresponsible for the baryon effect. The unchanged $\eta_{DMP}$ for the CSF run is because baryons have a similar increasing ($\sim 5$ per cent) effect on T and W, while both T and W seem to be unaffected by baryons for the AGN run.
For the baryon effect on $\eta_{total}$, the key difference is in $T_{total}$. The drops of $T_{total}$ in both hydro-dynamical runs are possibly caused by collisional gas, of which thermal energy is either dissipated due to turbulences and frictions, or locked up into stars.
The classification of relaxed and unrelaxed clusters
----------------------------------------------------
{width="\textwidth"}
Separating out relaxed clusters from unrelaxed ones is not an easy task. @Neto2007 adopted $2T/|U| < 1.35$, $\Delta_r < 0.07$ and $f_s < 0.1$ to select relaxed galaxy clusters. They found $\sim 50$ per cent of halos at $M_{vir} = 10^{15} {~h^{-1}\>{\rm M_\odot}}$ are relaxed. However, they did not take the surface pressure energy $E_s$ into account in their virial ratio calculation. [@Shaw2006] adopted a slightly narrower limit ($\beta = 0.2$, equivalent to $|1 - \eta| < 0.2$) to select virial equilibrium halos with $E_s$ in their $\eta$. With only this criterion, they excluded 3.4 per cent of 2159 halos ($M_{halo} \gtrsim 3 \times 10^{13} {~h^{-1}\>{\rm M_\odot}}$) as un-virialized ones. [@Power2012] picked out dynamically relaxed halos with a slightly smaller $\Delta_r < 0.04$ at $z$ = 0. From this we conclude that there is no consistency in the literature about parameter for relaxed halos. In Fig. \[fig:td\], we investigate relations between these parameters: $\eta$ vs $\zeta$ (left column), $\Delta_r$ (middle column), $f_s$ (right column), which are normally used for classifying cluster dynamical states. From top to bottom, we show results from the AGN, the CSF and the DM runs, respectively. Symbol color encodes the cluster velocity dispersion $\sigma$, indicated in the top colorbar. Dashed vertical lines show $\eta = 1$, where clusters are in dynamical equilibrium. Grey regions indicate limits inside which galaxy clusters are relaxed.
In agreement with Fig. \[fig:be\], there is a good linear correlation between $\eta$ and $\zeta$ shown in the left column of Fig. \[fig:td\]. This is because $\sigma$ in $\zeta$ is equivalent to the square root of T in $\eta$, while $\sigma_t$ is similar to a square root of W. For all three versions of simulations, $\zeta$ is around 0.65 at $\eta = 1$. After excluding some noisy data points with $\eta < 0.8$, we find very similar slopes after linear fitting. Thus, we simply use all data points at the same time to fit, which results in black dotted lines with a slope of 0.312. This leads us to propose $\zeta$ as a proxy for $\eta$, which can be deduced from observation. All particles are used to calculate $\sigma$ and $\zeta$ here. Thus, to apply this relation on observations, one needs to consider the bias of using galaxies as the velocity dispersion tracer, which has been investigated and corrected in [@Munari2013], while for cluster mass M in $\sigma_t$, one can use lensing mass from observation. Using simulations with mock observation images, [@Puchwein2007] have shown that the recovered lensing mass does not depend on the assumption of hydrostatic equilibrium. Similar to our proposal, [@Puchwein2007] also suggested to use the difference between dynamically recovered mass from X-ray and lensing mass to distinguish dynamical states.
Although smaller $\Delta_r$ tends to have $\eta$ closer to 1, there are clusters that have larger $\Delta_r$ with $\eta \rightarrow 1$. Similarly the same is true for $f_s$. The virial equilibrium implies that $<d^2 I / dt^2> = 0$, time-averaged over a period that is long compared to the local dynamical timescale [@Shaw2006; @Poole2006]. Therefore, we expect a roughly symmetric distribution around zero due to those halos that are oscillating around the virial equilibrium position. These halos with large $\Delta_r$ and $f_s$ but $\eta \rightarrow 1$ could be still in the process of settling down to dynamical relaxation, but with a glimpse of equilibrium.
For our limited cluster mass range, we do not see a clear mass dependence on these parameters in Fig. \[fig:td\]. However, $\sigma$ shows a weak dependence on these parameters, especially in the left column, where higher $\sigma$ value tends to have both higher $\eta$ and $\zeta$ values. However, this trend is not clear for $\Delta_r$ and $f_s$.
From Fig. \[fig:td\], there is *no bimodal distribution* in any of the runs for either single or combined parameters. Data points from all three simulations have a similar distribution, other than a weak decrease of $\eta$ and a weak increase of $f_s$ from the DM run to the two hydro-dynamical runs.
Applying the selection criteria from [@Neto2007], we find that 70 (78 and 78) out of 123 halos from the DM run (from the CSF run and the AGN run, respectively) are dynamically relaxed. This gives a similar relaxation fraction as [@Neto2007]. One can visually find out that most of unrelaxed clusters are cut out by limits from $\Delta_r$ and $f_s$, which is also in agreement with [@Neto2007]. [@Power2012] suggested a smaller value of $\Delta_r
\approx 0.04$ to select dynamically relaxed halos [see also @Maccio2007]. Observational results suggest a much lower relaxation fraction: $\sim 28$ per cent from SDSS survey [@Wen2013]; $\sim 16$ per cent from X-ray selected clusters [@Mantz2015]. Thus, we apply restricted criteria to select out relaxed clusters: $0.85 < \eta < 1.15$ [@Knebe2008]; $\Delta < 0.04$ [@Power2012]; $f_s < 0.075$. By applying these thresholds, we select out 41, 43 and 48 dynamical relaxed clusters from the DM, the CSF and the AGN runs respectively. This gives a relaxation fraction of $\sim 35$ per cent. 29 ($\sim 65$ per cent) of these relaxed clusters are cross identified in all three runs; 34 ($\sim 75$ per cent) of them are cross identified in both the CSF and the AGN runs. In agreement with the baryon effect on individual parameters, most of halos have their dynamical relaxation states unchanged. Although AGN feedback impacts on substructures as well as $f_s$, it plays a minor role in changing the dynamical state of clusters.
Discussion and Conclusions {#concl}
==========================
Using our simulated galaxy cluster catalogue of 123 galaxy clusters from , we investigated the dynamical state of clusters in the DM (dark-matter-only) run; the CSF (gas cooling, star formation and supernova feedback) run and the AGN (with also AGN feedback) run. These three sets of simulations allow us to explore how baryons affect cluster dynamical states. We examined four parameters: the virial ratio $\eta$, the velocity dispersion deviation $\zeta$, the centre of mass offset $\Delta_r$ and the substructure mass fraction $f_s$, which are normally used to separate dynamically relaxed clusters from unrelaxed ones. The main results are summarised as follows.
1. The radial profiles of $\eta$ and $f_s$ become relatively constant at outer radius ($R \gtrsim 0.6 R_{vir}$). However, $\Delta_r$ does not show such features. It means that we can expect $\eta_{500} \approx \eta{vir}$ and $f_{s, 500} \approx f_{s, vir}$. However, this is not applicable for $\Delta_r$.
2. The baryon models (both with and without AGN feedback) have a weak effect on $\eta$, which is $\sim 10$ per cent lower in the two hydro-dynamical compared to the DM run. This is mainly caused by the drop of kinetic energy T with gas dynamics. Therefore, $\eta_{DMP}$ shows very similar results between all three runs.
Baryon models have no impact on $\Delta_R$ for the CSF run; this is also true for the AGN run at smaller masses, but there is a slightly smaller $\Delta_r$ in the AGN run than in the DM run at the higher mass end.
$f_s$ is about 40 (20) per cent higher in the CSF run than in the DM run at smaller (higher) masses, while $f_s$ from the AGN run is 20 per cent lower than from the CSF run.
3. There is good linear correlation between $\eta$ and $\zeta$ for all three runs, which encourages us to use $\zeta$ as an indicator of $\eta$, which can not be easily measured from observation. Using this relation, one can deduce the virial ratio for observed galaxies.
4. For all the investigated parameters, there is no clear bimodal distribution between relaxed and unrelaxed clusters.
5. With more restricted thresholds for $\eta, \Delta_r$ and $f_s$, we find that $\sim 35$ per cent of our sample clusters are relaxed, in which $\sim$ 65 per cent are cross identified in all three runs. This means that baryons play a minor role in regulating cluster dynamical states.
Using controlled cluster simulations, [@Poole2006] quantified the effects of mergers on the dynamical state of galaxy clusters and showed that dark matter typically relaxes slightly later than gas. A recent work by [@Zhang2016], who also used controlled cluster simulations but only with adiabatic gas, investigated baryon effects on merger times. They found merger timescale can be shortened by a factor of up to 3 for clusters with gas fractions of 0.15, compared to the one without gas. This indicates that clusters with baryons will virialize faster than ones without baryons, which is similar to the finding in [@Poole2006]. With galaxy clusters from cosmology simulations, we only find that baryons decrease the virial ratio by $\sim 10$ per cent from the DM run, which makes the mean of $\eta$ in the two hydro-dynamical runs much closer to 1. Because clusters in cosmological volume can never be isolated because mergers and in-falling material are ongoing, their dynamical states can hardly be exactly in dynamical relaxed. We further note here that the relaxation fraction seems to be unaffected ($\lesssim 5$ per cent) by baryons. This could because 1) our cluster sample is not large enough; 2) this relaxation fraction depends on the arbitrary selection limits. The total baryon mass fraction is normally around 10 - 15 per cent within galaxy clusters [e.g. @Sun2012; @Gonzalez2013; @Lagana2013; @Borgani2006; @Planelles2013]. It is interesting to see that $\eta$ is dragged down around a similar fraction by baryons, while its value from dark matter component is almost untouched. Another unchanged quantity is the linear relation between $\eta$ and $\zeta$, which urges us to propose a simple fitting function for observers to get $\eta$ from observed galaxy clusters. However, there is no bimodal distribution between relaxed and unrelaxed galaxy clusters. It makes a tough task for choosing the limits for these parameters to select out galaxy clusters in dynamical equilibrium.
Using different wavelength tracers to determine dynamical states of galaxy clusters can give different answers. Using photometric data of the Sloan Digital Sky Survey, [@Wen2013] derived the asymmetry, the ridge flatness and the normalized deviation of a smoothed optical map, which is coming from the brightness distribution of member galaxies. With their defined relaxation parameter from the upper three quantities, they found that 28 per cent of 2092 clusters are dynamically relaxed. In X-ray observation, the power ratio and the centroid shift are normally used to select out dynamically relaxed clusters [e.g. @Boehringer2010; @Rasia2013a]. In addition, [@Mantz2015] proposed the symmetry-peakiness-alignment criterion for classifying cluster dynamical states. With their criterion, they report a relaxation fraction of 16 per cent for their 361 X-ray selected clusters. Combining different wavelength results could give accurate answers. For example, [@Ge2016] has investigated the dynamical state of two paired clusters under optical, X-ray and radio emissions; [@Rossetti2016] characterized the dynamical states of galaxy clusters detected with the Sunyaev-Zeldovich (SZ) effect by the Planck and compare them with their dynamical states derived from X-ray surveys. They found a slightly higher relaxation fraction from the X-ray sample ($\sim$74 per cent) than from SZ sample ($\sim$ 52 per cent), which could due to different selection effects.
The reliability and agreement between these tracers from different wavelength observations, between different methods, as well as the consistency with theoretical predictions are still unclear. We will address these questions with our galaxy cluster sample in the next paper.
Acknowledgements {#acknowledgements .unnumbered}
================
All the figures in this paper are plotted using the python matplotlib package [@Hunter:2007]. This research has made use of NASA’s Astrophysics Data System (ADS), the arXiv preprint server and Wikipedia. Simulations have been carried out at the CINECA supercomputing Centre in Bologna, with CPU time assigned through ISCRA proposals and through an agreement with the University of Trieste. WC acknowledges the supports from University of Western Australia Research Collaboration Awards PG12105017, PG12105026, from the Survey Simulation Pipeline (SSimPL; [`http://www.ssimpl.org/`]{}) and from iVEC’s Magnus supercomputer under National Computational Merit Allocation Scheme (NCMAS) project gc6. WC, CP, AK, GFL, and GP acknowledge support of ARC DP130100117. CP, AK, and GFL acknowledge support of ARC DP140100198. CP acknowledges support of ARC FT130100041. SB and GM acknowledge support from the PRIN-INAF12 grant ’The Universe in a Box: Multi-scale Simulations of Cosmic Structures’, the PRINMIUR 01278X4FL grant ’Evolution of Cosmic Baryons’, the INDARK INFN grant and ’Consorzio per la Fisica di Trieste’. AK is supported by the [*Ministerio de Economía y Competitividad*]{} (MINECO) in Spain through grant AYA2012-31101 as well as the Consolider-Ingenio 2010 Programme of the [*Spanish Ministerio de Ciencia e Innovación*]{} (MICINN) under grant MultiDark CSD2009-00064. He further thanks Luna for lunapark. GP acknowledges support from the ARC Laureate program of Stuart Wyithe.
\[lastpage\]
[^1]: It is publicly available at https://github.com/ilaudy/PIAO
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the paper, the authors establish some asymptotic formulas and double inequalities for the factorial $n!$ and the gamma function $\Gamma$ in terms of the tri-gamma function $\psi''$.'
address:
- 'Department of Mathematics, Valahia University of Târgovişte, Bd. Unirii 18, 130082 Târgovişte, Romania'
- 'College of Mathematics, Inner Mongolia University for Nationalities, Tongliao City, Inner Mongolia Autonomous Region, 028043, China; Department of Mathematics, College of Science, Tianjin Polytechnic University, Tianjin City, 300387, China; Institute of Mathematics, Henan Polytechnic University, Jiaozuo City, Henan Province, 454010, China'
author:
- Cristinel Mortici
- Feng Qi
title: 'Asymptotic formulas and inequalities for gamma function in terms of tri-gamma function'
---
[^1]
Introduction and motivation
===========================
We recall that the classical Euler’s gamma function may be defined by $$\label{gamma-dfn}
\Gamma(z)=\int^\infty_0t^{z-1} e^{-t}\operatorname{d}t$$ for $\Re(z)>0$, that the logarithmic derivative of $\Gamma(x)$ is called psi or di-gamma function and denoted by $$\psi(x)=\frac{\operatorname{d}}{\operatorname{d}x}\ln\Gamma(x)=\frac{\Gamma'(x)}{\Gamma(x)}$$ for $x>0$, that the derivatives $\psi'(x)$ and $\psi''(x)$ for $x>0$ are respectively called tri-gamma and tetra-gamma functions, and that the derivatives $\psi^{(i)}(x)$ for $i\in\mathbb{N}$ and $x>0$ are called polygamma functions.
We also recall from [@mpf-1993 Chapter XIII] and [@widder Chapter IV] that a function $f(x)$ is said to be completely monotonic on an interval $I$ if it has derivatives of all orders on $I$ and satisfies $0\le(-1)^{n}f^{(n)}(x)<\infty$ for $x\in I$ and all integers $n\ge0$. The class of completely monotonic functions may be characterized by the celebrated Bernstein-Widder Theorem [@widder p. 160, Theorem 12a] which reads that a necessary and sufficient condition that $f(x)$ should be completely monotonic in $0\le x<\infty$ is that $$\label{berstein-1}
f(x)=\int_0^\infty e^{-xt}\operatorname{d}\alpha(t),$$ where $\alpha(t)$ is bounded and non-decreasing and the integral converges for $0\le x<\infty$.
In [@Sevli-Batir-Modelling Theorem 2.1], it was proved that the function $$F_{\alpha}(x) =\ln \Gamma(x+1) -x\ln x+x-\frac{1}{2}\ln x-\frac12\ln(2\pi)-\frac{1}{12}\psi'(x+\alpha)$$ is completely monotonic on $(0,\infty)$ if and only if $\alpha\ge\frac12$ and that the function $-F_{\alpha}(x)$ is completely monotonic on $(0,\infty)$ if and only if $\alpha=0$. Consequently, the double inequality $$\label{double-ineqSevli-BatirModel}
\frac{x^x}{e^x}\sqrt{2\pi x}\,\exp \biggl(\frac{1}{12}\psi'\biggl(x+\frac{1}{2}\biggr) \biggr) <\Gamma(x+1) <\frac{x^x}{e^x}\sqrt{2\pi x}\,\exp \biggl(\frac{1}{12}\psi'(x) \biggr)$$ was derived in [@Sevli-Batir-Modelling Corollary 2.1]. These results were also established in [@AMIS042013A.tex] and its preprint [@Merkle-Convexity2Complete-Mon.tex] independently from a different origin and by a different motivation. For some more information on bounding the gamma function $\Gamma$, please refer to the newly published paper [@Bukac-Sevli-Gamma.tex], the survey articles [@bounds-two-gammas.tex; @Wendel2Elezovic.tex-JIA; @Wendel-Gautschi-type-ineq-Banach.tex], and plenty of references collected therein.
The goal of this paper is to discover best asymptotic formulas and double inequalities for the factorial $n!=\Gamma(n+1)$ and the gamma function $\Gamma(x)$ in terms of the tri-gamma function $\psi'\bigl(x+\frac12\bigr)$. These results have something to do with the function $F_{\alpha}(x)$ and the double inequality .
An asymptotic formula and a double inequality for $n!$
======================================================
In this section, we establish a best asymptotic formula and a double inequality for the factorial $n!=\Gamma(n+1)$ in terms of the tri-gamma function $\psi'\bigl(x+\frac12\bigr)$.
\[Qi-Moticic-best-approx\] As $n\to \infty$, the asymptotic formula $$\label{a}
n! \sim \frac{n^n}{e^n}\sqrt{2\pi n}\,\exp \biggl(\frac{1}{12}\psi'\biggl(n+\frac{1}{2}\biggr) \biggr)$$ is the most accurate one among all approximations of the form $$\label{aa}
n! \sim \frac{n^n}{e^n}\sqrt{2\pi n}\,\exp \biggl(\frac{1}{12}\psi'(n+a)\biggr),$$ where $a\in\mathbb{R}$.
For $n\ge 1$, define a sequence $w_{n}$ by $$n!=\Gamma(n+1)=\sqrt{2\pi}\,n^{n+1/2}e^{-n}\exp \biggl(
\frac{1}{12}\psi'(n+a) \biggr) \exp w_{n}.$$ Taking into account $$\label{psi-poly-recur}
\psi^{(k)}(z+1)=\psi^{(k)}(z)+(-1)^k\frac{k!}{z^{k+1}}$$ for $k=1$, see [@aa p. 260, 6.4.6], yields $$w_{n+1}-w_{n}=1+\ln(n+1)-\biggl(n+\frac{3}{2}\biggr) \ln(n+1) +\biggl(n+\frac{1}{2}\biggr) \ln n+\frac{1}{12(n+a)^{2}}$$ and $$w_{n+1}-w_{n}=\biggl(-\frac{1}{6}a+\frac{1}{12}\biggr) \frac{1}{n^{3}} +\biggl(\frac{1}{4}a^{2} -\frac{3}{40}\biggr) \frac{1}{n^{4}}+O\biggl(\frac{1}{n^{5}}\biggr).$$ Hence, we have $$\lim_{n\to\infty}\bigl\{n^{3}\bigl[w_{n+1}-w_{n}\bigr]\bigr\}=\frac{1}{12}-\frac{1}{6}a.$$ Lemma 1.1 in [@m1; @Mortici-aar.tex] states that if the sequence $\{\omega_n:n\in\mathbb{N}\}$ converges to $0$ and $$\lim_{n\to\infty}n^k(\omega_n-\omega_{n+1})=\ell\in\mathbb{R}$$ for $k>1$, then $$\lim_{n\to\infty}n^{k-1}\omega_n=\frac{\ell}{k-1}.$$ Consequently, the sequence $w_n$ converges fastest only if $a=\frac12$.
\[ii-thm\] For every integer $n\ge 1$, we have $$\label{ii}
\exp \biggl(\frac{1}{240n^{3}}-\frac{11}{6720n^{5}}\biggr)
<\frac{e^nn!}{n^{n}\sqrt{2\pi n}\exp \bigl( \frac{1}{12}\psi'\bigl(n+\frac{1}{2}\bigr)\bigr)}
<\exp \frac{1}{240n^{3}}.$$
The double inequality may be rewritten as $$\label{f}
f(n)=\ln \Gamma(n+1)-\biggl(n+\frac{1}{2}\biggr)
\ln n+n-\frac{1}{2}\ln (2\pi) -\frac{1}{12}\psi'\biggl(n+\frac{1}{2}\biggr) -\frac{1}{240n^{3}} \le0$$ and $$\begin{gathered}
\label{g}
g(n)=\ln \Gamma(n+1)-\biggl(n+\frac{1}{2}\biggr)\ln n+n-\frac{1}{2}\ln (2\pi) \\
-\frac{1}{12}\psi'\biggl(n+\frac{1}{2}\biggr) -\frac{1}{240n^{3}}+\frac{11}{6720n^{5}}
\ge0.\end{gathered}$$ Employing the recurrence formula applied to $k=1$ and straightforwardly computing reveal that $f(n+1)-f(n)=u(n)$ and $g(n+1)-g(n)=v(n)$, where $$\begin{aligned}
u(x) &=1+\ln(x+1) -\biggl(x+\frac{3}{2}\biggr) \ln(x+1) +\biggl(x+\frac{1}{2}\biggr) \ln x\\
&\quad+\frac{1}{12\bigl(x+\frac{1}{2}\bigr)^{2}} -\frac{1}{240(x+1)^{3}}+\frac{1}{240x^{3}}\end{aligned}$$ and $$\begin{aligned}
v(x) &=1+\ln(x+1) -\biggl(x+\frac{3}{2}\biggr) \ln(x+1) +\biggl(x+\frac{1}{2}\biggr) \ln x+\frac{1}{12\bigl(x+\frac{1}{2}\bigr)^{2}} \\
&\quad-\frac{1}{240(x+1)^{3}}+\frac{1}{240x^{3}}+\frac{11}{6720(x+1)^{5}}-\frac{11}{6720x^{5}}.\end{aligned}$$ It is not difficult to verify that $$u''(x) = \frac{13x+74x^{2}+232x^{3}+391x^{4}+330x^{5}+110x^{6}+1}{20x^{5}(x+1)^{5}(2x+1)^{4}} >0$$ and $$v''(x) =-\frac{Q(x)}{1120x^{7}(x+1)^{7}(2x+1)^{4}}<0,$$ where $$\begin{aligned}
Q(x) &=825x+5499x^{2}+21325x^{3}+52589x^{4} \\
&\quad+83867x^{5}+83881x^{6}+47936x^{7}+11984x^{8}+55.\end{aligned}$$ This shows that $u(x)$ is strictly convex and $v(x)$ is strictly concave on $(0,\infty)$. Further considering $\lim_{x\to\infty}u(x) =\lim_{x\to\infty}v(x) =0$, we obtain that $u(x)>0$ and $v(x)<0$ on $(0,\infty)$. Consequently, the sequence $f(n)$ is strictly increasing and $g(n)$ is strictly decreasing while they both converge to $0$. As a result, we conclude that $f(n)<0$ and $g(n)>0$ for every integer $n\ge 1$. The proof of Theorem \[ii-thm\] is complete.
An asymptotic series and a double inequality for $\Gamma$
=========================================================
We now discover an asymptotic series and a double inequality for the gamma function $\Gamma(x)$ in terms of the tri-gamma function $\psi'\bigl(x+\frac12\bigr)$.
\[Gamma-asymp-thm\] As $x\to\infty$, we have $$\begin{gathered}
\Gamma(x+1)\sim\sqrt{2\pi}\,x^{x+1/2}\exp \biggl(\frac{1}{12}\psi'\biggl(x+\frac{1}{2}\biggr)-x+\frac{1}{240}\frac1{x^{3}}\\*
-\frac{11}{6720}\frac1{x^{5}}+\frac{107}{80640}\frac1{x^{7}} -\frac{2911}{1520640}\frac1{x^{9}}+\dotsm\biggr).\end{gathered}$$
Motivated by the inequality , we now consider a new function $h(x)$ defined by $$\Gamma(x+1) =\sqrt{2\pi}\,x^{x+1/2}e^{-x}\exp \biggl(
\frac{1}{12}\psi'\biggl(x+\frac{1}{2}\biggr) \biggr) \exp h(x) ,$$ that is, $$h(x) =\biggl[ \ln \Gamma(x+1) -\biggl(x+\frac{1}{2}\biggr) \ln x+x-\ln \sqrt{2\pi}\,\biggr] -\frac{1}{12}\psi'\biggl(x+\frac{1}{2}\biggr).$$ Using the formulas $$\ln \Gamma(x+1) -\biggl(x+\frac{1}{2}\biggr) \ln x+x-\ln \sqrt{2\pi}\,=\sum_{m=1}^\infty \frac{B_{2m}}{2m(2m-1) x^{2m-1}}$$ and $$\psi'(x) =\frac{1}{x}+\frac{1}{2x^{2}}+\sum_{m=1}^\infty \frac{B_{2m}}{x^{2m+1}}=\sum_{m=1}^{\infty} \frac{B_{m-1}}{x^{m}},$$ see [@aa p. 257, 6.1.40] and [@aa p. 260, 6.4.11], figures out $$\label{h(x)-diff}
h(x) =\sum_{m=1}^\infty \frac{B_{2m}}{2m(2m-1)x^{2m-1}}-\sum_{m=1}^\infty \frac{B_{m-1}}{12\bigl(x+\frac{1}{2}\bigr)^{m}},$$ where $B_{k}$ for $k\ge0$ denote Bernoulli numbers which may be generated by $$\frac{z}{e^z-1}=\sum_{k=0}^\infty B_k\frac{z^k}{k!}=1-\frac{z}2+\sum_{k=1}^\infty B_{2k}\frac{z^{2k}}{(2k)!}, \quad \vert z\vert<2\pi.$$ Making use of $$\begin{gathered}
\sum_{k=1}^{m}\frac{a_{k}}{\bigl( x+\frac{1}{2}\bigr) ^{k}}
=\sum_{k=1}^{m}a_{k}\biggl( 1+\frac{1}{2x}\biggr) ^{-k}\frac{1}{x^{k}}\\
=\sum_{k=1}^{m}a_{k}\Biggl[\sum_{i=0}^\infty\binom{-k}{i} \frac{1}{2^{i}x^{i}}\Biggr]\frac{1}{x^{k}}
=\sum_{k=1}^{m}\sum_{i=0}^\infty \frac{a_{k}}{2^{i}}\binom{-k}{i}\frac{1}{x^{k+i}}\end{gathered}$$ in , where $a_k$ is any sequence and $$\binom{-k}{i}=\frac{1}{i!}\prod_{\ell=0}^{i-1}(-k-\ell),$$ we obtain that $$h(x) =\frac{1}{240}\frac1{x^{3}}-\frac{11}{6720}\frac1{x^{5}}+\frac{107}{80640}\frac1{x^{7}} -\frac{2911}{1520640}\frac1{x^{9}}+ O\biggl(\frac1{x^{11}}\biggr).$$ The proof of Theorem \[Gamma-asymp-thm\] is complete.
\[ii-thm-continou\] For every integer $x\ge 1$, we have $$\label{ii-continuous}
\exp \biggl(\frac{1}{240x^{3}}-\frac{11}{6720x^{5}}\biggr)
<\frac{e^{x}\Gamma(x+1)}{x^{x}\sqrt{2\pi x}\exp \bigl( \frac{1}{12}\psi'\bigl(x+\frac{1}{2}\bigr) \bigr)}
<\exp \frac{1}{240x^{3}}.$$
Let $f(x)$ and $g(x)$ for $x\in[1,\infty)$ be defined by and respectively. Making use of inequalities $$\begin{aligned}
\frac{1}{12x}-\frac{1}{360x^{3}}+\frac{1}{1260x^{5}}
&<\ln \Gamma(x+1) -\biggl(x+\frac{1}{2}\biggr) \ln x+x-\frac{1}{2}\ln (2\pi) \\
&<\frac{1}{12x}-\frac{1}{360x^{3}}+\frac{1}{1260x^{5}}-\frac{1}{1680x^{7}}\end{aligned}$$ and $$\frac{1}{x}+\frac{1}{2x^{2}}+\frac{1}{6x^{3}}-\frac{1}{30x^{5}}
<\psi'(x)
<\frac{1}{x}+\frac{1}{2x^{2}}+\frac{1}{6x^{3}}-\frac{1}{30x^{5}}+\frac{1}{42x^{7}},$$ which may be deduced from [@Koumandos-jmaa-06 Theorem 2 and Corollary 1], finds that $f(x)<a(x)$ and $g(x)>b(x)$, where $$\begin{aligned}
a(x) &=\frac{1}{12x}-\frac{1}{360x^{3}}+\frac{1}{1260x^{5}} -\frac{1}{12}\Biggl[ \frac{1}{x+\frac{1}{2}}+\frac{1}{2\bigl(x+\frac{1}{2}\bigr)^{2}}\\
&\quad +\frac{1}{6\bigl(x+\frac12\bigr)^{3}}-\frac{1}{30\bigl(x+\frac{1}{2}\bigr)^{5}}\Biggr] -\frac{1}{240x^{3}}\\
&= -\frac{A(x-1)}{5040x^{5}(2x+1)^{5}}\\
&<0,\\
b(x) &=\frac{1}{12x}-\frac{1}{360x^{3}}+\frac{1}{1260x^{5}}-\frac{1}{1680x^{7}} -\frac{1}{12}\Biggl[ \frac{1}{x+\frac{1}{2}}+\frac{1}{2\bigl(x+\frac{1}{2}\bigr)^{2}}\\
&\quad +\frac{1}{6\bigl(x+\frac12\bigr)^{3}}-\frac{1}{30\bigl(x+\frac{1}{2}\bigr)^{5}} +\frac{1}{42\bigl(x+\frac{1}{2}\bigr)^{7}}\Biggr] -\frac{1}{240x^{3}}+\frac{11}{6720x^{5}}\\
&= \frac{B(x-1)}{20160x^{7}(2x+1)^{7}}\\
&>0,\\
A(x)&=3760x+6565x^{2}+5310x^{3}+1980x^{4}+264 x^{5}+785,\\
B(x) &= 93268x+263179x^{2}+382830x^{3} \\
&\quad+315336 x^{4}+147504x^{5}+35952x^{6}+3424x^{7}+12547.\end{aligned}$$ The proof of Theorem \[ii-thm-continou\] is thus complete.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The work of the first author was supported in part by the Romanian National Authority for Scientific Research, CNCS-UEFISCDI, under Grant No. PN-II-ID-PCE-2011-3-0087.
[99]{}
M. Abramowitz and I. A. Stegun (Eds), *Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables*, National Bureau of Standards, Applied Mathematics Series **55**, 10th printing, Dover Publications, New York and Washington, 1972.
S. Koumandos, *Remarks on some completely monotonic functions*, J. Math. Anal. Appl. **324** (2006), no. 2, 1458–1461; Available online at <http://dx.doi.org/10.1016/j.jmaa.2005.12.017>.
D. S. Mitrinović, J. E. Pečarić, and A. M. Fink, *Classical and New Inequalities in Analysis*, Kluwer Academic Publishers, 1993.
C. Mortici, *A quicker convergence toward the $\gamma$ constant with the logarithm term involving the constant $e$*, Carpathian J. Math. **26** (2010), no. 1, 86–91.
F. Qi, *A completely monotonic function involving the gamma and tri-gamma functions*, Appl. Math. Inf. Sci. (2014), in press.
F. Qi, *A completely monotonic function involving the gamma and tri-gamma functions*, available online at <http://arxiv.org/abs/1307.5407>.
F. Qi, *Bounds for the ratio of two gamma functions*, J. Inequal. Appl. **2010** (2010), Article ID 493058, 84 pages; Available online at <http://dx.doi.org/10.1155/2010/493058>.
F. Qi and Q.-M. Luo, *Bounds for the ratio of two gamma functions: from Wendel’s asymptotic relation to Elezović-Giordano-Pečarić’s theorem*, J. Inequal. Appl. 2013, **2013**:542, 20 pages; Available online at <http://dx.doi.org/10.1186/1029-242X-2013-542>.
F. Qi and Q.-M. Luo, *Bounds for the ratio of two gamma functions—From Wendel’s and related inequalities to logarithmically completely monotonic functions*, Banach J. Math. Anal. **6** (2012), no. 2, 132–158.
F. Qi and Q.-M. Luo, *Complete monotonicity of a function involving the gamma function and applications*, Period. Math. Hungar. **??** (2014), in press; Available online at <http://dx.doi.org/10.1007/???>.
F. Qi and C. Mortici, *Some best approximation formulas and inequalities for Wallis ratio*, available online at <http://arxiv.org/abs/1312.3782>.
H. Şevli and N. Bat[i]{}r, *Complete monotonicity results for some functions involving the gamma and polygamma functions*, Math. Comput. Modelling **53** (2011), 1771–1775; Available online at <http://dx.doi.org/10.1016/j.mcm.2010.12.055>.
D. V. Widder, *The Laplace Transform*, Princeton University Press, Princeton, 1946.
[^1]: This paper was typeset using -LaTeX
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present $BV$ photometry of the Sculptor dwarf galaxy to $V=22$. These data give evidence for a bimodality in Sculptor’s metallicity distribution based on a discontinuity in the luminosities of horizontal branch (HB) stars and by the presence of two distinct red giant branch (RGB) bumps. A consistent picture of the evolved stars in Sculptor is given by the presence of (1) a metal-poor population of $\feh \sim -2.3$ with an exclusively blue HB and that corresponds to the blueward side of the Sculptor RGB and the more luminous RGB bump, and (2) a less metal-poor population of $\feh \sim -1.5$ required to explain the less luminous red HB, the red side of the RGB, and a second, less luminous RGB bump. Best fits to the HB populations are obtained with enhanced oxygen abundances, $\ofe \sim
+0.5$. Variations in the global HB and RGB morphology of Sculptor can be explained by differences in the radial distribution of these two populations. The presence of these two populations shows that the Sculptor dwarf galaxy has an [*internal*]{} second parameter problem.
author:
- 'S. R. Majewski, M. H. Siegel, Richard J. Patterson, and R. T. Rood'
title: An Internal Second Parameter Problem in the Sculptor Dwarf Spheroidal Galaxy
---
\#1\#2\#3\#4\#5\#6\#7
to\#2
------------------------------------------------------------------------
\#1
Introduction
============
The Sculptor dwarf spheroidal (dSph) galaxy was the first Galactic dSph to be identified (Shapley 1938) and has a long observational history. Its variables have been tabulated by a large number of studies (Baade & Hubble 1939; Thackeray 1950; van Agt 1978; Goldsmith 1993, Hereafter G93; Ka[ł]{}uzny 1995, hereafter K95), and the period distribution of RR Lyrae stars suggests a metallicity spread (G93, K95). While a range of abundances is generally accepted, impressions of the Sculptor horizontal branch (HB) morphology outside of the instability strip have varied depending on photometric depth, field size and filter systems employed to construct the color magnitude diagram (CMD). Kunkel & Demers’ (1977, KD77) Sculptor CMD (324 stars to $V=20.6$) yielded 43 red HB (RHB) out of 49 HB stars and a deficit of stars with $B-V<+0.3$ as well as a red giant branch (RGB) well described by a metal-poor population ($\feh =-1.9$). Norris & Bessell (1978) re-analyzed the CMD in combination with two spectra to argue for a Sculptor metallicity spread of $-2.2 \le \feh \le -1.5$, and Smith & Dopita (1983) confirmed an inhomogeneous metallicity distribution function (MDF) via narrow-band photometry. Da Costa’s (1984, D84) deep, but small area photometry to the Sculptor MSTO did not provide strong constraints on either the HB or RGB; however, it did show an abundance spread similar to previous results (confirmed by Da Costa 1988) and a predominantly [*red*]{} HB (7 of 10 HB stars). The conclusion of these studies was that Sculptor is a “second parameter () object” that shows a rather red global HB for its mean abundance (D84).
More recently, however, Schweitzer (1995, SCMS) produced a CMD with 1043 stars that reveals a prominent [*blue*]{} HB (BHB) with more stars than KD77 and D84. In the first wide-field, [*CCD*]{} survey of Sculptor, K95 reported the usual metallicity spread based on the RGB (as did SCMS), but also substantiated the large BHB population and derived a more moderate Sculptor HB morphology index of $(B-R)/(B+V+R) = -0.15$ (see also the Grebel 1994 CMD). Because their $VI$ photometry of $>6000$ stars with $V<21.5$ covered a much larger area than previous results, K95’s significantly increased BHB:RHB ratio suggests differences in the spatial distribution of BHB and RHB stars. This might be due to abundance gradients in the dSph. However, since the most metal rich population in Sculptor’s RGB has $\feh \sim -1.5$, which would normally give a uniform to blue HB, the variation in the RHB population must be due to spatial variation in the [**]{} effect.
In this letter, we present $BV$ photometry of the Sculptor dSph galaxy. We find the usual evidence for RGB stars ranging from metal-poor ($\feh
\sim -1.5$) to very metal-poor ($\feh \sim -2.3$); however, on the basis of two [*distinct*]{} HBs and two distinct RGB bumps, Sculptor’s MDF may be better characterized as [*bimodal*]{}. This bimodality gives rise to one population with a effect, and a second one with likely very little HB . Differences in radial distributions for these two populations can account for the variation in HB morphology within Sculptor and among previous surveys of this galaxy.
Observations and Reduction
==========================
We observed Sculptor on UT 23 July and 1–2 August, 1991 with the Las Campanas 1-m Swope telescope using the thinned, $1024^2$ TEK2 CCD camera. Five overlapping, 104 wide pointings were arranged in a 2$\times$2 grid with a center frame overlapping the other four to lock together the photometry. Each field was typically observed with one $B$ and $V$ exposure of 1800 and 900 sec length, respectively. The data were reduced with the IRAF package CCDRED and photometered with the DAOPHOT II and ALLFRAME programs (Stetson 1987, 1994). Detections were matched using DAOMASTER and then calibrated to observed Graham (1982) standard stars using our own code. This code compares calibrated magnitudes of stars in common on different CCD frames and determines minor frame-to-frame systematic errors (e.g., due to shuttering errors, transient transparency changes, errors in the photometric transformation). Because of photometric conditions, the derived mean residuals for each frame ($\le 0.1$ mag on the basis of $\ge689$ comparison stars) were used as offsets and applied iteratively with new color determinations until convergence. Our resulting photometric precision is $(\sigma_{B},\sigma_{V})=(0.05,0.05)$ mag at the HB.
Horizontal Branch
=================
Our $(B-V,~V)_0$ (Figure 1) and $(B-V,~B)_0$ (not shown) CMDs show an HB that appears to be kinked over the RR Lyrae gap. All tests of the photometry pipeline have shown this “kink” to be real, and a hint of this HB “kink” can be seen in KD77. Similarly kinked HBs have been noted previously in the CMDs of some “bimodal” Galactic globular clusters (GGC), e.g., NGC 6229 (Borissova 1997), NGC 2808 (Ferraro 1990), and NGC 1851 (Walker 1992), to which the Sculptor CMD bears some resemblance. Indeed, our derived $B:V:R$ (blue:variable:red HB) ratio of (0.42:0.19:0.39) resembles those of bimodal GGCs (see Borissova 1997 for a summary). Stetson (1996) make a poignant comparison of the bimodal NGC 1851 CMD to those of the similar metallicity, “ GGC pair” NGC 288 and NGC 362; that NGC 1851 has [*both*]{} an RHB like NGC 362 [*and*]{} a BHB like NGC 288 suggests that NGC 1851 has an internal problem. Stetson use this fact to argue as unlikely that the effect is due to differences in age, helium abundance or \[CNO/Fe\] within NGC 1851. Despite the similarities of our Sculptor HB to the HBs of bimodal GGCs, there are two key reasons why the Stetson analysis does not apply here: (1) From the RGB width we [*know*]{} that Sculptor has an abundance spread. (2) There is no [*a priori*]{} reason to assume that all of the stars in Sculptor are coeval. Bearing this in mind we now explore the origin of the bimodality of the Sculptor HB.
The $V$ magnitude difference from the red edge of the Sculptor BHB to the blue edge of the RHB is $\sim0.15 \pm 0.02$ mag. If the bimodality is completely due to differences in \[Fe/H\], typical values for $dM_V/d{\rm \feh}$ suggest an \[Fe/H\] difference of 0.5 to 1.0 dex. This is consistent with reported \[Fe/H\] spreads from fitting isochrones to the Sculptor RGB.
The situation is, however, more complex because we are comparing HB stars at different colors, and both the luminosity of the theoretical ZAHB (zero-age-HB) and bolometric correction vary with position along the HB. Moreover, the HB is strongly affected by the oxygen abundance. At a constant core mass () increasing \[O/Fe\] increases $L_{\rm HB}$. However increasing \[O/Fe\] also leads to a decrease in . All other things being equal, a decrease in leads to a decrease in $L_{\rm HB}$. The net result is that the ZAHB variation with \[O/Fe\] can be rather complex. In the Galaxy it is generally thought that for metallicities appropriate for Sculptor, \[O/Fe\] is constant with a value in the range +0.3–0.5. There is no reason to assume that Sculptor has undergone the same chemical enrichment history as the Galaxy so we consider all $ 0.0 \le \ofe \le +0.5$ possible. Most recent HB models have an assumed , relation. The only available models that allow us to explore the composition parameters independently are those of Rood (unpublished). To convert and to $M_V$ and $B-V$ we use the results of Kurucz (1979) and Bell & Gustaffson (1978) blended to reproduce observed HBs of GGCs smoothly. Throughout this paper, we assume $(m-M)_0 = 19.71$ and $\ebmv = 0.02$ (K95) for Sculptor.
Figure 2 shows the observed CMD of the Sculptor HB with superimposed ZAHBs terminated at the red end at a mass of 0.85with a cross mark indicating a mass of 0.80. These are roughly the maximum possible masses for 12 and 15 Gyr populations, respectively. Since all stars undergo some mass loss the ZAHB population will not actually reach these two points. Evolution and observational scatter will carry some stars redward, but for practical purposes the end of the ZAHB should mark the redward extent of the HB.
We start with the hypothesis that the Sculptor BHB is a low metallicity population and the RHB a higher metallicity population, both consistent with the spread of the RGB. The fairly uniform distribution across the RGB suggests comparable numbers in each group. The size of the observational error would obscure obvious bimodality on the RGB.
The BHB can be fit reasonably with $\feh = -2.3$ and $0.0 \le \ofe \le
+0.5$. Indeed, the BHB rather resembles that of the low metallicity GGC M92 (see Figure 1). The RHB can be fit with oxygen enhanced models with $-1.9 \le \feh \le -1.5$. The odd behavior of the ZAHB level with for the $\ofe = +0.5$ ZAHBs is due to approximations used for . Independently of such modeling details, one can expect the variation of ZAHB level with to be less for oxygen enhanced models than for scaled solar abundances. The models with $\ofe = 0.0$ cannot fit the RHB: at $\feh = -1.9$ the ZAHB does not extend far enough to the red; at $\feh = -1.5$ the level of the ZAHB is too low.
One could conceivably produce the observed bimodality using one composition with $-1.9 \le \feh \le -1.5$ and $\ofe=+0.5$. Such a mono-compositional bimodality is observed in GGCs but modeling it requires the [*ad hoc*]{} introduction of bimodality in some underlying parameter (Catelan 1998). However, in Sculptor a composition spread is observed, and a bimodal composition is quite natural, e.g., arising from two bursts of star formation. Hence, it seems undesirable to us to discard the “natural explanation” in favor of the yet to be determined mechanism that produces bimodal HBs in GGCs.
It is clear from the $\ofe = +0.5$ ZAHB (Figure 2a) that even if most of the low metallicity population is found on the BHB, some could be found in the RR Lyrae strip and on the RHB. We suspect that this is a small fraction of the low metallicity population, because the red end of the BHB veers away from the ZAHB suggesting that the ZAHB is populated only for $(B-V)_0 \lesssim 0.15$. In analogy to M92 we suspect that $\ga 90$% of the low metallicity population is found on the BHB and that its age is similar to that of M92.
Similarly, from Figure 2 we see that the higher metallicity population could contaminate the BHB. The RHB population does drop as one approaches the RR Lyrae strip. But there is precedent from the bimodal HB GGCs that such a population could increase further to the blue. There is reason to think this is not true for Sculptor. First, if there is significant high metallicity contamination of the BHB, where are the low metallicity stars we infer must be present from the RGB spread? Second, the BHB morphology is more like that of M92 than the blue-HB-tails of clusters with higher metallicity–M13, NGC 288, etc. These arguments in themselves are not compelling, but fit the overall scenario we develop here.
Normal GGCs with the metallicity we suggest for the Sculptor RHB have uniform HBs. This means that the Sculptor high metallicity population suffers from a “too red” problem like, e.g., GGCs NGC 362 and NGC 7006 and the extreme halo cluster Pal 14. While the case for age as the in GGCs has been hard to establish (e.g., Stetson 1996; Catelan 1998; VandenBerg 1998; but see counter views by Chaboyer 1996, Sarajedini 1997), there is good reason to think that a higher metallicity population in a low density system like Sculptor might be younger. Thus, we hypothesize that the RHB arises from a population several Gyr younger than the BHB. Indeed, D84 has suggested multiple age components ($\delta({\rm age}) \sim3$ Gyr) from his study of turnoff stars.
If bimodal, Sculptor’s two HB populations probably overlap significantly in the instability strip. The distribution of RR Lyrae periods in Sculptor (G93, K95) shows a large range, consistent with a large spread in metallicity. The periods of RRab stars at the blue fundamental edge of the instability strip (those with the shortest periods) are well correlated with the metallicity (Sandage 1993a). In Sculptor, the shortest period RRab (ignoring two stars with very discrepant periods) has a period of 0.474 days (K95), implying a metallicity of $\feh=-1.6$. While the red fundamental edge is not as useful a metallicity indicator, the existence of RRab stars with $P\gtrsim0.8$ days indicates the presence of another population with $\feh<-2.0$. In addition, G93 and K95 both note a correlation of average magnitude with period in Sculptor RRab stars. Because $\left<M_V\right>$ is a function of $\feh$, the spread in $\left<M_V\right>$ also implies a metallicity spread. The intensity weighted average $\left<V\right>$ magnitude for the majority of the RRab stars lies in the range $20.1 < \left<V\right> < 20.25$ (K95), or $ 0.24 < \left<M_V\right> < 0.64$, which corresponds to $-2.3
< \feh < -1.3$ (Sandage 1993b).
Red Giant Branch
================
Our analysis so far points to a bimodality of populations in the Sculptor HB. However, such bimodality is also suggested in the giant branch, where two distinct RGB bumps can be seen (Figure 1): one near $(B-V, V)_0 = (0.8, 19.3)$ and one near $(0.8, 20.0)$. The former RGB bump lies toward the blue side of the RGB, near the expected locus for metal poor stars, while the latter RGB bump lies toward the red side of the RGB, near the expected locus for more metal rich stars. To illustrate the differences, we fit a mean RGB locus to the entire Sculptor RGB, divide the RGB in half, and plot (Figure 3) RGB luminosity functions for all stars within $\Delta (B-V) \sim 0.125$ left and right of the mean RGB locus. We isolate the redward RGB bump at $V_0 \sim
20.0$. The blueward bump is less clearly defined but probably is $19.0
\le V \le 19.4$. The extreme magnitude differences between the RGB bumps again argues for a metallicity separation of order a dex. We can use the absolute magnitudes of the RGB bumps to obtain a global metallicity (\[M/H\]) for the two bump populations (Ferraro 1999): Using $(m-M)_0 = 19.71$ (K95), we find \[M/H\] $\sim -2.1$ and $\sim -1.3$. Ferraro (1999, Figure 11a) also give relations for the RGB bump dependence on the magnitude difference between the bump and HB. If we adopt $V = 20.2$ for the height of the BHB population and assign this to the metal-poor RGB, we obtain $V_{bump} - V_{HB} \sim -0.9$; this implies an abundance $\feh \lesssim -2.4$, on the Zinn (1985) scale. For the RHB population, if we adopt $V_{HB}=20.35$ and assign to this the other RGB bump, we obtain $V_{bump} - V_{HB} = -0.35$; this is the difference expected for $\feh \lesssim -1.6$.
The presence of the distinct RGB bumps, their estimated $M_V$, and their location relative to the HB suggest a bimodal MDF with $\feh \sim -2.3$ and $\feh \sim -1.6$.
Discussion
==========
From analysis of the RGB and HB, a consistent scenario can be assembled. In Figure 1 we show representative RGB, AGB, and HB fiducials for the metal-poor ($\feh=-2.23$), BHB cluster M92 and the less metal poor ($\feh=-1.44$) cluster Pal 14. These clusters bracket the Sculptor RGB, while each cluster separately approximates the BHB and RHB, respectively. Apart from the fact that Pal 14 may be a little metal-rich by a few 0.1 dex, the two clusters provide a reasonable bimodal paradigm for the Sculptor MDF.
Our bimodal interpretation of Sculptor differs somewhat from previous studies that argue for an abundance [*spread*]{}. It should be noted that a true bimodality in the RGB of Sculptor in the form of two distinct RGB sequences would be masked somewhat by observational scatter and the superposition of the asymptotic giant branch for the more metal rich population. The presence of two [*distinct*]{} RGB bumps, rather than a slanting RGB bump “continuum,” is evidence for bimodality in Sculptor. We note that a suggestion of bimodality (or [ *tri*]{}modality) was made previously by Grebel (1994).
In §1 we argued that disparate HB morphologies found among different surveys of the Sculptor CMD derived from radial differences in global HB morphology. Figure 4 provides evidence that this is the case: The global HB index increases by 0.4 from the center to the $\sim 500"$ radius accessible with our catalogue. We have argued for a [ *bimodal*]{} MDF. Accordingly, the radial gradient in Figure 4 is not likely due to a radial abundance gradient, or the gradual diminishing of a effect. Rather, the cumulative evidence suggests that the HB radial dependence is due to changes in proportions of two nearly mono-metallic populations with radius. Indeed, the relative densities of the blue:red half of the RGB track those of the BHB:RHB very well (Figure 4). The spatial distribution of the \[BHB, blue RGB, metal-poor\] population appears to be more extended than that of the \[RHB, red RGB, less metal poor\] population, which shows a higher core concentration. Spatial differences in the Sculptor HB were suggested previously by Light (1988) and Da Costa (1996) and are explored further by Hurley-Keller (1999). Da Costa (1996) also point out radial HB index gradients (with a similar sense) in the Leo II and And I dSphs, and adopt the same interpretation of mixing variations in bimodal HB populations.
The existence of bimodal, $+$ non- populations within dSphs would be significant since, unlike bimodal GGCs such as NGC 1851, in dSphs it is (now) entirely plausible to consider multiple star formation bursts with age as the . In Sculptor’s case, it is likely that the $\feh \sim -2.3$ population formed in an earlier, more extended burst. If the presence of these two distinct populations is born out, the (relatively nearby) Sculptor dSph could well prove to be a Rosetta stone of the HB and the adamantine question.
We thank Eva Grebel for helpful discussions and the referee for useful suggestions.
Baade, W. & Hubble, E. 1939, PASP, 51, 40 Bell, R. A. & Gustaffson, B. 1978, A&AS, 34, 229 Borissova, J., Catelan, M., Spassova, N. & Sweigert, A. V. 1997, AJ, 113, 692 Catelan, M., Borissova, J., Sweigert, A. V., & Spassova, N. 1998, ApJ, 494, 265 Chaboyer, B., Demarque, P. & Sarajedini, A. 1996, , 459, 558 Da Costa, G. S. 1984, , 285, 483 (D84) Da Costa, G. S. 1988, in IAU Symp. 126, The Harlow Shapley Symposium on Globular Cluster Systems in Galaxies, eds. J. E. Grindlay & A. G. D. Philip (Dordrecht: Kluwer), 217 Da Costa, G. S., Armandroff, T. E., Caldwell, N. & Seitzer, P. 1996, , 112, 2576 Ferraro, F. R., Clementini, G., Fusi Pecci, F., Buonanno, R. & Alcaino, G. 1990, A&AS, 84, 59 Ferraro, F. R., Messineo, M., Fusi Pecci, F., De Palo, M. A., Straniero, O. & Chieffi, A. 1999, in preparation Goldsmith, C. G. 1993, in New Perspectives on Stellar Pulsation and Pulsating Variable Stars, eds. J. M. Nemec & J. M. Matthews (Cambridge: CUP), 358 Graham, J. A. 1982, , 94, 244 Grebel, E. K., Roberts, W.J. & van de Rydt, F. 1994, in Third CTIO/ESO Workshop on The Local Group, eds. A. Layden, R. C. Smith & J. Storm (Munich: ESO), 148 Hurley-Keller, D., Mateo, M. & Grebel, E. K. 1999, in preparation Harris, W. E. 1996, , 112, 1487 Holland, S. & Harris, W. E. 1992, , 103, 131 Ka[ł]{}uzny, J., Kubiak, M., Szymanśki, M., Udalski, A., Krzemiński, W. & Mateo, M. 1995, A&AS, 112, 407 Kunkel, W. E. & Demers, S. 1977, , 214, 21 (KD77) Kurucz, R. L. 1979, ApJS, 40, 1 Light, R. M. 1988, Ph.D. thesis, Yale University Norris, J. & Bessell, M. S. 1978, , 225, L49 Sandage, A. 1970, , 162, 841 Sandage, A. 1993a, , 106, 687 Sandage, A. 1993b, , 106, 703 Sarajedini, A., Chaboyer, B. & Demarque, P., , 109, 1321 Shapley, H. 1938, Harvard Bull., 908, 1 Smith, H. S. & Dopita, M. A. 1983, , 271, 113 Schweitzer, A. E., Cudworth, K. M., Majewski, S. R. & Suntzeff, N. B. 1995,, 110, 2747 (SCMS) Stetson, P. B. 1987, , 99, 191 Stetson, P. B. 1994, , 106, 250 Stetson, P. B., VandenBerg, D. A. & Bolte, M. 1996, AJ, 96, 909 Thackeray, A. 1950, Observatory, 70, 144 van Agt, S. L. Th. J. 1978, Pub. David Dunlop Obs., 3, 205 VandenBerg, D. A. 1998, in IAU Symp. 189, Fundamental Stellar Properties: The Interaction between Observations and Theory, eds. T. R. Bedding, A. J. Booth & J. Davis (Dordrecht: Kluwer), 439 Zinn, R. 1993, in The Globular Cluster-Galaxy Connection, ASP Conf. Ser. 48, eds. G. H. Smith & J. P. Brodie (San Francisco: ASP), 38
Captions
========
[fig]{}. 1. – $(B-V,~V)_0$ CMD for the Sculptor dwarf galaxy with overlaid fiducials for M92 ([*dashed line*]{}; from Sandage, 1970) and Pal 14 ([*solid line*]{}; from Holland & Harris 1992). We adopt cluster distance moduli and reddenings from Harris (1996) and Holland & Harris (1992), respectively. The right panel highlights the RGB and red bump region.
[fig]{}. 2. – Fits of model ZAHBs to the HB of Sculptor. Panel (a) shows oxygen enhanced ($\ofe=+0.5$) models, and (b) shows models with solar $\ofe$. In each panel, the [*solid line*]{} shows the model with $\feh=-2.3$, the [*dashed line*]{} shows $\feh=-1.9$, and the [ *dotted line*]{} shows $\feh=-1.5$. The ZAHBs are terminated at the red end at a mass of 0.85and a cross mark indicates a mass of 0.80(see text).
[fig]{}. 3. – Differential (right ordinate) and cumulative (left ordinate) RGB luminosity functions for stars within 0.125 mag in $(B-V)_0$ color to either the blue ([*dot-dash curves*]{}) or red ([ *solid curves*]{}) of the mean RGB locus. The [*dot-dash curves*]{} are offset vertically by $+1.0$ for the cumulative and by $+80$ for the differential luminosity function. Breaks in the slope of the cumulative distributions (indicated by [*thin solid lines*]{}) point to locations of RGB bumps, marked by vertical lines.
[fig]{}. 4. – Radial dependence of HB ([*filled circles*]{}) and RGB ([*open circles*]{}) morphology from our catalogue. The RR Lyrae counts in the same areas are from K95. The values for $(B:V:R)$ for the HB and $B+R$ for the RGB in each annulus are given for each point.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study codes on graphs combined with an iterative message passing algorithm for quantization. Specifically, we consider the binary erasure quantization (BEQ) problem which is the dual of the binary erasure channel (BEC) coding problem. We show that duals of capacity achieving codes for the BEC yield codes which approach the minimum possible rate for the BEQ. In contrast, low density parity check codes cannot achieve the minimum rate unless their density grows at least logarithmically with block length. Furthermore, we show that duals of efficient iterative decoding algorithms for the BEC yield efficient encoding algorithms for the BEQ. Hence our results suggest that graphical models may yield near optimal codes in source coding as well as in channel coding and that duality plays a key role in such constructions.'
author:
- |
Emin Martinian\
\[-5pt\] Massachusetts Institute of Technology\
\[-5pt\] Cambridge, MA 02139\
\[-5pt\] [email protected]\
- |
Jonathan S. Yedidia\
\[-5pt\] Mitsubishi Electronic Research Labs\
\[-5pt\] Cambridge, MA 02139\
\[-5pt\] [email protected]
bibliography:
- 'paper.bib'
title: Iterative Quantization Using Codes On Graphs
---
Introduction
============
Researchers have discovered that error correction codes defined on sparse graphs can be iteratively decoded with low complexity and vanishing error probability at rates close to the Shannon limit. Based on the close parallels between error correction and data compression, we believe that similar graphical codes can approach the fundamental limits of data compression with reasonable complexity. Unfortunately, the existing suboptimal channel decoding algorithms for graphical codes generally fail unless the decoder input is already near a codeword. Since this is usually not the case in source coding, either a new type of graph or a new suboptimal algorithm (or both) is required.
The advantages of graphical codes for error correction are compelling. Without coding, an energy per bit ($E_b/N_0$) of 9.6 dB is required to achieve a bit error rate (BER) of $10^{-5}$ with binary pulse amplitude modulation on an additive white Gaussian noise channel. Using a 256-state convolutional code and Viterbi (, maximum-likelihood) decoding, on the order of 7 dB of coding gain can be achieved and the resulting energy per bit required is on the order of 3 dB from the Shannon limit ($E_b/N_0$ = 0.2 dB) for binary rate-1/2 codes. Very long turbo codes have been shown to approach to within 0.1 dB of the Shannon limit [@elet:Brink:2000], and very long irregular low density parity check (LDPC) codes have been shown [@comlet:2001:chung] to approach even closer to the Shannon limit. Thus LDPC codes can provide on the order of a 10 dB gain over uncoded systems and on the order of a 3 dB gain over non-iterative approaches.
Before developing iterative quantization techniques it is worth investigating the potential gains of such an approach over existing systems. For asymptotically high rates, when compressing a continuous source with finite moments relative to mean square error (MSE) distortion, entropy coded scalar quantization (ECSQ) is 1.53 dB from the rate-distortion limit [@Gray_1998]. For moderate rates the gap is larger: in quantizing a Gaussian source relative to MSE distortion, ECSQ systems are 1.6–3.4 dB away from the rate-distortion limit. For these parameters, trellis coded quantization (TCQ) using a 256-state code with optimal quantization has a gap of 0.5–1.4 dB [@Marcellin_1990]. For higher rates, sources with larger tails (, a source with a Laplacian distribution), or sources with memory, the gaps are larger. Thus for memoryless sources quantized at moderate rates, new codes have the potential to improve performance by the noticeable margin of a few decibels. More generally, the codes on graphs paradigm may prove valuable in a variety of scenarios involving speech, audio, video and other complicated sources.
To illustrate possible approaches to developing graphical codes we focus on the binary erasure quantization (BEQ) problem which is the dual of the binary erasure channel (BEC) coding problem. First we describe the BEQ problem model in . Next, in we present our main result for the BEQ: duals of low density parity check codes can be analyzed, encoded, and decoded by dualizing the corresponding techniques for the BEC. Specifically, by dualizing capacity achieving codes for the BEC we obtain rate-distortion approaching codes for the BEQ. Finally, we close with some concluding remarks in .
Quantization Model {#sec:quantization-model}
==================
Vectors and sequences are denoted with an arrow (, $\genericS{x}$). Random variables or random vectors are denoted using the sans serif font (, $\genericRV{x}$ or $\genericRVS{x}$). We consider the standard (memoryless) data compression problem and represent an instance of the problem with the tuple $(\srcAlph,p_{\rvSrc}(\src),\dist{\cdot,\cdot})$ where $\srcAlph$ represents the source alphabet, $p_{\rvSrc}(\src)$ represents the source distribution, and $\dist{\cdot,\cdot}$ represents a distortion measure. Specifically, a source $\rvsSrc$ consists of a sequence of $n$ random variables $\rvSrcC{1}$, $\rvSrcC{2}$, $\ldots$, $\rvSrcC{n}$ each taking values in $\srcAlph$ and generated according to the distribution $p_{\rvsSrc}(\sSrc) =
\prod_{i=1}^n p_{\rvSrc}(\srcC{i})$. A rate $\rate$ encoder $\encoder{\cdot}$ maps $\rvsSrc$ to an integer in $\{1, 2, \ldots, 2^{n\rate}\}$, and the corresponding decoder $\decoder{\cdot}$ maps the resulting integer into a reconstruction $\rvsRecon$. Distortion between the source $\rvsSrc$ and the reconstruction $\rvsRecon$ is measured via $D =
\frac{1}{n}\sum_{i=1}^n \dist{\rvSrcC{i},\rvReconC{i}}$.
Shannon derived the minimum possible rate required by any data compression system operating with distortion $D$. The so-called rate-distortion function is given by the formula $$\label{eq:rd-formula}
R(D) =
\min_{p_{\rvRecon|\rvSrc}(\recon|\src):E[\dist{\rvSrc,\rvRecon}]
\leq D} I(\rvSrc;\rvRecon)$$ where $I(\cdot;\cdot)$ denotes mutual information and $E[\cdot]$ denotes expectation.
Binary Erasure Quantization
---------------------------
To highlight connections between error correction and data compression, we consider the binary erasure quantization (BEQ) problem where the source vector consists of ones, zeros, and “erasures” represented by the symbol $\erasure$. Neither ones nor zeros may be changed, but erasures may be quantized to either zero or one. Practically, erasures may represent source samples which are missing, irrelevant, or corrupted by noise and so do not affect the distortion regardless of the value they are assigned. Formally, the BEQ problem with erasure probability $e$ corresponds to
$$\begin{aligned}
\srcAlph &= \{0,1,\erasure\}\\
p_{\rvSrc}(\src) &= \frac{1-e}{2} \cdot\delta(\src) + \frac{1-e}{2}
\cdot\delta(\src-1) + e \cdot\delta(\src-\erasure)\\
\dist{a,b} &=
0 \textnormal{ if } a = \erasure \textnormal{ or } a = b,
\textnormal{ and } 1 \textnormal{ otherwise.}\end{aligned}$$
It is straightforward to show that for $D=0$ the distribution $$p_{\rvRecon|\rvSrc}(\recon|\src) =
\delta(\recon-\src) \textnormal{ if } \src \in \{0,1\},
\textnormal{ and }
\frac{1}{2} \cdot\delta(\recon) + \frac{1}{2}
\cdot\delta(\recon-1) \textnormal{ if } \src = \erasure.$$ optimizes and yields the value of the rate-distortion function at $D=0$: $$\label{eq:bec-rd-formula}
R_{\mathrm{BEQ}}(D=0) = 1-e.$$
Codes For Erasure Quantization {#sec:codes-eras-quant}
==============================
It is well-known that the encoder for a quantizer serves a similar function to the decoder for an error correcting code in the sense that both take a vector input (, a source to quantize or channel output to decode) and map the result to bits (, the compressed source or the transmitted message). The decoder for a quantizer can similarly be identified with the encoder for an error correcting code in the sense that both take bits as input and produce a vector (, a source reconstruction or a channel input). Thus it is natural to investigate whether swapping the encoder and decoder for a good error correcting code such as a low density parity check (LDPC) code produces a good quantizer.
LDPC Codes Are Bad Quantizers
-----------------------------
One benefit of studying the BEQ problem is that it demonstrates why low density parity check (LDPC) codes are inherently unsuitable for quantization. Specifically, consider an LDPC code like the one illustrated in using Forney’s normal graph notation [@it:2001:forney]. If all the variables connected to a given check are not erased, then there is an even chance that no code symbol can match the source in that position and thus the distortion will be positive regardless of the code rate. Thus, as stated in and proved in , successful decoding is asymptotically unlikely unless the density of [*every*]{} parity check matrix for the code increases logarithmically with the block length.[^1]
\[th:ldpc-codes-bad\] Let $\cbook_{(n)}$ be a sequence of linear codes of length $n$ and fixed rate $R$ such that the probability that binary erasure quantization using $\cbook_{(n)}$ of a random source sequence with $e \cdot n$ erasures will succeed with zero distortion is bounded away from 0 as $n \rightarrow \infty$. Then regardless of the values of $R$ and $e$, the degree of the parity-check nodes in any parity-check graph representation of $\cbook_{(n)}$ must increase at least logarithmically with $n$.
The poor performance of LDPC codes for quantization may seem surprising in light of their excellent properties in channel coding, but it has long been recognized that good codes for error correction and quantization may be different. The former is essentially a packing problem where the goal is to place as many codewords as possible in a given space such that the codewords are far apart and can be distinguished despite noise. The latter is a covering problem where the goal is to place as few codewords as possible in a given space such that every point in space is near at least one codeword.
From any good error correcting code (respectively data compression code) it is easy to obtain another code which is almost as good at error correction (resp. data compression) but terrible at source coding (resp. channel coding). For example, removing half the codewords has an asymptotically negligible effect on error correction since it only decreases the rate by $1/n$ and only increases robustness. But, removing half the codewords can dramatically hinder source coding since half the time the source may be very far from the nearest codeword. Conversely, doubling the number of codewords has an asymptotically negligible effect on data compression since the rate only increases by $1/n$ while the distortion may decrease slightly. But doubling the number of codewords can be catastrophic for error correction if it drastically reduces the distance between codewords.
Dual LDPC Codes
---------------
Many researchers have explored duality relationships between error correction and source coding. Such work demonstrates that often a good solution for one problem can be obtained by dualizing a good solution to the other. Continuing in this tradition, we study the properties of dual LDPC codes for binary erasure quantization.
Formally, a length $n$ binary linear code $\cbook$ is a subspace of the $n$ dimensional vector space over the binary field and the dual code ${\cbook^{\perp}}$ is the subspace orthogonal to $\cbook$. For LDPC codes, the code $\cbook$ is usually specified by the parity check matrix $\mat{H}$ representing the constraint that $\mat{H}\genericS{x}^{T} =
0$ if and only if $\genericS{x}$ is a codeword. To obtain the dual code ${\cbook^{\perp}}$ we can recall that the generator ${\mat{G}^{\perp}}$ of ${\cbook^{\perp}}$ is exactly $\mat{H}$. If the code $\cbook$ is represented by a normal graph as in , then the graph of the dual code $\cbook$ can be obtained by swapping $+$ and $=$ nodes [@it:2001:forney]. In dualizing the code graph in this manner it may be useful to note that while the graph of $\cbook$ obtained from $\mat{H}$ represents a syndrome former for $\cbook$, the dualized graph represents an encoder for ${\cbook^{\perp}}$.
For example, is obtained by dualizing the code graph in . Notice that while the original code cannot quantize the source with distortion 0, the dual code can. Intuitively, the advantage of a low density encoder structure is that it provides a simple representation of a basis which can be used to construct the desired vector. In the following sections, we investigate the properties of dual LDPC codes for quantization with both optimal quantization and iterative quantization.
Optimal Quantization/Decoding and Duality
-----------------------------------------
The following theorem (proved in ) demonstrates the dual relationship between channel decoding and source quantization using optimal decoding/quantization algorithms.
\[thrm:dual-optim-quant\] A channel decoder for the code $\cbook$ can correctly decode every received sequence with the erasure pattern[^2] $\genericS{e}$ if and only if a quantizer for the code ${\cbook^{\perp}}$ can successfully quantize every[^3] source sequence with the erasure pattern $\genericS{e}^{\perp} = 1 -
\genericS{e}$.
From this result we immediately obtain the following Corollary.
Let $\cbook_{(n)}$ be a sequence of linear codes which achieves the capacity of a binary erasure channel with erasure probability $e$ using optimal decoding. The sequence ${\cbook^{\perp}}_{(n)}$ obtained by taking the duals of $\cbook_{(n)}$ achieves the minimum rate for $D=0$ for the BEQ with erasure probability $e^{\perp} = 1-e$ using optimal quantization.
The statement and proof of the two preceding results contain a curious duality between erased/known symbols in source coding and known/erased symbols in channel coding. A similar duality exists between a likelihood ratio, $\lambda$, and its Fourier transform $\Lambda = \frac{1 - \lambda}{1 +
\lambda}$ used in dualizing the sum-product algorithm [@it:2001:forney pp. 545–546]. Specifically, the Fourier transform maps known/erased likelihood ratios to erased/known likelihood ratios.
Iterative Decoding/Quantization and Duality
-------------------------------------------
In the following we first review the intuition behind iterative erasure decoding algorithms and describe the particular decoding algorithm we consider in . Next we outline the intuition behind a similar approach for iterative quantization and precisely describe our quantization algorithm in . Finally, we show that these algorithms are duals.
### Iterative Erasure Decoding
Many iterative message-passing decoding algorithms are essentially based on the following idea. The outgoing message on edge $i$ of a + node is the modulo-2 sum of all incoming messages (excluding edge $i$) with the proviso that if any incoming message (excluding edge $i$) is $*$ then the outgoing message is also $*$. For an = node, the outgoing message on edge $i$ is $*$ only if all other incoming messages are $*$, otherwise the outgoing message is the same as the known incoming message or messages. These message-passing rules can be interpreted as determining the outgoing message on edge $i$ by applying the following “sum” and “product” formulas to all other incoming messages. [^4]
$$\label{eq:bec-rules}
\begin{array}{c||c|c|c}
+ & 0 & 1 & * \\ \hline\hline
0 & 0 & 1 & * \\ \hline
1 & 1 & 0 & * \\ \hline
* & * & * & * \\
\end{array}
\textnormal{\normalsize{ \ \ and \ \ }}
\begin{array}{c||c|c|c}
\times & 0 & 1 & * \\ \hline\hline
0 & 0 & \textnormal{\#} & 0 \\ \hline
1 & \textnormal{\#} & 1 & 1 \\ \hline
* & 0 & 1 & * \\
\end{array}$$
It is well-known that such algorithms yield optimal decoding on a tree and also perform well on graphs with cycles provided appropriate scheduling and initialization rules are selected. Initializing all messages to $*$ and using sequential or parallel schedules are common choices. For the purpose of proving theorems, we consider a sequential schedule in the algorithm of .
### Iterative Erasure Quantization
The message-passing rules in can also be applied to the BEQ problem for graphs without cycles provided some form of tie-breaking is used. Specifically, some variables will receive erasure messages even after the algorithm has completed. Such variables can be arbitrarily chosen to be either 0 or 1 and still produce a valid quantization. For example in quantizing the source $(*,*,1)$ with a (3,2) single parity check code, both $(1,0,1)$ and $(0,1,1)$ are equally valid results and this tie can broken arbitrarily.
On a graph with cycles, however, generalizing this approach by initializing all unknown messages to $*$ usually fails. For example, on the dual of a Gallager code or a code like the one represented in such an initialization rule leads to all messages being erased at every step of the algorithm. To perform effective tie-breaking, we need to distinguish between variables which can be arbitrarily set to 0 or 1 and variables which have not yet received enough information to be determined.
One way to distinguish between these cases is to denote the former as erasures with the symbol $*$ and the latter as null messages with the symbol $\varnothing$, and initialize all messages to $\varnothing$. With this initialization, we can use the following “sum” and “product” rules: [^5]
$$\label{eq:beq-rules}
\begin{array}{c||c|c|c|c}
+ & 0 & 1 & * & \varnothing \\ \hline\hline
0 & 0 & 1 & * & \varnothing \\ \hline
1 & 1 & 0 & * & \varnothing \\ \hline
* & * & * & * & * \\ \hline
\varnothing & \varnothing & \varnothing & * & \varnothing
\end{array}
\textnormal{\normalsize{ \ \ and \ \ }}
\begin{array}{c||c|c|c|c}
\times & 0 & 1 & * & \varnothing \\ \hline\hline
0 & 0 & \textnormal{\#} & 0 & 0 \\ \hline
1 & \textnormal{\#} & 1 & 1 & 1 \\ \hline
* & 0 & 1 & * & \varnothing \\ \hline
\varnothing & 0 & 1 & \varnothing & \varnothing
\end{array}$$
Specifically, the outgoing message from a + node in a graph like is computed by combining incoming messages from all other edges with the $+$ rule. The outgoing message from an = node is computed by combining incoming messages from all other edges with the $\times$ rule. Whenever an = node has all incoming messages being $*$, the value of the node is arbitrary. This tie can be broken by arbitrarily choosing a value of 0 or 1 provided the tie is broken consistently. Essentially, the requirement of consistent tie-breaking can be interpreted as a constraint on the message-passing schedule: tie-breaking information for a given tie should be propagated through the graph before other ties are broken.
In order to provide a precise algorithm for the purpose of proving theorems, we consider the in based on applying the rules in with a sequential schedule and all tie-breaking collected into step 8.
### Iterative Algorithm Duality
Our main results regarding iterative quantization are the following three theorems stating that works correctly, can be analyzed in the same manner as the algorithm, and works quickly:
\[thrm:eq-works\] For any linear code with generator matrix $\mat{G}$, $(\mat{G},\genericS{z})$ either fails in step 5 or else returns $\genericS{w}$ such that $\genericS{w}\mat{G}$ matches $\genericS{z}$ in all unerased positions.
\[thrm:algs-fail-same\] Consider a linear code with parity check matrix $\mat{H}$ and its dual code with generator matrix ${\mat{G}^{\perp}}= \mat{H}$. The algorithm $(\mat{H},\genericS{y})$ fails in step 5 if and only if the algorithm $({\mat{G}^{\perp}},\genericS{z})$ fails in step 5 where $\genericS{y}$ has erasures specified by $\genericS{e}$ and $\genericS{z}$ has erasures specified by $\genericS{e}^{\perp} = 1 - \genericS{e}$.
\[thrm:e-quant-runs-quick\] The algorithm $(\mat{G},\genericS{z})$ runs in time $\bigO(n\cdot d)$ where $n$ is the length of $\genericS{z}$ and $d$ is the maximum degree of the graph corresponding to $\mat{G}$.
These results (proved in ) imply that the parallel structure between erasure decoding and erasure quantization allows us to directly apply virtually every result from the analysis of one to the other. For example, these theorems combined with the analysis/design of irregular LDPC codes achieving the capacity of the binary erasure channel [@it:2002:oswald] immediately yield the following Corollary:
There exists a sequence of linear codes which can be efficiently encoded and decoded that achieves the rate-distortion function for binary erasure quantization.
Concluding Remarks {#sec:concluding-remarks}
==================
In this paper we demonstrated how codes on sparse graphs combined with iterative decoding can achieve the Shannon limit for binary erasure quantization. The main contribution of our algorithm is in recognizing the role of tie-breaking, scheduling, and initialization in iterative quantization. The key insight in our analysis is the strong dual relationship between error correction and quantization for codes on graphs and their associated decoding/quantization algorithms (both optimal and iterative). We conjecture that the main task in designing iterative message-passing algorithms for more general quantization problems lies in designing appropriate tie-breaking, scheduling, and initialization rules for such scenarios and exploiting similar dual relationships to channel decoding.
Proofs {#app:proofs}
======
Consider quantizing a random source and choose some $c > 0$ and let $d$ be the smallest integer such that at least $c \cdot n$ parity checks have degree at most $d$. For each such parity check, the probability that all variables in the check are not erased is at least $(1-e)^d$. Hence the probability that the check cannot be satisfied is at least $(1/2) \cdot(1-e)^d$. Since there are $c \cdot n$ such checks, the probability that at least one check cannot be satisfied is $$\begin{aligned}
\Pr[\textnormal{encoding failure}] &\geq
1 - \left[1 - (1/2)\cdot\left(1-e\right)^d\right]^{c\cdot n} \\
&\geq
1 - \left[1 - \left(1/2-e/2\right)^d\right]^{c\cdot n} \\
&= 1 - \exp \left\{c\cdot n \ln \left[1 -
\left(1/2-e/2\right)^d\right] \right\} \\
& \geq
1 - \exp \left\{-c\cdot n \cdot
\left(1/2-e/2\right)^d \right\} \\
&= 1 - \exp \left\{- \exp \left[ \ln c + \ln n + d \ln
\left(1/2-e/2\right)\right] \right\}.\end{aligned}$$ Hence for the probability of decoding failure to become small, $d$ must grow at least logarithmically with $n$ for every $c > 0$. Note that this argument applies to any parity-check graph representation of the code.
We will show that unique channel decoding is possible if and only if the matrix equation $\mat{M} \genericS{x} = \genericS{y}$ has a solution (where $\mat{M}$ will be defined shortly). Similarly, we will show that source quantization is possible for every $\genericS{z}$ if and only if the matrix equation $\genericS{w} \mat{M} = \genericS{z}$ has a solution for every $\genericS{z}$. By demonstrating that both conditions are satisfied if and only if the same matrix $\mat{M}$ has rank $n$, we will prove the desired result.
Assume that all erasures occur in the last $\weight{\genericS{e}}$ positions (, $\genericS{e} = 0^{n-\weight{\genericS{e}}}\
1^{\weight{\genericS{e}}} $). [^6] This incurs no loss of generality since the coordinates of $\cbook$ can always be permuted accordingly and the theorem applied to the permuted code and its permuted dual code. Let $\genericS{x}$ represent the transmitted signal and let $\genericS{y}$ denote the received signal. Optimal decoding corresponds to finding a vector which is a codeword of $\cbook$ and consistent with the unerased received values. The requirement that $\genericS{x}$ is a codeword corresponds to the equation $\mat{H} \genericS{x} = 0$ where $\mat{H}$ is the parity check matrix of $\cbook$. The requirement that $\genericS{x}$ is consistent with the received unerased data corresponds to the equation $(\mat{I}_{n-\weight{\genericS{e}}}\ \mat{0}) \genericS{x} =
\vecIToJ{y}{1}{n-\weight{\genericS{e}}}$ where $\mat{I}_{t}$ represents a $t$-by-$t$ identity matrix and $\vecIToJ{y}{i}{j}$ represents the sub-vector $(y_i, y_{i+1}, \ldots, y_j)$. Thus successful decoding is possible if and only if the equation $$\label{eq:bec-dec-cond}
\left(\begin{array}{cc}
\mat{I}_{n-\weight{\genericS{e}}} & \mat{0}\\
\multicolumn{2}{c}{\mat{H}}
\end{array}\right)
\genericS{x} =
\left(\begin{array}{c}
\vecIToJ{y}{1}{n-\weight{\genericS{e}}}\\
0
\end{array}
\right)$$ has a unique solution. According to well-known properties of linear algebra, uniqueness is equivalent to the matrix in having full column rank (, rank $n$). Note that existence of a solution is guaranteed since a codeword was sent and no errors occurred.
Let $\genericS{z}$ represent the source to be quantized with erasure pattern $\genericS{e}^{\perp}$. Since we assumed that all erasures in $\genericS{e}$ occurred in the last $\weight{\genericS{e}}$ samples, the dual erasure pattern $\genericS{e}^{\perp}$ has all erasures occurring in the first $\weight{\genericS{e}^{\perp}}$ positions (, $\genericS{e}^{\perp} =
1^{n-\weight{\genericS{e}}}\ 0^{\weight{\genericS{e}}}$). Optimal decoding corresponds to finding a vector $\genericS{w}$ which is a codeword of $\cbook$ and consistent with the unerased received values. The former requirement corresponds to the equation $\genericS{v}{\mat{G}^{\perp}}= \genericS{w}$ where ${\mat{G}^{\perp}}$ is the generator matrix of ${\cbook^{\perp}}$ and $\genericS{v}$ is a binary vector of appropriate dimension. The latter requirement corresponds to the equation $\genericS{u}\left(\mat{I}_{\weight{\genericS{e}^{\perp}}}\ \mat{0}\right) +
\genericS{w} = \genericS{z}$ where $\genericS{u}$ is a binary vector chosen to ensure that the first $\weight{\genericS{e}^{\perp}}$ positions (, the erased positions) match regardless of $\genericS{w}$. Thus successful decoding is possible for every $\genericS{z}$ if and only if a solution exists for $$\label{eq:beq-dec-cond}
\left(\begin{array}{cc}
\genericS{u} & \genericS{v}
\end{array}
\right)
\left(\begin{array}{cc}
\mat{I}_{\weight{\genericS{e}^{\perp}}} & \mat{0}\\
\multicolumn{2}{c}{{\mat{G}^{\perp}}}
\end{array}\right)
= \genericS{z}$$ for every $\genericS{z}$. According to well-known properties of linear algebra, existence of a solution for every $\genericS{z}$ is equivalent to the matrix in having full column rank (, rank $n$). Note that uniqueness of a solution is neither guaranteed not required since quantization is successful if at least one solution is found.
Noting that ${\mat{G}^{\perp}}= \mat{H}$ and $\weight{\genericS{e}^{\perp}}
= n-\weight{\genericS{e}}$ completes the proof since these conditions imply that the matrices in and are identical.
For the algorithm to exit the while loop and reach step 8, every unerased element of $z_i$ must have been erased in step 3 and assigned a reserved message variable. After a variable is reserved all its checks must be erased. Since checks can never changed from erased to unerased, a reserved variable can never again be selected in step 2 and thus a variable can never be reserved more than once.
Thus after the while loop, each unerased position in $\genericS{z}$ has a corresponding reserved variable. Hence there exists an assignment of the message variables which result in a codeword matching $\genericS{z}$ in the unerased positions. This assignment could be computed via brute-force by solving he corresponding system of linear equations, but in we show that this step can be computed more efficiently.
The proof relies on the following invariant for steps 1–7 of both algorithms: $$\label{eq:iter-inv}
\forall j, \ \textnormal{$y_j$ is erased if and only if $z_j$ is
unerased}.$$ This condition is trivially true before the algorithm begins and forms the base case for a proof by induction. We assume that holds at iteration $i$ of steps 1–7 and show that it must also hold at iteration $i+1$.
First, implies that the outcome of step 1 is the same for each algorithm. Next, since ${\mat{G}^{\perp}}= \mat{H}$ the tests in step 1 and step 2 of $(\mat{H},\genericS{y})$ and $({\mat{G}^{\perp}},\genericS{z})$ yield the same result. Finally, at step 3, $y_j$ is unerased while $z_j$ is erased. Therefore, by induction, condition is true at every iteration and $(\mat{H},\genericS{y})$ fails at step 5 if and only if $({\mat{G}^{\perp}},\genericS{z})$ fails at step 5.
The while loop executes at most $n$ times. Therefore step 1 requires at most $\bigO(n)$ operations. Consider storing the number of variables with exactly one unerased check in a data structure which supports insertion and removal in constant time (, a hash table). We can initialize the data structure with $\bigO(d\cdot n)$ operations. Removing an element in steps 2 and 3 and updating the data structure to account for step 3 requires $\bigO(d)$ operations. Thus steps 1 through 8 require $\bigO(d\cdot n)$ operations and all that remains is to bound the running time of step 9.
Denote the first reserved variable by $v_{j(1)}$, the second reserved variable by $v_{j(2)}$ and so on to $v_{j(\weight{e})}$. As described in step 9, we first assign a value to $v_{j(\weight{e})}$ and work backward. Specifically, we set $v_{j(i)}$ to the modulo-2 sum of $z_{j(i)}$ and all message variables connected to $z_{j(i)}$ (except $v_{j(i)}$). This is possible for $z_{j(\weight{e})}$ since no other reserved variable could be connected to $z_{j(\weight{e})}$.[^7] Similarly, $z_{j(\weight{e}-1)}$ must be connected to only unreserved variables as well as perhaps to $v_{j(\weight{e})}$ and therefore a value can be determined for $v_{j(\weight{e}-1)}$. Thus, by induction we can determine every $v_j$.
Adding up the operations computed for each step yields a running time of $\bigO(d\cdot n)$.
#### Acknowledgment
\
The authors wish to thank G. D. Forney, Jr. for many helpful comments on the manuscript.
[^1]: We may expect the density of a code to increase as it approaches the [*capacity*]{} or [*rate-distortion function*]{}, but a code whose density also increases with [*block length*]{} seems undesirable.
[^2]: If symbol $i$ is erased (resp. unerased) then $e_i = 1$ (resp. $e_i = 0)$ in our notation for erasure patterns.
[^3]: Note that some source sequences (, the all zero sequence) can be successfully quantized using ${\cbook^{\perp}}$ regardless of $\genericS{e}^{\perp}$. Similarly, a system which decodes ambiguous received sequences to the all zero sequence may succeed even when many erasures occur. Thus to obtain the desired equivalence between correct decoding and successful quantization we define correct decoding (resp. successful quantization) as being able to deduce the transmitted codeword (resp. a codeword matching non-erased positions of the source) for every possible received sequence (resp. source) with the the erasure pattern $\genericS{e}$ (resp. $\genericS{e}^{\perp}$).
[^4]: In the product rule for erasure decoding, the symbol \# denotes a contradiction which is impossible if only erasures and no errors occurred.
[^5]: In the product rule for erasure quantization, the symbol \# denotes a contradiction. If a contradiction is generated then quantizing the given source with no distortion is impossible and the algorithm fails.
[^6]: We use $\weight{\genericS{a}}$ to denote the number of non-zero values in $\genericS{a}$ (, the weight of $\genericS{a}$) and $b^c \stackrel{\Delta}{=} \underbrace{(b\ b \ldots b)}_{\textnormal{c
times}}$.
[^7]: If $z_{j(\weight{e})}$ was connected to another reserved variable $v$, that would imply $v$ was reserved when connected to $z_{j(\weight{e})}$ which was unerased as well as $z$ which must also have been unerased. This contradicts step 2 in .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate experimentally that interaction between nonlocal solitons in nematic liquid crystals (NLC) can be controlled by the degree of nonlocality. For a given beam width, the degree of nonlocality can be modulated by changing the pretilt angle $\theta_0$ of NLC molecules through bias voltage $V$. As $V$ increases (so does $\theta_0$), the degree of nonlocality decreases. When the degree of nonlocality is below a critical value, the solitons behave in the way like their local counterpart, i.e., in-phase solitons attract while out-of-phase solitons repulse each other. Such a voltage-controlled interaction between the solitons can be readily implemented in experiments.'
author:
- Wei Hu
- Tao Zhang
- Qi Guo
- Li Xuan
- Sheng Lan
title: 'Nonlocality-controlled interaction of spatial solitons in nematic liquid crystals '
---
The interaction properties of two spatial optical solitons depend on the phase difference between them, their coherence,[@Stegeman1999; @TSKu2005] and the nonlinear nonlocality of the materials in which the solitons propagate. For a local Kerr-type nonlinearity, two coherent bright solitons attract (or repulse) each other when they are in-phase (or out-of-phase). On the other hand, if the solitons are mutually incoherent[@Anderson-pre-85] or the nonlinear nonlocality of the materials is strong enough,[@Snyder1997; @OL2002; @xie-oqe-05] the soliton interaction is always attractive, independent of the phase difference. Very recently, Ku et al.[@TSKu2005] demonstrated both theoretically and experimentally that the interaction of the solitons can be controlled by varying their total coherence. Thus, the two out-of-phase solitons can repulse or attract each other, depending on whether the coherence parameter is below or above a threshold. Similar dependence of the interaction behavior on nonlocality was also theoretically predicted by Rasmussen et al.[@Rasmussen2005] based on the (1+1)-Dimensional mode of the nematic liquid crystal (NLC), i.e., there exists a critical degree of nonlocality above which the two out-of-phase solitons will attract each other.
The NLC with a pretilt angle induced by an external low-frequency electric field has been confirmed[@PRL2003; @PRL2004] to be a typical material with the strongly nonlocal (referred also as highly nonlocal in some papers) nonlinearity.[@Snyder1997] In the previous works concerning the single soliton propagation[@PRL2003; @PRL2004; @APL2000] and the soliton interaction[@Rasmussen2005; @OL2002; @Fratalocchi-mclc-2004] in the NLC, however, the peak of the pretilt angle was always made to be $\pi/4$ in order to maximize the nonlinearity. As will be seen later, the degree of nonlocality can only be modulated by changing the beam width in this case, which is not convenient in practice. Recently, Peccianti et al. have shown that[@OL2005] the nonlocality can be varied by changing the pretilt angle via a bias voltage. In this letter, we use the (1+2)-D model with an arbitrary pretilt angle[@OL2005] to describe the (1+2)-D soliton interaction in the NLC. By defining a general characteristic length of the nonlinear nonlocality for the NLC, a voltage-controlled degree of nonlocality is shown to be achieved conveniently. In experiments, we observe the nonlocality-controllable (through the change of the bias voltage) transition from attraction to repulsion of the two out-of-phase solitons in the NLC.
Let us consider the (1+2)-D model of light propagation in a cell containing the NLC. The configuration of the cell and the coordinate system are the same as in the previous works.[@PRL2003; @PRL2004; @APL2000; @OL2005; @Fratalocchi-mclc-2004; @peccianti-jnopm-o3] The optical field polarized in $x$-axis with envelope $A$ propagates in $z$-direction. An external low-frequency electric field $E_{RF}$ is applied in $x$-direction to control the initial tilt angle of the NLC. The evolution of the paraxial beam $A$ and the tilt angle $\theta$ can be described by the system[@peccianti-jnopm-o3; @PRL2003] $$\begin{aligned}
2i k\partial_z A +\nabla_\perp^2 A + k_0^2 \epsilon_a^{op} \sin(\theta +
\theta_0) \sin(\theta-\theta_0) A = 0,\label{nlse0}
\\
2K \nabla_\perp^2\theta +\epsilon_0 \left(\epsilon_a^{RF} E_{RF}^2+
\epsilon_a^{op}
\frac{|A|^2}{2}\right)\sin(2\theta) = 0, \label{roe0}\end{aligned}$$ where $\theta_0$ is the peak-tilt of the NLC molecules in the absence of the light, $K$ is the average elastic constant of the NLC, $\nabla_\perp^2
=\partial_x^2 +\partial_y^2$, $k^2=k_0^2(n_\perp^2+
\epsilon_a^{op}\sin^2 \theta_0)$, $\epsilon_a^{op}=
n_\parallel^2-n_\perp^2$ and $\epsilon_a^{RF}=
\epsilon_0(\epsilon_\parallel-\epsilon_\perp$) characterize respectively the optical and low-frequency dielectric anisotropies with respect to the major axis of the NLC molecules (director). In Eq. (\[roe0\]), the term $\partial_{z}^2\theta$ has been canceled out because the dependence of $\theta$ on $z$ is proven to be negligible.[@Rasmussen2005; @peccianti-jnopm-o3] In the absence of the laser beam, the pretilt angle $\hat{\theta}$ depends only on $x$[@peccianti-jnopm-o3] $$2K \partial_x^2 \hat{\theta} + \epsilon_0 \epsilon^{RF}_a E_{RF}^2
\sin(2\hat{\theta}) = 0. \label{pretilt}$$
Furthermore, the system that describes $A$ and the optically induced angle perturbation $\Psi$ \[$\theta =\hat{\theta}+(\hat{\theta}/\theta_0)\Psi$\] can be simplified as[@peccianti-jnopm-o3; @OL2005] $$\begin{aligned}
2i k\partial_z A +\nabla_\perp^2 A + k_0^2 \epsilon_a^{op}
\sin(2 \theta_0) \Psi A = 0,\label{nlse1}
\\
\nabla_\perp^2\Psi -
%\epsilon^{RF}_a E_{RF}^2 \frac{\sin(2\theta_0)}{2K\theta_0} \left[1- 2 \theta_0 \cot(2\theta_0) \right]
\frac{1}{w_m^2} \Psi
+ \frac{\epsilon_0\epsilon_a^{op}}{4K}\sin(2\theta_0)|A|^2 = 0, \label{roe1}\end{aligned}$$where a parameter $w_m>0$ for $|\theta_0|\leq \pi/2$, which reads $$w_m(\theta_0) =\frac{1} {E_{RF}(\theta_0)}\left\{\frac{2\theta_0 K}
{\epsilon^{RF}_a \sin(2\theta_0)\left[1- 2 \theta_0 \cot(2\theta_0)
\right]}\right\}^{1/2}.\label{wm}$$ Introducing the normalization that $X=x/w_0$, $Y=y/w_0$, $Z=z/(2kw_0^2)$, $a=A/A_0$, and $\psi=\Psi/\Psi_0$, where $A_0=
4\sqrt{\pi K/\epsilon_0}/k_0\epsilon_a^{op} w_0^2$, $\Psi_0=
\sin(2\theta_0)/k_0^2w_0^2\epsilon_a^{op}$, and $w_0$ the initial beam width, we have the dimensionless system $$\begin{aligned}
i\partial_Z a + \nabla_{XY}^2 a +\gamma \psi a=0, \label{nlse2}
\\
\nabla_{XY}^2\psi -\alpha^2 \psi +4\pi|a|^2=0, \label{roe2}\end{aligned}$$ where $\nabla_{XY}^2 =\partial_X^2 +\partial_Y^2$, and $$\label{alpha}
\gamma=\sin^2(2\theta_0), ~~\alpha=w_0/w_m.
%&=& \epsilon^{RF}_a E_{RF}^2 w_0^2 \frac{\sin(2\theta_0)} {2\theta_0 K}
% \left[1- 2 \theta_0 \cot(2\theta_0) \right].$$ For a symmetrical geometry,[@explain-1] Eq. (\[roe2\]) has a particular solution $\psi(x,y) =(4\pi/\alpha^2)\int
R(x-x^\prime,y-y^\prime)|a(x^\prime,y^\prime)|^2 d x^\prime d
y^\prime$, and $R(x,y) =(\alpha^2/2\pi)K_0(\alpha\sqrt{x^2+y^2})$, where $K_0$ is the zero-th order modified Bessel function.
![(a) The characteristic length $w_m$ and the pretilt angle $\theta_0$ of the NLC versus the bias voltage $V$. (b) The characteristic length $w_m$ (a solid curve) and the critical power of a single soliton (circles) versus the pretilt angle $\theta_0$. The parameters are for a $80\mu m$-thick cell filled with the NLC (TEB30A).[@explain-4] \[1\]](fig-1.eps){width="8cm"}
We define $w_m$ in Eq. (\[wm\]) as the general characteristic length of the nonlinear nonlocality for the NLC,[@explain-2] then it is obvious that the factor $\alpha$ in Eq. (\[alpha\]) indicates the degree of nonlocality, as defined in Ref. for the specific case of Gaussion response function. A monotonous function of $\theta_0$ on $E_{RF}$ is given by Eq. (\[pretilt\]), and it can be approximated as[@OL2005] $\theta_0\approx(\pi/2)[1-(E_{FR}/E_{RF})^3]$ when $E_{RF}$ is higher than the Fre$\grave{e}$derichsz threshold $E_{FR}$. Therefore, we can clearly observe from Eq. (\[wm\]) that $w_m$ is determined only by $E_{RF}$ (or by the bias $V$), or equivalently by the peak-pretilt angle $\theta_0$ for a given NLC cell configuration, as shown in Fig.\[1\]. When the bias is properly chosen so that $\theta_0=\pi/4$,[@PRL2003] $w_m$ is fixed and $\alpha$ can be modulated only by changing $w_0$. This is the case discussed in Ref.. For a given beam width, however, $\alpha$ can be changed continuously by $w_m$ through continuously varying the bias. As a result, the voltage-controlled degree of nonlocality through the medium of the pretilt $\theta_0$ of the NLC can be achieved conveniently. With the decrease of the bias, $\theta_0$ goes from $\pi/2$ to 0, then $\alpha$ varies from $\infty$ to 0 for a fixed $w_0$ and the degree of nonlocality increases from locality to strong nonlocality. In addition, we can see the factor $\gamma$ in Eq. (\[alpha\]) stands for the nonlinear couple between $A$ and $\Psi$. It reaches maximum when $\theta_0=\pi/4$, which takes major responsibility for the lowest critical power[@Snyder1997; @cao-wlxb-05] of a single soliton nearby $\theta_0=\pi/4$, as shown in Fig.\[1\](b). As the pretilt angle approaches $\pi/2$, the critical power increases sharply, while $w_m$ decreases to zero and $\alpha$ increases to $\infty$ for the given $w_0$ (the degree of nonlocality decreases, moving towards locality).
![Numerical simulation results of the interactions of the in-phase and the out-of-phase solitons. The width for each soliton is $4\mu$m, and the input power is $1.1m$W. The separation and the relative angle between two solitons are respectively $12\mu$m and $0.57^o$($\tan 0.57^o =0.01$). \[2\]](fig-2.ps){width="8cm"}
To show the influence of the pretilt angle $\theta_0$ (or equivalently the degree of nonlocality for a fixed $w_0$) on the interaction between the two solitons, we have carried out numerical simulations directly based on Eqs. (\[nlse0\])-(\[pretilt\]). The simulation results show that there exist critical values for the degree of nonlocality below (or above) which two out-of-phase solitons will repulse (or attract) each other. The critical values depend on the initial separation and relative angle in the ($y,z$)-plane between the solitons. These results agree with the prediction based on the (1+1)-D model.[@Rasmussen2005] The critical degree of nonlocality for the two parallel solitons is very weak so that the corresponding critical pretilt angle $\theta_{0c}$ is very close to $\pi/2$, leading to a very high critical power for the soliton state \[see Fig. \[1\](b)\]. However, the use of a relative angle will significantly increase the critical degree of nonlocality and make $\theta_{0c}$ not be close to $\pi/2$. Hence, the critical powers for different pretilt angles around $\theta_{0c}$ do not differentiate too much. This makes it possible to observe the soliton states at a fixed input power for different pretilt angles (or different bias voltages).[@explain-3] Figure \[2\] presents the simulation results of two solitons with a relative angle of $0.57^o$ for different values of $\theta_0$. We can see that for $\theta_0 \leq \pi/4$, the nonlocality is strong enough to guarantee the attraction of both the in-phase and the out-of-phase solitons. However, the degree of nonlocality becomes lower than the critical degree of nonlocality when $\theta_0=0.45\pi$. In this case, the out-of-phase solitons begin to repulse each other and the in-phase solitons remain attraction.
![Scheme of the experimental setup. NA, neutral attenuator; BS, beam splitters; M, plate mirror; PP, parallel-face plate for adjusting the phase difference; O, 10$\times$ microscope objective; LC, liquid crystal cell; MS, microscope; F, laser-line filter; BP, beam profiler. \[3\]](fig-3.eps){width="8cm"}
The experimental setup is illustrated in Fig. \[3\]. The laser beam from an argon-ion laser is split into two beams, then they are combined together with a small separation through the other beam-splitter and launched into a $80\mu m$-thick NLC cell by a 10$\times$ microscope objective. The beam width at the focus $w_0$, the separation $d_s$ and relative angle $\beta$ between the two beams are measured by an edged-scanning beam profiler when the NLC cell is removed. The phase difference between the two beams is adjusted by the rotation of a $1.8mm$-thick parallel-face plate, and measured through the interference pattern by the beam profiler located on the other branch after the second beam-splitter. The cell is filled with the NLC TEB30A (from SLICHEM China Ltd.), whose $n_\parallel = 1.6924$, $n_\perp = 1.5221$, $K\approx 10^{-11}N$, $\epsilon_a^{op}=0.5474$, and $\epsilon_a^{RF}=9.4\epsilon_0$. The Fre$\grave{e}$dericksz threshold $V_t \approx 1.14$V for the $80\mu
m$-thick cell. The launched power for each beam is fixed to 7mW when the bias is changed, and the other parameters for the beams in the NLC are $w_0=3.2\mu$m, $d_s=10\mu$m, and $\tan\beta=0.011$. When the phase difference is adjusted to $0$ or $\pi$, we record the beam traces for the different biases by the CCD camera, as shown in Fig.\[4\].
![Photos of the beam traces for the soliton pair propagation in the NLC cell. The biases applied on the LC are $1.0$V, $1.4$V and $2.4$V, corresponding to the pretilt angles of $0.01\pi$, $0.25\pi$, and $0.45\pi$, respectively. \[4\]](fig-4.eps){width="10cm"}
Let us compare the photos for the in-phase and the out-of-phase solitons when the bias $V=1.4$V ($\theta_0 \approx \pi/4$). They are almost the same for both cases. It means for $\theta_0 = \pi/4$ the degree of nonlocality is strong enough to eliminate the dependence of the interactions on the phase difference between the solitons. In this case, $w_m \approx 25.3\mu$m, which is bigger than the separation of the two beams, and $\alpha =0.126$ for the $3.2\mu$m-width solitons.
For the bias $V=1.0$V sightly lower than the threshold $V_t = 1.14$V (a small tilt angle in the sample educes some reorientation at the bias voltage lower than the threshold[@OL2005]), the pretilt angle $\theta_0$ is nearly zero and nonlocality is much stronger than that when $V=1.4$V. For this reason, a second cross point is observed for both the in-phase and the out-of-phase solitons.
When the bias $V$(pretilt angle $\theta_0$) increases, the degree of nonlocality $1/\alpha$ and the characteristic length $w_m$ decrease. For $V=2.4$V ($\theta_0 \approx 0.45\pi$), we have $w_m \approx
11\mu$m, which approximately equals to the separation between the two solitons. In this case, we observe the attraction of the in-phase solitons and the repulsion of the out-of-phase solitons. We also see the two in-phase solitons fused into one soliton, which is qualitatively same with the numerical simulation result in Fig. \[2\].
In conclusion, we have investigated theoretically and experimentally the interactions of the nonlocal spatial solitons in the NLC when the applied bias is adjusted. Given is a general definition of the characteristic length of the nonlinear nonlocality for the NLC, which is the function of the bias through the medium of the pretilt angle. Hence, the voltage-controllable degree of nonlocality in the NLC can be implemented expediently. We experimentally observe the transition from attraction to repulsion of the two out-of-phase solitons in the NLC as the degree of nonlocality decreases via increasing the bias. Such a voltage-controllable soliton interaction might have its potential applications in developing all-optical signal processing devices.
This work was supported by the Natural Science Foundation of Guangdong Province of China (Grant No. 04105804), and partially supported by the National Natural Science Foundation of China (Grant Nos. 60278013, 10474023, and 60277033).
[Reference]{} About the dependence on the phase-difference and the coherence, see, G. I. Stegeman and M. Segev, Science [**286**]{}, 1518 (1999) and references therein.
T.S. Ku, M.-F. Shih, A. A. Sukhorukov, and Y. S. Kivshar, , 063904 (2005).
D. Anderson and M. Lisak, , 2270 (1985).
A. W. Snyder and D. J. Mitchell, Science [**276**]{}, 1538 (1997).
M. Peccianti, K. Brzdakiewicz, and G. Assanto, 27, 1460 (2002).
Y. Xie and Q. Guo, Opt. Quant. Electron. [**36**]{}, 1335 (2004).
P. D. Rasmussen, O. Bang, and Wieslaw Kr$\acute{o}$likowski, , 066611(2005).
C. Conti, M. Peccianti, and G. Assanto, , 073901(2003).
C. Conti, M. Peccianti, and G. Assanto, , 113902(2004).
M. Peccianti, A. De Rossi, G. Assanto, A. De Luca, C. Umeton, and I. C. Khoo, , 7(2000).
A. Fratalocchi, M. Peccianti, C. Conti, and G. Assanto, Mol. Cryst. Liq. Cryst. [**421**]{}, 197 (2004).
M. Peccianti, C. Conti, and G. Assanto, , 415(2005).
M. Peccianti, C. Conti, G. Assanto, A. De Luca, and C. Umeton, J. Nonl. Opt. Phys. Mat. [**12**]{}, 525 (2003).
When the optical beam is small enough, the bourndary efffect of the NLC cell can be neglected, and the configuration can approximate symmetry.
The charateristic length $w_m$ defined in Eq. (\[wm\]) is general. When $\theta_0=\pi/4$, $w_m$ defined by us will be reduced to the charateristic value $R_c$ defined in Refs. and .
Q. Guo, B. Luo, F. Yi, S. Chi, and Y. Xie, Phys. Rev. E [**69**]{}, 016602 (2004); N. Cao and Q. Guo, Acta Phys. Sin. [**54**]{}, 3688 (2005).
The bias $V \approx
E_{RF}(d+\Delta)$,[@OL2005] where $d$ is the NLC cell thickness and $\Delta$ is taken to approximately be $4\mu$m to consist with experimnental data for our sample. $w_m(\theta_0)$ and $w_m(V)$ are from Eq. (\[wm\]). The critical power is a numerical result of Eqs. (\[nlse0\])-(\[pretilt\])
Rigorously speaking, except one of them, they are quasi-solitons (breathers) rather than solitons because the fixed input power can only equal exactly one of the critical powers for different biases, but approximate the others in the case under consideration.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
S. Beaulieu$^{1,2^\ast}$, A. Comby$^{1}$, A. Clergerie$^{1}$, J. Caillat$^{3}$, D. Descamps$^{1}$,\
N. Dudovich$^{4}$, B. Fabre$^{1}$, R. Géneaux$^{5}$, F. Légaré$^{2}$, S. Petit$^{1}$, B. Pons$^{1}$,\
G. Porat$^{4}$, T. Ruchon$^{5}$, R. Taïeb$^{3}$, V. Blanchet$^{1}$ & Y. Mairesse$^{1}$\
\
\
\
\
\
\
\
title: 'Attosecond-resolved photoionization of chiral molecules'
---
Chiral light-matter interactions have been investigated for two centuries, leading to the discovery of many chiroptical processes used for discrimination of enantiomers. Whereas most chiroptical effects result from a response of bound electrons, photoionization can produce much stronger chiral signals that manifest as asymmetries in the angular distribution of the photoelectrons along the light propagation axis. Here we implement a self-referenced attosecond photoelectron interferometry to measure the temporal profile of the forward and backward electron wavepackets emitted upon photoionization of camphor by circularly polarized laser pulses. We found a delay between electrons ejected forward and backward, which depends on the ejection angle and reaches 24 attoseconds. The asymmetric temporal shape of electron wavepackets emitted through an autoionizing state further reveals the chiral character of strongly-correlated electronic dynamics.
Bolts and nuts are amongst the most common chiral objects in our macroscopic world. Their chiral nature is used to convert rotation to directional translation: rotating the nut on a bolt induces its translation forward or backward, depending on the rotation direction. A very similar effect occurs in the microscopic world when enantiopure chiral molecules are photoionized by circularly polarized radiation [@powis08]. The ejected photoelectrons tend to go forward or backward relative to the light propagation axis, depending on the helicity of the ionizing light and the handedness of the molecules [@ritchie76; @bowering01]. As a result, the photoelectron angular distribution shows an asymmetry, called photoelectron circular dichroism (PECD). PECD is one of the most sensitive probes of static [@nahon15] and dynamical [@comby16; @beaulieu16-1] molecular chirality, producing signals that are up to two orders of magnitude larger than most circular dichroisms. From a classical point of view, PECD can be seen as the result of the combined action of the chiral molecular potential and the circular ionizing electric field on the outgoing electron trajectories [@beaulieu16]. Quantum mechanically, it arises from the interference between partial waves of different parity constituting the outgoing photoelectron wavepacket [@powis08]. Both interpretations show that subtle differences in the ionization dynamics can have dramatic consequences. Consequently, PECD has been proposed as a possible hypothesis to explain the homochirality of terrestrial life [@meierhenrich08]: the asymmetric electron ejection induces an asymmetric recoil of the ions, which can lead to enantiomeric separation when accumulated over hundreds of millions of years [@tia13].
![Schematic view of the two timescales of photoionization of two camphor enantiomers. (1S)-(-)-camphor is shown in (**A**) and (**B**), (1R)-(+)-camphor in (**C**) and (**D**). In direct photoionization (A), (C) the forward (f) and backward (b) electron wavepackets may be delayed by a few attoseconds because of the asymmetric scattering of the outgoing electron in the chiral molecular potential. In the case of autoionization (B),(D) the dynamics of the autoionizing state can lead to different temporal structures of photoelectron wavepackets in the forward and backward directions, on the femtosecond timescale.[]{data-label="fig0"}](Figure0.pdf){width="80.00000%"}
Photoionization was considered to be instantaneous from an experimental point of view until attosecond technology made it possible to measure the underlying ultrafast electron dynamics. Delays of a few attoseconds were measured between electrons originating from different atomic orbitals [@klunder11], from distinct bands of a solid [@cavalieri07], associated with different vibrational states of a molecular ion [@haessler09] or from differing spin-orbit states [@jordan17]. The direction of the electron emission also influences the photoionization dynamics: delays have been observed between electrons ejected at different angles [@heuser16; @hockett16; @baykusheva17] or from different sides of an asymmetric molecule [@chacon14; @cattaneo16-1].
The photoionization process involves more complicated dynamics when autoionization occurs. In that case, the photoabsorption promotes the system into a metastable bound state which is coupled to equienergetic continuum states through configuration interaction. This coupling leads to autoionization of the metastable state. The interference of direct and indirect photoionization channels produces characteristic (Fano) spectral profiles [@fano61], associated with complex temporal dynamics. Pioneering attosecond experiments in rare gases recently showed the possibility of measuring the buildup of Fano lineshapes in the temporal domain [@kaldun16] as well as the spectral phase across the resonance [@kotur16; @gruson16], allowing the reconstruction of the temporal profile of the electron wavepacket [@gruson16].
![Principle and implementation of photoelectron interferometry. (**A**) Schematic experimental setup. Two phase-locked femtosecond laser pulses with linear or circular polarization are focused into a jet of enantiopure camphor molecules in the interaction zone of a velocity map imaging (VMI) spectrometer. The photoelectrons are accelerated by a set of electrodes onto microchannel plates imaged by a phosphor screen and a CCD camera, which records the 2D projection of the 3D angular distribution of the photoelectron spectrum with an energy resolution around 0.2 eV at 2 eV. (**C**) Ionization scheme. The 400 nm pulse (40 fs duration, intensity $\sim 5\times 10^{12}$ W/cm$^2$) induces multiphoton above-threshold ionization of the molecules. The first ionizing transition lies in the vicinity of an autoionizing resonance. The 800 nm pulse (30 fs, intensity $\sim 5\times 10^{11}$ W/cm$^2$) produces additional transitions, leading to the creation of sidebands between the ATI peaks. (**B**) Typical measured photoelectron angular distribution and evolution of the second sideband as a function of delay between the two fields. The oscillations in the upper and lower half of the distribution are out of phase, reflecting the up-down asymmetry of the total ionizing electric field. (**D**) Angle-integrated photoelectron spectrum, constituted of ATI peaks and n-th order sidebands (SBn). The full blue lines are the angularly-integrated spectrally-resolved $2\omega$-oscillation phases for each sideband and the red dots are the spectrally-averaged $2\omega$-oscillation phases for non-resonant SB2 and SB3.[]{data-label="fig1"}](Figure1.pdf){width="80.00000%"}
Here, we aimed to determine whether or not the electrons ejected forward and backward from a sample of randomly oriented enantiopure chiral molecules were temporally synchronized. Answering this question, for both direct (Fig. \[fig0\](A,C)) and indirect (Fig. \[fig0\](B,D)) photoionization pathways, is a challenging task. Up to now, attosecond delay measurements in the gas phase have been conducted on rare gas atoms or di- and triatomic molecules, which were used as benchmark systems. These experiments have revealed a strong influence of the weak probe field on the outcome of the measurement. It is thus necessary to perform accurate theoretical calculations to calibrate these measurement-induced delays [@klunder11; @ossiander17]. Such theoretical calculations for large and low-symmetry chiral molecules (*e.g.* camphor, C$_{10}$H$_{16}$O) are currently far from reach.
In order to directly access the delays between forward and backward electron emission, without any measurement-induced effects, we implemented a self-referenced photoelectron interferometry technique using photoionization by two phase-locked laser fields to detect differential attosecond photoionization delays with a resolution of 2 as. By independently controlling the chirality of the ionizing and probe light pulses, we fully decoupled the intrinsic photoionization delays from the measurement-induced delays.
#### Photoelectron interferometry {#photoelectron-interferometry .unnumbered}
When an intense femtosecond laser pulse ionizes an atom or a molecule, multiple photons can be absorbed above the ionization threshold (Above-Threshold Ionization, ATI). In the spectral domain, ATI produces a comb of photoelectron peaks separated by the laser photon energy [@agostini79]. Each peak is characterized by a spectral width $\delta \omega$, and the overall ATI spectrum extends over a width $\Delta \omega$. In the time domain, the ATI process leads to the emission of attosecond electron bursts (Fig. \[fig0\](A-C)) of duration $\delta t$, which form a train. The overall duration of the train $\Delta t$ is set by the laser pulse duration, typically a few tens of femtoseconds. The ATI emission can last longer if an autoionizing state is populated: the lifetime of the autoionizing state increases the electron wavepacket duration (Fig. \[fig0\](B-D)). Characterizing the temporal dynamics of the ionization process requires measuring the process on two timescales $\delta t$ and $\Delta t$. The femtosecond structure of the wavepacket is encoded in the spectral intensity and phase within the bandwidth of each ATI peak: $\Delta t$ is related to $\delta \omega$. On the other hand, the attosecond sub-structures are encoded in the relative amplitude and phase between the different ATI peaks: $\delta t$ is associated to $\Delta \omega$. In order to have a complete picture of the temporal dynamics of the ionization process, it is thus necessary to measure the spectral phase of the ATI peaks, both within their bandwidth ($\delta \omega$) and from one peak to the next ($\Delta \omega$). This is possible using photoelectron interferometry [@paul01-1; @zipp14; @gruson16].
We first present the basic concepts of photoelectron interferometry and highlight the rich spectroscopic information that it provides about the ionized target. For now, we are leaving aside the chiral character of the experiment (by integrating over the photoelectron ejection angles). The principle of the measurement is described in Fig. \[fig1\]. The target molecule we chose was camphor, a bicyclic ketone which has been extensively studied in PECD experiments performed in single-photon [@powis08; @nahon16], in multiphoton [@lux12; @lehmann13] and in ATI regimes [@lux15]. Camphor has a first ionization potential of 8.76 eV and Rydberg states starting around 6.2 eV: it is ionized by $2+n$ resonance-enhanced multiphoton transitions when using a 400 nm ultraviolet (UV) field, with $n$ being the order of the ATI peak. In the presence of a weak infrared (IR) 800 nm field (frequency $\omega$), new peaks, called sidebands, appear between the ATI comb. Two quantum paths lead to the same sideband: addition of an IR photon to ATI peak $n$, or subtraction of one IR photon from ATI peak $n+1$. These two paths interfere and the sideband amplitude oscillates as a function of the relative delay between the UV and IR fields, at $2\omega$ frequency [@schumacher94; @zipp14; @gong17]. The phase of the sideband oscillations encodes the relative phase between the two neighboring ATI peaks and thus the temporal properties of the photoemission process.
The photoelectrons were collected by a velocity map imaging spectrometer (VMI), which measures the angle-resolved photoelectron spectrum (Fig. \[fig1\](A,B)). The superposition of 800 nm and 400 nm pulses produces an electric field which is stronger in the upper or lower direction, depending on the relative delay between the two fields. As a consequence, the electrons ejected up and down are modulated in opposite phase [@zipp14; @skruszewicz15] (Fig. \[fig1\](B)). We measured the sideband phase independently on the angularly-integrated upper and lower half of the photoelectron image and averaged the phase obtained from the upper half with the $\pi$-shifted phase obtained from the lower half. More details about the experimental setup, raw VMI images and their inversion [@garcia04] as well as data analysis are given in Supplementary Material (SM).
Figure \[fig1\](D) shows the sideband oscillation phase $\varphi_{2\omega}$ as a function of the photoelectron kinetic energy $E$. The first sideband (SB1), which encodes the phase difference between ATI peaks 1-2, presents an abrupt $\pi$-phase jump around 1.9 eV. This is the signature of a resonance associated with one or the other of the two contributing ATI peaks. The second and third sidebands, which are built respectively upon ATI peaks 2-3 and 3-4, have a smooth phase variation across their bandwidth (*i.e.* without any trace of resonance). We can thus conclude that the resonance occurs in the formation of the first ATI peak (ATI1), and does not propagate to the higher ATI peaks, as confirmed by theoretical calculations of resonant photoelectron interferometry presented in the SM.
We proceeded to investigate the chiral (enantio-specific) photoionization dynamics in the two different regimes identified above: direct – attosecond – ionization (SB2 and SB3), and indirect – femtosecond – ionization in the vicinity of an autoionizing resonance (SB1).
#### Attosecond delays in non-resonant photoionization {#attosecond-delays-in-non-resonant-photoionization .unnumbered}
We started by analyzing the direct photoionization dynamics which occur on the attosecond timescale and are encoded in the relative phase between the different ATI peaks. It can be obtained by extracting the oscillation phases of the signals averaged over the bandwidth of each sideband, $\varphi_{2\omega}(\overline{E})$. Neglecting the variations over the spectral width is equivalent to assuming that the photoionization process is strictly periodic from one laser cycle to the next. The spectral homogeneity of the sideband phases shown in Fig. \[fig1\](D) indicates that this assumption is reasonable for SB2 and SB3, but not for the resonant SB1, which is discussed later. This scheme is similar to the conventional RABBIT analysis (Reconstruction of Attosecond Beatings By Interference of Two-photon transitions) [@paul01-1], here extended to the case of multiphoton ionization [@zipp14; @gong17]. Simulations presented in the SM validate the analogy between the two techniques.
![Forward/backward differential delays in non-resonant photoionization of camphor. (**A**) When the UV is linearly polarized and the IR is left-circularly polarized, the differential delay $ \Delta \tau^{f/b}_{cc}$ is zero for SB2 and SB3. (**B**) Resolving angularly the differential delay shows that it reaches non-zero values for electrons ejected close to the laser polarization plane (for SB2). The inset depicts the angular distribution of the normalized SB2 intensity. (**C**) When the UV field is circularly polarized and the IR one is linearly polarized, the differential delay $ \Delta \tau^{f/b}_{W}$ is non zero for SB2. (D) Angular resolution of the differential delays from SB2. Error bars are defined as the 95$\%$ confidence interval.[]{data-label="fig2"}](Figure2.pdf){width="80.00000%"}
Measuring the phase $\varphi_{2\omega}(\overline{E})$ is equivalent to measuring the time delay $\tau$ that maximizes the signal of each sideband. This delay can be decomposed as the sum of three contributions, that reflect the three steps of the sideband creation [@dahlstrom12]:
$$\tau = \tau_{light} + \tau_{W} + \tau_{cc}
\label{eq}$$
The ionization is triggered by absorption of light at a well defined time $\tau_{light}$. Next, the electron scatters in the molecular potential and acquires a delay $\tau_{W}$ – the Wigner delay [@wigner55]. The electron also interacts with the weak IR field, which induces continuum-continuum transitions from the main ATI peaks to the sidebands, introducing an additional delay $\tau_{cc}$. Whereas $\tau_{light}$ and $\tau_{cc}$ are induced by the measurement, $\tau_{W}$ is intrinsic to the probed system and is the physical quantity of interest. It represents the delay between an electron scattering in a given potential and in a reference potential, as introduced by Wigner in 1955 [@wigner55].
To resolve the enantiosensitivity of Wigner delays, we turned to chiroptical measurements comparing the sideband oscillation phases for electrons emitted in the forward vs. backward directions, and extracting the difference $\Delta \tau^{f/b} = \tau^{f} - \tau^{b} $. This procedure naturally eliminates $\tau_{light}$, which is strictly common to the forward and backward electrons. Further decoupling is achieved by using different combinations of linearly and circularly polarized light. Indeed, the forward/backward (f/b) symmetry can only be broken by the chiral nature of the interaction, that is, if a circularly polarized light pulse is used. We can thus selectively break the f/b symmetry only in the ionization step by using a circularly polarized UV field and a linear IR field. In that case $\Delta \tau^{f/b}_{cc}$ = 0 and $\Delta \tau^{f/b}=\Delta \tau^{f/b}_{W}$. Alternatively, we can render the Wigner delay f/b symmetric by using linearly polarized UV for ionization and circularly polarized IR probing photons to obtain $\Delta \tau^{f/b}_{W}=0$ and $\Delta \tau^{f/b}=\Delta \tau^{f/b}_{cc}$.
The photoelectron images were separated in four quadrants, and the signal was angularly-averaged in each quadrant. A Fourier analysis of the $2\omega$-oscillations was conducted to determine the delay that maximized each sideband, in each quadrant. We calculated the difference between the delays measured in the forward and backward directions $\Delta \tau^{f/b}$. This procedure was repeated for left and right helicities, and for five consecutive measurements of each enantiomer. The delays measured from opposite helicities or opposite enantiomers have opposite signs, revealing that the differential f/b ionization delay is a genuine chiral observable. In order to extract the most accurate value of the differential delay, we averaged the results obtained from the (+) and (-) enantiomers: $\Delta \tau^{f/b}=(\Delta \tau^{f/b}_{(+)} -\Delta \tau^{f/b}_{(-)})/2$ (see SM).
First, we used a linearly polarized UV ionizing field and a circularly polarized IR measurement field. The Wigner component $\tau_W$ of the sideband delay was thus f/b symmetric ($\Delta \tau^{f/b}=\Delta \tau^{f/b}_{cc}$). The results shown in Fig. \[fig2\](A) reveal that the differential f/b continuum-continuum induced delay is zero (within the 2 as accuracy of the present measurement). This means that the laser-induced transitions produce essentially the same delay on electrons emitted in the forward and backward directions, without any significant sensitivity to the chiral character of the ionic potential. However, a weak influence of the chiral potential is still noticeable in this polarization configuration: the intensity of SB2 averaged over all UV-IR delays, shows a f/b asymmetry (PECD) on the order of 0.5$\%$ (twice smaller than when the UV is circularly polarized).
In order to find signatures of chirality in the photoionization delays, we resolved the angular dependence of the photoionization dynamics [@aseyev03; @heuser16; @hockett17]. We integrated the photoelectron signal in slices of 10$^\circ$ around different ejection angles $\alpha$ from the polarization plane of the light, and measured the associated delays ($\Delta \tau^{f/b}_{cc}$). The results for SB2 are shown in Fig. \[fig2\](B). For electrons emitted beyond 70$^\circ$, the signal was too low to extract reliable values. A weak but non-zero $\Delta \tau^{f/b}_{cc}$ is measured when electrons are ejected close to the polarization plane of the IR laser, reaching 5 $\pm 2$ as at $\alpha=25^\circ$. This delay tends to vanish for higher ejection angles but the error bars become larger due to the lower level of signal. Measurements on SB3 show a zero delay whatever the ejection angle (see SM). In the commonly accepted intuitive picture of multicolor photoionization, the linear UV field induces bound-free transitions starting from the molecular core region, and the IR field subsequently drives continuum-continuum transitions while the electron is escaping from the core region. The continuum-continuum transitions are thus rather insensitive to the details of the molecular potential. This interpretation was recently confirmed by the observation of a zero continuum-continuum delay between electrons escaping the two sides of an oriented asymmetric molecule in theoretical calculations [@chacon14]. The zero overall delay we measured between forward and backward electrons on SB2 and SB3 is consistent with this picture. Nevertheless, the angle-resolved measurements demonstrate that the continuum-continuum transitions can be slightly influenced by the core (chiral) region of the potential.
Second, we broke the f/b symmetry in the ionization step by switching the polarization state of the ionizing UV field to circular while using a linearly polarized IR field. The photoionization process is here intrinsically f/b asymmetric, while the continuum-continuum coupling is f/b symmetric: $\Delta \tau^{f/b} = \Delta \tau^{f/b}_{W}$. The magnitude of the temporally-averaged PECD on the SB2 was larger than in the previous configuration, reaching 1$\%$. The measurements (Fig. \[fig2\](C)) show a differential delay of $\Delta \tau^{f/b}_{W} = 7 \pm 2$ attoseconds for the second sideband (SB2). Our experiment is thus able to reveal a small f/b asymmetry in the Wigner delay in the photoionization of chiral molecules by circularly polarized light. The $\Delta \tau^{f/b}$ vanishes for SB3 because of the decrease of both the f/b asymmetric character of the photoionization and the absolute Wigner time delay with increasing photoelectron kinetic energy.
The evolution of the differential Wigner delays with photoelectron ejection angle are shown in Fig. \[fig2\](D). Close to $\alpha$ = 0$^\circ$, $\Delta \tau^{f/b}_{W}$ is null, which is not surprising since the PECD also vanishes in the laser polarization plane. For electrons emitted in the 60-70$^\circ$ slice, $\Delta \tau^{f/b}_{W}$ reaches 24 as. This angle-resolved analysis shows that while the average difference between forward and backward electron ionization times is only 7 as, it strongly varies with ejection angle and can reach higher values for electrons emitted away from the laser polarization plane. Repeating this analysis for SB3 shows that the differential Wigner time delay remains zero within the error bars for all photoelectron ejection angles (see SM). By accessing the angular dependence of the emission time, our measurements give access to the phase properties of the photoionization matrix elements. The determination of the underlying scattering phase shifts has been a long-standing quest of photoionization experiments, and our results show that their energy-derivative, *i.e.* the Wigner delays, are accessible with high accuracy using a relatively simple setup.
The differential Wigner delay is a signature of the asymmetric scattering process which is at the heart of the photoelectron circular dichroism phenomenon. Wigner delays are determined by the energy derivative of the scattering phase. We performed a theoretical analysis of the photoionization of camphor molecules (see SM for details about the theoretical model). Our calculations confirm the existence of asymmetric Wigner delays, even in a randomly oriented ensemble of molecules. The theoretical forward/backward differential Wigner delay $\Delta \tau^{f/b}_{W}$ is of the order of 5 as for 2 eV electrons, which agrees with the present experimental observation. Interestingly, the evolution of $\Delta \tau^{f/b}_{W}$ with respect to the photoelectron energy, shown in the SM, shows rich spectroscopic features that are not visible in the photoelectron spectrum and cannot be easily distinguished in the PECD signal. Thus $\Delta \tau^{f/b}_{W}$ is a remarkable chiral observable, which enables tracking of subtle features of the molecular potential such as, for example, the differential Cooper minima, surviving the molecular orientation averaging (see SM). This observation opens prospects for highly sensitive experiments, for instance through molecular-frame measurements [@tia17], as well as accurate testing of advanced quantum theories of molecular photoionization.
The differential f/b analysis and the control of the chiral symmetry breaking of the interaction enabled us to fully decouple the different components of the photoionization delays, without the need for any theoretical input, and to reveal a tiny but measurable delay in the direct photoionization. We proceeded to use photoelectron interferometry to extract more complex dynamics occurring when the chiral molecules are photoionized in the vicinity of an autoionizing resonance.
#### Resonant photoionization {#resonant-photoionization .unnumbered}
Continuum resonances play an essential role in the photoionization of most polyatomic molecules. They can arise from the shape of the molecular potential, in a single-electron picture (shape resonances) [@piancastelli99], or from multi-electron dynamics involving electron correlations and couplings between different channels [@fano61]. In both cases, spectrally localized scattering phase jump(s) are expected, reflecting the modification of the ionization dynamics. For instance, in chiral molecules, the PECD was recently shown to be enhanced in the vicinity of a Fano resonance [@catone12]. Here we used the photoelectron interferometry technique to directly track the asymmetric ionization dynamics of chiral molecules in the vicinity of an autoionization resonance.
![Phase-resolved resonant photoionization in camphor. (**A**),(**C**) Spectral amplitudes (black) and forward and backward spectral phases ($\varphi^{f}(E)$, blue and $\varphi^{b}(E)$, red) of SB1 in (1R)-(+)-camphor, using left-circularly polarized UV - linearly polarized IR (A) and linearly polarized UV - left-circularly polarized IR (C). (**B**),(**D**) forward backward asymmetry of the spectral phase ($\Delta \varphi^{f/b}$) in camphor, using left-circularly polarized UV - linearly polarized IR in (B) and linearly polarized UV - left-circularly polarized IR (D). In (B) and (D), the dots represent the mean values of the forward backward asymmetry of the spectral phase while the solid lines show the error bars, which are defined as the 95$\%$ confidence interval.[]{data-label="fig3"}](Figure3.pdf){width="80.00000%"}
As shown in Fig. \[fig1\](D), the SB1 presents a sharp $\pi$-phase jump around 1.9 eV, which reflects the presence of a resonance on the first ATI peak. We resolved this phase jump in the forward (blue) $\varphi^F$ and backward (red) $\varphi^B$ directions, using circularly polarized UV (Fig. \[fig3\](A)) or IR (Fig. \[fig3\](C)) light. When the UV light is circularly polarized, the spectral phase exhibits a weak $\sim$ 0.75 rad bump centered around 2.1 eV. A significant difference was seen between the forward and backward spectral phases. This difference ($\Delta \varphi^{f/b} = \varphi^F - \varphi^B$) (Fig. \[fig3\](B)) shows a good mirroring between the two enantiomers.
The case where the f/b symmetry is broken by the weak IR pulse (linear UV and circular IR) is more intriguing. The spectral phases show a steep $\sim \pi$ jump around 1.9 eV, in opposite directions for forward and backward electrons. After the jump, the phases become nearly identical, as they are separated by $\sim 2 \pi$ (Fig. \[fig3\](C)). The f/b differential phases ($\Delta \varphi^{f/b}$) obtained in the two enantiomers are almost exactly opposite (Fig. \[fig3\](D)). To a first approximation, the presence of this huge asymmetry is unexpected. The circularly polarized field acts during the continuum-continuum transitions, which should be affected mostly by the long-range (non-chiral) part of the molecular potential and should, therefore, be f/b symmetric [@chacon14]. Our measurement demonstrates that in the vicinity of a resonance, the f/b symmetry can also be broken during the continuum-continuum transitions. This finding is in agreement with a recent theoretical investigation of photoelectron interferometry, which demonstrated that the simple separation of the measured delay ($\tau$) in a sum of the contributions from Wigner ($\tau_W$) and continuum-continuum ($\tau_{cc}$) delays did not hold anymore in the presence of a resonance [@argenti17]. Indeed, Argenti *et al.* demonstrated that $\tau_W$ and $\tau_{cc}$ are entangled in resonant photoionization. The measured delay ($\tau$) is representative of the two-color photoionization process, and our results show that the circular polarization of the weak IR field is sufficient to induce a major symmetry breaking, in the presence of a resonance.
#### Asymmetric electron wavepackets {#asymmetric-electron-wavepackets .unnumbered}
We could retrieve the temporal profiles of the two-color forward and backward wavepackets by Fourier-transforming their measured spectral amplitudes and phases. In order to extract the angular dependence of these temporal profiles, we analyzed the oscillations of SB1 as a function of the electron ejection angle ($\alpha$), as done previously in the non-resonant case (see raw data in SM). Figure \[fig4\] shows the resulting angle-resolved wavepackets in the temporal domain. In both polarization configurations, the wavepacket shows a single temporal peak when the electrons are ejected close to the propagation axis of the light (90$^\circ$), and a double peak structures when the electrons are ejected near the laser polarization plane (0$^\circ$). The latter are signatures of the temporal interference between the direct non-resonant and the resonant components of the autoionizing wavepackets [@gruson16]. The effect of the resonance appears more confined around the laser polarization direction when the UV field is linearly polarized (Fig. \[fig4\](B)), probably because of a stronger anisotropy of the resonant excitation compared to the circularly polarized case.
![Angle-resolved temporal profile of the autoionizing photoelectron wavepacket. The contourplot depicts the temporal profile of the wavepacket as a function of the electron ejection angle. The upright external plots show forward (blue) and backward (red) cuts of the wavepackets along specific angles. (**A**): left-circularly polarized UV, linearly polarized IR. (**B**): linearly polarized UV, left-circularly polarized IR. []{data-label="fig4"}](Figure4.pdf){width="100.00000%"}
The chiral nature of the photoionization process can be investigated by comparing cuts of the temporal profile of electrons ejected at positive (forward, blue) and negative (backward, red) angles ($\alpha$). When the UV field is circularly polarized, the two bumps of the forward electron wavepacket emitted around 30$^\circ$ maximize $\sim 400$ as after the backward wavepacket (Fig. \[fig4\](A)). Interestingly, a similar delay is measured around 60$^\circ$, where the wavepacket shows a single peak structure. As the ejection angle further increases, the ordering between forward and backward emission reverses, with a $\sim - 250$ as delay around 80$^\circ$. These subtle features obey chiral inversion when switching from one enatiomer to the other, as shown in the SM. In the other polarization configuration (Fig. \[fig4\](B)), the two bumps from the forward and backward wavepackets are synchronized in time for electrons ejected close to the laser polarization plane ($\alpha$ = 0$^\circ$). However their relative yield is strongly f/b asymmetric. This means that in the vicinity of resonances, where $\tau_{W}$ and $\tau_{cc}$ are entangled [@argenti17], the perturbative IR pulse can be used to break the f/b symmetry and to subsequently tailor asymmetric electronic wavepackets, both in time and space. At larger emission angles, where the dynamics are governed by a single -non-resonant- pathway, the single-peak wavepackets become forward-backward symmetric. This analysis provides a deep insight into the angular-dependence of the multielectron dynamics governing autoionization.
Comparing the two polarization configurations used in the measurements (Fig. \[fig4\]) shows that the wavepacket asymmetry is in fact much stronger when it is the weak IR field that is circularly polarized. We attribute this result to the sequential nature of the resonant photoionization process. The linearly polarized UV photons populate a quasi-bound state embedded in the continuum, which can relax by ionization, releasing electrons at the energy of the first ATI peak (autoionization, Fig. \[fig1\] (C)). However another process could lead to ionization of the quasi-bound state: the absorption of one IR photon releasing an electron with the energy of the first sideband. This can be seen as a classic PECD experiment, starting from a highly excited quasi-bound state. Recent experiments showed that PECD could be observed when bound states excited by linear photons were ionized by circularly polarized photons [@comby16]. The present scheme extends this scenario to quasi-bound states. On the other hand, when the UV photons are circularly polarized, they can induce an asymmetric wavepacket in the excited states, a phenomenon called PhotoeXcitation Circular Dichroism (PXCD) [@beaulieu16-2]. The ionization of such a wavepacket by linearly polarized light produces f/b asymmetries, but they were observed to be weaker than the PECD from excited states. This could explain why we observe a weaker wavepacket asymmetry when the IR photons are linearly polarized.
#### Time-frequency analysis {#time-frequency-analysis .unnumbered}
The temporal profile of the wavepackets only provides spectrally integrated information about the rich ongoing dynamics. The spectral origin of temporal asymmetries can be revealed using a time-frequency analysis [@busto17]. The Wigner-Ville distribution (WVD) is particularly interesting because it encodes the quantum interference between different components of a wavepacket. For a wavefunction $\Psi(t)$, the WVD ($W(\Omega,t)$) is defined as:
$$W(\Omega,t)=\int \Psi(t-\tau/2) \Psi^*(t+\tau/2) e^{i\Omega\tau} d\tau$$
where $\Omega$ is the angular frequency, and $t$ is the time. Figure \[fig5\] shows the WVD of the electron wavefunctions emitted around $\alpha$ = +10$^\circ$ (forward, $\Psi^f(t)$) and $\alpha$ = -10$^\circ$ (backward, $\Psi^b(t)$)) from the laser polarization plane. The distributions were calculated by averaging the wavepacket from (+)-camphor and the mirrored wavepacket from (-)-camphor. A strong negative lobe is present around time $t=0$ at the energy of the resonance (1.9 eV), revealing the quantum interference between the direct and indirect ionization components. Hyperbolic fringes converging to the energy of the resonance are observed in the leading or falling edge of the distribution, depending on the f/b emission direction. The asymmetry of these fringes reflects an asymmetric destructive interference between wavepacket components. In order to isolate the asymmetric part of the wavefunction, we calculate the WVD of the differential wavefunction $\Delta\Psi^{fb}(t)=\Psi^f(t)-\Psi^b(t)$ (Fig. \[fig5\](C)). They are strikingly simple, with a temporally long and spectrally narrow signal at the energy of the resonance, and almost no negative components. We conclude that while the forward and backward wavepackets are each formed by the coherent superposition of a resonant and a non-resonant contribution, the chiral character of the wavepacket appears to be strongly dominated by a single - resonant - pathway, which leads to the disappearance of the signature of quantum mechanical interference in the WVD. The WVD thus provides unique insight into the origin of the asymmetric shaping of photoelectron wavepackets during resonant photoionization of chiral molecules.
![Wigner-Ville distributions of the autoionizing photoelectron wavepackets. The distributions were calculated on an average wavepacket obtained by summing the wavepacket from (1R)-(+)-camphor and the mirrored wavepacket from (1S)-(-)-camphor. (**A**) shows the WVD of a forward wavepacket emitted around $\alpha$ = 10$^\circ$ from the linearly polarized UV field, and (**B**) the WVD of a backward wavepacket emitted around $\alpha$ = -10$^\circ$ (**C**) is the WVD of the forward-backward differential wavepacket.[]{data-label="fig5"}](Figure5.pdf){width="100.00000%"}
#### Conclusions and Perspectives {#conclusions-and-perspectives .unnumbered}
Our results show that using circularly polarized photons to drive photoionization of chiral molecules induces asymmetric delays in the photoemission, on both femtosecond and attosecond timescales. In direct non-resonant photoionization, the forward/backward asymmetry in the Wigner time delays are on the order of few attoseconds. In the vicinity of an autoionizing resonance driven by electronic correlation, we have found that the emitted two-color electron wavepacket is strongly asymmetric, demonstrating the chiral character of this multielectronic effect. By using the synergies of molecular and light chirality in the vicinity of resonances, we demonstrated tailoring of the shape of the released electron wavepackets, both in time and space, which is a new scheme for multidimensional attosecond quantum control. The high accuracy of the measurements can also be used as a powerful benchmarking tool for quantum theories of molecular photoionization. Finally, our approach could be generalized to a broad variety of systems to shed light on the ultrafast symmetry breakings which are at the heart of very recent technological and scientific breakthroughs, such as in superconducting chiral nanotubes [@qin17] and chiral spintronics devices [@naaman15].
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank R. Bouillaud and L. Merzeau for technical assistance, and Anne L’Huillier for fruitful discussions. This project has received funding from the European Research Council (ERC) under the European Unions Horizon 2020 research and innovation program no. 682978 - EXCITERS, and from LASERLAB-EUROPE, grant agreement no. 284464, EC’s Seventh Framework Programme. We acknowledge the financial support of the French National Research Agency through ANR-14-CE32-0014 MISFITS, ANR-14-CE32-0010 XSTASE and IdEx Bordeaux LAPHIA (ANR-10-IDEX-03-02). JC and RT acknowledge financial support from the LABEX PlasaPar project, the program “Investissements d’avenir” under the reference ANR-11-IDEX-0004-02, the program ANR-15-CE30-0001-01-CIMBAAD. S.B. acknowledge the NSERC Vanier Scholarship. The data presented in the study are available from the corresponding author on reasonable request.
Supplementary materials {#supplementary-materials .unnumbered}
=======================
Supplementary Text\
Figs. S1 to S15\
References *(1-62)*
**Supplementary Materials**
Experimental details
====================
Experimental setup
------------------
A scheme of the experimental setup is shown in figure \[setup\]. It consists of a standard Mach-Zehnder interferometer. The incoming beam is delivered by the Aurore laser system at CELIA, which provides 800 nm 25 fs pulses at 1 kHz, with up to 7 mJ energy. In one of the two arms, we frequency double the pulses by using a type-I 200 $\mu$m thick BBO crystal. The remaining 800 nm is filtered out by reflection on two dichroic mirrors. In the 800 nm arm, a delay stage is installed to temporally overlap the two beams. The control of the attosecond delay is achieved by a pair of wedges rather than by translating mirrors. This enables us to convert a rather large translation motion of one wedge into small delays (1 micron for 67 attoseconds, while a 1 micron translation of the mirrors would induce a 6.7 femtoseconds delay), releasing some constraints on the translation stage accuracy and repeatability.
Motorized quarter-wave plates are placed in both arms of the interferometer, allowing to fully control the polarization of each color independently. After the quarter-wave plates, all reflections are at $\sim 0^\circ$ to avoid polarization state artifacts. The 400 and 800 nm beams are recombined using a dichroic mirror and are focused into the interaction region of the Velocity Map Imaging Spectrometer. To compensate chromatic aberration induced by the lens, we have installed a lens telescope in the 800 nm arm, which allows us to focus both 400 nm and 800 nm at the exact same position in the spectrometer. In all presented experiments, the 400 nm/800 nm delay was scanned over a range of $\sim$ 6700 as by steps of 133 as.
![Scheme of the optical setup. BS is a 80/20 beamsplitter; $\lambda /2$ and $\lambda /4$ are half- and quarter-wave plates, respectively; BBO is a type-I 200 $\mu$m thick $\beta$-Barium borate crystal; DM are dichroic mirrors; FM is a flip mirror and BD is a beam dump. The setup is not at scale.[]{data-label="setup"}](SI_Setup.png)
The enantiopure chiral samples (Sigma Aldrich, 98$\%$ purity for (1R)-(+)-camphor and 95$\%$ purity for (1S)-(-)-camphor) were heated in an oven at 60$^\circ$C, carried to the VMI by a 80$^\circ$C heated line, and conducted under vacuum by a 100$^\circ$C heated tube to a 250 $\mu m$ nozzle located 7 cm away from the interaction zone. No carrier gas was used. The pressure in the interaction chamber was typically 2$\times 10^{-7}$ mbar. The velocity distribution of the photoionized electron was projected onto a set of dual microchannel plates and imaged by a phosphor screen and a CCD camera.
Theory
======
Equivalence between ATI and RABBIT electron interferometry
-----------------------------------------------------------
The measurement of photoelectron wavepacket spectral amplitudes and phases is usually performed using the RABBIT technique [@paul01-1]. In RABBIT, the ionizing radiation is a comb of XUV harmonics. The photoionization is driven by single photon absorption since the photon energy of the harmonics is usually greater than the ionization potential of the target. An additional weak IR pulse is used to produce photoelectron sidebands. The oscillation phase of those sidebands as a function of the XUV-IR attosecond delay provides the information which is needed to reconstruct the temporal dynamics of the photoionization.
The scheme used in our experiments is similar but the single-photon ionization by high-harmonics is replaced by Above-Threshold Ionization (ATI) by a UV (400 nm) laser field [@zipp14]. This choice was motivated by two key elements: (i) the polarization state of the UV pulse is easily tunable from linear to circular, and (ii) ATI produces much simpler photoelectron spectra in large molecules with many molecular orbitals lying energetically close to each other. In this section of the SM, our goal is to show that our ATI electron interferometry scheme allows us to retrieve equivalent information than in a RABBIT experiment.
As introduced by Gruson *et al.* [@gruson16], spectrally resolved RABBIT (Rainbow-RABBIT) enables the direct measurement of spectral phase jumps in the vicinity of a resonance. In this case, the *n*-th order harmonic of the XUV comb photoionizes the target close to an autoionizing resonance. The spectral amplitude and phase of the corresponding photoelectron peak are thus strongly modulated by the presence of the resonance. The adjacent harmonics (*(n-1)*-th and *(n+1)*-th orders) photoionize the target in a flat region of the continuum (*i.e.* away from resonances). The spectral amplitude and phase of the associated photoelectron peak are thus unstructured. The absorption/emission of an additional IR photon from both peaks leads to the formation of a sideband. The sideband results from the interference of the resonant photoelectron peak (*n-th*) with the non-resonant adjacent peak. The spectral phase of the sideband thus directly provides the spectral phase of the resonant photoelectron peak, since it is heterodyned by a peak with a flat phase.
Our case is quite similar, but the ATI peaks share some common transition pathways. If a resonance is present at the energy of the (*n-th*) ATI peak, its signature may be carried to the (*(n+1-th)*) ATI peak, which could not be used as a phase reference to retrieve the phase jump. It is thus important to answer the following question: does the effect of the resonance propagate through the ATI peaks, or is it localized on the resonant ATI peak only?
![Photoelectron interferometry in the RABBIT (**A**) and ATI (**B**) regimes. The photoelectron spectra are shown in black. The spectrally-resolved $2\omega$-oscillation phase of the sidebands in the RABBIT regime (A) is shown in red, and in the ATI regime (B) is shown in blue.[]{data-label="TDSE_LCPMR"}](TDSE_LCPMR.png)
To do so, we solved the time dependent Schrödinger Equation for a 1-D model atom, represented by a short-range Gaussian potential where we added symmetric barriers in order to induce a shape resonance into the continuum, as in [@gaillac16]. The parameters of the potential were adjusted to approximately match the ionization potential of the molecule as well as the energy of the resonance encountered in the experiment. The propagation was performed on a spatial grid using the Crank-Nicolson algorithm in the velocity gauge. The photoelectron spectrum was extracted from the final wave function thanks to the window operator.
We used two kinds of electric field, both composed of harmonics of the fundamental IR field, whose wavelength is set at $806.4$ nm ($\hbar\omega=1.5375$ eV). In the first case, dubbed “ATI”, we ionized the system with only the second harmonic ($2\omega$) at an intensity of $I_{2\omega}= 7 \times10^{12}$ W/cm$^{-2}$ and a pulse duration of FWHM $\approx20$ fs in intensity. In the second case, called “RABBIT”, the ionizing field was an attosecond pulse train of the same duration, made from harmonics 6, 8, 10, 12 and 14 of the IR frequency $\omega$ in order to reach the same energies as in the “ATI” case. For the sake of simplicity, we considered that all harmonics were synchronized and of the same intensity $I_H= 3 \times10^{8}$ W/cm$^{-2}$.
In order to observe sidebands, we added a weak IR field ($I_{\omega}= 8 \times10^{10}$ W/cm$^{-2}$), with the same duration, and we varied its phase $\delta\equiv\omega\tau$. Following the procedure introduced in [@gruson16], the photoelectron spectra $S(\delta)$ as a function of delay $\delta$ were analysed for each energy, using the following fit:
$$\label{fit}
S(\delta)= \overline{S}\left[1-B_2cos(2\delta+\varphi_2)-B_4cos(4\delta+\varphi_4)\right],$$
where $B_2$ and $B_4$ are proportionality coefficients and $\varphi_2$ and $\varphi_4$ are the phases of the $2\omega$- and $4\omega$-oscillations.
Note that in the “ATI” case, we analyzed the spectra in only one direction, i.e. $x>0$. This is compulsory as the parity of the sideband final state is different, whether the path leading to the sideband comes from the ATI peak below or above. Therefore, there would be no sideband oscillation if the spectra were taken over the all $x$ space. This is not the case in the “RABBIT” case, because the sideband parity is always “even”, as invariably reached by a two-photon transition.
As shown in figure \[TDSE\_LCPMR\], both RABBIT and ATI techniques show a large spectral phase jump across the bandwidth of the SB1, while it is mostly flat for SB2 and SB3. This results confirm that the effect of the resonance does not propagate above the resonant ATI peak. This observation validates that our experimental interferometric ATI scheme is well suited for the measurement of the spectral amplitude and phase of photoelectron wavepackets. Beyond this similarity, we observe some differences in the evolution of the spectral phases: the phase jumps are slightly different, and a linear phase appears on each higher ATI sidebands as well as between the ATI sidebands. These are the signature of differences between the single photon and multiphoton ionization dynamics.
Calculation of differential Wigner delays in one-photon ionization of camphor molecules
---------------------------------------------------------------------------------------
Atomic units will be used throughout this section unless otherwise stated.
### The theoretical approach
In the framework of the dipolar approximation, ionization probabilities and cross sections depend on the matrix element which, when defined in the molecular frame, reads $$d^{(mol)}_{\bf{k'}}=<\Psi_{\bf{k'}}^{(-)}|{\bf r.E}|\Psi_i>
\label{dipole_MF}$$ where ${\bf E}$ is the electric field associated to the incident radiation, $\Psi_i$ is the initial molecular state, and $\Psi_{\bf{k'}}^{(-)}$ is the stationnary ingoing scattering state describing an electron ejected in the direction $\bf{\hat{k'}}$ in the molecular frame. ${\bf r.E}$ is the dipolar interaction term, expressed in the length gauge, where ${\bf r}$ stands for the vectorial coordinate of the active electron. In this respect, we work in a single active electron picture where inner electrons remain frozen throughout the interaction and are accordingly described by the same molecular orbitals in both $\Psi_i$ and $\Psi_{\bf{k'}}^{(-)}$.
The optimization of molecular geometry and $\Psi_i$ result from Density Functional Theory calculations using the GAMESS-US package [@schmidt93] with the 6-311++G\*\* underlying Gaussian basis [@krishnan80] and CAM-B3LYP exchange-correlation functional [@yanai04].
The electric field, circularly polarized, is naturally defined in the laboratory frame so that the dipolar interaction term must be transposed into the molecular frame as $$\begin{aligned}
{\bf r.E}=x \pm i y &=& \sqrt{\frac{2\pi}{3}} r (a Y^{-1}_1(\Omega) + b Y^{1}_1(\Omega) ) \nonumber \\
&=&\sqrt{\frac{2\pi}{3}} r \sum_\nu{ (a {\cal D}_{\nu -1}^{(1)} + b {\cal D}_{\nu 1}^{(1)} ) Y^{1}_\nu(\Omega') }
\label{transform_rE}\end{aligned}$$ where we have assumed $||{\bf E}||=1$ and $Y^m_l$ are usual spherical harmonics defined either in the laboratory frame (with argument $\Omega$) or in the molecular frame (with argument $\Omega'$). The coefficients ($a,b$) are (0,-2) in the $x + i y$ case and (2,0) in the $x - i y$ one. ${\cal D}_{m n'}^{(l)} \equiv {\cal D}_{m n'}^{(l)}(\alpha,\beta,\gamma)$ are Wigner matrix elements [@amati60] related to the rotation, in terms of the Euler angles $(\alpha,\beta,\gamma)$, that brings the laboratory frame in coincidence with the molecular one.
The multi-center scattering state $\Psi_{\bf{k'}}^{(-)}$ is expanded in terms of partial waves [@dill74] $$\Psi_{\bf{k'}}^{(-)}({\bf r})=\sum_{l,m} {i^l e^{-i \sigma_l} \Psi_{k'lm}^{(-)}({\bf r}) Y_l^{m*}(\bf{\hat{k'}}) }
\label{partialwaves}$$ where $\sigma_l$ is the Coulomb phase shift for electron wavevector $k'$ and angular momentum $l$. The complex $\Psi_{k'lm}^{(-)}({\bf r})$ states, which fulfill appropriate boundary conditions, are related to real states $\Psi_{k'lm}({\bf r})$ through the transformation $$\Psi_{k'lm}^{(-)}({\bf r})=\sum_{l'm'} {({\bf I}+i{\bf K})_{lm,l'm'} \Psi_{k'l'm'}({\bf r}) }
\label{normalizationS}$$ where ${\bf I}$ is the unitary matrix and ${\bf K}$ is the so-called ${\bf K}$-matrix (see below). The $\Psi_{k'lm}$ are real solutions of the Schrödinger equation $H \Psi_{k'lm} = \epsilon \Psi_{k'lm}$, with $\epsilon=k'^2/2$.
To solve the Schrödinger equation, we employ an approximate form of the potential felt by the electron ejected from neutral camphor, in terms of so-called ElectroStatic Potential ESP-charges [@besler90], which basically consist of non-integer charges $Z_i^{eff}$ located on the nuclei of the molecule, so that
$$V({\bf r})=-\sum_i {\frac{Z_i}{|{\bf r} - {\bf R}_i|} + \int{\frac{\rho({\bf r'})}{|{\bf r} - {\bf r'}|} }} \sim -\sum_i {\frac{Z_i^{eff}}{|{\bf r} - {\bf R}_i|}}
\label{eqESP}$$
on Van der Walls surfaces surrounding the molecule. $Z_i$ and ${\bf R}_i$ are the real charges and locations of the nuclei while $\rho({\bf r})$ is the core electron density which does not include the density associated to the active electron. Therefore $\lim_{r \rightarrow \infty}V({\bf r}) = -1/r$. Subsequently to a single-center decomposition of $V({\bf r})$ onto spherical harmonics [@decleva94; @abusamha10], the coupled-channel Schrödinger problem is solved using the renormalized Numerov method of Johnson [@johnson78]. The ${\bf K}$-matrix elements are defined in the asymptotic $r$-region according to $$\Psi_{k'lm}({\bf r}) \sim \frac{1}{\sqrt{\pi k'}r} \sum_{l''m''} {(\sin(\theta_{l''})\delta_{l'm',l''m''} +
\cos(\theta_{l''})K_{l'm',l''m''})Y_{l''}^{m''}({\bf \hat{r}}) }
\label{Kmatrix}$$ where $\theta_{l''}=k'r-l''\pi/2-(1/k')\ln(2k'r)+\arg\Gamma[l''+1-i/k']$.
Once the $\Psi_{k'l'm'}$ states and ${\bf K}$-matrix elements are known, we are able to compute the dipolar amplitudes $a_{klm\nu}$ $$a_{klm\nu}=<\Psi_{klm}^{(-)} | r Y^\nu_1 | \Psi_i >
\label{amplitudes}$$ using eq. (\[normalizationS\]). According to eqs. (\[dipole\_MF\])-(\[partialwaves\]), the dipole in the molecular frame is thus simply $$d^{(mol)}_{\bf{k'}}=\sqrt{\frac{2\pi}{3}} \sum_{lm\nu} {(-i)^l e^{i\sigma_l} a_{k'lm\nu} (a {\cal D}_{\nu -1}^{(1)} + b {\cal D}_{\nu 1}^{(1)} ) Y^{m}_l(\bf{\hat{k'}}) }.
\label{dipole_MF_bis}$$ Performing the inverse rotation on $Y^{m}_l$ to pass from the molecular frame to the laboratory one, we obtain the dipole in this latter $$d^{(lab)}_{\bf{k}}=\sqrt{\frac{2\pi}{3}} \sum_{lm\nu\mu} {(-i)^l e^{i\sigma_l} a_{klm\nu} (a {\cal D}_{\nu -1}^{(1)} + b {\cal D}_{\nu 1}^{(1)} ) {\cal D}_{m \mu}^{(l)*} Y^{\mu}_l(\bf{\hat{k}}) }
\label{dipole_LF}$$ which can be evaluated for any direction of electron ejection $\bf{\hat{k}}=(\theta,\varphi)$ in the laboratory.
Introducing the simple notation $d^{(lab)}_{\bf{k}}=|d^{(lab)}_{(k,\theta,\varphi)}|e^{i\varphi(k,\theta,\varphi)}$, the Wigner delay can be computed in the case of an oriented molecule as [@wigner55; @chacon14; @goldberger62; @baykusheva17] $$\tau_W(k,\theta,\varphi)=\frac{\partial}{\partial E} \varphi(k,\theta,\varphi) \equiv \frac{1}{k}\frac{\partial}{\partial k} \varphi(k,\theta,\varphi).
\label{delay_oriented}$$ However this delay is defined for a given orientation $\hat{R}$ while the experiment deals with samples of randomly oriented molecules. We thus have to introduce an orientation-averaged delay $\bar{\tau}_W(k,\theta,\varphi)$, within which the contribution of a particular orientation $\hat{R}$ is weighted by its contribution to the total electron production $${\cal N}(k,\theta,\varphi)=\int{d\hat{R} |d^{(lab)}_{\bf{k}}(\hat{R})|^2},
\label{calN}$$ according to [@baykusheva17] $$\bar{\tau}_W(k,\theta,\varphi) = \int{d\hat{R} \tau_W(\hat{R};{\bf k}) \frac{|d^{(lab)}_{\bf k}(\hat{R})|^2}{{\cal N}(\bf{k})}}.
\label{delay_av}$$ In practice, the integrations on $\hat{R}$, with $d{\hat R}=\frac{1}{8\pi^2}d\alpha \sin(\beta)d\beta d\gamma$, are performed by (Simpson) numerical quadratures with angular spacing $\Delta \alpha=\Delta\beta= \Delta\gamma$ small enough to ensure convergence of the computed $\bar{\tau}_W$ values.
The one-photon counterpart to the measured differential Wigner delay is then evaluated by subtracting the $\bar{\tau}_W$ values in the forward and backward directions: $$\Delta \bar{\tau}_W^{f/b}(k)= \bar{\tau}_W(k,0,\varphi) - \bar{\tau}_W(k,\pi,\varphi),
\label{deltatauf/b}$$ the axis $z$ of quantization being collinear with the direction of propagation of the incident direction. The delay difference is also inspected between the left and right directions: $$\Delta \bar{\tau}_W^{l/r}(k)= \bar{\tau}_W(k,\pi/2,\pi) - \bar{\tau}_W(k,\pi/2,0).
\label{deltatauLR}$$ Note that similar quantities can be evaluated for a fixed orientation, $\Delta \tau_W^{f/b}(\hat{R};k)= \tau_W(\hat{R};k,0,\varphi) - \tau_W(\hat{R};k,\pi,\varphi)$ and $\tau_W^{l/r}(\hat{R};k)= \tau_W(\hat{R};k,\pi/2,\pi) - \tau_W(\hat{R};k,\pi/2,0)$.
Finally, our calculations also enable to estimate the PhotoElectron Circular Dichroism (PECD) in the forward/backward and left/right directions as $$\begin{aligned}
PECD^{f/b}(k)&=& 2\frac{{\cal N}(k,0,\varphi) - {\cal N}(k,\pi,\varphi)}{{\cal N}(k,0,\varphi) + {\cal N}(k,\pi,\varphi)} \label{PECDf/b}\\
PECD^{l/r}(k)&=& 2\frac{{\cal N}(k,\pi/2,\pi) - {\cal N}(k,\pi/2,0)}{{\cal N}(k,\pi/2,\pi) + {\cal N}(k,\pi/2,0)} \label{PECDlr}.\\\end{aligned}$$
### Results
All the following results have been obtained using $\Delta \alpha=\Delta\beta= \Delta\gamma =\pi/32$ which guaranteed convergence of all computed observables.
We first present in Fig. \[fig\_si\_1\] the differential electron production ${\cal N}$, defined in eq. (\[calN\]), along the forward, backward, left and right directions of electron ejection, for (1S)-(-)-camphor irradiated by a left circularly polarized radiation. While ${\cal N}$’s are strictly identical in the left and right directions, they significantly differ in the forward and backward directions for electron energies $E$ less than 30 eV. This is a feature commonly observed in photoionization of chiral systems [@ritchie76; @nahon16], which gives rise to a sizeable $PECD^{f/b}$ while $PECD^{l/r}=0$ for all $E$ (see inset of Fig. \[fig\_si\_1\]). Interestingly the $PECD^{f/b}$ changes sign about $E=10$ eV. Since our calculations assume that the nuclei are frozen and consider only the single outermost molecular orbital, such a sign change illustrates kinetic energy effects, i.e. the energy dependence of multiple electron scattering off the chiral potential [@ritchie76; @nahon16; @beaulieu16]. The $PECD^{f/b}$ exhibits a maximum value of $\sim 8\%$ about $E=2$ eV, and a shoulder shows up in the $E=6-10$ eV region.
The differential Wigner delays, $\Delta \bar{\tau}_W^{f/b}$ and $\Delta \bar{\tau}_W^{l/r}$, defined in eqs. (\[deltatauf/b\]) and (\[deltatauLR\]) respectively, are displayed as a function of $E$ in Fig. \[fig\_si\_2\]. As for the PECD, we find that $\Delta \bar{\tau}_W^{l/r}=0$ whatever is $E$. By contrast, $\Delta \bar{\tau}_W^{f/b} \ne 0$ and presents sizeable values in the attosecond range. The forward/backward differential delay is maximum at low $E$, as intuitively expected since low energy electrons spend more time in the chiral potential. However, it consists in this energy region of a small fraction only of the typical Wigner delay, which is illustrated in the inset of Fig. \[fig\_si\_2\]. In practice, $\bar{\tau}_W^f \sim \bar{\tau}_W^b \sim \bar{\tau}_W^l \sim \bar{\tau}_W^r$ and the ratio $|\Delta \bar{\tau}_W^{f/b}/\bar{\tau}_W^{f,b}|$ maximizes in the intermediate energy region centered about $\sim 10$ eV. Kinetic energy effects also show up in $\Delta \bar{\tau}_W^{f/b}$ through a sign change about $E=4$ eV. But more interestingly, the differential Wigner delay presents rich spectroscopic features in terms of local maxima/minima for $E > 1$ eV. These features are totally absent in the electron production signals of Fig. \[fig\_si\_1\] and almost invisible in the usual PECD (inset of Fig. \[fig\_si\_1\]). In fact, differentiating the PECD with respect to $E$ partially allows retrieving the structures strongly marked in $\Delta \bar{\tau}_W^{f/b}$, but with very small amplitudes. This procedure would thus be difficult on experimental data. By contrast, our present experimental investigation shows that measuring differential delays is now feasible, and will surely be addressed by highly sensitive setups in the next future. In this respect, it seems that the differential delay $\Delta \bar{\tau}_W^{f/b}$ is a very valuable chiral observable which encodes subtle features of the underlying chiral potential, beyond usual (PECD) signatures.
We now investigate the origin of the features observed in the differential Wigner delays. $\Delta \bar{\tau}_W^{f/b}$ involves delays averaged on the molecular orientations. We thus looked at the delays for all underlying orientations, $\Delta \tau_W^{f/b}(\hat{R};E)$, and observed that some of them have very important values. For instance, focusing on the local maximum of $\Delta \bar{\tau}_W^{f/b}$ centered about $E=2.5$ eV in Fig. \[fig\_si\_1\], we found that that the orientation defined by ($\alpha=22.5^\circ,\beta=45^\circ,\gamma=0^\circ$) yields $\Delta \tau_W^{f/b}(\hat{R};E) \sim 2$ fs at this energy (see Fig. \[fig\_si\_3\](A)). This important delay difference stems from a phase jump of $\sim\pi$ in the forward direction, in a narrow $E$-range where the phase in the backward direction behaves smoothly and thus leads to a vanishing Wigner delay (see Fig. \[fig\_si\_3\](B)). Simultaneously the dipole modulus has $\sim 0$ amplitude in the forward direction (see Fig. \[fig\_si\_3\](E)). In other words, the important delay in the forward direction is nothing else than the signature of a Cooper minimum [@cooper62; @cloux15] induced by the potential shape in the ($\theta=0,\varphi$)-electron direction for ($\alpha=22.5^\circ,\beta=45^\circ,\gamma=0^\circ$) molecular orientation. This large delay difference, in the fs range, survives orientation averaging, and leads to the maximum of $\Delta \bar{\tau}_W^{f/b}$ in the as range at $E \sim 2.5$ eV (see Fig. \[fig\_si\_3\]).
![(**A**) Wigner delays, $\tau_W$, in the forward and backward directions for ($\alpha=22.5^\circ,\beta=45^\circ,\gamma=0^\circ$); the (red) continuous line refers to the forward direction and the (black) dashed one to the backward direction. Associated variations of the phase (**B**) and square modulus (**C**) of the dipoles. (**D-F**) graphs are similar to (A-C), but for ($\alpha=180^\circ,\beta=157.5^\circ,\gamma=0^\circ$).[]{data-label="fig_si_3"}](Bernard_3.png){width="80.00000%"}
In our calculations, which are based on the single-active electron approximation and employ a simplified form of the molecular potential in terms of ESP-charges, all local variations of $\Delta \bar{\tau}_W^{f/b}$ are due to such differential Cooper minima. This is illustrated in Figs. \[fig\_si\_3\](D-F) for the oscillation appearing in Fig. \[fig\_si\_2\] about $E=8.5$ eV. In this case the important negative delay computed in the forward direction for ($\alpha=180^\circ,\beta=157.5^\circ,\gamma=0^\circ$) contributes positively to the averaged $\Delta \bar{\tau}_W^{f/b}$ because $\sin(\beta)<0$. It is important to note that in the experiment, the multiphoton excitation/ionization preferentially selects some molecular orientations within the sample of ramdomly oriented molecules. Our calculations suggest that this photoselection can have a huge impact on the measured quantities. In the future, it would thus be of great interest to study the differential Wigner time delays within the molecular-frame of chiral molecules.
It has to be noted that not only differential Cooper minima but also differential resonances are amenable to important local variations of $\Delta \tau_W^{f/b}(\hat{R};E)$ because of underlying $\sim \pi$ phase jumps. However, our calculations are based on the single-active electron approximation, which inhibits the occurrence of autoionizing resonances. Shape resonances do not show up either, at least using our simplified description of the ionic potential in terms of ESP-charges. Nevertheless, it is clear that fine differential features of chiral molecular potentials can, in general, be probed by measurements and computations of $\Delta \bar{\tau}_W^{f/b}$.
Data analysis and error bars
============================
Raw VMI images
--------------
For each time delay, each enantiomer and each polarization state, we record 2D projections of the 3D-PAD, averaged over $\mathrm{\sim 5x10^4}$ laser shots. As mentioned in the main text, these photoelectron distributions are up-down asymmetric. We thus separate the up- and the down- part of the signals, and symmetrize them in order to be able to perform the pBasex inversion. A typical raw VMI image (before any symmetrization or anti-symmetrization) in linear scale (Fig. \[raw\](A)) and in logarithmic scale (Fig. \[raw\](B)) are presented below. For this example, the UV was linearly polarized, the IR was circularly polarized and the target molecules were (1S)-(-)-camphor.
![Raw experimental VMI images. (**A**) In linear scale and (**B**) in logarithmic scale. These images have been recorded using circularly polarized UV and linearly polarized IR fields, in 1S-(-)-camphor. []{data-label="raw"}](raw.png){width="80.00000%"}
Resonant photoionization of chiral molecules
--------------------------------------------
This subsection describes the procedure for extracting the error bars of Fig. 4 (B) and (D) of the main paper.
![Forward/Backward asymmetry of the spectral phase ($\Delta \varphi^{f/b}$) across the resonant sideband. In (**A**) ((1S)-(-)-camphor) and (**B**) ((1R)-(+)-camphor), the 400 nm is circularly polarized and the 800 nm is linearly polarized. In (**C**) ((1S)-(-)-camphor) and (**D**) ((1R)-(+)-camphor), the 400 nm is linearly polarized and the 800 nm is circularly polarized. For the (1S)-(-)-camphor panels (A,C), the blue dots represent the $\Delta \varphi^{f/b}$ extracted for different scans and the red lines/dots are the mean values. For the (1R)-(+)-camphor panels (B,D), the red dots represent the $\Delta \varphi^{f/b}$ extracted form different scans and the blue lines/dots are the mean values.[]{data-label="femto"}](SI_AllScan_Femto.png){width="80.00000%"}
For each polarization configuration and for each enantiomer, at least five consecutive delay scans are performed for left and right polarizations. The photoelectron images obtained using left (L) and right (R) polarizations are used to produce a sum image (L+R) and a difference image (L-R). Each photoelectron image is averaged over $\sim 5\cdot10^4$ laser shots. For a non-chiral sample, or non-chiral ionizing light, (L-R) is expected to be zero, while for a chiral sample photoionized with a chiral light it is expected to be non-zero and antisymmetric with respect to the laser propagation axis. On the contrary, L+R images are symmetric with respect to the laser propagation axis. In order to eliminate artifacts due to the imperfect nature of the waveplates and detector, we antisymmetrize the (L-R) images and symmetrize the (L+R) images.
The (L-R) and (L+R) images are decomposed as a sum of Legendre polynomials [@garcia04]. When the laser-molecule interaction is cylindrically symmetric with respect to the detection plane (for instance when both the pump and probe pulse are linearly or circularly polarized), this decomposition procedure enables reconstructing the 3D photoelectron angular distribution from its 2D projection onto the VMI detector. In the present experiment, a combination of circular and linear polarization is used and the cylindrical symmetry condition required to rigorously retrieve the 3D distribution from its projection is lost. This can lead to deviations between the Legendre decomposition and the 3D distribution, in particular in terms of energy smearing. However, in our measurements we detect very sharp structures in the Legendre decomposition: the phase jump across SB1 extends over about 0.1 eV. This indicates that there is no major blurring of the reconstructed distribution, and that it must not deviate much from the 3D distribution. The full 3D distribution could be retrieved by repeating the experiment for different directions of the linearly polarized field and performing a tomographic reconstruction [@wollenhaupt09]. This procedure would be prohibitively long with a 1kHz laser system, but will be achievable in a reasonable time using fiber-based lasers at few 100 kHz rate.
We analyze the oscillations of the photoelectron image associated with left polarization (L) by summing the (L-R) and (L+R) decompositions. The R image, obtained by the difference between (L+R) and (L-R), would give perfectly antisymmetric results with respect to the light propagation axis.
In order to study the statistical dispersion of our data, the phase of the sidebands oscillations is measured for each delay scan, in each quadrant. The forward and backward phases from a given (upper or lower) hemisphere are subtracted to calculate $\Delta \varphi^{f/b}$. Two values are thus obtained for each scan, giving us a total of at least 10 sets of measurements. The $\Delta \varphi^{f/b}$ extracted for each independent measurement as well as their average values are shown in Fig \[femto\].
Using Student’s statistics, we calculate the 95 $\%$ confidence intervals based on the $n$ individual measurement of $\Delta \varphi^{f/b}$. The confidence intervals are defined as:
$$\bigg[\overline{\Delta \varphi^{f/b}} - t \sqrt{\frac{S}{n}} , \overline{\Delta \varphi^{f/b}} + t \sqrt{\frac{S}{n}} \bigg]$$
where $\overline{\Delta \varphi^{f/b}}$ is the mean value of all individually measured $\Delta \varphi^{f/b}_i$, $t$ is called the quantile and is a quantity that depends on the confidence interval and the number of individual measurements ($n$), and $S$ is the unbiased estimator of variance, defined as :
$$S = \frac{1}{n-1} \sum_{i=1}^{n} \big( \Delta\varphi^{f/b}_i - \overline{\Delta \varphi^{f/b}} \big)^2$$
The mean values of the measured f/b asymmetry of the spectral phase, $\overline{\Delta \varphi^{f/b}}$, as well as the 95 $\%$ confidence intervals are presented in figure 4 (B) and (D) of the main paper. We can conclude, based on statistical variability, that our data are accurate and statistically meaningful.
Fourier analysis of the sidebands oscillation
---------------------------------------------
The sideband signal as a function of delay is Fourier transformed and shows a single peak at an energy corresponding to a $2\omega_{IR}$ oscillation frequency (Fig. \[fft\]). This ensures that we are in a RABBIT-like regime and that no higher-order phenomenon affects our phase measurement. Indeed, because ATI peaks also oscillate due to multiple quantum path interferences (*e.g.* for ATI1 interference between \[2+1\] (3x400nm) and \[2+2’\] (2x400nm + 2x800nm) paths), one could wonder if this amplitude modulation would affect the phase of the sideband oscillation. A detailed investigation of RABBIT performed by Antoine Camper *et al.* [@camper13] demonstrated that when the Fourier transform of the sideband oscillations only showed a single $2\omega_{IR}$ peak (no $4\omega_{IR}$ or higher orders), the oscillation phase of the sideband was solely determined by the phase difference between adjacent peaks.
![Fourier analysis of the temporal oscillation of the SB1 (black), SB2 (red) and SB3 (blue). The photoelectron signal have been integrated over the spectral bandwidth of each SB prior to the Fourier analysis.[]{data-label="fft"}](SI_FFT.png){width="60.00000%"}
Angular-resolved extraction of spectral phase across SB1
--------------------------------------------------------
In this subsection, we present the oscillation of the SB1 signal as a function of photoelectron energy and ejection angle. In Fig. \[Raw\_Oscill\] (A), we clearly see the steep $\sim \pi$ phase-jump between 1.87 eV and 1.92 eV for electron ejected between 0-10$^\circ$. The magnitude of the phase jump across the resonance is still $\sim \pi$, but is much smoother for electron ejected between 20-30$^\circ$. The phase jump disappear for ejection angle between 50-60$^\circ$. One can notice that the signal-to-noise ratio is very good for electron ejected closer to the laser polarization axis, where the photoelectron signal is maximum. The behavior of the spectral phase as a function of photoelectron ejection angle results in a complex spatio-temporal shaping of the photoelectron wavepacket (see Fig. 4 of the main paper).
![Raw oscillations of the SB1 signal as a function of photoelectron energy and ejection angle. The different colored lines represent different photoelectron energy. The electrons are ejected between 0-10$^\circ$ in (**A**), between 20-30$^\circ$ in (**B**) and between 50-60$^\circ$ in (**C**). These data were taken with linearly polarized UV and circularly polarized IR fields, in 1S-(-)-camphor.[]{data-label="Raw_Oscill"}](Raw_Oscill.png){width="80.00000%"}
Enantiomeric mirroiring in the reconstruction of resonant photoelectron wavepackets
-----------------------------------------------------------------------------------
In this subsection, we present the reconstruction of the autoionizing photoelectron wavepackets for both polarization configurations as well as for both camphor enantiomers. We present the wavepackets along the same ejection angles as shown in the Fig.4 of the main paper. Our aim is to show that even very subtle effects are mirroired when switching the enantiomer.
For example, in figure \[WP\_enan1\], the very small delays between forward and backward wavepacket emitted along $\alpha=30^\circ$ (Fig. \[WP\_enan1\] (B,F)), $\alpha=60^\circ$ (Fig. \[WP\_enan1\] (C,G)), $\alpha=80^\circ$ (Fig. \[WP\_enan1\] (D,H) show a striking mirroiring when switching between enantiomers. It demonstrate the validity of these measurements. A nice enantiomeric mirroiring can also be observed for the case where the UV is linearly polarized and the IR circularly polarized (Fig. \[WP\_enan2\]).
![Temporal profile of the resonant electron wavepackets emitted in different directions in (1S)-(-)-camphor (top, **A-D**) and (1R)-(+)-camphor (bottom, **E-H**), using a circularly polarized UV and a linearly polarized IR field. []{data-label="WP_enan1"}](WP_enan1.png){width="80.00000%"}
![Temporal profile of the resonant electron wavepackets emitted in different directions in (1S)-(-)-camphor (top, **A-D**) and (1R)-(+)-camphor (bottom, **E-H**), using a circularly polarized UV and a linearly polarized IR field. []{data-label="WP_enan2"}](WP_enan2.png){width="80.00000%"}
Attosecond delays in non-resonant photoionization of chiral molecules
---------------------------------------------------------------------
We will describe the procedure to extract the f/b asymmetry of the attosecond delays in non-resonant photoionization, *i.e.* for SB2 and SB3. To do so, we have used the same data sets as in the last section. For each measurement, we calculated the $\Delta \tau^{f/b}$, by using the subtracted forward and backward $2\omega_{IR}$ oscillation phases, which has been extracted from signals integrated over the sideband bandwidth, as in the conventional RABBIT technique. The individual $\Delta \tau^{f/b}$ for SB2 and SB3 are shown as blue dots in figure \[atto2\] (A) for (1S)-(-)-camphor and as red dots in figure \[atto2\] (B) for (1R)-(+)-camphor. In both cases, the black dots are the mean values. The value lies close to zero for both enantiomers and both sidebands. When changing the enantiomer, the results should be perfectly mirrored (opposite). Any deviation from perfect mirroring could be ascribed to statistical errors (different S/N ratio for each enantiomer), to systematic error (detector inhomogeneity or polarization state artifact, for example) or to different enantiopurity for each sample. To account for those unwanted detrimental effects, and to estimate more precisely the $\Delta \tau^{f/b}$, we redefine $\Delta \tau^{f/b}$ as:
$$\Delta \tau^{f/b} = \frac{1}{2}\big( \Delta \tau_{(+)}^{f/b} - \Delta \tau_{(-)}^{f/b} \big)
\label{enantiomeric}$$
where $\Delta \tau_{(+)}^{f/b}$ and $\Delta \tau_{(-)}^{f/b}$ are the experimentally measured $\Delta \tau^{f/b}$ for (1R)-(+)-camphor and (1S)-(-)-camphor, respectively. Note that after this procedure, we have removed all the unwanted experimental induced errors and the $\Delta \tau^{f/b}$ are perfectly opposite for (1R)-(+)- and (1S)-(-)-camphor. After doing this treatment, we calculate the 95 $\%$ confidence interval error bars using the Student’s statistics procedure described in the previous section. The $\Delta \tau^{f/b}$ for (1R)-(+)-camphor (green dots) and (1S)-(-)-camphor (orange dots) are presented on figure \[atto2\] (C). Note that for this case, the UV is linearly polarized and the IR is circularly polarized.
![Forward/Backward asymmetry in the attosecond photoionization delay ($\Delta \tau^{f/b}$) for the two non-resonant sidebands, when using linearly polarized UV and circularly polarized IR. In (**A**)/(**B**) we present the $\Delta \tau^{f/b}$ extracted from each independent scans that we have performed in (1S)-(-)-camphor/(1R)-(+)-camphor, respectively. In (A), the blue dots represent the $\Delta \tau^{f/b}$ extracted from different scans and the black dots are the mean value. In (B), the red dots represent the $\Delta \tau^{f/b}$ extracted from different scans and the black dots are the mean value. In (**C**), we present the enantiomeric $\Delta \tau^{f/b}$, as defined in equation \[enantiomeric\].[]{data-label="atto2"}](SI_AllScan_Atto2.png){width="80.00000%"}
One can see that for both non-resonant SB2 and SB3, the $\Delta \tau^{f/b}$ is found to be zero. Again, only the weak IR field is circularly polarized, so the f/b asymmetry is solely broken during ’measurement’ step. Our data suggest that for non-resonant photoionization, the ’measurement’ step induces the same delay in the forward than in the backward direction, with respect to the light propagation direction.
![Forward/Backward asymmetry of the attosecond photoionization delay ($\Delta \tau^{f/b}$) for the two non-resonant sidebands, when using circularly polarized UV and linearly polarized IR. In (**A**)/(**B**) we present the $\Delta \tau^{f/b}$ extracted from each independent scans that we have performed in (1S)-(-)-camphor/(1R)-(+)-camphor, respectively. In (A), the blue dots represent the $\Delta \tau^{f/b}$ extracted from different scans and the black dots are the mean value. In (B), the red dots represent the $\Delta \tau^{f/b}$ extracted from different scans and the black dots are the mean values. In (**C**), we present the enantiomeric $\Delta \tau^{f/b}$, as defined in equation \[enantiomeric\].[]{data-label="atto1"}](SI_AllScan_Atto1.png){width="80.00000%"}
We now turn our attention to the case when the UV is circularly polarized and the IR is linearly polarized. In that case, because the IR is linear, it cannot induce any f/b asymmetry in the ’measurement’ step, which means that the measured delays will only reflect the f/b asymmetry of the Wigner time delay. Our experimental data reveal a $\Delta \tau^{f/b} = \mathrm{7 \pm 2} as$ for (1S)-(-)-camphor ($\Delta \tau^{f/b} = \mathrm{- 7 \pm 2}$ as for (1R)-(+)-camphor) for the SB2 and a null $\Delta \tau^{f/b}$ for the higher energy SB3.
We have performed an angular-resolved analysis of the differential photoionization delay of the SB3 by slices of 10$^\circ$ instead of by quadrant, in order to see if it vanishes for all ejection angles. Figure \[Angular\_SB3\](A) shows the evolution of the differential Wigner delay with electron ejection angle for SB3. The differential delay remains 0 within the error bars for all ejection angles, which confirms the validity of our initial conclusion. One can notice that the error bars on this measurement are larger than those on SB2. This is caused by the low level of the signal, which is typically one order of magnitude lower for SB3 than for SB2. Repeating the measurements with higher accuracy (active stabilization of the delay line, higher repetition rate of the laser, longer pulses with narrower spectrum...) could enable us to reduce the error bars and reveal possible small but non-zero values. Figure \[Angular\_SB3\](B) shows the evolution of the differential continuum-continuum delay with electron ejection angle for SB3, when the UV field is linearly polarized and the IR field is circularly polarized. The $\Delta \tau_{cc}^{f/b}$ also remains 0 within the error bars for all ejection angles.
![Angle-resolved f/b differential delay for SB3. In (**A**) when using circularly polarized UV and linearly polarized IR and in (**B**) when using linearly polarized UV and circularly polarized IR.[]{data-label="Angular_SB3"}](Angular_SB3.png){width="80.00000%"}
Sampling and phase measurements of the oscillating signals
----------------------------------------------------------
The temporal resolution of our experiment is determined by the accuracy with which we are able to measure the phase of the oscillating (sine) function. It depends on the way we sample the sine function as well as on the algorithm used to extract the phase. From the sampling point of view, the experimental acquisition time sets a limit and imposes the total number of samples that can be acquired in a reasonable time. In our experimental conditions we recorded 5 oscillations with 10 points per oscillations, and sequentially repeated the measurement for 5 consecutive delay scans in each enantiomer. The overall acquisition time was roughly 10 hours. Recording 5 oscillations fulfills the standard set by IEEE to properly sample a sine function [@sedlacem05].
For the phase measurement, we compared two methods: Discrete Fourier Transform and 4-parameter sine fitting. The method can significantly influence the accuracy of the results, as demonstrated for instance in [@sedlacem05]. The 4-parameter fitting followed the procedure recommended by IEEE (IEEE-STD-1057).
Figure \[dft\_vs\_sine\] shows the $\Delta \tau^{f/b}$ extracted from both methods as a function of electron ejection angle. The error bars are calculated by Student statistical analysis of the consecutive measurements. The results from the two methods are consistent within the error bars. The DFT provides smaller error bars, and we thus chose this procedure for our data analysis.
![Comparaison between Discrete Fourier Transform (DFT) and 4-parameter Sine Fit methods to extract the angle-resolved $\Delta \tau^{f/b}$ for sideband 2. An offset of 2 degree was added to the point extracted with the DFT method in order to clearly see the difference between the DFT and Sine Fit methods.[]{data-label="dft_vs_sine"}](DFT_vs_Sine.png){width="80.00000%"}
[10]{}
I. Powis, [*Advances in Chemical Physics*]{}, S. A. Rice, ed. (John Wiley & Sons, Inc., 2008), p. 267–329.
B. Ritchie, [*Phys. Rev. A*]{} [**13**]{}, 1411 (1976).
N. Böwering, [*et al.*]{}, [*Physical Review Letters*]{} [**86**]{}, 1187 (2001).
L. Nahon, G. A. Garcia, I. Powis, [*Journal of Electron Spectroscopy and Related Phenomena*]{} [**204**]{}, 322 (2015).
A. Comby, [*et al.*]{}, [*The Journal of Physical Chemistry Letters*]{} [**7**]{}, 4514 (2016).
S. Beaulieu, [*et al.*]{}, [*Faraday Discussions*]{} [**194**]{}, 325-348 (2016).
S. Beaulieu, [*et al.*]{}, [*New Journal of Physics*]{} [**18**]{}, 102002 (2016).
U. Meierhenrich, Amino [Acids]{} and the [Asymmetry]{} of [Life]{}, Advances in [Astrobiology]{} and [Biogeophysics]{} (Springer, Berlin, Heidelberg, 2008).
M. Tia, [*et al.*]{}, [*The Journal of Physical Chemistry Letters*]{} [ **4**]{}, 2698 (2013).
K. Klünder, [*et al.*]{}, [*Physical Review Letters*]{} [**106**]{}, 143002 (2011).
A. L. Cavalieri, [*et al.*]{}, [*Nature*]{} [**449**]{}, 1029 (2007).
S. Haessler, [*et al.*]{}, [*Physical Review A*]{} [**80**]{}, 011404(R) (2009).
I. Jordan, [*et al.*]{}, [*Physical Review A*]{} [**95**]{}, 013404 (2017).
S. Heuser, [*et al.*]{}, [*Physical Review A*]{} [**94**]{}, 063409 (2016).
P. Hockett, E. Frumker, D. M. Villeneuve, P. B. Corkum, [*Journal of Physics B: Atomic, Molecular and Optical Physics*]{} [**49**]{}, 095602 (2016).
D. Baykusheva, H. J. Wörner, [*The Journal of Chemical Physics*]{} [ **146**]{}, 124306 (2017).
A. Chacon, M. Lein, C. Ruiz, [*Physical Review A*]{} [**89**]{}, 053427 (2014).
L. Cattaneo, J. Vos, M. Lucchini, C. Cirelli, U. Keller, [*International Conference on Ultrafast Phenomena*]{} (Optical Society of America, 2016), p. UM2B.3.
U. Fano, [*Phys. Rev.*]{} [**124**]{}, 1866 (1961).
A. Kaldun, [*et al.*]{}, [*Science*]{} [**354**]{}, 738 (2016).
M. Kotur, [*et al.*]{}, [*Nature Communications*]{} [**7**]{}, 10566 (2016).
V. Gruson, [*et al.*]{}, [*Science*]{} [**354**]{}, 734 (2016).
M. Ossiander, [*et al.*]{}, [*Nature Physics*]{} [**13**]{}, 280 (2017).
P. Agostini, F. Fabre, G. Mainfray, G. Petite, N. K. Rahman, [*Phys. Rev. Lett.*]{} [**42**]{}, 1127 (1979).
P. M. Paul, [*et al.*]{}, [*Science*]{} [**292**]{}, 1689 (2001).
L. J. Zipp, A. Natan, P. H. Bucksbaum, [*Optica*]{} [**1**]{}, 361 (2014).
L. Nahon, [*et al.*]{}, [*Physical Chemistry Chemical Physics*]{} [**18**]{}, 12696 (2016).
C. Lux, [*et al.*]{}, [*Angewandte Chemie International Edition*]{} [**51**]{}, 5001 (2012).
C. S. Lehmann, N. B. Ram, I. Powis, M. H. M. Janssen, [*The Journal of Chemical Physics*]{} [**139**]{}, 234307 (2013).
C. Lux, M. Wollenhaupt, C. Sarpe, T. Baumert, [*[ChemPhysChem]{}*]{} [**16**]{}, 115 (2015).
D. W. Schumacher, F. Weihe, H. G. Muller, P. H. Bucksbaum, [*Physical Review Letters*]{} [**73**]{}, 1344 (1994).
X. Gong, [*et al.*]{}, [*Physical Review Letters*]{} [**118**]{}, 143203 (2017).
S. Skruszewicz, [*et al.*]{}, [*Physical Review Letters*]{} [**115**]{}, 043001 (2015).
G. A. Garcia, L. Nahon, I. Powis, [*Review of Scientific Instruments*]{} [ **75**]{}, 4989 (2004).
J. M. Dahlström, A. [L’Huillier]{}, A. Maquet, [*Journal of Physics B: Atomic, Molecular and Optical Physics*]{} [**45**]{}, 183001 (2012).
E. Wigner, [*Physical Review*]{} [**98**]{}, 145 (1955).
S. Aseyev, Y. Ni, L. Frasinski, H. Muller, M. Vrakking, [*Physical Review Letters*]{} [**91**]{}, 223902 (2003).
P. Hockett, [*Journal of Physics B: Atomic, Molecular and Optical Physics*]{} [**50**]{}, 154002 (2017).
M. Tia, [*et al.*]{}, [*The Journal of Physical Chemistry Letters*]{} [ **8**]{}, 2780 (2017).
M. Piancastelli, [*Journal of Electron Spectroscopy and Related Phenomena*]{} [**100**]{}, 167 (1999).
D. Catone, [*et al.*]{}, [*Physical Review Letters*]{} [**108**]{}, 083001 (2012).
L. Argenti, [*et al.*]{}, [*Physical Review A*]{} [**95**]{} 043426 (2017).
S. Beaulieu, [*et al.*]{}, [*[arXiv:1612.08764]{} \[physics\]*]{} (2016).
D. Busto, [*et al.*]{},[*arXiv:1709.07639v1 \[physics.atom-ph\]*]{} (2017).
F. Qin, [*et al.*]{}, [*Nature Communications*]{} [**8**]{}, 14465 (2017).
R. Naaman, D. H. Waldeck, [*Annual Review of Physical Chemistry*]{} [**66**]{}, 263 (2015).
R. Gaillac, M. Vacher, A. Maquet, R. Taïeb, J. Caillat, [*Phys. Rev. A*]{} [**93**]{}, 013410 (2016).
M. Schmidt, [*et al.*]{}, [*J. Comput. Chem.*]{} [**14**]{}, 1347 (1993).
R. Krishnan, J. S. Binkley, R. Seeger, J. A. Pople, [*The Journal of Chemical Physics*]{} [**72**]{}, 650 (1980).
T. Yanai, D. P. Tew, N. C. Handy, [*Chemical Physics Letters*]{} [**393**]{}, 51 (2004).
D. Amati, [*Il Nuovo Cimento (1955-1965)*]{} [**16**]{}, 1160 (1960).
D. Dill, J. L. Dehmer, [*The Journal of Chemical Physics*]{} [**61**]{}, 692 (1974).
B. H. Besler, K. M. Merz, P. A. Kollman, [*Journal of Computational Chemistry*]{} [**11**]{}, 431 (1990).
P. Decleva, A. Lisini, M. Venuti, [*Journal of Physics B: Atomic, Molecular and Optical Physics*]{} [**27**]{}, 4867 (1994).
M. Abu-Samha, L. B. Madsen, [*Phys. Rev. A*]{} [**81**]{}, 033416 (2010).
B. R. Johnson, [*The Journal of Chemical Physics*]{} [**69**]{}, 4678 (1978).
M. L. Goldberger, K. M. Watson, [*Phys. Rev.*]{} [**127**]{}, 2284 (1962).
J. W. Cooper, [*Phys. Rev.*]{} [**128**]{}, 681 (1962).
F. Cloux, B. Fabre, B. Pons, [*Physical Review A*]{} [**91**]{}, 023415 (2015).
M. Wollenhaupt, [*et al.*]{}, [*Applied Physics B*]{} [**95**]{}, 647 (2009).
A. Camper, T. Ruchon, [*Proceedings of UVX*]{} [**1**]{}, 01014 (2013).
M. Sedlacek, M. Krumpholc, [*Metrology and Measurement Systems*]{} [**12**]{}, s427 (2005).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The dynamic electric dipole polarizability function for the magnesium atom is formed by assembling the atomic electric dipole oscillator strength distribution from combinations of theoretical and experimental data for resonance oscillator strengths and for photoionization cross sections of valence and inner shell electrons. Consistency with the oscillator strength (Thomas-Reiche-Kuhn) sum rule requires the adopted principal resonance line oscillator strength to be several percent lower than the values given in two critical tabulations, though the value adopted is consistent with a number of theoretical determinations. The static polarizability is evaluated. Comparing the resulting dynamic polarizability as a function of photon energy with more elaborate calculations reveals the contributions of inner shell electron excitations. The present results are applied to calculate the long-range interactions between two and three magnesium atoms and the interaction between a magnesium atom and a perfectly conducting metallic plate. Extensive comparisons of prior results for the principal resonance line oscillator strength, for the static polarizability, and for the van der Waals coefficient are given in an Appendix.'
author:
- 'James F. Babb'
title: An empirically constructed dynamic electric dipole polarizability function of magnesium and its applications
---
Introduction
============
Magnesium is an abundant element currently of interest in several applications. Analysis of photo association spectroscopy for the Mg dimer [@TieKotJul02; @KnoRuhTie13] indicates that the $s$-wave scattering length for collisions between two ground state Mg atoms is positive [@TieKotJul02], with the accuracy of the determination affected by the remnant uncertainty in the value of the atom-atom van der Waals constant [@KnoRuhTie13]. (A Bose-Einstein condensate of Mg atoms has not been created experimentally, to date.) However, a determination of the leading term in the long-range interaction of electronically excited states of the Mg dimer, which is related to the principal resonance line oscillator strength, using molecular spectroscopy has been elusive [@KnoRuhTie14]. Mg atoms were investigated as a possible sympathetic cooling agent in collisions with NH [@SolZucHut09; @GonMarMay11] and Lonij *et al.* [@LonKlaHol11] theoretically explored the interaction of an Mg atom with a wall for applications to atom interferometry. And, while Mg is abundant in the solar system, it is interesting to note that the Mg principal resonance line was recently detected in the observation of the exoplanet HD 209458b using transit spectroscopy [@VidMadHui13]. Modeling Mg absorption in exoplanet atmospheres depends proportionally on the principal resonance line absorption oscillator strength [@BouLecVid14], the value of which in turn affects the use of that line as a probe of escaping atoms in exoplanet atmospheric spectroscopy [@BouLecVid15].
An analysis of the dynamic electric dipole polarizability function of Mg is valuable for several reasons. First, because the principal resonance line oscillator strength is an important contributor to the function, it is possible to determine a value that achieves consistency with oscillator strength sum rules. Secondly, it is desirable to have an independent assessment of the completeness of existing elaborate calculations of the function itself, for which extensive tabulations from two different types of calculations are available [@DerPorBab10; @JiaMitChe15].
Calculations of dynamic electric dipole polarizabilities are of intrinsic theoretical interest due to the challenges inherent in treating correlations and excitations of all electrons quantum-mechanically at different photon energies [@Amu90; @AmuCheYar12]. Such calculations are necessarily important benchmarks for theoretical methods applied to photoabsorption [@Ber02], photodetachment [@MasSta00], blackbody radiation shifts [@MidFalLis12; @SheLemHin12; @BelSheLem12] and AC Stark shifts [@StaBudFre06], magic wavelengths [@BarStaLem08], and parity non conservation amplitudes [@Khr91], as well as being helpful in the ongoing development of density functional theory (DFT) methods for dispersion forces (cf. [@BasHesSal08; @TkaSch09; @TkaDiSCar12; @TaoPerRuz13; @TouRebGou13]). In addition, for metals experimental data at a wide spectrum of photon energies are relatively scarce though X-ray data, and sometimes optical data [@GoeHoh95; @SarBeiShe06], are available. There is recent progress for systematically measuring static polarizabilities [@MaIndZha15]. Many theoretical approaches are available, but their reliability in calculating dynamic polarizabilities can be difficult to gauge without critical evaluation, but critical evaluations are limited to the static polarizabilities [@Sch06; @ThaLup06; @MitSafCla10]. Nevertheless, dynamic polarizabilities are of great utility in calculating coefficients appearing in certain potential energies, particularly van der Waals constants (for investigations of ultra-cold collisions, for photo association spectroscopy, and for ultra-cold gas studies) and Lennard-Jones constants (for atom-surface interactions, where recent applications include tests for gravity-related new physics at submillimeter distances [@DimGer03; @HarObrMcG05; @MurHau09], optical clocks [@DerObrDzu09; @DzuDer10], atom-graphene interactions [@AroKauSah14], noncontact van der Waals friction [@JenLacDeK15], and interactions between nanostructures [@TaoPer14]).
In a previous paper treating the sodium atom [@KhaBabDal97], a semi-empirical theory utilizing oscillator strength sum rules and input data from experiments and calculations predicted a value of the van der Waals coefficient [@KhaBabDal97], which was found to be in harmony with subsequent experimentally determined fits from photo-association spectroscopy data [@vanVer99; @KnoSchSce11] and *ab initio* theoretical methods [@DerJohSaf99]. In the case of sodium, the availability of precise measurements of the principal resonance line oscillator strength from photo-association spectroscopy and of the static electric dipole polarizability from atom interferometry augmented the semi-empirical analysis [@KhaBabDal97]. In the case of magnesium, such data are not available. Therefore, in this paper, for Mg, the electric dipole oscillator strength distribution is composed using extant data on electric dipole oscillator strengths, photoabsorption cross sections, and energies obtained experimentally and theoretically. As will be shown, consistency with the oscillator strength (Thomas-Reiche-Kuhn) sum rule requires a value of the principal resonance line oscillator strength that is several percent lower than values listed in critical tabulations by Morton [@Mor03] and by Kelleher and Podobedova [@KelPod08]. Other evidence for the value adopted is given. The static electric dipole polarizability is evaluated and compared with other values. The dynamic polarizability function is calculated and compared with previous results obtained by Porsev et al. [@DerPorBab10] using configuration interaction and many-body perturbation theory with core contributions (CI-MBPT) [@PorDer06] and by Jiang *at al.* [@JiaMitChe15] using the configuration interaction with semi-empirical core-valence interaction (CICP) method. The present dynamic polarizability function is used to evaluate the van der Waals constant, Axilrod-Teller-Muto constant, and atom-surface interactions. Results from the literature for the principal resonance line oscillator strength, static electric dipole polarizability, and van der Waals constant are collected and compared in the Appendix.
Dipole oscillator strength sum rules
====================================
The absorption oscillator strength from the ground state $|0\rangle$ with eigenvalue $E_0$ to an excited state $|n\rangle$ with eigenvalue $E_n$ is $$f_n = {\textstyle{\frac{2}{3}}}(E_n-E_0) \left|\left\langle 0 \left| \sum_{i=1}^N {\mathbf r}_i \right| n\right\rangle\right|^2,$$ where ${\mathbf r}_i$ is the position vector of electron $i$, and $N$ is the number of electrons. Atomic units are used throughout unless otherwise specified.
Denoting by $\allsum_n$ the sum-integral (the sum over all discrete transitions excluding the initial state and the integration over all continuum states), the resultant sum rules are $$S(0) = \allsum_n f_n = N ,$$ with $N=12$ for Mg, $$S(-1) = \allsum_n f_n / (E_n-E_0) =
{\textstyle{\frac{2}{3}}} \left\langle 0 \left| \left(\sum_{i=1}^N {\mathbf r}_i \right)^2\right| 0 \right\rangle ,$$ and $$S(-2) =\allsum_n f_n / (E_n-E_0)^2 = \alpha(0) ,$$ where $\alpha(0)$ is the static electric dipole polarizability. The dynamic electric dipole polarizability function is $$\alpha(\omega) = \allsum_n \frac{f_n}{(E_n-E_0)^2 -\omega^2 },$$ where $\omega$ is the photon energy. By direct integration the $S(-1)$ sum rule is related to the atom-wall interaction coefficient $C_3$ [@KhaBabDal97] $$\label{atom-wall-formula}
C_3 = {\textstyle{\frac{1}{8}}} S(-1) = \frac{1}{4\pi}\int_0^\infty d\omega \,\alpha(i \omega),$$ the van der Waals coefficient is $$\label{vdW-formula}
C_6 = \frac{3}{\pi}\int_0^\infty d\omega\,[\alpha (i \omega)]^2 ,$$ and the Axilrod-Teller-Muto coefficient is $$\label{Axilrod-Teller-formula}
C_9 = \frac{3}{\pi} \int_0^\infty d\omega\;[\alpha (i\omega)]^3.$$
Oscillator strength distribution {#OSD}
================================
A magnesium atom has twelve electrons. Their configuration is $(1s^2 2s^2 2p^6 3s^2 )\;{}^1S_0$.
Discrete transitions {#discrete}
--------------------
Sources with tabulations of values, experimental and theoretical, for the absorption oscillator strengths are Mitchell [@Mit75a], Mendoza and Zeippen [@MenZei87a], Ray and Mukherjee [@RayMuk89], Jönsson and Fischer [@JonFis97], Hamonou and Hibbert [@HamHib08], and Derevianko and Porsev [@DerPor11].
There are numerous theoretical determinations of the oscillator strength for the principal resonance line $(3s^2)\; {}^1S - (3s3p) \;{}^1P^o$. A detailed survey is given in the Appendix. Reliable theoretical calculations range from 1.709 to 1.76, and there were at least ten experimental determinations as of 2003 [@Mor03]. In a critical review, Morton [@Mor03] adopted a value of $1.83\pm 0.03$ based on a weighted mean of the ten experimental values. A long-standing discrepancy between experimental and theoretical trends was noted previously [@Fis75; @VicSteLau76; @JonFisGod99; @ZatBarGed09]. The tendency of high-level theoretical results to be less than 1.8 was noted recently by Zatsariny *et al.* [@ZatBarGed09] who calculated a nonrelativistic value of 1.738 and pointed out that a “very extensive and essentially converged multiconfiguration Hartree-Fock (MCHF)” *ab initio* calculation by Jönsson, Fischer, and Godefroid [@JonFisGod99] found 1.717. Jönsson, Fishcher, and Godefroid [@JonFisGod99], using the observed transition energy to evaluate the oscillator strength, obtained 1.710 and also pointed out (see their Table 11) that (as of 1999) theoretical values were consistently smaller than the experimental ones (to date) by about 5 percent. Recently, Derevianko and Porsev [@DerPor11] quote for the matrix element governing the line strength a value $4.03 \pm 0.02$ with error of 0.5% based on their calculations from 2001 [@PorKozRak01]; the corresponding oscillator strength using the experimental transition energy [@KelPod08] is $1.73 \pm 0.02$. The original NIST (NBS) tabulation of 1969 [@WieSmiMil69] adopts $1.8\pm 0.18$ from an average of the experiments of Refs. [@Lur64; @SmiGal66] and the calculation of Weiss [@Wei67]. The value 1.8 (cited as a private communication from A. W. Weiss) is given with 3 percent error $(\pm 0.05)$ at the 90 percent confidence level in the 2008 NIST revised tabulation [@KelPod08].
I adopt the value 1.75, which is the lower limit of the value 1.80(5) from Kelleher and Podobedova [@KelPod08] and the upper limit of the value $1.73(2)$ recommended by Derevianko and Porsev.
Mitchell [@Mit75a] used the anomalous dispersion (hook) method to measure the second $(3s$–$4p)$ through sixth $(3s$–$8p)$ resonance transition oscillator strengths and found, respectively, $0.107\pm 0.0019$, $(2.27\pm 0.12)\times10^{-2}$, $(8.53\pm 0.46)\times10^{-3}$, $(4.11\pm 0.36)\times10^{-3}$, and $(2.34\pm 0.15)\times10^{-3}$, all determined relative to the value of 1.72 for the principal transition that Victor and Laughlin [@VicLau73] calculated using a semi-empirical model potential method. Other theoretical determinations of $f$ values for the second and higher resonance transitions are those of Saraph [@Sar76], Mendoza and Zeippen [@MenZei87a], Chang and Tang [@ChaTan90] and Zatsarinny *et al.* [@ZatBarGed09]. I adopt the values of Chang and Tang, who calculated for the principal to sixth resonance lines, respectively, oscillator strengths 1.75, 0.111, 0.024, 0.0091, 0.0043, and 0.0024. These five discrete (second to sixth resonance) transitions contribute $0.151$ to the $S(0)$ sum and $2.81$ to the $S(-2)$ sum for the $3s$ shell.
Including the principal resonance line, the *discrete transition* contribution from the $3s$ shell to $S(0)$ is 1.90 and to $S(-2)$ is $71.4$.
Continuum transitions
---------------------
A number of sources exist for the continuum oscillator strengths corresponding to the ejection of a $3s$ electron [@DitMar53; @BurSea60; @BatAlt73; @ParReeTom76; @DesMan83; @PreBurGar84; @RadJoh85; @YehLin85; @FisSah87; @MenZei87b; @MocSpi88b; @Alt89; @VerYakBan93; @ChiHua94; @FunYih01; @KimTay00; @WehLukJur07; @HauKamKos88; @WanWanZho10; @PinBalAbd13; @LeeBalAbd15]. The photoionization cross sections calculated using a variational MCHF method by Fischer and Saha [@FisSah87] are in good agreement with the experimental results of Wehlitz *et al.* [@WehLukJur07]. The threshold cross section of 2.5 Mb calculated by Fischer and Saha [@FisSah87] is slightly larger than both the value $2.1\pm0.3$ Mb measured by Fung *et al.* [@FunYih01] and the value $2.36\pm 0.02$ Mb that Parkinson, Reeves, and Tomkins [@ParReeTom76] found by extrapolation from the measured discrete oscillator strengths of Ref. [@Mit75a]. Wehlitz *et al.* [@WehLukJur07] normalized their own measurements to the value 2.1 Mb at threshold. At a photon energy of 30 eV, a recent calculation by Pindzola *et al.* [@PinBalAbd13] using a time-dependent close-coupling method with an effective core potential gives $\sigma_{3s} (30\,\mathrm{eV})$ $=$ $0.217\;\mathrm{Mb}$, while Verner *et al.* [@VerYakBan93] calculate $0.255\;\mathrm{Mb}$. At 80 eV, the “complete” experiment of Haussman *et al.* [@HauKamKos88], for which the total absorption cross section of Ref. [@HenLeeTan82] was used for normalization, yields $\sigma_{3s} (80\,\mathrm{eV})$ $=$ $0.080\pm 0.011\;\mathrm{Mb}$ from the main transition compared to the value $0.087$ from Verner *et al.* Haussman *et al.* also measured an additional $0.014 \pm 0.004\;\mathrm{Mb}$ contribution from satellites. The measurements of Wehlitz *et al.* [@WehLukJur07] are adopted for energies from threshold to 11.6 eV and the $3s$ cross section data were extended to higher energies using the results of Verner *et al.* [@VerYakBan93].
The contributions to the three sum rules $S(0)$, $S(-1)$, and $S(-2)$ from the continuum are, respectively, 0.261, 0.162, and 0.277. Combining these with the discrete contributions from the Sec. \[discrete\] the total valence shell contributions are 2.16, 11.8, and 71.7, which are listed in Table \[osc\]. The calculated valence shell contribution of 2.16 to the $S(0)$ sum confirms that the contribution from the $3s$ electron to $S(0) $ is greater than 2 [@Stw71; @ResCurBro08], indicating configuration interaction of the valence electrons with the core electrons. Maeder and Kutzelnig [@MaeKut79] obtained 2.06 using a model potential including core-valence correlation.
The excitation and ejection of $K$ shell electrons was considered by Verner *et al.* [@VerYakBan93], Kutzner *et al.* [@KutMayTho02], and Hasoğlu *et al.* [@HasAbdNab14]. Hasoğlu *et al.* used $R$-matrix methods to calculate excitation of $K$ shell electrons to $np$ states resonant below the threshold. The resonances are estimated to contribute 0.03 to the value of $S(0)$. The relativistic random-phase approximation modified to include relaxation effects (RRPAR) calculations of Kutzner *et al.* are in good agreement with the cross sections calculated by Verner *et al.* from threshold to high energies. Banna *et al.* [@BanSlaMat82] measured the shake-up peak just above the threshold, but absolute cross sections are not available. The cross sections of Verner are adopted from the threshold for ejection of a $1s$ electron at 1310.9 eV. The $K$ shell contribution yields 1.56 to $S(0)$ and it is negligible for $S(-1)$ and $S(-2)$. The $1s$ contribution to $S(0)$ found here for Mg is comparable to that found for Na [@KhaBabDal97].
Deshmukh and Manson [@DesMan83], Nasreen, Manson, and Deshmukh [@NasManDes89], Kutzner *et al.* [@KutMayTho02], and Verner *et al.* [@VerYakBan93] calculated the partial cross sections for the ejection of a $2s$ electron. The cross section value 0.3 Mb at the threshold energy of 94.0 eV is adopted [@NasManDes89; @KutMayTho02] and linearly joined to the results of Verner *et al.* at 270 eV, which are used for higher energies. For the $2s$ shell the contribution to the sums $S(0)$, $S(-1)$, and $S(-2)$ are, respectively, 1.02, 0.1, and 0.02.
The remaining oscillator strength must come from the $2p$ shell and the expected contribution to $S(0)$ is 7.26.
Deshmukh and Manson [@DesMan83], Altun [@Alt89], Nasreen, Manson, and Deshmukh [@NasManDes89], Kutzner, Maycock, and Thorarinson [@KutMayTho02], investigated photoionization of a $2p$ electron. There a number of resonances corresponding to excitation to autoionizing states [@Alt89] and they contribute significantly to the oscillator strength. Measurements of the cross section by Haussman *et al.* [@HauKamKos88] at 80 eV found that the resonances constitute 25% of the $2p$ shell photoionization cross section. The cross sections averaged over the resonances from the correlated length gauge many-body perturbation theory (MBPT) calculations of Altun (Figure 9 of Ref. [@Alt89]) are adopted from 63.29 eV to 344.89 eV, giving $\sigma_{2p} (80\,\mathrm{eV})$ $=$ $6.2\;\mathrm{Mb}$ which is slightly larger than the reference value used by Haussman *et al.* The calculations of Verner *et al.* are used from the threshold energy of 54.9 eV up to 63.29 eV and for energies above 344.89 eV. These data yield for the $2p$ shell the contributions to the sums $S(0)$, $S(-1)$, and $S(-2)$, respectively, 7.33. 1.61, and 0.45. If the calculations of Altun are multiplied by a factor of 0.985, the cross section at 80 eV becomes $6.1$ Mb and the sums $S(0)$, $S(-1)$, and $S(-2)$ are calculated to be, respectively, 7.26, 1.60, and 0.44 and the values are listed in Table \[osc\].
The contributions to the sums $S(0)$, $S(-1)$, and $S(-2)$ from the excitation of the $1s$, $2s$, $2p$, and $3s$ electrons are summarized in Table \[osc\].
------- ------ ------ -------
1s 1.56 0.02 $...$
2s 1.02 0.10 0.02
2p 7.26 1.60 0.44
3s 2.16 11.8 71.7
Total 12.0 13.5 72.2
------- ------ ------ -------
Discussion
----------
Stwalley [@Stw71], Pal’chikov and Ovsiannikov [@PalOvs04], Ovsiannikov *et al.* [@OvsPalKat06], and Sarkisov *et al.* [@SarBeiShe06] constructed oscillator strength distributions of Mg for calculations of dynamic polarizabilities, considering valence transitions. Sarkisov *et al.* [@SarBeiShe06] included an estimate of $2p$ excitations.
From their tabulated data, the results of Pal’chikov and Ovsiannikov [@PalOvs04] and Ovsiannikov *et al.* [@OvsPalKat06] indicate a total discrete contribution of $1.9$ to the $S(0)$ sum rule for $3s$ discrete transitions in agreement with the present result. They used a value of $1.73$ for the principal resonance line oscillator strength, which offsets in the sum rule their slightly larger value of $0.122$ for the second resonance line oscillator strength, compared to the present adopted values of, respectively, $1.75$ and $0.111$. For the $S(-2)$ sum, they find for the second to sixth resonance transitions a contribution of $3.12$ to $S(-2)$, to be compared to the present value of $2.81$. The difference between their value and the present value is due primarily to the different values of the second resonance line oscillator strength. Their total value (valence electrons) for $S(-2)$ is 71.39, while the present value for the $3s$ shell is 71.7.
Including all shells, the present value of $\alpha(0)$ is 72.2 from the $S(-2)$ sum rule, see Table \[osc\]. It lies only 0.2 above the range of values $71.3(7)$ recommended by Porsev and Derevianko [@PorDer06] and it is compared with a number of other theoretical calculations in the Appendix.
Note that if the value 1.8 is adopted for the principal oscillator strength [@KelPod08], without any other adjustments to the adopted data, the present $S(0)$ sum becomes 12.05 and $\alpha(0)$ becomes $73.4$, which is far beyond the value recommended by Porsev and Derevianko. The value 1.83 for the oscillator strength adopted by Morton [@Mor03] is more difficult to reconcile within the present analysis. The $S(0)$ sum becomes 12.08 and the value of $\alpha(0)$ becomes 74.6. Sarkisov *et al.* [@SarBeiShe06] use the oscillator strength data from Morton [@Mor03] and estimate the $2p$ and $3s$ continuum contributions using the data from Verner *et al.* [@VerYakBan93] and find $\alpha(0)=73.6$. Stwalley’s early calculation [@Stw71] used the value $1.82(5)$ for the principal resonance line oscillator strength (see his Ref. 7 for sources) and obtained an estimate $\alpha(0)=75.0 \pm 3.0$.
A recent experiment [@MaIndZha15] using a pulsed cryogenic molecular beam electric deflection method obtained a value $\alpha(0) = 59(15)$, which is not sufficiently accurate to discriminate between theoretical calculations.
Reshetnikov *et al.* [@ResCurBro08] explored the relationship between the uncertainty in $\alpha(0)$ and the uncertainty in the lifetime of the first resonance transition in two-valence electron atoms and ions. Their formalism allows a valence shell contribution to $N$ that is not exactly 2, as found here and in Ref. [@MaeKut79]. In terms of the the valence contribution, $N_e$, the principal resonance line oscillator strength $f_{3s,3p}$, and the excitation energies of the first and second resonance transitions, respectively, $E_{3s,3p}$ and $E_{3s,4p}$, they give $$\alpha(0) = \frac{f_{3s,3p}}{E^2_{3s,3p}} + \frac{N_e-f_{3s,3p}}{2 E^2_{3s,4p}}$$ and an uncertainty estimate for the polarizability $$\Delta \alpha(0) =\frac{N_e-f_{3s,3p}}{2 E^2_{3s,4p}}.$$ Using the present adopted value $f_{3s,3p}=1.75$, calculated value $N_e=2.16$, and transition energies [@KelPod08] $E_{3s,3p}= 0.159\,705$ and $E_{3s,4p}=0.224\,840$, yields an estimate $\alpha(0)=72.7 \pm 4$. Likewise, using their formula for estimating the uncertainty of $f_{3s,3p}$, given the present calculated value $\alpha(0)=72.2$, yields $f_{3s,3p}= 1.79 \pm 0.07$. The formulae of Reshetnikov *et al.* [@ResCurBro08] demonstrate that the present results are mutually consistent, but the estimates obtained are not sufficiently precise to allow selection of a particular value of $f_{3s,3p}$ from the many available values, see Appendix.
The availability of a more accurate measurement of $\alpha(0)$ and a definitive measurement of the principal resonance line lifetime would significantly improve the present model [@KhaBabDal97]. Nevertheless, the values adopted here, in particular $f_{3s,3p}=1.75$, generate sum rules that are consistent and not in contradiction with other major studies, while a value of $f_{3s,3p} \geq 1.8$ is inconsistent.
Dynamic electric dipole polarizability function {#alpha-omega}
===============================================
The dynamic electric dipole polarizability function at imaginary frequencies is constructed using the discrete and continuum oscillator strength data as assembled in Sec. \[OSD\].
The continuum oscillator strength distribution is given in terms of the photoionization cross section $\sigma(E)$ by $$\frac{df}{dE} = \frac{\sigma(E)}{2\pi^2 \alpha_{\mathrm{fs}}}, \quad E > 0.281 ,$$ with $\alpha_{\mathrm{fs}}$ the fine structure constant, and the dynamic dipole polarizability at imaginary energy is $$\label{eq-alpha-omega}
\alpha(i\omega) = \sum_n \frac{f_n}{(E_n- E_0)^2 + \omega^2}+ \int dE\,\frac{df/dE}{E^2 + \omega^2} .$$
The function $\alpha (i\omega)$ resulting from the analysis in Sec. \[OSD\] is shown at low energies in Fig. \[fig-pol-low\].
![The dynamic dipole polarizability function $\alpha(i\omega)$ from the present calculations. \[fig-pol-low\]](pol-low.pdf)
It may be compared with the calculations of Derevianko *et al.* [@DerPorBab10] and those of Jiang *et al* [@JiaMitChe15]. The present function $\alpha (i\omega)$ was evaluated at the fifty energies corresponding to the energies $\omega_k$ of a 50-point quadrature, as listed in Table A of Ref. [@DerPorBab10], and the energies for a 40-point quadrature listed in Table C of Jiang *et al.* In Fig. \[fig-allthree\] the data are plotted. Agreement is very good between the present results and the CI-MBPT results of Ref. [@DerPorBab10]. There are noticeable discrepancies between the present results and the CICP results of Ref. [@JiaMitChe15].
![For the dynamic dipole polarizability function $\alpha(i\omega)$, comparison between the present values (line), the configuration interaction with semi-empirical core-valence interaction (CICP) values from Ref. [@JiaMitChe15] (circles), and the configuration interaction and many-body perturbation theory with core interactions (CI-MBPT) values from Ref. [@DerPorBab10] (squares). \[fig-allthree\]](allthree.pdf)
To further investigate the discrepancies, in Fig. \[fig-alpha-diffs\] the percentage difference between the values from the functions given in Ref. [@DerPorBab10] or Ref. [@JiaMitChe15] and the present values are shown.
![For the dynamic dipole polarizability function $\alpha(i\omega)$, percent difference comparison between the present values and the CI-MBPT values from Ref. [@DerPorBab10] and between the CICP values from Ref. [@JiaMitChe15] . The quantity plotted is $100\times [( \textrm{other})- (\textrm{present} )]/\textrm{present}$, where “other” is either Ref. [@DerPorBab10] or [@JiaMitChe15]. The filled squares are the percent differences of the present values from the CI-MBPT values and the filled circles are the percent differences of the present values from the CICP values. The plotted values show that the present $\alpha(i\omega)$ is within several percent of the CI-MBPT values across the range of energies and show that the CICP values are larger for photon energies between roughly 1 and 200 atomic units. \[fig-alpha-diffs\]](compare.pdf)
The present model and the calculations of Ref. [@DerPorBab10] agree within several percent at all energies. The present values are larger than those of Ref. [@DerPorBab10] at low energies, due to the larger principal oscillator strength adopted herein. The several percent agreement across all energies is satisfactory and a more detailed analysis might await an accurate experimental value for $\alpha(0)$ and a definitive measurement of the principal resonance line oscillator strength. In contrast, the differences between the present model and the model of Ref. [@JiaMitChe15] are made apparent in Fig. \[fig-alpha-diffs\] showing that the CICP model of Ref. [@JiaMitChe15] yields larger values for $\alpha(i\omega)$ in the energy range of 1 to 200 au. The difference may arise due to the choice of “effective” core oscillator strengths in the CICP model [@MitBro03; @JiaMitChe15]. The percent difference peaks at about 16 % around 5–6 $e^2/a_0$ (135–160 eV) placing the missing oscillator strength of the CICP model in the inner $s$ shells, where the “effective” oscillator strengths are placed to model inner shell absorption [@MitBro03; @JiaMitChe15].
Applications
============
van der Waals coefficient $C_6$ {#C6}
-------------------------------
The long-range potential energy between two Mg atoms separated by a distance $R$ is $-C_6/R^6$, where $C_6$ is given by Eq. (\[vdW-formula\]). For the van der Waals coefficient, I find $C_6 = 642.4$ by evaluating Eq. (\[vdW-formula\]) using the quadrature method of Ref. [@DerPorBab10], $$C_6 \approx \frac{3}{\pi} \sum_{j=1}^{50} w_j \alpha^2(i\omega_j),$$ using the present values of $\alpha(i\omega)$ evaluated at the energies $\omega_j$ and with the weights $w_j$ given in Table A of Ref. [@DerPorBab10]. Porsev and Derevianko [@PorDer02] quote accuracy of 2% or better for their value $C_6=627(12)$ obtained using a semi-empirical hybrid relativistic many-body perturbation theory (CI-MBPT) approach. The present value lies just above their range, mainly corresponding to their principal resonance line oscillator strength of 1.73, compared to the present adopted value of 1.75, thus, $642 \times (1.73/1.75)^2 \approx 627$. The present result improves upon the earlier empirical estimate of $683(35)$ by Stwalley [@Stw71]. A more detailed survey and comparison of other determinations of $C_6$ is given in the Appendix.
Atom-wall coefficient $C_3$ {#C3}
---------------------------
The long-range potential energy of an Mg atom at distance $z$ from a perfectly conducting wall is $-C_3/z^3$, where $C_3$ is given by Eq. (\[atom-wall-formula\]). Mitroy and Bromley [@MitBro03] calculated $C_3 = 1.704$ using the CICP approach, while the CI-MBPT value is $1.666$ [@DerPorBab10]. Lonij *et al.* [@LonKlaHol11] gave an approximate value of $1.51$ using a limited 4-parameter model for the dynamic polarizability. The value of $C_3$ is known to be sensitive to the completeness of the description of the core electrons [@FroFisJon98; @DerJohSaf99; @LonKlaHol10].
The present value is $C_3=1.69$ using $C_3={\textstyle{\frac{1}{8}}} S(-1)$ and the value of $S(-1)$ from Table \[osc\] and 1.687 using Eq. (\[atom-wall-formula\]) and the quadrature from Ref. [@DerPorBab10], $$C_3 \approx \frac{1}{4\pi} \sum_{j=1}^{50} w_j \alpha(i \omega_j) .$$ Both of the present values (sum rule and quadrature) are in larger than that of Ref. [@DerPorBab10]. The slightly larger value of $C_3$ from the CICP calculations, Ref. [@JiaMitChe15], is consistent with the relatively larger values of $\alpha(i\omega)$, as discussed in Sec. \[alpha-omega\] and shown in Figs. \[fig-allthree\] and \[fig-alpha-diffs\].
Other properties
----------------
The Axilrod-Teller-Muto coefficient $C_9$, Eq. (\[Axilrod-Teller-formula\]), characterizes the mutual long-range interaction potential of three atoms. The value obtained by Mitroy and Bromley [@MitBro03] is $33\,380$ and that obtained by Porsev *et al.* [@DerPorBab10] is $33\,241$. Using the dynamic polarizability function I evaluated, Eq. (\[Axilrod-Teller-formula\]), using the quadrature of Ref. [@DerPorBab10], $$\label{C9-approx}
C_9 \approx \frac{3}{\pi} \sum_{j=1}^{50} w_j \alpha^3 (i\omega_j),$$ and obtained $C_9=34\,480$.
The larger value for $C_9$ found here mainly reflects the larger principal oscillator strength 1.75 adopted compared to the principal oscillator strengths found in Refs. [@MitBro03] and [@DerPorBab10]. The oscillator strength appears as a cubic power in Eq. (\[C9-approx\]) through $\alpha(i\omega)$, see Eq. (\[eq-alpha-omega\]). For example, comparing to Ref. [@DerPorBab10], which used a principal oscillator strength of 1.73, scaling the present value I obtain $34\,450 \times (1.73/1.75)^3 = 33\,310$, which is within $0.2$ % of the value of Ref. [@DerPorBab10].
Conclusion
==========
Experimental and theoretical data were assembled and used to formulate the dynamic polarizability function for Mg. I find that consistency in the sum rules can be achieved using the adopted value of the principal resonance line oscillator strength to be $1.75$; lower than the curated values of 1.83 [@Mor03] and 1.8 [@KelPod08], but in agreement with theoretical calculations. Comparisons of the dynamic dipole polarizability functions from the present work and those calculated using the CI-MBPT approach and the CICP approach were presented. Good agreement (within several percent) was found with the CI-MBPT results over all photon energies providing an independent confirmation of the CI-MPBT approach for Mg [@DerPorBab10]. For the CICP method the differences were more pronounced, approaching 16% at energies around 5–6 $e^2/a_0$, or about 135–160 eV, indicating that the “effective" oscillator strengths of Refs. [@MitBro03] and [@JiaMitChe15] may not completely model oscillator strengths corresponding to the core electrons. To improve the present model it would be valuable to have more accurate experimental measurements of the polarizability and a definitive measurement of the principal oscillator strength. Where sufficient and reliable data is available, the present methodology can be applied to other atoms. *Note added in proof: A recent calculation \[Y. Singh, B. K. Sahoo, and B. P. Das, Phys. Rev. A **88**, 062504 (2013)\], using a relativistic coupled cluster theory with all singly and doubly excited configurations, found a polarizability value of 72.54(50). I thank Dr. Singh for communicating this result.*
Discussions with C. Ballance, T. G. Lee, T. Gorczyca, C. F. Fischer, S. Manson, and M. Bromley are gratefully acknowledged. This work was supported in part by grants for ITAMP from the National Science Foundation to the Smithsonian Institution and to Harvard University.
Values from the literature
==========================
In this Appendix, values for Mg of the principal resonance transition oscillator strength, of the static electric dipole polarizability, and of the van der Waals constant are collected from the literature. Some earlier collections include Refs. [@Mit75a; @MenZei87a; @MocSpi88b; @RayMuk89] for the principal oscillator strength, Refs. [@ReiMey76; @MaeKut79; @RayMuk89; @MitBro03; @HamHib08; @MitSafCla10] for the polarizability, and Ref. [@MitBro03] for the van der Waals constant.
------------------ --------- ------------------------------- -----------------------
MCDF-CV 1.709 Jönsson and Fischer [@JonFis97] (1997)
MCHF 1.717 Jönsson, Fischer, & Godefroid [@JonFisGod99] (1999)
MP 1.717 Victor and Laughlin [@VicLau73] (1973)
MCHF+BP 1.719 Fischer, Tachiev, & Irimia [@FroFisTac06] (2006)
RCI+Breit 1.722 Cheng *et al.* [@CheGaoQin11] (2011)
CIDF-CP 1.72 Stanek, Glowacki, & Migdalek [@StaGloMig96] (1996)
CI+MBPT[^1] 1.724 Savukov and Johnson [@SavJoh02] (2002)
CI[^2] 1.725 Mengali and Moccia [@MenMoc96b] (1996)
CI[^3] 1.73 Nesbet and Jones [@NesJon77] (1977)
CI+MBPT[^4] 1.73(2) Derevianko and Porsev [@DerPor11] (2011)
CICP 1.732 Mitroy and Bromley [@MitBro03] (2003)
CI 1.735 Hamonou and Hibbert [@HamHib08] (2008)
CI[^5] 1.737 Weiss [@Wei67] (1967)
MCHF-CV 1.738 Zatsarinny *et al.* [@ZatBarGed09] (2009)
MCHF[^6] 1.747 Fischer [@Fis75] (1975)
CI-frozen core 1.75 Chang and Tang [@ChaTan90] (1990)
CI-CV[^7] 1.755 Moccia and Spizzo [@MocSpi88b] (1988)
CI-frozen core 1.76 Saraph [@Sar76] (1976)
NIST adopted[^8] 1.8 Kelleher and Podobedova [@KelPod08] (2008)
Experiment[^9] 1.83(3) Morton [@Mor03] (2003)
------------------ --------- ------------------------------- -----------------------
For the oscillator strength, as discussed in the Sec. \[OSD\], it was noted in several recent papers [@JonFisGod99; @ZatBarGed09] that in general the most sophisticated theoretical calculations lie several percent below the published experimental values, see also earlier similar comments in Refs. [@VicLau73], [@VicSteLau76], and [@MenZei87a]. Also, it was noted that the “best” calculations lie below [@ZatBarGed09] the adopted value of $1.8$ in the NIST tabulation [@KelPod08]. As shown in Table \[osc-table\], the configuration interaction (CI), multi-configuration Dirac-Fock (MCDF), and multi-configuration Hartree-Fock with Breit-Pauli interactions (MCHF+BP) calculations, are all in general agreement; the latter two methods include relativistic effects. In addition, the semi-empirical model potential (MP) calculation of Victor and Laughlin [@VicLau73] agrees with the MCHF calculation. In the Table, where treatment of core-valence correlation was included the suffix CV is appended. In addition, the relativistic configuration interaction with the Breit interaction (RCI+Breit ) of Ref. [@CheGaoQin11] are in close agreement with the configuration interaction Dirac-Fock with core polarization (CIDF-CP) calculation of Ref. [@StaGloMig96]. CI+MBPT methods use complete relativistic CI calculations for the valence electrons in a frozen core combined with MBPT to account for core-valence interactions. The CI frozen core calculations of Saraph [@Sar76] and Chang and Tang [@ChaTan90] and the CI-CV calculations of Moccia and Spizzo [@MocSpi88b] are in close agreement with values between 1.75 and 1.76. There is substantial theoretical evidence for a value of the principal oscillator strength around 1.75.
Many calculations of the static polarizability $\alpha(0)$ are available. There are several good tables containing summaries of other earlier works. In particular, from 1976, Reinsch and Meyer [@ReiMey76] and from 1991, Archibong and Thakkar [@ArcTha91], see also Schwerdtfeger [@Sch06]. Thakkar and Lupinetti [@ThaLup06] recommend a theoretical value of $71.22\pm 0.36$, which includes a relativistic correction of $-0.35$, and Chu and Dalgarno [@ChuDal04] recommend 71.
Table \[alpha\] lists some of the results from the literature. The quantum defect theory (QDT) value from Chernov *et al.* [@CheDorKre05] models the response of the valence electrons only. The pseudopotential (PP) [@MaeKut79] and model potential (MP) calculations [@VicSla74; @Pat00] model the response of the valence electrons with inclusion of effective potentials to treat the core electrons. An effective core potential is used similarly in the configuration interaction core potential (CICP) calculation of Müller, Flesch, and Meyer [@MulFleMey84]. The CICP calculation of Mitroy and Bromley [@MitBro03] utilizes a model that treats core excitation using effective oscillator strengths designed to reproduce the core polarizability. The multi-reference configuration interaction (MRCI) calculation of Partridge *et al.* [@ParBauPet90] gives 71.2. They also calculated a CI-CV value of 70.3 (not listed in Table \[alpha\]), which is in good agreement with the similar calculation of 70.9 by Hamanou and Hibbert [@HamHib08], but the average value of 74.37 from the CI-CV calculations by Moccia and Spizzo [@MocSpi88c] is significantly larger. The coupled cluster double-excitation with contributions of single and triple excitations \[CCD+ST (CCD)\] model of Castro and Canuto [@CasCan93] yields a value of 70.89, somewhat lower than the MRCI value and the fourth order many body perturbation theory \[MBPT(4)\] value of 71.7 calculated by Archibong and Thakkar [@ArcTha91], while the coupled cluster with single and double excitation-effective Hamiltonian (CCSD-EH) approach of Stanton [@Sta94] yields 72.2 using basis sets from Ref. [@WidJoaPer91], denoted WMR in the Table. The pseudo-natural orbital coupled electron pair approximation (PNO-CEPA) calculation of Reinsch and Meyer [@ReiMey76] is close to the CI+MBPT calculation of Porsev and Derevianko [@PorDer06]. Two time-dependent density functional theory (TDDFT) calculations are included in Table \[alpha\]. Using time-dependent DFT with a self-interaction correction (TDDFT-SIC), Chu and Dalgarno [@ChuDal04] obtained 71.8 and using the symmetry adapted perturbation theory codes, SAPT(DFT), Patkowski *et al.* [@PatPodSza07] obtained 73.27 for the polarizability. An extensive table of values for the polarizability of Mg resulting from various density functional theory functionals is given in Ref. [@BasHesSal08]. The R-matrix calculation of Robb [@Rob75] is 75(5); the relatively large value results because core-valence correlation effects were not included [@MaeKut79]. Excluding the relatively large value from Ref. [@MocSpi88c], the CI, MRCI, CICP, and MBPT calculations fall in the range from 70.74 to 71.7. Reshetnikov *et al.* [@ResCurBro08] use a semi-empirical method that utilizes a sum rule with constraints and error bars determined using measured the lifetime and excitation energies. Their value is $74.4(2.7)$, with the accuracy limited by the available input data. The recent experiment of Ma *et al.* [@MaIndZha15] using a cryogenic molecular beam found a value of $59(15)$, which is not yet sufficiently accurate to test the calculations against.
--------------------------------------- ----------- ------------------------- -----------------------
Experiment (cryogenic molecular beam) 59(15) Ma *et al.* [@MaIndZha15] (2015)
QDT 69.54 Chernov *et al.* [@CheDorKre05] (2005)
PP 70.5 Maeder and Kutzelnigg [@MaeKut79] (1979)
CICP 70.74(71) Müller, Flesch, & Meyer [@MulFleMey84] (1984)
CCD+ST (CCD) 70.89 Castro and Canuto [@CasCan93] (1993)
CI 70.90 Hamanou and Hibbert [@HamHib08] (2008)
MRCI 71.2 Partridge *et al.* [@ParBauPet90] (1990)
CICP 71.35 Mitroy and Bromley [@MitBro03] (2003)
CI+MBPT 71.3(7) Porsev and Derevianko [@PorDer06] (2006)
PNO-CEPA 71.32 Reinsch and Meyer [@ReiMey76] (1976)
MBPT(4) 71.70 Archibong and Thakkar [@ArcTha91] (1991)
TDDFT-SIC 71.8 Chu and Dalgarno [@ChuDal04] (2004)
MP 72.0 Patil [@Pat00] (2000)
MP 72.1 Victor and Slavsky [@VicSla74] (1974)
CCSD-EH (WMR) 72.2 Stanton [@Sta94] (1994)
SAPT (DFT) 73.27 Patkowski *et al.* [@PatPodSza07] (2007)
CI-CV[^10] 74.37 Moccia and Spizzo [@MocSpi88c] (1988)
Sum rule 74.4(2.7) Reshetnikov *et al.* [@ResCurBro08] (2008)
R-matrix[^11] 75(5) Robb [@Rob75] (1975)
--------------------------------------- ----------- ------------------------- -----------------------
Table \[vdW\] presents a collection of van der Waals constant values from the literature and a significant range is apparent, though the CI+MBPT and CICP calculations, which include models of core electron excitations, are in good agreement. The value $620(5)$ from the model potential (MP) calculation of Santra, Christ, and Greene [@SanChrGre04] is close close to the pseudopotential (PP) calculation of 618.4 from Maeder and Kutzelnigg [@MaeKut79], both of which include effective potentials to account for the presence of core electrons, but don’t fully include their excitations. The MP calculation from Patil [@Pat00], however, is significantly larger, at 648. Three DFT calculations are listed in Table \[vdW\]. Hult et al. [@HulRydLun99] introduced a local dielectric function and cutoff on the interaction volume and obtained 615. In contrast, Chu and Dalgarno used time-dependent DFT with a self-interaction correction (TDDFT-SIC) and an empirical correction to obtain 626 with an estimated uncertainty of 1%. Patkoswki *et al.* [@PatPodSza07] used the symmetry adapted perturbation theory codes, SAPT(DFT), and obtained 635. The calculation of $C_6$ by Stanton [@Sta94] used a quadrature and values of the dynamic polarizability at imaginary frequencies calculated using the CCSD-EH coupled cluster approach with the basis sets from [@WidJoaPer91]. The CI-CV calculations of Moccia and Spizzo [@MocSpi88c] in the velocity gauge (VG) and in the length gauge (LG) are substantially larger than the other listed calculations. Robb [@Rob75] estimated his R-matrix calculation to be accurate to 10 %. Stwalley [@Stw71] used an empirically constructed polarizability function to calculate $C_6$. The large value for $C_6$ corresponds to the choice of 1.82 for the principal oscillator strength. In Ref. [@KnoRuhTie13] it was found that a 2% uncertainty in $C_6$ leads to an uncertainty of no more than 0.3 nm in the scattering length for ${}^{24}\textrm{Mg}_2$.
---------------- --------- -------------------------- -----------------------
DFT 615 Hult *at al.* [@HulRydLun99] (1999)
PP 618.4 Maeder and Kutzelnigg [@MaeKut79] (1979)
MP[^12] 620(5) Santra, Christ, & Greene [@SanChrGre04] (2004)
TDDFT-SIC[^13] 626(6) Chu and Dalgarno [@ChuDal04] (2004)
CI+MBPT 627(12) Porsev and Derevianko [@PorDer06] (2006)
CICP 629.5 Mitroy and Bromley [@MitBro03] (2003)
MP 632.27 Victor and Slavsky [@VicSla74] (1974)
SAPT(DFT) 635 Patkowski *et al.* [@PatPodSza07] (2007)
CCSD-EH (WMR) 648 Stanton [@Sta94] (1994)
MP 648 Patil [@Pat00] (2000)
CI-CV (VG) 658.1 Moccia and Spizzo [@MocSpi88c] (1988)
CI-CV (LG) 670.9 Moccia and Spizzo [@MocSpi88c] (1988)
Empirical 683(35) Stwalley [@Stw71] (1971)
R-matrix 689(70) Robb [@Rob75] (1975)
---------------- --------- -------------------------- -----------------------
[100]{}
E. Tiesinga, S. Kotochigova, and P. S. Julienne, Phys. Rev. A [**65**]{}, 042722 (2002).
H. Knöckel, S. Rühmann, and E. Tiemann, J. Chem. Phys. [**138**]{}, 094303 (2013).
H. Knöckel, S. Rühmann, and E. Tiemann, Eur. Phys. J. D [**68**]{}, 293 (2014).
P. Soldan, P. S. Zuchowski, and J. M. Hutson, Discuss. Faraday Soc. [**142**]{}, 191 (2009).
M. L. González-Martínez and J. M. Hutson, [Phys. Rev. A]{} [**[84]{}**]{}, 052706 ([2011]{}).
V. P. A. Lonij, C. E. Klauss, W. F. Holmgren, and A. D. Cronin, J. Phys. Chem. A [**115**]{}, 7134 (2011).
A. Vidal-Madjar, C. M. Huitson, V. Bourrier, J.-M. Désert, G. Ballester, A. Lecavelier des Etangs, D. K. Sing, D. Ehrenreich, R. Ferlet, G. Hébrard, and J. C. McConnell, Astron. Astroph. [**560**]{}, A54 (2013).
V. Bourrier, A. Lecavelier des Etangs, and A. Vidal-Madjar, Astron. Astroph. [**565**]{}, A105 (2014).
V. Bourrier, A. Lecavelier des Etangs, and A. Vidal-Madjar, Astron. Astroph. [**573**]{}, A11 (2015).
A. Derevianko, S. Porsev, and J. Babb, At. Data Nucl. Data Tables [**96**]{}, 323 (2010).
J. Jiang, J. Mitroy, Y. Cheng, and M. Bromley, At. Data Nucl. Data Tables [ **101**]{}, 158 (2015).
M. Y. Amusia, [*Atomic Photoeffect*]{} (Plenum, New York, 1990), trans. K. T. Taylor.
M. [Amusia]{}, L. [Chernysheva]{}, and V. [Yarzhemsky]{}, [*Handbook of Theoretical Atomic Physics Data for Photon Absorption, Electron Scattering, and Vacancies Decay*]{} (Springer, Berlin, 2012).
J. Berkowitz, [*Atomic and Molecular Photoabsorption: [Absolute]{} Total Cross Sections*]{} (Academic Press, San Diego, 2002).
M. Masili and A. F. Starace, [Phys. Rev. A]{} [**[62]{}**]{}, [033403]{} ([2000]{}).
T. Middelmann, S. Falke, C. Lisdat, and U. Sterr, Phys. Rev. Lett. [**109**]{}, 263004 (2012).
J. A. Sherman, N. D. Lemke, N. Hinkley, M. Pizzocaro, R. W. Fox, A. D. Ludlow, and C. W. Oates, Phys. Rev. Lett. [**108**]{}, 153002 (2012).
K. Beloy, J. A. Sherman, N. D. Lemke, N. Hinkley, C. W. Oates, and A. D. Ludlow, Phys. Rev. A [**86**]{}, 051404 (2012).
J. E. Stalnaker, D. Budker, S. J. Freedman, J. S. Guzman, S. M. Rochester, and V. V. Yashchuk, Phys. Rev. A [**[73]{}**]{}, [043416]{} ([2006]{}).
Z. W. Barber, J. E. Stalnaker, N. D. Lemke, N. Poli, C. W. Oates, T. M. Fortier, S. A. Diddams, L. Hollberg, C. W. Hoyt, A. V. Taichenachev, and V. I. Yudin, Phys. Rev. Lett. [**100**]{}, 103002 (2008).
I. B. Khriplovich, [*Parity Non-conservation in Atomic Phenomena*]{} (Gordon and Breach, New York, 1991).
R. Bast, A. Hesselmann, P. Sałek, T. Helgaker, and T. Saue, [ChemPhysChem]{} [**9**]{}, 445 (2008).
A. Tkatchenko and M. Scheffler, Phys. Rev. Lett. [**102**]{}, 073005 (2009).
A. Tkatchenko, R. A. DiStasio, R. Car, and M. Scheffler, Phys. Rev. Lett. [ **108**]{}, 236402 (2012).
J. Tao, J. P. Perdew, and A. Ruzsinszky, Int. J. Mod. Phys. B [**27**]{}, 1330011 (2013).
J. Toulouse, E. Rebolini, T. Gould, J. F. Dobson, P. Seal, and J. G. Angyan, J. Chem. Phys. [**138**]{}, 194106 (2013).
D. Goebel and U. Hohm, Phys. Rev. A [**52**]{}, 3691 (1995).
G. S. Sarkisov, I. L. Beigman, V. P. Shevelko, and K. W. Struve, Phys. Rev. A [**73**]{}, 042501 (2006).
L. Ma, J. Indergaard, B. Zhang, I. Larkin, R. Moro, and W. A. de Heer, Phys. Rev. A [**91**]{}, 010501 (2015).
P. Schwerdtfeger, in [*Atoms, Molecules, and Clusters in Electric Fields*]{}, edited by G. Maroulis (Imperial College Press, London, 2006), Chap. 1, p. 1.
A. J. Thakkar and C. Lupinetti, in [*Atoms, Molecules, and Clusters in Electric Fields*]{}, edited by G. Maroulis (Imperial College Press, London, 2006), Chap. 14, pp. 505–530.
J. Mitroy, M. S. Safronova, and C. W. Clark, J. Phys. B [**43**]{}, 202001 (2010).
S. Dimopoulos and A. A. Geraci, Phys. Rev. D [**68**]{}, 124021 (2003).
D. M. Harber, J. M. Obrecht, J. M. McGuirk, and E. A. Cornell, Phys. Rev. A [**72**]{}, 033610 (2005).
B. Murphy and L. V. Hau, Phys. Rev. Lett. [**102**]{}, 033003 (2009).
A. Derevianko, B. Obreshkov, and V. A. Dzuba, Phys. Rev. Lett. [**103**]{}, 133201 (2009).
V. A. Dzuba and A. Derevianko, J. Phys. B: At. Mol. Opt. Phys. [**43**]{}, 074011 (2010).
B. Arora, H. Kaur, and B. K. Sahoo, J. Phys. B [**47**]{}, 155002 (2014).
U. D. Jentschura, G. Łach, M. De Kieviet, and K. Pachucki, Phys. Rev. Lett. [**114**]{}, 043001 (2015).
J. Tao and J. P. Perdew, J. Chem. Phys. [**141**]{}, 141101 (2014).
P. Kharchenko, J. F. Babb, and A. Dalgarno, Phys. Rev. A [**55**]{}, 3566 (1997).
F. A. van Abeelen and B. J. Verhaar, Phys. Rev. A [**59**]{}, 578 (1999).
S. Knoop, T. Schuster, R. Scelle, A. Trautmann, J. Appmeier, M. K. Oberthaler, E. Tiesinga, and E. Tiemann, Phys. Rev. A [**83**]{}, 042704 (2011).
A. [Derevianko]{}, W. R. [Johnson]{}, M. S. [Safronova]{}, and J. F. [Babb]{}, Phys. Rev. Lett. [**82**]{}, 3589 (1999).
D. C. [Morton]{}, Astrophys. J., Suppl. Ser. [**149**]{}, 205 (2003).
D. E. Kelleher and L. I. Podobedova, J. Phys. Chem. Ref. Data [**37**]{}, 267 (2008).
S. G. Porsev and A. Derevianko, Zh. Eksp. Teor. Fiz. [**129**]{}, 227 (2006), \[JETP **102**, 195 (2006)\].
C. J. Mitchell, J. Phys. B [**8**]{}, 25 (1975).
C. [Mendoza]{} and C. J. [Zeippen]{}, Astron. Astroph. [**179**]{}, 339 (1987).
D. Ray and P. K. Mukherjee, J. Phys. B [**22**]{}, 2103 (1989).
P. Jönsson and C. F. Fischer, J. Phys. B [**30**]{}, 5861 (1997).
L. Hamonou and A. Hibbert, J. Phys. B [**41**]{}, 245004 (2008).
A. Derevianko and S. G. Porsev, [*Adv. At. Molec. Opt. Phys.*]{}, edited by E. Arimondo, P. R. Berman, and C. C. Lin (Academic Press, Amsterdam, 2011), Vol. 60, Chap. 9, pp. 415 – 459.
C. F. Fischer, Can. J. Phys. [**53**]{}, 338 (1975).
G. A. [Victor]{}, R. F. [Stewart]{}, and C. [Laughlin]{}, Ap. J., Suppl. [**31**]{}, 237 (1976).
P. Jönsson, C. F. Fischer, and M. R. Godefroid, J. Phys. B [**32**]{}, 1233 (1999).
O. Zatsarinny, K. Bartschat, S. Gedeon, V. Gedeon, V. Lazur, and E. Nagy, Phys. Rev. A [**79**]{}, 052709 (2009).
S. G. Porsev, M. G. Kozlov, Y. G. Rakhlina, and A. Derevianko, Phys. Rev. A [**64**]{}, 012508 (2001).
W. L. Wiese, J. R. Fuhr, and B. M. Miles, [*Atomic Transition Probabilities, Vol. II: Sodium through Calcium, NSRDS-NBS Vol. 22*]{} (US GPO, Washington, D.C., 1969).
A. Lurio, Phys. Rev. [**136**]{}, A376 (1964).
W. W. Smith and A. Gallagher, Phys. Rev. [**145**]{}, 26 (1966).
A. W. Weiss, J. Chem. Phys. [**47**]{}, 3573 (1967).
G. Victor and C. Laughlin, Nucl. Instrum. Methods [**110**]{}, 189 (1973).
H. E. Saraph, J. Phys. B [**9**]{}, 2379 (1976).
T. N. Chang and X. Tang, J. Quant. Spectrosc. Radiat. Transfer [**43**]{}, 207 (1990).
R. W. [Ditchburn]{} and G. V. [Marr]{}, Proc. Phys. Soc., London, Sect. A [**66**]{}, 655 (1953).
A. [Burgess]{} and M. J. [Seaton]{}, Mon. Not. R. Astron. Soc. [**120**]{}, 121 (1960).
G. N. [Bates]{} and P. L. [Altick]{}, J. Phys. B [**6**]{}, 653 (1973).
W. H. Parkinson, E. M. Reeves, and F. S. Tomkins, J. Phys. B [**9**]{}, 157 (1976).
P. C. Deshmukh and S. T. Manson, Phys. Rev. A [**28**]{}, 209 (1983).
J. M. Preses, C. E. Burkhardt, W. P. Garver, and J. J. Leventhal, Phys. Rev. A [**29**]{}, 985 (1984).
V. Radojevi ć and W. R. Johnson, Phys. Rev. A [**31**]{}, 2991 (1985).
J. Yeh and I. Lindau, At. Dat. Nucl. Dat. Tables [**32**]{}, 1 (1985).
C. F. Fischer and H. P. Saha, Can. J. Phys. [**65**]{}, 772 (1987).
C. [Mendoza]{} and C. J. [Zeippen]{}, Astron. Astroph. [**179**]{}, 346 (1987).
R. Moccia and P. Spizzo, J. Phys. B [**21**]{}, 1133 (1988).
Z. Altun, Phys. Rev. A [**40**]{}, 4968 (1989).
D. A. [Verner]{}, D. G. [Yakovlev]{}, I. M. [Band]{}, and M. B. [Trzhaskovskaya]{}, At. Dat. Nucl. Dat. Tables [**55**]{}, 233 (1993).
H.-C. Chi and K.-N. Huang, Phys. Rev. A [**50**]{}, 392 (1994).
H. S. Fung and T. S. Yih, Nucl. Phys. A [**684**]{}, 696C (2001).
D.-S. Kim and S. S. Tayal, J. Phys. B: At. Mol. Opt. Phys. [**33**]{}, 3235 (2000).
R. Wehlitz, D. Lukić, and P. N. Juranić, J. Phys. B [**40**]{}, 2385 (2007).
A. Hausmann, B. Kämmerling, H. Kossmann, and V. Schmidt, Phys. Rev. Lett. [**61**]{}, 2669 (1988).
G. Wang, J. Wan, and X. Zhou, J. Phys. B [**43**]{}, 035001 (2010).
M. S. Pindzola, C. P. Ballance, S. A. Abdel-Naby, F. Robicheaux, G. S. J. Armstrong, and J. Colgan, J. Phys. B [**46**]{}, 035201 (2013).
T. G. Lee, C. P. Ballance, S. A. Abdel-Naby, J. L. King, T. W. Gorczyca, and M. S. Pindzola, J. Phys. B [**48**]{}, 065201 (2015).
B. L. Henke, P. Lee, T. J. Tanaka, R. L. Shimabukuro, and B. K. Fujikawa, At. Dat. Nucl. Dat. Tables [**27**]{}, 1 (1982).
W. C. Stwalley, J. Chem. Phys. [**54**]{}, 4517 (1971).
N. Reshetnikov, L. J. Curtis, M. S. Brown, and R. E. Irving, Phys. Scr. [ **77**]{}, 015301 (2008).
F. Maeder and W. Kutzelnigg, Chem. Phys. [**42**]{}, 95 (1979).
M. Kutzner, V. Maycock, J. Thorarinson, E. Pannwitz, and J. A. Robertson, Phys. Rev. A [**66**]{}, 042715 (2002).
M. F. Haso[ğ]{}lu, S. A. Abdel-Naby, E. Gatuzz, J. Garc[í]{}a, T. R. Kallman, C. Mendoza, and T. W. Gorczyca, Astrophys. J., Suppl. Ser. [**214**]{}, 8 (2014).
M. Banna, A. Slaughter, R. Mathews, R. Key, and S. Ballina, Chem. Phys. Lett. [**92**]{}, 122 (1982).
G. Nasreen, S. T. Manson, and P. C. Deshmukh, Phys. Rev. A [**40**]{}, 6091 (1989).
V. G. Pal’chikov and V. D. Ovsiannikov, Quantum Electronics [**34**]{}, 412 (2004).
V. D. Ovsyannikov, V. G. Pal’chikov, H. Katori, and M. Takamoto, Quantum Electron. [**36**]{}, 3 (2006).
J. Mitroy and M. W. J. Bromley, Phys. Rev. A [**68**]{}, 052714 (2003).
S. G. Porsev and A. Derevianko, Phys. Rev. A [**65**]{}, 020701 (2002).
C. Froese Fischer, P. Jönsson, and M. Godefroid, Phys. Rev. A [**57**]{}, 1753 (1998).
V. P. A. Lonij, C. E. Klauss, W. F. Holmgren, and A. D. Cronin, Phys. Rev. Lett. [**105**]{}, 233202 (2010).
E.-A. Reinsch and W. Meyer, Phys. Rev. A [**14**]{}, 915 (1976).
C. Froese Fischer, G. Tachiev, and A. Irimia, At. Dat. Nucl. Dat. Tables [ **92**]{}, 607 (2006).
C. Cheng, X. Gao, B. Qing, X.-L. Zhang, and J.-M. Li, Chin. Phys. B [**20**]{}, 033103 (2011).
M. Stanek, L. Glowacki, and J. Migdalek, J. Phys. B [**29**]{}, 2985 (1996).
I. M. Savukov and W. R. Johnson, Phys. Rev. A [**65**]{}, 042503 (2002).
S. Mengali and R. Moccia, J. Phys. B [**29**]{}, 1613 (1996).
R. K. Nesbet and H. W. Jones, Phys. Rev. A [**16**]{}, 1161 (1977).
E. F. Archibong and A. J. Thakkar, Phys. Rev. A [**44**]{}, 5478 (1991).
X. Chu and A. Dalgarno, J. Chem. Phys. [**121**]{}, 4083 (2004).
V. E. Chernov, D. L. Dorofeev, I. Y. Kretinin, and B. A. Zon, Phys. Rev. A [ **71**]{}, 022505 (2005).
G. A. Victor and D. B. Slavsky, J. Chem. Phys. [**61**]{}, 3484 (1974).
S. Patil, Eur. Phys. J. D [**10**]{}, 341 (2000).
W. M[ü]{}ller, J. Flesch, and W. Meyer, J. Chem. Phys. [**80**]{}, 3297 (1984).
H. Partridge, C. W. Bauschlicher Jr., L. G. M. Pettersson, A. D. McLean, B. Liu, M. Yoshimine, and A. Komornicki, J. Chem. Phys. [**92**]{}, 5377 (1990).
R. Moccia and P. Spizzo, J. Phys. B [**21**]{}, 1145 (1988).
M. A. Castro and S. Canuto, Phys. Lett. A [**176**]{}, 105 (1993).
J. F. Stanton, Phys. Rev. A [**49**]{}, 1698 (1994).
P. Widmark, B. Joakim Persson, and B. Roos, Theo. Chim. ACTA [**79**]{}, 419 (1991).
K. Patkowski, R. Podeszwa, and K. Szalewicz, J. Phys. Chem. A [**111**]{}, 12822 (2007).
W. D. [Robb]{}, Chem. Phys. Lett. [**34**]{}, 479 (1975).
R. Santra, K. V. Christ, and C. H. Greene, Phys. Rev. A [**69**]{}, 042510 (2004).
E. Hult, H. Rydberg, B. I. Lundqvist, and D. C. Langreth, Phys. Rev. B [ **59**]{}, 4708 (1999).
[^1]: Mean of LG and VG line strength with experimental transition energy from [@KelPod08]
[^2]: Mean of LG 1.72 and VG 1.73
[^3]: Mean of LG 1.746 and VG 1.717
[^4]: Line strength with experimental transition energy from [@KelPod08]
[^5]: Mean of LG 1.773 and VG 1.701
[^6]: Mean of LG 1.757 and VG 1.736
[^7]: Mean of LG 1.76 and VG 1.75
[^8]: Cited as Weiss, private communication. Using the CI value for the line strength from Weiss (1967) and the measured transition energy yields 1.77.
[^9]: Weighted average of ten experimental values as of 2003.
[^10]: Mean of LG 74.7 and VG 74.03
[^11]: Does not include core valence according to Ref. [@MaeKut79]
[^12]: Does not include core contributions
[^13]: Listed in Table VII as “corrected,” corresponding to an empirical rescaling.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Pandemic and epidemic diseases such as CoVID-19, SARS-CoV2, and Ebola have spread to multiple countries and infected thousands of people. Such diseases spread mainly through person-to-person contacts. Health care authorities recommend contact tracing procedures to prevent the spread to a vast population. Although several mobile applications have been developed to trace contacts, they typically require collection of privacy-intrusive information such as GPS locations, and the logging of privacy-sensitive data on a third party server, or require additional infrastructure such as WiFi APs with known locations. In this paper, we introduce CONTAIN, a privacy-oriented mobile contact tracing application that does not rely on GPS or any other form of infrastructure-based location sensing, nor the continuous logging of any other personally identifiable information on a server. The goal of CONTAIN is to allow users to determine with complete privacy if they have been within a short distance, specifically, Bluetooth wireless range, of someone that is infected, and potentially also when. We identify and prove the privacy guarantees provided by our approach. Our simulation study utilizing an empirical trace dataset (Asturies) involving 100 mobile devices and around 60000 records shows that users can maximize their possibility of identifying if they were near an infected user by turning on the app during active times.'
author:
- Arvin Hekmati
- Gowri Ramachandran
- Bhaskar Krishnamachari
bibliography:
- 'sample-base.bib'
title: 'CONTAIN: Privacy-oriented Contact Tracing Protocols for Epidemics'
---
<ccs2012> <concept> <concept\_id>10002978.10003029.10011150</concept\_id> <concept\_desc>Security and privacy Privacy protections</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002951.10003227.10003245</concept\_id> <concept\_desc>Information systems Mobile information processing systems</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
Introduction
============
Viruses such as CoVID-19 and SARS-CoV2 have infected hundreds of people throughout the world, resulting in multiple fatalities. The World Health Organization has issued guidelines to limit virus outbreak by tracking people who may have come in contact with an infected individual [@world2014contact]. Such measures allow government and health care officials to drastically reduce the spread of the virus to a large community. The infected individual is required to share all his or her travel details with the health care authorities to reliably track and quarantine people who may contract the virus due to their physical proximity with the infected individual.
The contact tracking process typically involves gathering privacy sensitive information from the infected individual. Privacy-savvy individuals may not be willing to share all the information, which may hamper the contact tracing process while exposing the vast population to continued spread. A privacy-sensitive contact tracking approach would encourage the people to participate in the contact tracing process confidently, which is the focus of this work.
Several contact tracking applications involving mobile applications, wireless technologies, and GPS have been presented in the literature [@danquah2019use; @farrahi2014epidemic; @10.1145/2820783.2820880; @8422886; @prasad2017enact; @7763193; @DBLP:journals/corr/ReddyKRC15; @YONEKI201483]. Such approaches either expect the infected individual to self-report their contact information or rely on external infrastructures such as wireless access points and GPS. On the one hand, the approaches requiring self-reporting are not reliable since the individuals may not honestly report the information to the government authorities. On the other hand, the infrastructure-dependent methods work only when the mobile device is in the proximity of the infrastructure; note that the GPS reception is not reliable in indoor environments. EPIC [@8422886] and ENACT [@prasad2017enact] were developed to trace contacts using a mobile application in a privacy-preserving manner. While they avoid the use of GPS, these frameworks rely on a wireless access point. Mobile phones report the access point identifier and the timestamp to a server, which is then matched against the infected user’s information to detect contact proximity.
In this work, we present CONTAIN (an acronym coined from “CONtact TrAcINg"), which is a mobile application to enable privacy-sensitive contact tracing for epidemics. We show how to leverage the users’ mobile devices sending anonymous encrypted or random messages to each other via Bluetooth to allow users to determine with a 100% privacy if they have been in range of an infected user in the past, and when. There is no collection of privacy sensitive GPS, no reliance on external infrastructure for location tracking, no continuous logging of personally identifiable information from users. Specifically we introduce two privacy-friendly contact tracing protocols for CONTAIN; The first protocol uses symmetric key encryption, while the second protocol requires the secure generation of sufficiently large random numbers. In a way, they are duals of each other: the first protocol nodes upload the messages they have heard, while in the second protocol, the nodes upload the messages they have sent. Both protocols provide privacy for users that are not infected as such users never need to reveal their identity or contacts. Both protocols offer users with opt-in measures to get verified and declare if they are infected, while still maintaining some measure of privacy for the general public as they do not need to reveal their true identities or those of their contacts or the locations they have visited to everyone. And both protocols allow users to privately detect if they have had contacts with others that are infected.
Related Work {#sec:rw}
============
During the EBOLA outbreak in 2014, the World Health Organization (WHO) explained the importance of contact tracing and laid down protocols for tracing contacts [@world2014contact]. But, WHO did not employ any mobile application, and they have only provided guidelines for the health care workers to improve the efficiency of the contact tracing process. A number of mobile phone based contact tracing applications have been presented in the literature [@danquah2019use; @farrahi2014epidemic; @10.1145/2820783.2820880; @8422886; @prasad2017enact; @7763193; @DBLP:journals/corr/ReddyKRC15; @YONEKI201483].
Danquah *et al.* [@danquah2019use] presents an electronic system for contact tracing to address the shortcomings of paper-based contact tracing efforts in Port Loko District, Sierra Leone. An Ebola Contact Tracing application called ECT app was developed to manage the communication between contact tracing coordinators and contact tracers. Our proposal focuses on tracking individuals who may have come in contact with an infected patient. We could extend our end application to enable support for coordination and cooperation.
The importance and effectiveness of epidemic contact tracing via communication traces are discussed in [@farrahi2014epidemic]. A contact tracing data set was used to show when to initiate the tracing and explains how the choice between tracing of random people from the population and the infected people influences the effectiveness of the contract tracing process. But, Farrahi *et al.* [@farrahi2014epidemic] does not present any mechanisms to collect contract tracing data from the users in a privacy-preserving fashion.
Shahabi *et al.* [@10.1145/2820783.2820880] presents a framework called PLACE in a vision paper to trace the proximity of individuals based on location data. [@10.1145/2820783.2820880] discussed the importance of processing the information in a privacy-preserving manner, but it did not present a specific solution. Our proposal focuses on tracing contact using obfuscated Bluetooth beacons without relying on GPS as a primary source. Besides, GPS-based approaches do not provide reliable results in an indoor environment.
EPIC [@8422886] and ENACT [@prasad2017enact] are developed to trace contacts using a mobile application in a privacy-preserving manner. Both these frameworks rely on a wireless access point, wherein the mobile phones are expected to report the access point identifier and the timestamp to a server, which is then matched against the infected user’s information to detect contact proximity. Such an approach would only work on locations with wireless access points. Our proposal aims to trace contact with mobile phones alone.
Qathrady *et al.* [@7763193] presents an infection tracing framework for hospitals based on a centralized server. This scheme relies on static infrastructures deployed in the hospital environment, and it does not address privacy explicitly. Our proposal focuses on a location-agnostic mobile-based application.
Reddy *et al.* [@DBLP:journals/corr/ReddyKRC15] presents a dengue-monitoring application, which helps the population to report dengue to the health authorities along with their location information. Besides, it includes a symptoms checker to help the users. This application relies on the information provided by the users, and it does not record communication traces. Our proposed application aims to record physical contacts using Bluetooth technology.
Yoneki *et al.* [@YONEKI201483] presents EpiMap to create a map of infected people and their contacts using a mobile application called FluPhone, which collects Bluetooth proximity data and GPS location to trace connections in the case of infection. But, FluPhone does not present any privacy-preserving mechanisms. Our proposal aims to develop a mobile application that collects traces without violating the privacy of the users.
In summary, all the existing solutions either rely on wireless access points or lack support for tracing contacts in a privacy-preserving, opt-in fashion.
CONTAIN: Privacy-oriented Contact Tracing Protocols {#sec:contain}
===================================================
Design Goals
------------
The design goals of the CONTAIN framework are described in this section.
- The application should not rely on external infrastructure such as WiFi access points with known locations or the use of GPS for determining contact information as it may violate the location privacy of the user.
- The application should not disclose any personally identifiable information to other mobile devices.
- A user should be able to check if they have been near an infected user on their own, in a completely private manner, without being forced to reveal to anyone else their infection status.
- The anonymized information needed for other users to determine if they have been in contact with a user that is infected should be made available to the public through a trusted server only if the user can prove to the operator of the server (e.g. through a medical certificate) that he or she is infected.
![Components of the CONTAIN Application.[]{data-label="fig:comp"}](CONTAIN-Architecture.png)
System description
------------------
Figure \[fig:comp\] shows the components of the CONTAIN framework.
- **Mobile Device** Mobile devices (smartphones) have become prominent in the last decade. Almost every individual is carrying a highly capable mobile device (smartphone). The core processes of the CONTAIN application run on user’s smartphones. We describe below two different protocols that can be used for the CONTAIN application, one based on encrypted beacons (see Section \[sec:proto1\]) and one based on random beacons (see Section \[sec:proto2\]). The CONTAIN application does not rely on or require GPS (in a given implementation of it, it may be possible to additionally allow users to reveal GPS information to the application on an opt-in basis, but this is not an essential part or requirement of our design).
- **Bluetooth technology** Bluetooth is one of the widely used communication technology in contemporary smartphones. Today, this technology is used for short-range communication with hands-free earphones and external Bluetooth-based devices such as speakers and other audio systems. The communication range of Bluetooth is on the order of 10 meters or less [@8419192].
- **Verification server** The smartphone and the Bluetooth technology can be used to gather contact traces in a private manner, which is discussed in Section \[sec:proto1\] and \[sec:proto2\]. A third-party server is required to verify the infection status of a user (than claims to be infected) and then to make anonymous information from that infected user available for other users to use, in a private manner, to verify if they have been near the infected user.
The CONTAIN application requires Bluetooth technology in combination with computation resources and ability to access the verification server over the Internet (through any means such as Cellular or WiFi). Note that the contact tracing logic on the phone transmits and receives an anonymous encrypted or random Bluetooth beacon from other mobile devices running the CONTAIN application.
![Illustration of how the Privacy-Sensitive Contact Tracing App would work with our first protocol.[]{data-label="fig:proto1"}](CONTAIN-Protocol1.png)
Protocol 1: Encrypted Beacons {#sec:proto1}
-----------------------------
Figure \[fig:proto1\] shows the illustration of the first contact tracing protocol.
1. Each user periodically beacons, using Bluetooth, a message, $\mathcal{M}$, consisting of a unique name or ID, a time-stamp, and a random number (salt) that changes over time. Each message is encrypted using a symmetric key $enc(\mathcal{M})$. Note that the users do not share this key with others.
2. Other users that hear the encrypted message beacon log it locally.
3. Either periodically or in a batch, on an opt-in basis, each user can upload all the encrypted beacon messages they have heard to a common verification server.
4. If a user $i$ becomes infected, they inform the server, on an opt-in basis, with evidence that they are infected, such as a medical report or “infection certificate". The above step of uploading encrypted beacon messages could also be taken in conjunction with this step.
5. The verification server proceeds to make all the messages uploaded by user $i$ publicly available, but still in the encrypted form.
6. Each other user $j$ can privately check these now publicly available encrypted messages to see if they can decrypt any of them. Note that this verification process could be automated. If any user manages to decrypt the message, then he or she is at risk as he or she has been near an infected person. In this case, the user should then proceed to get tested themselves.
![Illustration of how the Privacy-Sensitive Contact Tracing App would work with our second protocol.[]{data-label="fig:proto2"}](CONTAIN-Protocol2.png)
Protocol 2: Random Beacons {#sec:proto2}
--------------------------
Figure \[fig:proto2\] shows the illustration of the second contact tracing protocol.
1. Each user beacons a sufficiently large random number and logs this random number locally. Here, a large number is chosen to minimize the chance of collisions with beacons generated by other users.
2. Others that hear the beacon log this number locally.
3. Either periodically or in a batch, on an opt-in basis, each user can upload all the random beacon messages they have transmitted to a common verification server.
4. If a user $i$ becomes infected, on an opt-in basis, they inform the verification server with evidence that they are infected, such as a medical report or “infection certificate”. The above step of uploading the random beacon messages could also be done in conjunction with this step.
5. The verification server then proceeds to make all the random numbers uploaded by this user publicly available.
6. Other users periodically check the server and see if any of the numbers they have logged locally match those that were revealed by an infected user; if so they should proceed to get tested.
The fourth step in both protocols requires a trusted third party that a) operates the “publicizing” server, and b) can verify through a medical certificate that the user whose uploaded anonymous data is being made public has truly tested positive for being infected. This party must be trusted and perform the verification correctly to minimize false alarms. It may be necessary for this role to be fulfilled by an organization (whether public or private) with relevant experience and reputation. For example, it could be a healthcare organization or government entity, or medical insurance provider.
Both protocols require the ability to beacon information over short range Bluetooth links. Today’s mobile phones, for security, privacy and energy efficiency reasons, generally require applications that do such beaconing to run in the foreground; regrettably, this could make the protocols less effective than they could be. But perhaps mobile device manufacturers could be prevailed upon to provide greater access to such an app in the public interest. Alternatively, at least at relatively more crowded venues such as airports, campuses, shopping areas or public transportation, users can be encouraged to download and turn on these apps.
Anonymizing an Infected User\[sec:anonInfected\]
------------------------------------------------
Note that the information revealed by the verification server does not explicitly contain anything that would reveal to others who the infected user is; it only allows other users to determine if they have been near an infected user and when. In some cases, if a user knows that during a certain time they were only in contact with (within bluetooth range) of a specific person they know, they may be able to infer the identity of the infected person from the time information provided. One way the infected user could potentially avoid even this minimal chance of disclosure would be to only release information from times when they know they were in the presence of many other devices (which could be determined based on the timestamps of logged messages), and of course they could also opt-out of notifying other parts.
While it may be assumed the infected user must make themselves known to the owner/operator of the verification server since they must provide the medical certificate that they are infected, there is an additional step that could be taken to anonymize them further. This would involve asking the infected user to use a public-key pair that does not tie them to a real-world identity, and have the medical authority digitally sign an infection certificate (using their own key pair) that only includes the infected user’s public key. The infected user can then provide this digital certificate to the verification server. From this point on the verification server can verify using the public key of the medical authority that the anonymous user contacting it (whom they know only through its public key) is indeed infected, and the verification server can thus ensure that only the (anonymous) data uploaded by a user that is truly infected is being made publicly available.
Anonymizing Encounter Time \[sec:anonET\]
-----------------------------------------
We would like to clarify one additional aspect that is \*not\* guaranteed to be kept private by either protocol. If an infected user $i$ has opted-in to notify the verification server and share the data it logged, then any user $j$ that has actually been near the user $i$ will know the time of encounter, which will a) de-anonymize the location of the encounter (since the user may remember where they were at that encounter time; or *a fortiori* in protocol 1 they could also inject their own GPS location explicitly into the message that is encrypted to keep track of where they were during each encounter), and b) de-anonymize the identify of the infected user $i$ to $j$ (if user $j$ knows and remembers or has some other way of identifying who they were near at the time of encounter).
Fortunately, there is a solution to this problem. If it is desired to further strengthen the system to provide for an additional layer of privacy to avoid the above shortcoming, one possible solution would be to use protocol 2 (random beacons) and eliminate the public notification step so that the verification server does \*not\* publish all the random numbers sent by the infected user. Instead, require the user interested in checking if they were near an infected user to upload all the random numbers they have logged along with an anonymizing “ID” to the server (using TOR or something similar to ensure an anonymous upload). Then the server will verify if there is any overlap between the random numbers from the infected user and random numbers uploaded the user that is checking. If so, it publishes only a notification that the user corresponding to that anonymous ID has been (or not been) in contact with an infected user.
Privacy Analysis {#sec:privacy}
================
In this section we formally identify and prove the privacy guarantees we can provide. The properties and their proofs assume that the trusted verification server behaves correctly in all cases.
We first list the key privacy properties that both our protocols guarantee below:
- **P1.** Beacons emitted by a user do not reveal any personally identifying information or location information about that user to other users.
- **P2.** Users that are not infected are not required to upload any information to the verification server.
- **P3.** Users will not be notified by the verification server about or receive any data from the verification server that can help them verify potential contact with users that are not infected
- **P4.** Users that are infected must opt-in in order for other users to determine if they have been near that infected user.
- **P5.** Users can check if they have been near an (opted-in) infected user without revealing any personally identifiable information about themselves.
Further, there are two additional privacy properties that can be shown to hold with the appropriate additional modifications as discussed in section \[sec:anonInfected\] (for both protocols) and section \[sec:anonET\] (for protocol 2 only):
- **P6.** Users that are infected can keep themselves anonymous even from the operator of the verification server.
- **P7.** A user that finds out it has been contact with an infected user will not be able to determine when or where exactly the contact happened.
**Proof of P1:** We assume that the underlying Bluetooth protocol itself is making user of ephemeral, time-varying ID’s and not sending static, identifiable MAC addresses (such schemes are already implemented by privacy-sensitive mobile device operating systems such as Apple’s iOS [@AStudyofMACAddressRandomizationinMobileDevicesandWhenitFails]). In protocol 1, the beacons are encrypted using a symmetric key that is not revealed to any other user. The use of a salt that can be changed each time further makes it so that there is no way to connect a particular individual or device with the logged data since it will be different each time the salt is changed. In protocol 2, the beacons consist of random numbers that can also change each time, so again, there is no identifying information being transmitted. Neither protocol requires collection of any location information. In protocol 1, the user could chose to encrypt their own GPS data if they so wish, but this information could only be checked by them on their own since it is encrypted with a key that only they possess.
**Proof of P2:** This property is a direct consequence of the two protocols. At no step is a device required to upload any information if they are not infected. Although step 3 in each protocol does allow each user to upload their random logged info periodically, they are not required to do so before they are infected, and even then it is on an opt-in basis.
**Proof of P3:** Since users that are not infected do not upload any information to the verification server in either protocol, there is no way for other users to be notified about or learn anything about them.
**Proof of P4:** For other users to determine if they have been near an infected user, the verification server needs to make available logs from that infected user. However, in both protocols, the logs are provided by infected users only on an opt-in basis.
**Proof of P5:** For both protocols, the verification server can make the anonymous information they get from the infected users available on a public website. Users can use an anonymous web browser such as TOR to download that information and check if they have been near any of the infected user in a completely private manner, without revealing any of their identifying information to the operator of the verification server or any other party.
**Proof of P6:** To allow infected users to keep themselves anonymous from the verification, the certificate from the medical authority should include only the public key of the infected user, as described in the additional step of section \[sec:anonInfected\]. If that additional measure is taken, then the infected users gain an additional level of anonymity and privacy.
**Proof of P7:** This is addressed by the solution presented in Section \[sec:anonET\]. By restricting to protocol 2 and asking users that want to check if they are infected to anonymously provide their logged data to the verification server, it can be ensured that the users that are checking will be notified that there was an encounter but not when or where the encounter happened.
Evaluation {#sec:eval}
==========
Most mobile operating systems today require apps that use Bluetooth communication to run in the foreground. Thus it requires active user intervention to be useful. In this section we evaluate through trace-based simulations how well CONTAIN would work as a function of how long the app is turned on by the user each day, and when it is turned on. We use a dataset named Asturies provided by CRAWDAD.org for our simulation [@oviedo-asturies-er-20160808]. The dataset contains the encounter of mobile devices with each other at different timestamps. We split the day time into 24 slots and mapped one hour of the dataset to each time slot. Users are considered to turn on the CONTAIN mobile app to participate in the experiment according to three different scenarios, which are described in Section \[Evaluation:Methodology\].
During the experiments, we varied both the number of initial infected users that are chosen randomly and also the probability that the disease gets transferred between people to see their effects on the results. Moreover, we considered several time slots that users turn on their Bluetooth-based CONTAIN app to evaluate the effectiveness of our protocols in different scenarios. Each experiment is repeated 1000 times to capture the randomness of the simulations.
Methodology {#Evaluation:Methodology}
-----------
We conducted two sets of experiments. First we designed an experiment to study how the infection rate grows among people according to the initial number of infected users and the contagiousness probability.
Second, we designed an experiment with three different scenarios in terms of when users schedule the time slots that they turn on their Bluetooth-based CONTAIN app.
- **Random:** In this scenario, users turn on their Bluetooth randomly in 12 contiguous hours that correspond to the most active time of the day.
- **Decentralized:** In this scenario, each user turns on his or her Bluetooth and runs the CONTAIN app when they are in a crowded place such as train stations, offices, shopping malls, etc. in a decentralized manner. For this scenario, we sort the time slots for each user according to the number of devices that the user connected to in that time slot. The second criterion for sorting is the crowdedness of the area that the user is. We measure the crowd level of an area by counting the number of records that we have in each time slot. Then, during an experiment where the user turns on for $k$ slots, it does so for the first $k$ slots in the sorted order.
- **Centralized:** In this scenario, we assume that there is some centralized coordination mechanism that ensures that all users turn on their Bluetooth and run the CONTAIN app at specific hours of the day. In this scenario, we sort the time slots according to first the average number of active mobile devices, and second the number of records that we have in each hour in the dataset. Thus the app is activated by everyone during the hours with the most active mobile devices.
Results
-------
Figure \[fig:infected\] shows the results from our first set of experiments, showing the number of users who get infected by infectious diseases such as CoVID-19 and SARS-CoV2 with 90% confidence interval versus contagiousness probability. These are also the number of initial users who are assumed to be infected at the beginning of the simulation. As you can see, as we have more number of initially infected users and also higher contagiousness probability, more people get infected.
![Number of Infected Users vs Contagiousness Probability and Initially Infected Users[]{data-label="fig:infected"}](Infected.png)
Figure \[fig:test\_req\] represents the number of users who have been in contact with or in the proximity of infected persons. In this experiment, we assumed the contagiousness probability is 2%, and the initial number of infected users is 2. From Figure \[fig:test\_req\], it is clear that as more people cooperate and turn on their Bluetooth for more number of time slots, more number of potentially infected users could be identified reliably.
![Number of Test Required Users vs Number of Active Bluetooth Hours[]{data-label="fig:test_req"}](Test_Req.png)
As we can see in Figure \[fig:infected\], when the number of infected people grows so fast, it is essential to identify and notify potentially infected people to get tested. Besides, this helps the individuals who may have come in contact with the infected individual to take corrective actions, including self-isolation, to prevent him or her from spreading the infection to others.
In this experiment, we have, on average, 100 devices active per week and around 60000 records for those devices per week, which indicates a dense environment with lots of interactions among people. That is why such a small contagiousness probability and also initially infected people results in a considerable number of infected users. Consequently, as we can see in Figure \[fig:test\_req\], our application informs many more people in case more number of them actively use the Bluetooth-based CONTAIN application. In this way, people can be alerted to get tested, take care of themselves and not spread the virus among other people by restricting their activities and avoiding crowded areas.
It is also interesting and important to note that when people act in a decentralized manner and turn on their Bluetooth-based app only when they go to a crowded environment, we will also have better performance as compared to the centralized and random scenarios. It is also interesting to see that decentralized scenario results in more number of people being detected as potentially infected and required to do the test as compared to the centralized scenario. This is because, if each person turns his or her Bluetooth in the places where lots of other people are available (this is evident from the dataset that we used in our study), much more number of encounters will be recorded as compared to the centralized rule, which does not fit everyone’s activity.
Conclusion {#sec:conc}
==========
In this paper, we have presented CONTAIN, which is a privacy-friendly and Bluetooth-based mobile application for contact tracing. Unlike the existing contact tracing application, the proposed contact tracing protocols have shown to be infrastructure independent. We have presented two contact tracing protocols, both of which do not reveal any personally identifiable information. Through a simulation study, we have shown how the contact tracing can effectively mitigate the spread of infectious diseases.
We plan to investigate how privacy-oriented approaches influence adoption in our future work. Besides, we will also develop methods to overcome system challenges that prevent Android and iOS users from running Bluetooth-based applications in the background. And, the existing contract tracing projects do not provide any open-source software for experimentation purposes, with the exception of FluPhone [@YONEKI201483]. We plan to develop and release open-source software implementations of CONTAIN for both Android and iOS devices, and perform a real world pilot study in the future.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A class of semi-linear hyperbolic systems in 1+1 dimensions was investigated several years ago by Ori and Gorbonos [@gorbonos]. This class, to which we shall refer as [*flux-conserving systems*]{}, exhibits a variety of interesting mathematical properties. Here we demonstrate how the formalism of flux-conserving systems can be applied to the problem of black-hole evaporation in 1+1-dimensions. More specifically, we show how the semiclassical CGHS [@CGHS] field equations may be approximated by a certain flux-conserving system.'
author:
- |
Amos Ori\
Department of Physics, Technion-Israel Institute of Technology,\
Haifa 32000, Israel
title: 'Flux-conserving hyperbolic systems and two-dimensional evaporating black holes'
---
Introduction
============
The investigation of gravitational collapse and black-hole (BH) dynamics often leads to systems of partial differential equations (PDEs) in 1+1 dimensions. This happens, for example, in the study of four-dimensional (4D) BHs when spherical symmetry is assumed. The usage of spherical symmetry for dimensional reduction in BH physics has been a traditional research strategy in General Relativity for many decades—since the discovery of the Schwarzschild and Reissner-Nordstrom solutions almost a century ago.
About two decades ago Callan, Giddings, Harvey and Strominger (CGHS) [@CGHS] introduced a two-dimensional toy model for investigating the semiclassical evaporation of black holes, with the hope of addressing the information puzzle. For appropriate initial conditions, this model describes the formation of a BH as well as its subsequent evaporation. Mathematically, this model again reduces to a system of PDEs in 1+1 dimensions.
Among the family of 4D spherically-symmetric solutions, an interesting subclass is the one in which the matter content is made of (at most) two radial fluxes of null fluids. The null fluxes are assumed to be pressure-less and non-interacting (except gravitationally), hence each flux is conserved separately. If only one such null fluid is present, the system admits an exact solution, known as the Vaidya solution [@vaidya]. Obviously, if both fluxes vanish the system reduces to the Schwarzschild solution. It is useful to extend this subclass by allowing the presence of a spherically-symmetric electric field, in addition to the null fluids. This gives rise to the mass-inflation model [@poisson], the charged Vaidya solution [@vaidya2] (in the case where one flux vanishes), and, obviously, to the Reissner-Nordstrom solution (in the special case where both fluxes vanish). All these solutions played a central role in studying the non-trivial internal structure of spherically-symmetric charged BHs [@graves], and particularly in investigating the singularity which develops at the inner horizon [@Hiscock; @poisson; @Ori1].
A few years ago Ori and Gorbonos [@gorbonos] explored a simple class of semi-linear hyperbolic systems in 1+1 dimensions, to which I will refer as [*flux-conserving systems*]{}. Each such system is characterized by a single function of one variable, denoted by $h(\bar{R})$ below. This class, which was first introduced in Ref. [@Ori-charged], exhibits several interesting mathematical properties. In particular, in any such system one can identify two conserved null fluxes. In the special case where one of these fluxes vanish, the solution takes a Vaidya-like form and the system then reduces to an ordinary differential equation (ODE), which may remarkably simplify the analysis. The flux-conserving systems contain, as special cases, the conventional Vaidya [@vaidya] and mass-inflation [@poisson] solutions (as well as the Schwarzschild and Reissner-Nordstrom solutions), which all correspond to the same function $h(\bar{R})$ [@gorbonos]. It also includes the [*classical*]{} CGHS solution, though the semiclassically-corrected CGHS system does not belong to this class.
The original motivation for introducing these semi-linear systems in Ref. [@gorbonos] was to provide a simpler (and at the same time wider) mathematical context for investigating the formation of null weak singularities at the inner horizon of BHs. This goal was undertaken by Gorbonos and Wolansky [@second], who demonstrated that for a certain (piecewise-linear) class of functions $h(\bar{R})$ a generic null weak singularity indeed forms.
The present paper will focus on a different aspect of these semi-linear systems: the very conservation of the individual fluxes, and its possible application to analyzing the evolution of dynamical BHs. As an example, consider a flux-conserving system which describes a BH bombarded by ingoing radiation (with no outflux). In such a system the influx at past null infinity directly determines the rate of change of the BH mass. In the traditional framework of conventional 4D spherically-symmetric BHs this is illustrated by the ingoing Vaidya solution, in which the information is fully encoded in the mass-function $M(v)$. The same applies to (Vaidya-like solutions for) evolving BHs in [*any*]{} type of flux-conserving system.
More specifically, our main goal in this paper is to demonstrate how the mathematical concept of flux-conserving systems may be used to analyze the evaporation of two-dimensional semiclassical BHs.
As was already mentioned above, the classical CGHS system is a flux-conserving one [@gorbonos], but this situation changes when the semiclassical effects are included. At first glance, an evaporating BH (say, in the Hartle-Hawking state) may actually be regarded as an archetype of a system which violates flux-conservation: No influx arrives from past null infinity, and yet negative influx penetrates into the BH and causes its evaporation. We shall show here, however, that by a re-definition of the field variables, the CGHS field equations may be transformed into a form which is approximately flux-conserving. More specifically, as long as the evaporating BH is still macroscopic, one may expand the semiclassical field equations in a small quantity (proportional to the number $N$ of quantum fields and inversely proportional to the variable $R$ defined below). The zero order in this expansion yields the classical system, whereas the next order represents the leading semiclassical effects. We shall show that up to (and including) this first order, in our new variables the transformed field equations constitute a flux-conserving system. This provides a powerful analytic tool for analyzing the structure of the semiclassical spacetime of a CGHS evaporating BH.
We start with a very brief review of the formalism of flux-conserving systems (Sec. II) and the CGHS model (Sec. III)—introducing our own notation for the latter, and focusing on the properties which are relevant to the present analysis. Then in Sec IV we perform the field redefinitions, which eliminate the first-order derivative terms from the evolution equations. Finally in Sec. V we explore the large-$R$ approximate system (which incorporates the leading-order semiclassical effects), and demonstrate that it constitutes a flux-conserving system. We then describe how the CGHS evaporating-BH spacetime may be obtained (approximately) from an ingoing Vaidya-like solution of this system.
Flux conserving systems
=======================
By [*flux-conserving system*]{} we refer to a semi-linear hyperbolic system in $1+1$ dimensions, for two unknowns $\bar{R}(u,v)$ and $\bar{S}(u,v)$, which takes the form $$\bar{R}_{,uv}=e^{\bar{S}}F(\bar{R})\ ,\ \ \ \bar{S}_{,uv}=e^{\bar{S}}F'(\bar{R}),
\label{evf}$$ for any given function $F(\bar{R})$, where a prime denotes $d/d\bar{R}$.
This class of hyperbolic systems displays several unique mathematical properties, which were discussed in some detail in Ref. [@gorbonos]. [^1] Here we shall briefly mention some of these properties.
First, the system admits a gauge symmetry: If one transforms the coordinates as $u\rightarrow\tilde{u}(u),v\rightarrow\tilde{v}(v)$, then $\bar{R}$ behaves as a scalar (it is unchanged), but $\bar{S}$ transforms as $$\tilde{S}=\bar{S}-\ln\left(\frac{d\tilde{v}}{dv}\right)-\ln\left(\frac{d\tilde{u}}{du}\right)
\label{coo}$$ (that is, $e^{\bar{S}}$ transforms like a metric component $g_{uv}.$)
Second, we define the two flux functions $$\bar{T}_{uu}\equiv\bar{R}_{,u}\bar{S}_{,u}-\bar{R}_{,uu}\:,\;\;\bar{T}_{vv}\equiv\bar{R}_{,v}\bar{S}_{,v}-\bar{R}_{,vv}\label{eq:fluxes}$$ where a comma denotes partial dervative. It then follows from the hyperbolic system (\[evf\]) that $$\bar{T}_{uu,v}=\bar{T}_{vv,u}=0.$$ Namely, $\bar{T}_{uu}$ only depends on $u$, and $\bar{T}_{vv}$ anly depends of $v$. We refer to this property as [*conservation of fluxes*]{}.
Naturally the solutions of this system may be divided into three categories: (i) solutions with $\bar{T}_{uu}=\bar{T}_{vv}=0$, (ii) solutions in which only one flux vanishes; and (iii) solutions in which both fluxes are non-vanishing. We refer to case (i) as [*vacuum-like*]{}, and to case (ii) as [*Vidya-like*]{}. Case (iii) is the mathematical generalization of the 4D spherically-symmetric systems with two null fluids, for example the one studied in Ref. [@poisson].
Next we define the "generating function” $h(\bar{R})$ by $$F(\bar{R})=-h'(\bar{R}).
\label{hFromF}$$ More precisely this is a one-parameter family of functions, due to the arbitrary integration constant. It will often be convenient to express this free parameter explicitly in the form $$h(\bar{R})=h_{0}(\bar{R})-\bar{m},
\label{hint}$$ where $h_{0}(\bar{R})$ denotes a [*specific*]{} integral of (\[hFromF\]), and $\bar{m}$ is an arbitrary constant.
In the vacuum-like case (i) the space of solutions reduces to a one-parameter family of solutions (apart from gauge transformations), characterized by their “mass parameter” $\bar{m}$. It is convenient to express the vacuum-like solutions in an "Eddington-like” gauge, in which both $\bar{R}$ and $\bar{S}$ depend on $x\equiv v-u$ solely. For given $\bar{m}$, $\bar{R}(x)$ is then determined from the ODE $$\frac{d\bar{R}}{dx}=h_{0}(\bar{R})-\bar{m},$$ and then $\bar{S}(x)$ by $e^{\bar{S}}=h_{0}(\bar{R})-\bar{m}$. The Eddington-like gauge thus exhibits the static nature of the vacuum-like solutions. [^2]
Assume now that $h(\bar{R})$ is monotonically increasing in some range of $\bar{R}$. [^3] Then, at least for some range of the parameter $\bar{m}$, $h$ will vanish at a certain $\bar{R}$-value denoted $\bar{R}_{0}$, and will become negative at $\bar{R}<\bar{R}_{0}$. As was discussed in Ref. [@gorbonos], the surface $\bar{R}=\bar{R}_{0}$ functions as an event horizon, and the vacuum-like solution then corresponds to a BH. \[If $h(\bar{R})$ has an additional root at $\bar{R}<\bar{R}_{0}$, it will correspond to an inner horizon.\]
The Vaidya-like solution {#vaidya}
------------------------
Assume now that only one flux vanishes. For concreteness we shall consider here the *ingoing* Vaidya-like solution, in which $\bar{T}_{vv}$ is nonvanishing. Then, as was shown in Ref. [@gorbonos], each solution is characterized by a mass-function $\bar{M}(v)$. The unknown $\bar{R}(u,v)$ is determined by the ODE $$\bar{R}_{,v}=h_{0}(\bar{R})-\bar{M}(v).
\label{eq:Rv}$$ In turn the other unknown $\bar{S}(u,v)$ is given by $$e^{\bar{S}}=-\bar{R}_{,u}.
\label{eq:Sv}$$ The nonvanishing influx satisfies $\bar{T}_{vv}=\bar{M}_{,v}$. Note that the solution expressed in this from assumes a specific $v$-gauge (to which we may refer as [*Eddington-like*]{}), but the gauge freedom in $u$ is preserved.
We shall now derive an expression for $\bar{S}_{,v}$ which was not included in Ref. [@gorbonos]. Differentiating $\bar{S}$ from Eq. (\[eq:Sv\]) yields $\bar{S}_{,v}=\bar{R}_{,uv}/\bar{R}_{,u}$. Substituting now Eq. (\[evf\]) for $\bar{R}_{,uv}$ and Eq. (\[eq:Sv\]) again for $\bar{R}_{,u}$, we obtain the simple relation $$\bar{S}_{,v}=-F(\bar{R}).\label{eq:SvNew}$$ This relation may be useful for constructing approximate solutions for evaporating $1+1$-dimensional black holes.
The *outgoing* Vaidya-like solution—the analogous situation in which the nonvanishing flux is $\bar{T}_{uu}$ rather than $\bar{T}_{vv}$—proceeds in a similar manner, though with certain sign changes. One obtains in this case the expressions $\bar{R}_{,u}=-[h_{0}(\bar{R})-\bar{M}(u)]$, $e^{\bar{S}}=\bar{R}_{,v}$, $\bar{T}_{uu}=-\bar{M}_{,u}$, and $\bar{S}_{,u}=F(\bar{R})$. [^4]
The CGHS model
==============
The CGHS model [@CGHS] consists of gravity in 1+1-dimensions coupled to a dilaton field $\phi$ and to a large number $N$ of identical massless scalar fields $f_{i}$. CGHS provided an action which includes an effective semiclassical term derived from the trace anomaly. With appropriate initial conditions, this model describes the formation of a BH by a collapsing shell, as well as its subsequent evaporation. The field equations are most conveniently expressed in double-null coordinates $u,v$, in which the metric takes the form $ds^{2}=-e^{2\rho}dudv$. The system thus consists of two unknowns, $\phi(u,v)$ and $\rho(u,v)$. The field equations for $\phi$ and $\rho$ are derived in Ref. [@CGHS]. We find it convenient to transform to new variables $R\equiv e^{-2\phi}$ and $S\equiv2(\rho-\phi)$ (as in Ref. [@DO]). The evolution equations then take the form $$R,_{uv}=-e^{S}-K\rho,_{uv}\ ,\ \ \ S,_{uv}=K\rho,_{uv}/R,\label{evolution_eq_mixed}$$ where $K\equiv N/12$. [^5] The Energy-momentum fluxes $T_{,uu}$ and $T_{,vv}$ are given by $$T_{,uu}=R_{,u}S_{,u}-R_{,uu}\:,\;\; T_{vv}\equiv R_{,v}S_{,v}-R_{,vv}.
\label{Tww}$$
The system (\[evolution\_eq\_mixed\]) may look quite simple at first glance, but one should recall that $\rho=(S-\ln\tilde{R})/2$ must be substituted in order to close the system, which messes the right-hand side. Bringing the equations to their standard form, in which $R,_{uv}$ and $S,_{uv}$ are explicitly given in terms of lower-order derivatives, we end up with the system $$R,_{uv}=-e^{S}\frac{\left(2R-K\right)}{2\left(R-K\right)}-R,_{u}R,_{v}\frac{K}{2R\left(R-K\right)},
\label{StandardR}$$ $$S,_{uv}=e^{S}\frac{K}{2R\left(R-K\right)}+R,_{u}R,_{v}\frac{K}{2R^{2}\left(R-K\right)}.\label{StandardS}$$
The classical model is obtained by setting $K=0$, after which the evolution equations reduce to $R,_{uv}=-e^{S},S,_{uv}=0$, which are easily solved. In an Eddington-like gauge the general solution takes the form $$R=e^{(v-u)}+f_u(u)+f_v(v) \, \, , \, \, \, S=v-u,
\label{generalSolution}$$ where $f_u(u),f_v(v)$ are arbitrary functions.
Note that the classical model is a flux-conserving system, with $\bar{R}=R,\bar{S}=S$, and $h_0=\bar{R}$. As such it admits the trivial (vacuum-like) static BH solution $R=e^{(v-u)}+\bar m, S=v-u$, as well as Vaidya-like solutions. This straightforward flux-conserving form breaks down once the semiclassical terms are included.
Approximate flux-conserving formulation of CGHS dynamics
========================================================
Field redefinition
------------------
The presence of terms quadratic in first-order derivatives makes the field equations (\[StandardR\],\[StandardS\]) harder to analyze. We wish to transform the field variables so as to get rid of these terms. A straightforward algebra leads to the following transformed variables: [@bilal] $$W(R)\equiv\sqrt{R(R-K)}-K\ln(\sqrt{R}+\sqrt{R-K})+K\left(\frac{1}{2}+\ln2\right)\label{eq:Wdef}$$ and $$Z\equiv S+\Delta Z(R), \label{eq:Zdef}$$ where $$\Delta Z(R)\equiv\frac{2}{K}(R-W)-\ln R. \label{eq:dZdef}$$ We shall only be concerned here about the domain $R>K$ (this includes the entire domain-of-dependence portion of the spacetime, both outside and inside the BH, because the singularity is located at $R=K$ [@DO]). Since $dW/dR=\sqrt{1-K/R}$, the function $W(R)$ is monotonically increasing throughout this domain. Therefore, the inverse function $R(W)$ is mathematically well defined, though it cannot be expressed explicitly.
With these new variables the evolution equations take the schematically-simpler potential form: $$W_{,uv}=e^{Z}\,V_{W}(W)\;,\quad Z_{,uv}=e^{Z}\,V_{Z}(W),
\label{potentialForm}$$ with certain “potentials” $V_{W}(W)$ and $V_{Z}(W)$. But the simplification does not come without a cost: These potentials are explicitly obtained as functions of $R$ rather than $W$. One finds $$V_{W}=-\frac{R-K/2}{\sqrt{R(R-K)}}\, e^{-\Delta Z(R)}\label{eq:VWdef}$$ and $$V_{Z}=\frac{2}{K}\left[\frac{R-K/2}{\sqrt{R(R-K)}}-1\right]\, e^{-\Delta Z(R)}.\label{eq:VZdef}$$ Nevertheless, the potential form (\[potentialForm\]) is useful for various analytical considerations, despite its implicit character. Furthermore, we shall primarily be concerned here about the leading-order semiclassical effects, and as it turns out, these leading-order effects are fully encoded in the large-$R$ (or, equivalently, large-$W$) asymptotic form of the field equations. At the desired order the functions $V_{W}(W)$ and $V_{Z}(W)$ \[as well as $R(W)$\] all take simple explicit forms, as we shortly demonstrate.
Large-$R$ asymptotic behavior
-----------------------------
The semiclassical effects, responsible for the BH evaporation, are encoded in the $K$ terms in the field equations (\[StandardR\],\[StandardS\]) or (\[eq:Wdef\]-\[eq:VZdef\]). (Notice that $W$ and $Z$ respectively approach $R$ and $S$ as $K\to 0$, and the field equations then reduce to their classical form, namely $V_W=-1,V_Z=0$.) To depict the leading semiclassical effect we may formally decompose all relevant functions in powers of $K$. Doing so, one observes that this effectively becomes an expansion in $K/R$ (see below). Although $K$ itself may be a large number, as long as the BH is macroscopic (namely, its mass is $\gg K$), $K/R\ll 1$ is satisfied throughout the BH exterior (and also in the interior, except in the neighborhood of the $R=K$ singularity—a region which will not concern us here). Thus, to depict the leading-order semiclassical effect we decompose all relevant quantities to first order in $K/R$ (or, equivalently, $K/W$). For the transformation of variable we obtain $$W=R\left[1-\frac{K}{2R}\ln R+O\left(\frac{K}{R}\right)^{2}\right]$$ and $$\Delta Z=-\frac{K}{4R}+O\left(\frac{K}{R}\right)^{2}.$$ The inverse function $R(W)$ is given at this order by $$R=W\left[1+\frac{K}{2W}\ln W+O\left(\frac{K}{W}\right)^{2}.\right]$$ The potentials $V_{W}(W),V_{Z}(W)$ can now easily be expanded to first order in $K/W$, $$V_{W}=-1-\frac{K}{4W}+O\left(\frac{K}{W}\right)^{2}\, \, ,
\, \, \, \, \,
V_{Z}=\frac{1}{W}\left[\frac{K}{4W}+O\left(\frac{K}{W}\right)^{2}\right].\label{eq:VZaprx}$$ Recall that this leading-order semiclassical approximation is applicable as long as $K/R\approx K/W\ll 1$, which for a macroscopic BH holds everywhere in the exterior (and also throughout most of the BH interior).
The large-$R$ flux-conserving system
------------------------------------
We now proceed with the above large-$R$ approximation, and analyze the PDE system $$W_{,uv}=e^{Z}V_{W}(W)\;,\quad Z_{,uv}=e^{Z}V_{Z}(W)\label{eq:system}$$ with $$V_{W}=-1-\frac{K}{4W}\; ,\quad V_{Z}=\frac{K}{4W^{2}}\; ,
\label{eq:VWZ}$$ omitting all higher-order terms $\propto (K/W)^2$.
Since $V_{Z}=dV_{W}/dW$, this is a flux-conserving system: It is described by Eq. (\[evf\]), with $\bar{R}=W$, $\bar{S}=Z$, and $F=V_{W}$. Its generating function is $$h_{0}(W)=W+\frac{K}{4}\ln W.
\label{eq:h(W)}$$
In particular the system (\[eq:system\],\[eq:VWZ\]) admits Vaidya-like solutions. As it turns out [@OriPrep], at the leading order an evaporating BH may be approximated by an ingoing Vaidya-like solution with a linearly-decreasing mass function. [^6] Assuming that the BH was created by the collapse of a shell of mass $M_{0}$, which propagated along the ingoing null orbit $v=v_{0}$, the mass function takes the form $\bar{M}(v)=M_{0}-(K/4)(v-v_{0})$. By shifting the origin of $v$ such that $v_{0}=-(4/K)M_{0}$, we obtain $$\bar{M}(v)=-\frac{K}{4}v\equiv m_{v}(v).\label{eq:mv}$$ Note that $dm_{v}/dv=-K/4=-N/48$, which reflects the standard evaporation rate of a two-dimensional BH [@CGHS].
The determination of the evaporating-BH spacetime thus reduces within this approximation to analyzing the ODE (\[eq:Rv\]) which now reads $$W_{,v}=h_{0}(W)-\bar{M}(v)=(W+\frac{K}{4}\ln W)+\frac{K}{4}v.\label{eq:Wv}$$ The other unknown $Z$ is then obtained from Eq. (\[eq:SvNew\]), which reads $$Z_{,v}=1+\frac{K}{4W}.
\label{eq:Zv}$$ \[Alternatively $Z$ may be obtained from Eq. (\[eq:Sv\]).\] The dependence upon $u$ enters through the initial conditions at $v=v_{0}$ (the collapsing shell). The appropriate initial conditions at the shell will be discussed in Ref. [@OriPrep].
Let us denote the solution to the two ODEs (\[eq:Wv\],\[eq:Zv\]) by $W_{c}(u,v)$ and $Z_{c}(u,v)$. Although it constitutes an exact solution to the system (\[eq:system\],\[eq:VWZ\]), as it stands it does not yield an adequately precise description of the CGHS BH spacetime. Indeed an approximation was involved in the first place in deriving the flux-conserving system (\[eq:system\],\[eq:VWZ\]) from the original CGHS system. This approximation is fine, because the terms omitted were of order $(K/W)^{2}$, whereas our attempt here is to construct an approximate solution at order $K/W$. As it turns out [@OriPrep], however, the solution $(W_{c},Z_{c})$ fails to satisfy the proper initial conditions at the shell ($v=v_{0}$), and the discrepancy is at order $K/W$, which cannot be neglected here. Stated in other words, the flux-conserving solution which properly describes the evaporating BH is not a pure ingoing Vaidya-like solution, but one which also contains a weak outflux—namely, a small correction $\delta W(u) \propto K$ to $W=W_c$. Nevertheless, owing to the smallness of $\delta W$ (and to the trivial form of the field equations at the limit $K\to 0$), one can simply superpose the correction $\delta W(u)$ on top of the Vaidya-like solution $W_{c}$ (with $Z_{c}$ unchanged) [@OriPrep]. This procedure yields a sufficiently accurate solution which satisfies the appropriate initial conditions. [^7]
In a forthcoming paper [@OriPrep] we shall implement this formalism, and construct an approximate solution $R(u,v),S(u,v)$ to the CGHS field equations which satisfies the correct initial conditions (at the shell as well as at past null infinity), and is accurate at first order in $K/W$.
[33]{}
A. Ori and D. Gorbonos, [*J. Math. Phys.*]{} **48**, 092502 (2007).
C. G. Callan, S. B. Giddings, J. A. Harvey and A. Strominger, Phys. Rev. D 45, R1005 (1992).
P. C. Vaidya, Proc. Indian Acad. Sci. A33, 264 (1951).
E. Poisson and W. Israel, [*Phys. Rev.*]{} D **41**, 1796 (1990).
W. B. Bonnor and P. C. Vaidya, [*Gen. Relativ. Gravit.*]{} **1**, 127 (1970). Here we only refer to the subclass in which the null fluids themselves are uncharged.
J. C. Graves and D. R. Brill, [*Phys. Rev.*]{} **120**, 1507 (1960).
W. A. Hiscock, [*Phys. Lett.*]{} **A83**, 110 (1981).
A. Ori, [*Phy. Rev. Lett.*]{} **67**, 789 (1991).
A. Ori, Phys. Rev. D 63, 104016 (2001).
D. Gorbonos and G. Wolansky, [*J. Math. Phys.*]{} **48**, 092503 (2007).
D. Levanony and A. Ori, Phys. Rev. **D80**, 084008 (2009).
This is closely related to the field redefinition introduced by A. Bilal and C. Callan, [*Nucl. Phys.*]{} **B394**, 73 (1993). A. Ori, “Approximate solution to the CGHS field equations for two-dimensional evaporating black holes” (in preparation).
[^1]: The field variables are denoted $R,S$ therein, instead of $\bar{R},\bar{S}.$ Here we use the over-bar variables to distinguish them from the specific variables of the CGHS model presented below. Similarly, we use $\bar{T}$ below to distinguish it from the physical stress-energy tensor $T$, and the same for the mass parameter $\bar{m}$ and the mass function $\bar{M}(v)$.
[^2]: In cases where the vacuum-like solution corresponds to a “black hole” (see below), this staticity characterizes the external region. Inside the horizon, the vacuum-like solution is homogeneous rather than static. The functions $\bar{R},\bar{S}$ then depend on $t\equiv v+u$.
[^3]: This may always be arranged, as long as $h(\bar{R})$ is not constant, by transforming $\bar{R}\rightarrow-\bar{R}$ if necessary (while fixing $h$).
[^4]: The above relations for both the ingoing and outgoing Vaidya-like solutions hold outside the BH. In the internal region certain sign changes apply, like in the vacuum-like case considered above.
[^5]: We have set here the cosmological constant $\lambda=1$ by a trivial shift in $\rho$, as explained in Ref. [@DO]. Also we set the null solution $f_{i}=0$ for the scalar fields, because we are interested here in the domain of evaporation where no incoming matter waves are present. The system also contains two constraint equations, but they are not needed here because any solution of the evolution equations, with appropriate initial conditions, automatically satisfies the constraint equations.
[^6]: However, a weak outflux term $\propto (K/W)$ has to be added in order to match the initial conditions, as discussed below.
[^7]: In fact it is the superposed outgoing component $\delta W(u)$ which gives rise to the Hawking outflux at future null infinity.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Decades-long repeat observations of supernova SN1987A offer us unique, real-time insights into the violent death of a massive star and its long-term environmental effects, until its eventual switch-off.'
author:
- 'Richard de Grijs$^{1,2,3}$'
title: 'Turning off the lights: Supernova SN1987A 30 years on'
---
Kavli Institute for Astronomy and Astrophysics, Peking University, Yi He Yuan Lu 5, Hai Dian District, Beijing 100871, China; e-mail: [email protected]\
Department of Astronomy, Peking University, Yi He Yuan Lu 5, Hai Dian District, Beijing 100871, China\
Discipline Scientist, International Space Science Institute–Beijing\
On 23 February 1987, supernova SN1987A—the explosive last gasps of a dying massive star—suddenly illuminated the Large Magellanic Cloud. Such bright stellar explosions occur only once or twice per century in large spiral galaxies like our Milky Way and much less frequently in Magellanic-type ‘dwarf’ galaxies. At a distance of 163,000 light-years[@dg14], SN1987A was the closest recorded supernova since Kepler’s Supernova in 1604 and Cassiopeia A in the late-17$^{\rm th}$ Century, both of which occurred in the Milky Way. Fortunately, we now have access to cutting-edge astronomical facilities, thus allowing us to monitor the explosion and subsequent evolution of the entire SN1987A system in real time.
In a recent article in [*The Astrophysical Journal Letters*]{}, Fransson [*et al.*]{}[@fr15] present a 20-year timeline of changes in the appearance of the supernova’s complex system of rings. Two of these rings are most likely caused by mass outflows from its red supergiant progenitor star up to 20,000 years before the explosion[@cr91]. The brighter but smaller inner ring, on the other hand, was generated by interactions of the progenitor’s swept-up stellar wind with the ambient gas. Hubble Space Telescope observations have enabled us to witness the appearance and subsequent disappearance of ‘hot spots’ along this latter circumstellar ring. These are likely caused by interactions of the densest gas clumps in the circumstellar gas with the supernova’s outward-propagating blast wave, thus leading to the appearance of a ring-like shape. The expanding supernova debris left behind after the blast wave passed by is decelerated by a ‘reverse shock’, which is due to electrons cascading down to lower energy levels following collisional excitation of neutral hydrogen atoms, triggered when the debris crosses the shock front.
Until approximately 2009, the observational timeline of Fransson [ *et al.*]{}[@fr15] shows an exponentially increasing contribution to the hot spots from shocked emission and outward acceleration of these clumps of up to 700–1000 km s$^{-1}$. New hot spots and faint, diffuse emission have since appeared outside of the now-fading inner ring. These new hot spots could have been triggered by the expanding blast wave if the density of the interstellar clumps was sufficiently high for the pressure from the reverse shock to cause cooling by radiating away heat. This, in turn, would cause the clumps to collapse into an even smaller volume and emit optically visible radiation[@pun02]. Ionization of the original stellar-wind ejecta caused by X-ray emission originating from lower-density material behind the expanding shock wave may be responsible for the diffuse emission.
The authors[@fr15] interpret the decrease in optical emission in the inner ring since 2013 as evidence of its dissolution. They deduce that the most likely reason for this decrease is that the area affected by shocks generating bright optical radiation is getting smaller. Alternatively, the expected steeply outward-increasing radial density distribution of the stellar-wind material[@dew12; @fr13] implies a corresponding increasing density threshold for clumps to sustain such radiative shocks. Either of these scenarios will lead to instabilities in the hot spots, conduction by the ambient hot gas and, consequently, rapid dissolution of the densest clumps[@pun02]. Fransson [*et al.*]{}[@fr15] predict that the inner ring will likely be fully destroyed within the next 10 years.
{width="\columnwidth"}
\[F1\]
SN1987A thus provides unprecedented insights into the physical properties of its massive progenitor from a perspective that is unavailable elsewhere. The system’s impact goes well beyond this immediate application. We can, in principle, also use the maximum observed angular size of the ring, combined with the speed of light and accurate delay-time measurements, to independently obtain a reliable geometric distance to the supernova. However, since the ring is resolved, we need to consider whether the emission used to measure delay times at a variety of wavelengths actually originates from the same region(s) in the ring. For instance, the ring size measured using optical emission from doubly ionized oxygen atoms[@pl95] is significantly different from that traced by ultraviolet lines[@pa91]. It has been suggested that the latter originate from the ring’s inner edge, whereas the optical lines come from its main body. In this case, and using the proper geometry, including a finite ring thickness, the ultraviolet light curve could result in an underestimate of the light-travel time across the optical ring diameter of up to 7% and, thus, a similar underestimate of the distance[@gu98].
We also need to consider potential errors caused by a misinterpretation of the underlying physics. Most importantly, it is often assumed that the fluorescent, scattered emission from the interstellar gas commences as soon as the supernova’s energetic photons hit a gas cloud. It is possible, though, that there is a slight delay in the onset of fluorescent emission as the gas first recombines from highly ionized states, for instance. Neglecting this step will lead to ring-size and, hence, distance overestimates.
SN1987A will likely become more X-ray and less hot-spot dominated as the system continues to evolve[@fr15]. The high-resolution Hubble Space Telescope observations of Fransson [*et al.*]{}[@fr15] and their spectroscopic analysis will allow us to settle the system’s full evolution unequivocally and provide the tightest geometric constraints yet on the use of resolved supernovae as distance tracers. As the important first rung of the extragalactic distance ladder, the importance of reducing the systematic uncertainties in the distance to the Large Magellanic Cloud cannot be overstated[@dg14].
[61]{}
de Grijs, R., Wicker, J. E. & Bono, G. *Astron. J.* **147**, 122 (2014).
Fransson, C. *et al.* *Astrophys. J. Lett.* **806**, L19 (2015).
Crotts, A. P. & Heathcote S. R. *Nature* **350**, 683–685 (1991).
Pun, C. S. J. *et al.* *Astrophys. J.* **572**, 906–931 (2002).
Dewey, D., Dwarkadas, V. V., Haberl, F., Sturm, R. & Canizares, C. *Astrophys. J.* **752**, 103 (2012).
Fransson, C., *et al.* *Astrophys. J.* **768**, 88 (2013).
Plait, P. C., Lundqvist, P., Chevalier, R. A. & Kirshner, R. P. *Astrophys. J.* **439**, 730–751 (1995).
Panagia, N., Gilmozzi, R., Macchetto, F., Adorf, H.-M. & Kirshner, R. P. *Astrophys. J. Lett.* **380**, L23–L26 (1991).
Gould, A. & Uza, O. *Astrophys. J.* **494**, 118–124 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce a Unified Disentanglement Network (UFDN) trained on The Cancer Genome Atlas (TCGA). We demonstrate that the UFDN learns a biologically relevant latent space of gene expression data by applying our network to two classification tasks of cancer status and cancer type. Our UFDN specific algorithms perform comparably to random forest methods. The UFDN allows for continuous, partial interpolation between distinct cancer types. Furthermore, we perform an analysis of differentially expressed genes between skin cutaneous melanoma (SKCM) samples and the same samples interpolated into glioblastoma (GBM). We demonstrate that our interpolations learn relevant metagenes that recapitulate known glioblastoma mechanisms and suggest possible starting points for investigations into the metastasis of SKCM into GBM.'
author:
- |
Benjamin Kompa\
`[email protected]` Beau Coker\
`[email protected]`
bibliography:
- '6867bib.bib'
nocite: '\nocite{}'
title: Learning a Generative Model of Cancer Metastasis
---
Introduction
============
Deep learning is being applied to many difficult problems in genomics and medicine. Alipanahi et al. used deep learning to learn site specific binding patterns of DNA and RNA-binding proteins [@Alipanahi2015-as]. Zhou et al. were able to predict non-coding variants using deep learning [@Zhou2015-yn]. Google has even produced an improved variant caller known as DeepVariant [@Anderson2018-fa].
More specifically, deep learning has been applied to understanding cancer prognosis. Chaudhary et al. were able to robustly predict survival in liver cancer [@Chaudhary2018-pf]. Cruz-Roa et al. leveraged deep learning to quantify the extent of breast cancer tumors in imaging data [@Cruz-Roa2017-oi]. Other groups have trained networks to identify metastatic breast cancer and lymph node metastasis [@Wang2016-eg; @Bejnordi2017-zk].
Nevertheless, there is little work in machine learning being done on what changes are occurring at a gene expression level in cancer samples. Understanding the genomic basis of cancer will yield better treatments and prognosis for patients [@Al-Lazikani2012-rn]. There are significant questions remaining in oncology about the relationships between different cancer types. For instance, while there is an association between melanoma, a type of skin cancer, and glioblastoma, a type of brain cancer, little is known about the molecular underpinnings of this relationship [@Desai2008-cj; @Scarbrough2014-wj].
Recently, deep generative models such as variational auto encoders (VAEs) and generative adversarial networks (GANs) have made large advances in image, audio, and text generation [@Hsu2017-ov; @Larsen2015-fz; @Pu2016-ln]. VAEs and GANs learn generative distributions on lower-dimensional encodings of input data [@Way2018-yi]. VAEs have found genomic applications. Rampasek et al. applied VAEs to learn drug responses based on gene expression data [@Rampasek2017-kv]. Way et al. trained a VAE called Tybalt to encode The Cancer Genome Atlas (TCGA) [@Way2018-yi]. Huang et al. have developed a theory of cancer development as a progression along a low dimensional space, justifying exploration of cancer metastasis using machine learning algorithms that learn low dimensional representations [@Huang2009-wr].
A new VAE-GAN hybrid architecture known as the Unified Feature Disentanglement Network (UFDN) learns fundamental features that distinguish input domains [@Liu2018-jg]. For multiple input data types, such as photographs, sketches, and watercolor paintings, the UFDN learns an VAE encoding of the data domains and trains a discriminator in the latent space to discriminate between domain types. Then, the UFDN can subsequently encode data from one domain and decode the data into a different domain[@Liu2018-jg]. An additional GAN distinguishes between real/fake images in the pixel space to promote high quality decodings[@Liu2018-jg].
In this work, we apply this new UFDN architecture to TCGA RNA-Seq data and learn a latent space embedding that allows us to convert between different cancer types given gene expression data. Given gene expression levels in a cancer sample of domain $A$, we can predict gene expression levels as if that cancer sample were of domain $B$. This represents a generative, personalized model of metastasis. We can sample points in our latent space encoding and decode them into any new cancer domain.
Additionally, we can partially interpolate between cancer domains. UFDN decoding is not strictly binary—input data can be decoded into a mix of output domains. We investigate *partial interpolations* of one cancer type into another, mimicking the progressive nature of metastasis.
We analyze the performance of our TCGA-trained UFDN on two tasks: predicting whether a sample is from cancerous or normal tissue and predicting which cancer sub-type a sample consists of. Additionally, we investigate partial interpolations from skin cutaneous melanoma (SKCM) TCGA samples to glioblastoma (GBM) by looking at differential expression of genes. We compute metagenes that summarize gene expression changes using integrative non-negative matrix factorization. Finally, we analyze Gene Ontology (GO) term enrichment in highly activated metagenes for each interpolated dataset.
![The overall workflow of our project. We aimed to identify the crucial changes in gene expression as cancer metastasizes from the original location to a new location. We encoded RNA-Seq samples from skin cutaneous melanoma, decoded them into glioblastoma, and then applied 3 bioinformatics tools to analyze which sets of genes were changing between cancer types.[]{data-label="fig:figure1"}](images/figure1.png){width="\textwidth"}
Methods
=======
Data Preprocessing
------------------
The data consisted of 10,433 samples of RNA-Seq gene expression levels across 33 cancer types for 20,501 genes from TCGA obtained via the R Package `curatedTCGA` [@Cancer_Genome_Atlas_Research_Network2013-pt; @Ramos2018-cx]. For the purpose of this work, we only considered the RSEM (RNA-Seq by Expectation Maximization [@Li2011-ra]) normalized expression levels. We divided 70%, 20%, and 10% to train, test, and holdout datasets, respectively.
Tybalt demonstrated that preprocessing gene expression levels by scaling gene-wise expression levels (across all samples) to between 0 and 1 yields a trainable latent space [@Way2018-yi]. We adapted this procedure by first clipping expression levels to fall within 3 standard deviations from the mean of gene-wise expression levels followed by the same min-max normalization of Tybalt [@Way2018-yi].
UFDN
----
### Theory
Liu et al. develop a UFDN as a combination of an encoder $E$, a generator $G$, and two discriminators: $D_v$ in the latent space and $D_x$ in the pixel space [@Liu2018-jg]. In our application, pixel space is replaced by “gene expression space.” $E$ takes input data and encodes it in a latent space. In our UFDN, we encode gene expression using fully connected networks. $D_v$ learns to discriminate between domains, or cancer types. Then generator $G$ uses a latent space encoding and a domain vector $d_v$ to produce gene expression data in domain $v$ [@Liu2018-jg]. Our UFDN uses $d_v\in\mathbb{R}^{33}$ since there are 33 cancer types in TCGA.
We define a *partial interpolation* with parameter $p\in[0,1]$ of an input of domain $c$ to domain $\hat{c}$ to be the decoding of the input into into a composition of domains $c$ and $\hat{c}$, with weight $p$ given to domain $\hat{c}$. That is, the domain vector of the partial interpolation has components $d_{v_{\hat{c}}}=p$, $d_{v_c}=1-p$, and remaining components zero. For instance, a 0.25-GBM interpolation means an input has been decoded with $d_{v_{GBM}}=0.25$ and original domain entry is $0.75$.
In the pixel space (or gene expression space), $D_x$ learns to distinguish between samples that have been decoded to their original domain $c$ or a new domain $\hat{c}$ [@Liu2018-jg]. The network is trained by iterative stochastic gradient updates to $E$, $D_v$, and $D_x$. For a more detailed exposition of the architecture of and gradient updates for training the UFDN, please see Section 3 of Liu et al. 2018 [@Liu2018-jg].
The encoder $E$ and generator $G$ are single layer networks, each with 500 hidden units, that learn a 100 dimensional latent space. The feature space discriminator $D_v$ is a single layer network with 64 hidden units and the pixel space discriminator $D_x$ is a two layer network with 500 and 100 hidden units. All networks are fully connected with leaky relu activation functions. We use 50,000 iterations of Adam updates with a learning rate of $10^{-4}$.
### Classification Tasks
We attempted two classification tasks using the UFDN. The first was classifying a sample as tumor or normal. This is referred to as the cancer status task. The second task was predicting cancer domain, one of 33 sub-types in the TCGA.
In order to solve these tasks, we developed 3 algorithms using UFDN:
- UDFN-MSE: classify a sample’s type by encoding the sample and decoding it into all 33 domains, predicting the type of the domain with lowest reconstruction error as defined by mean square error (MSE).
- Unsupervised UFDN: Inspired by the unsupervised domain adaptation experiments from Liu et al.[@Liu2018-jg], this algorithm predicts cancer status by encoding a sample into the latent space, then decoding it into the mesothelioma domain, regardless of input domain. We trained a random forest classifier to predict cancer status on mesothelioma training data. Use the prediction of this classifier to predict cancer status in the original input domain. The motivation for this approach is that the classifier trained on mesothelioma data is strong but the test data of interest is of a different cancer type.
- Semi-supervised UFDN: A hybrid of the two above algorithms used to predict cancer status and type. First, predict cancer type using UDFN-MSE. Then, predict cancer status using a random forest classifier trained on that specific type’s status data.
Interpolation Analysis
----------------------
We encoded 95 samples of SKCM (skin cutaneous melanoma) from our test set partition of the TCGA into our latent space using our trained UFDN. Then, we interpolated the samples into glioblastoma (GBM) at four different fractions of interpolation: 25%, 50%, and 75%, and 100%. The 100% interpolation represents a prediction of gene expression levels of the SKCM samples as GBM per sample.
In order to analyze how gene expression changed between SKCM samples and these samples as GBM, we performed a differential expression analysis using `edgeR` [@Robinson2010-xv; @McCarthy2012-ge]. This is an R package that uses a negative binomial distribution model to analyze significant gene expression changes between two groups [@Robinson2010-xv; @McCarthy2012-ge]. Although normally `edgeR` works with raw read counts, more recently the package creator has stated that RSEM normalized reads are also suitable for use with `edgeR` [@Smyth2015-ei].
We applied the inverse transformation of our min-max normalization to our four interpolated datasets since our UFDN decodes gene expression levels to the range \[0,1\]. Then we used `edgeR` to find differentially expressed genes between SKCM samples and 100% GBM interpolated samples. A p-value threshold for differential expression was set at $p=.05/20501=2.438*10^{-6}$ to control for false discovery.
Analyzing every single gene the significantly changed between SKCM and GBM would be a challenge, so we used integrative Non-negative Matrix Factorization (IntNMF) to learn metagenes that summarized gene expression changes [@Chalise2017-zl]. IntNMF learns a reduced dimensionality representation across multiple datasets [@Chalise2017-zl]. IntNMF learns a shared basis matrix $W\in\mathbb{R}^{p\times k}$ and where $p$ is the number of features (here, the differentially expressed genes) and $k$ is the number of metagenes, $k<<p$. Each dataset $D_j$ is described by a learned matrix $H_j\in\mathbb{R}^{k \times n}$ where $n$ is the number of samples in the dataset [@Chalise2017-zl]. Each row of $H_j$ represents the linear combination of metagenes of $W$ that combine to reconstruct the original sample in $D_j$ [@Chalise2017-zl]. We chose $k=60$ based on an analysis of the reconstruction error $\sum_j ||D_j-WH_j||_F$, where $F$ is the Frobenius norm. We learned $W$ and $H_j$ for each dataset using the R package `IntNMF` [@Chalise2017-zl].
Every element $g$ of column $W^{(i)}$ is non-negative and represents the contribution of gene $g$ to the $i$-th metagene [@Chalise2017-zl]. Each element $s$ of the $n$-th row of $H_j$ represents the contribution of metagene $s$ to the $n$-th sample of the $j$-th dataset. We can analyze how these metagenes change over the different interpolation datasets in order to understand how gene expression is changing [@Chalise2017-zl].
Finally, to understand the broad composition of the metagenes discovered by IntNMF, we used Gene Ontology (GO) enrichment analysis. GO terms are an ontology of three categories: biological processes, molecular function, and cellular component. They link together information about the functions and relationships of genes and proteins. `topGO` is an R package that analyzes if GO terms, which have been mapped to genes, show up more often than expected in a set of genes and associated scores for each gene [@Alexa2010-od].
We used a Kolmogorov-Smirnov like test known as Gene Score Enrichment Analysis that calculates p-values of enrichment based on a score for each gene[@Alexa2010-od]. In our work, we did this test on each metagene derived from IntNMF with the score for gene $g$ as $W^{(i)}_g$ [@Alexa2010-od]. By looking at the top scoring GO terms for each metagene, we understand what sort of genes are changing as we interpolate between cancer types [@Alexa2010-od].
Code
----
All our code is available at <https://github.com/bkompa/UFDN-TCGA>. We used Liu et al.’s implementation of UFDN as a starting point but had to expand the architecture to work with an arbitrary number of domains. We wrote all other code used for analysis with the various packages mentioned above.
Results
=======
UFDN Training and Performance
-----------------------------
First, we validated that our UFDN learned a non-trivial latent space representation of TCGA RNA-Seq data. We projected both the TCGA data and latent space encodings into UMAP space [@McInnes2018-sz]. UMAP learns a Riemann manifold representation of the data [@McInnes2018-sz]. We used hyper-parameters `spread=2.0` and `min_dist=.01` to produce Figure \[fig:umap\]. We observed distinct clusters by cancer types for both the original data and encodings. We proceeded in our downstream analysis confident that our UFDN had learned how to discern between cancer types based on these UMAP projections.
![UMAP projections of the RNA-Seq TCGA data (Figure \[fig:umap\]A) and UFDN latent space encodings of said data (Figure \[fig:umap\]B). The full 20,501 dimensional representation of gene expression levels have more cancer specific clusters. The 100 dimensional latent space encodings of these samples still clustered in the UMAP space, though to a lesser extent.[]{data-label="fig:umap"}](images/umapFig.png){width=".8\textwidth"}
Next, we estimated the ability of our UFDN to take data from a source domain (original cancer type) and interpolate these data into a target domain (new cancer type). We considered the fraction of the $k$ nearest neighbors, in the training data, of the interpolated samples that were in the target domain as a measure of success. These decoding rates are shown in Figure \[fig:kNN\]. There were certain cancers that the UFDN was able to more robustly interpolate into. These included glioblastoma, acute myloid leukemia, mesothelioma, and prostate adenocarcinoma, among others. Difficult cancers to interpolate into were sarcomas, which are a heterogeneous subcategory of soft tissue cancers and cervical squamous cell carcinoma.
![The fraction of $k$ nearest neighbors that were in the target domain (the rows of the figures) after decoding from a source domain (the columns of the figures). Some domains were noticeably more difficult to interpolate into. Glioblastoma had strong interpolation results across $k\in[1,5,10,20]$.[]{data-label="fig:kNN"}](images/knn2.png){width=".8\textwidth"}
Algorithm Cancer Status Acc (Train/Test) Cancer Type Acc (Train/Test)
---------------------- -------------------------------- ------------------------------ --
Random Forests **99.60%/98.41%** **99.65%/95.20%**
UFDN-MSE — 96.51%/94.10%
Unsupervised UDFN 95.60%/86.14% —
Semi-supervised UDFN **99.60%/98.41%** 96.51%/94.10%
: Results on two classification tasks compared to a random forest baseline.
\[tbl:classification\]
Finally, we analyzed our UFDN’s performance on two classification tasks: cancer status prediction and cancer type prediction. Table \[tbl:classification\] reports the performances of our three UFDN classification algorithms as compared to a random forest baseline. The random forests had a maximum depth of 15 and were composed of 100 trees. The semi-supervised UFDN algorithm was able to match the performance of random forests on the cancer status task and was comparable on the cancer type task. Other UFDN algorithms were less successful compared to the baseline.
Gene Expression Changes
-----------------------
After interpolating 95 samples of SKCM from the test set into GBM, we analyzed which genes had significant changes in expression between the SKCM and 1.0-GBM samples. Using `edgeR`, we looked for genes that had differential expression that exceeded a significance threshold of $p=2.43*10^{-6}$. There were 10,557 genes that exceeded this threshold. Figure \[fig:edgeR\] shows the plot of average log fold change versus average log counts per million and highlights the differentially expressed genes between the two groups.
![Differential expressed genes at a Bonferroni corrected p-value of $2.43*10^{-6}$. These genes are shown in red, while non DE genes are in black. 10,557 genes were differential expressed between skin cancer samples and 1.0-glioblastoma interpolated skin cancer samples.[]{data-label="fig:edgeR"}](images/edgerFig.png){width=".7\textwidth"}
For the 10,557 differential expressed genes, we learned a shared basis $W$ using IntNMF. By varying the rank of that basis, we were able to decrease the reconstruction error across datasets SKCM, 0.25-GBM, 0.5-GBM, 0.75-GBM, and 1.0-GBM. Figure \[fig:reconstruction\] reports how $k$ affected the reconstruction error. We chose $k=60$ for subsequent analysis based on the inflection point of this reconstruction curve. Hutchins et al. suggest that this is an optimal way to select $k$ for NMF [@Hutchins2008-oc]. $k=60$ was also chosen for computational considerations. Optimizing $H_j$ and $W$ for $k=60$ took nearly 7 hours and increasing $k$ much more would significantly increase this considerable time requirement.
![Reconstruction error based on the Frobenius norm from IntNMF versus $k$, the rank of $W$ and $H_j$ in IntNMF, on the x-axis as the number of metagenes. For subsequent analysis, $k$ was chosen to be 60 as error is nearly at an inflection point and plateauing.[]{data-label="fig:reconstruction"}](images/reconstructionFig.png){width=".8\textwidth"}
Finally, we visualized the rows of $H_j$ for each dataset in $\{$SKCM, 0.25-GBM, 0.50-GBM, 0.75-GBM, 1.00-GBM$\}$. The columns of each heatmap in Figure \[fig:sample\_heatmap\] represent the relative activation of the respective metagene. As interpolation towards GBM increases, distinct metagenes increase their responsibility for reconstructing $H_j$. In SKCM, metagene 36 has the most representation in the data. For 0.25-GBM, 0.50-GBM, and 0.75-GBM, it was metagenes 15, 32, and 1, respectively.
In the 1.00-GBM heatmap (Figure \[fig:sample\_heatmap\] E), we saw the increased activation of metagene 23. When we took 33 samples of TCGA GBM data from the test set and learned the matrix $H_{GBM}$ that minimized reconstruction error $||D_{GBM}-WH_{GBM}||_F$ for the same, fixed, $W$ learned previously by IntNMF, we observed the same metagene 23 dominating (Figure \[fig:sample\_heatmap\] F).
![Heatmap visualization of the $H_j$ matrices for each interpolation of the SKCM test data set. No row or column reordering was done to keep consistent metagene order across datasets. A full interpolation of SKCM data into GBM data results in a consistent activation of metagene 23 (Figure \[fig:sample\_heatmap\]E). This is replicated in $H_{GBM}$ (Figure \[fig:sample\_heatmap\]F), which was optimized against the fixed $W$ basis learned for the other 5 datasets.[]{data-label="fig:sample_heatmap"}](images/sampleHeatmap.png){width="\textwidth"}
We proceeded to analyze the dominant metagene for every dataset $H_j$ for GO term enrichment. In the interest of space, we only report the top 15 most enriched GO terms for metagene 23 based on p-value. Table \[Table2\] reports the GO term as well as p-value for each term.
GO ID Term p-value
------------ --------------------------------------------- ----------
GO:0003676 Nucleic acid binding 5.20E-19
GO:0003735 Structural constituent of ribosome 2.70E-15
GO:0003723 RNA binding 3.90E-14
GO:0003677 DNA binding 1.60E-12
GO:0005198 Structural molecule activity 3.80E-12
GO:0000981 DNA-binding transcription factor activit... 4.70E-12
GO:0003700 DNA-binding transcription factor activit... 3.50E-11
GO:0140110 Transcription regulator activity 2.80E-09
GO:0008376 Acetylgalactosaminyltransferase activity 4.10E-08
GO:0043492 ATPase activity, coupled to movement of ... 1.00E-07
GO:0060089 Molecular transducer activity 1.30E-07
GO:0004126 Cytidine deaminase activity 2.10E-07
GO:0019239 Deaminase activity 4.50E-07
GO:0048020 CCR chemokine receptor binding 7.30E-07
GO:0008009 Chemokine activity 8.10E-07
: The top 15 Gene Ontology Terms enriched in metagene 23
\[Table2\]
Discussion
==========
Our UFDN was able to learn a biologically relevant latent space encoding of TCGA data. Classification task results in Table \[tbl:classification\] indicate that our UFDN was able to compete with random forests that were trained on all 20,501 gene expression features. This indicates our algorithm was able to learn an efficient, useful embedding of gene expression data. Figure \[fig:umap\] demonstrates that we learned an encoding space that clustered cancers of the same domain. This likely facilitates successful interpolation and classification between cancer domains. Additionally, our UFDN could robustly interpolate into many cancer domains. Although Figure \[fig:kNN\] demonstrates that not every cancer domain was easy for the UFDN to decode into, one thing to note is that almost every column (target domain) had at least one element with high decoding fraction. It’s possible to consider multiple interpolations and potentially from converting from domain A to B to C would have a higher success rate than going from A to C.
We learned 10,557 differentially expressed genes between SKCM and 1.0-GBM interpolated samples as demonstrated in Figure \[fig:edgeR\]. This reduction in dimensionality allowed us to make IntNMF computationally tractable. The lower number of genes considered in IntNMF, the faster the learning of the shared basis $W$ and dataset specific $H_j$. Analysis of the reconstruction error from IntNMF informed our choice of 60 metagenes (see Figure \[fig:reconstruction\]). In Figure \[fig:sample\_heatmap\], we investigated how linear combinations of these distinct metagenes reconstructed samples from many partially interpolated datasets. We observed unique metagenes increasing activation for each partial interpolation. This is an approximation of how gene expression profiles change during metastasis.
When we learned $H_{GBM}$, the representation of TCGA GBM samples with respect to the basis $W$, something remarkable happened. Note that $W$ was not informed by the TCGA dataset $GBM$ at all. $W$ was simply the shared basis trained by IntNMF on interpolation datasets SKCM (equivalently, 0.00-GBM), 0.25-GBM, 0.5-GBM, 0.75-GBM, and 1.0-GBM. Yet when $H_{1.0-GBM}$ and $H_{GBM}$ were compared side by side in Figure \[fig:sample\_heatmap\] E&F, their metagene activation profiles were dominated by the same metagene 23. Therefore, our interpolation from SKCM to GBM successfully recapitulated observed gene expression activity.
Furthermore, when we explored several of the GO terms identified by a GO term enrichment analysis, metagene 23 was enriched for terms related to glioblastoma. GO:0008376 represents a glycoprotein with a known association to glioblastoma [@Zhang2003-vf; @Kroes2007-an]. GO:0004126 refers to cytidine deaminase activity. Cytidine deaminase gene therapy has been identified as a potential treatment for glioblastoma[@Fischer2005-zd; @Miller2002-fa]. GO:0048020 and GO:0008009 are associated with chemokines, which are implicated in glioblastoma development [@Zhou2002-qk; @Rempel2000-nr]. Our metagenes learned glioblastoma-specific genes and our UFDN interpolated skin cancer samples to glioblastoma. Further analysis of the metagenes activated during interpolations 0.25-GBM, 0.50-GBM, and 0.75-GBM could provide starting points for the investigation of the metastasis pathway from SKCM to GBM. This could help explain the association between melanoma and glioblastoma that is not currently understood [@Desai2008-cj; @Scarbrough2014-wj].
Conclusion
==========
Our UFDN learned a biologically relevant latent space that facilitated meaningful interpolations between cancer domains. Our latent space can be used to generate more examples of transitions between cancers types. Our interpolations from SKCM to GBM have feasible biological interpretations and suggest possible gene expression changes during the mysterious transition from melanoma to glioblastoma.
Acknowledgements {#acknowledgements .unnumbered}
================
We acknowledge the helpful suggestions of Dr. Devavrat Shah and Flora Meng on this project. Kompa is also indebted to the feedback of Scott Nanda and Kathryn Almon.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a wave function representation for the canonical ensemble thermal density matrix by projecting the thermofield double state against the desired number of particles. The resulting canonical thermal state obeys an imaginary time-evolution equation. Starting with the mean-field approximation, where the canonical thermal state becomes an antisymmetrized geminal power wave function, we explore two different schemes to add correlation: by number-projecting a correlated grand-canonical thermal state, and by adding correlation to the number-projected mean-field state. As benchmark examples, we use number-projected configuration interaction and an AGP-based perturbation theory to study small molecules and model systems.'
author:
- Gaurav Harsha
- 'Thomas M. Henderson'
- 'Gustavo E. Scuseria'
bibliography:
- 'CanonTFD.bib'
title: 'Wave function methods for canonical ensemble thermal averages in correlated many-electron systems'
---
Introduction
============
Thermal properties of many-body systems can be computed either in the canonical ensemble or the grand-canonical ensemble. The choice of ensemble makes no practical difference in the final result in large systems. It does so, however, for a finite system. This is because the relative fluctuation in particle number in the grand-canonical ensemble scales as the inverse square root of particle number itself, i.e. $$\frac{
\sqrt{\langle N^2 \rangle_{gc} - \langle N \rangle_{gc}^2}
}{
\langle N \rangle_{gc}
} \sim \frac{1}{\sqrt{\langle N \rangle_{gc}}},$$ and vanishes in the limit $\langle N \rangle_{gc} \rightarrow \infty$, where $\langle \ldots \rangle_{gc}$ denotes the grand-canonical thermal expectation value.
A wide range of methods are available to study the thermal properties of quantum systems within the grand-canonical ensemble, e.g., thermal Hartree-Fock, [@mermin_stability_1963; @sokoloff_consequences_1967] perturbation theories, [@matsubara_new_1955; @santra_finite-temperature_2017; @hirata_converging_2018] path integral and Green’s function methods, [@zgid_finite_2017] finite-temperature Monte Carlo, [@zhang_finite_temperature_1999; @militzer_path_2000; @rubenstein_finite-temperature_2012; @schoof_ab_2015; @takai_finite-temperature_2016; @claes_finite-temperature_2017; @liu_ab_2018; @he_finite-temperature_2019; @petras_using_2020; @liu_unveiling_2020] density matrix renormalization group and density functional theory based methods, [@verstraete_matrix_2004; @feiguin_finite-temperature_2005; @stoudenmire_minimally_2010; @pittalis_exact_2011; @nocera_symmetry-conserving_2016; @ren_time-dependent_2018] as well as the more recently explored thermal equivalents of configuration interaction and coupled cluster, [@sanyal_thermal_1992; @sanyal_systematic_1993; @mandal_thermal_1998; @mandal_finite-temperature_2003; @hermes_finite-temperature_2015; @hummel_finite_2018; @white_time-dependent_2018; @harsha_thermofield_2019; @harsha_thermal_cc_2019; @shushkov_real-time_2019; @white_time-dependent_2019; @white_finite-temperature_2020] and algorithms for quantum computers. [@wu_variational_2018; @mcardle_variational_2019; @zhu_variational_2019; @motta_determining_2020]
In contrast, canonical ensemble techniques are scarce and even fewer are suitable for efficient application to correlated electronic systems. One way to enforce a fixed number of particles is by introducing a second Lagrange multiplier $\mu_2$ for the fluctuation, in much the same spirit as the chemical potential $\mu_1$ acts as a Lagrange multiplier to fix the number of particles. That is, one can either define a generalization of the density operator as $$\rho = \exp \left[ -\beta \left( H - \mu_1 (N - N_0) - \mu_2 (N^2 - N_0^2) \right) \right ],$$ where the parameters $\mu_1$ and $\mu_2$ enforce the constraints, $$\langle N \rangle = N_{0},
\quad \mathrm{and,} \quad
\langle N^2 \rangle = N_0^2,$$ or introduce corrections to the grand-canonical ensemble averages by subtracting contributions from wrong number sectors in the Hilbert space. [@kosov_calculations_2008] While it provides the convenience of using several available grand-canonical methods, such simultaneous optimization problems can be numerically tedious as the optimized values of $\mu_2$ are generally very large and ideally infinite, something which has also been observed in spin-projection. [@andrews_spin_1991] On the other hand, we can evaluate the ensemble averages in the appropriate number sector to begin with, e.g. in the minimally entangled typical thermal states algorithm, [@stoudenmire_minimally_2010; @binder_symmetric_2017] canonical ensemble perturbation theory, [@jha_finite-temperature_2020] and projection based techniques.[@tanabe_quantum_2005; @esashika_effects_2005; @nakada_new_2006; @magnus_quantum_2017]
For a wide variety of problems which involve isolated finite systems with a fixed number of particles, the canonical ensemble is more appropriate. Examples of such systems include molecules in a warm gaseous phase (of interest in geochemistry), [@guillot_interiors_1999] ultra-cold chemical systems, [@balakrishnan_perspective:_2016; @bohn_cold_2017] quantum wires with number conserving Majorana modes, [@diehl_topology_2011; @ortiz_many-body_2014; @iemini_localized_2015] and superconductivity in small grain systems.[@mastellone_small_1998] Besides, the canonical ensemble provides a potential computational advantage over grand canonical alternatives since it eliminates the need for finding the appropriate chemical potential. Evidently, a robust and convenient framework to study canonical-ensemble finite-temperature properties of finite many-electron systems is desirable.
In this manuscript, we leverage the thermofield dynamics[@matsumoto_thermo_1983; @semenoff_functional_1983; @umezawa_methods_1984; @evans_heisenberg_1992] to construct a number-projected thermal wave function, called the canonical thermal state, which provides an exact wave function representation of the canonical ensemble density matrix. It obeys an imaginary-time Schrödinger equation which can be solved at various levels of approximation, and at the level of mean-field, reduces to a number-projected BCS wave function, also known as the antisymmetrized geminal power (AGP) state.[@coleman_structure_1965] A similar number-projected BCS theory for the canonical thermal state was also proposed by the authors of Refs. . Mean-field description, however, misses out on a lot of important physics. Here, we provide a recipe to generalize correlated ground-state theories (e.g., perturbation theory, CI, CC, etc.) to finite-temperature. Moreover, the identification of the mean-field state as an AGP allows us to exploit the newly developed tools for efficient evaluation of the thermal expectation values via AGP density matrices. [@khamoshi_efficient_2019] We restrict our discussion to electronic systems, but generalization to other fermionic and bosonic systems is straightforward.
Thermofield dynamics
====================
Thermofield dynamics is conventionally formulated for the grand-canonical ensemble, where it constructs a wave function representation of the thermal density operator by introducing a conjugate copy of the original system such that the ensemble thermal averages can be expressed as an expectation value over the thermal state, $$\langle \mathcal{O} \rangle = \mathrm{Tr}\left( e^{-\beta(H - \mu N)} \mathcal{O} \right)
= \frac{\langle \Psi(\beta) \vert \mathcal{O} \vert \Psi (\beta) \rangle}{\langle \Psi (\beta) \vert \Psi (\beta) \rangle},$$ where the thermal state $\vert \Psi (\beta) \rangle$ is given by
$$\begin{aligned}
\vert \Psi (\beta) \rangle &= e^{-\beta (H - \mu N) / 2} \vert \mathbb{I} \rangle,
\\
\vert \Psi (0) \rangle &= \vert \mathbb{I} \rangle = \prod_{p} \Big ( 1 + c_p^\dagger \tilde{c}_p^\dagger \Big ) \vert -; - \rangle.
\label{inf-temp-gcan-thermal-state}
\end{aligned}$$
Here $\beta$, $\mu$, $H$ and $N$ are the inverse temperature, chemical potential, the Hamiltonian, and the number operator respectively. The identity state $\vert \mathbb{I} \rangle$ is the exact infinite-temperature thermal state and is an extreme BCS state with Cooper pairs formed by pairing physical particles with the corresponding conjugate particles. The norm of the state gives the partition function. The product in Eq. \[inf-temp-gcan-thermal-state\] runs over all spin-orbitals $p$ and $\vert -; -\rangle$ denotes the vacuum state for both the physical and conjugate systems. By its definition, the thermal state obeys imaginary-time evolution equations, one each for $\beta$ and $\mu$,
$$\begin{aligned}
\frac{\partial}{\partial \beta} \vert \Psi (\beta) \rangle &= -\frac{1}{2} H \vert \Psi(\beta) \rangle,
\\
\frac{\partial}{\partial \mu} \vert \Psi (\beta) \rangle &= \frac{\beta}{2} N \vert \Psi(\beta) \rangle,
\end{aligned}$$
where we have assumed that $[H, N] = 0$, as physical electronic systems are number-conserving.
Like the ground state, finding $\vert \Psi (\beta) \rangle$ exactly is possible only for very small systems with a few electrons, and suitable approximations are generally required. The simplest approximation is the mean-field approach, where $H$ is replaced with a one-body mean-field Hamiltonian $H_0$. In the basis where $H_0 = \sum_p {\epsilon}_p c^\dagger_p c_p$, the resulting mean-field thermal-state is a BCS state of the form $$\begin{aligned}
\vert 0 (\beta, \mu) \rangle &= e^{-\beta (H_0 - \mu N) / 2} \vert \mathbb{I} \rangle,
\nonumber
\\
&= \prod_p \Big( 1 + e^{-\beta ({\epsilon}_p - \mu) / 2} c^\dagger_p \tilde{c}^\dagger_p\Big) \vert -;- \rangle.\end{aligned}$$ Higher order approximations are generally formulated with the mean-field state as the reference, $$\vert \Psi (\beta) \rangle \simeq \Omega(\beta, \mu) \, \vert 0 (\beta, \mu) \rangle,$$ which resembles the interaction picture approach. We exploited this theory in Refs. to formulate finite-temperature versions of configuration interaction and coupled cluster theory. We recommend these articles and references therein for further details on thermofield theory.
Canonical ensemble theory
=========================
The canonical ensemble thermal state can be constructed by projecting grand-canonical state against the desired particle number $N_0$, $$\vert \Psi (\beta) \rangle_{c}
= \mathcal{P}_{N_0} \vert \Psi (\beta) \rangle_{gc},$$ where $\mathcal{P}_{N_0}$ projects $\vert \Psi(\beta) \rangle_{gc}$ onto the Fock-space with $N_0$ electrons. The particle-conserving property of $H$ implies that $[H, \mathcal{P}_{N_0}] = 0$, and the resulting canonical thermal state obeys an imaginary-time evolution equation analogous to its grand-canonical counterpart, $$\frac{d}{d \beta} \vert \Psi (\beta) \rangle_{c} = -\frac{1}{2} H \vert \Psi(\beta) \rangle_{c}.
\label{imag-time-evol-eq}$$ Like the grand-canonical theory, a series of approximations can be introduced, from a simple mean-field theory to higher order theories that add correlation effects on it.
Mean-field formalism
--------------------
The imaginary-time evolution equation can be integrated within the mean-field approximation, $H \approx H_0$. As for the grand-canonical theory, using an $H_0$ that carries no implicit temperature dependence, and working in a basis where it is diagonal, the mean-field state becomes
\[mean-field-eqns\] $$\begin{aligned}
\vert \Psi_{0} (\beta) \rangle_{c}
&= \mathcal{P}_{N_0} \vert 0(\beta, \mu=0) \rangle,
\label{mean-field:eqn-a}
\\
&=
\mathcal{P}_{N_0} \prod_{p} \Big ( 1 + e^{-\beta {\epsilon}_p / 2} c_p^\dagger \tilde{c}_p^\dagger \Big ) \vert -; - \rangle,
\label{mean-field:eqn-b}
\\
&=
\mathcal{P}_{N_0} \prod_{p} \Big ( 1 + \eta_p P^\dagger_p \Big ) \vert -; - \rangle,
\label{mean-field:eqn-c}
\\
&=
\frac{1}{N_0!}\left( \Gamma_\beta^\dagger \right)^{N_0} \vert -;- \rangle
=
\vert \Psi_{AGP} (\beta) \rangle,
\label{mean-field:eqn-d}
\end{aligned}$$
where $\eta_p = e^{-\beta {\epsilon}_p / 2}$, and we have identified $P^\dagger_p = c^\dagger_p \tilde{c}^\dagger_p$ as the pair-creation operator. As already noted, the un-projected product state in Eq. \[mean-field:eqn-b\] is a BCS state and its number-projected version is well known as AGP, with the geminal creation operator $\Gamma_\beta^\dagger$ defined as $$\Gamma_\beta^\dagger = \sum_p \eta_p P_p^\dagger.$$ Identification of the mean-field state as an AGP is interesting and, with recent developments on efficient evaluation of overlaps and expectation values, as well as geminal based correlated wave function theories, [@khamoshi_efficient_2019; @henderson_geminal-based_2019; @henderson_correlating_2020; @dutta_geminal_2020] provides a good starting point to include correlation effects. An improved mean-field description can also be obtained by optimizing both the energy levels ${\epsilon}$ and the one-electron basis to find an $H_0$ that minimizes the Helmholtz free energy, in much the same way as Mermin’s thermal Hartree Fock theory in Ref. , and as discussed in Refs. .
Correlated thermal state
------------------------
A plethora of approximate wave function methods are available to study ground-state properties of correlated electronic systems. As we have shown in Refs. , the thermofield formalism allows for a direct generalization of these methods to finite-temperature. Since physical electronic systems conserve the number of particles, i.e. $[H, \mathcal{P}_{N_0}] = 0$, we face two options while constructing a correlated approximation to the canonical thermal state: *projection after correlation* (PAC), and *correlation after projection* (CAP). In PAC, we first construct an approximate grand-canonical thermal state by adding correlation on a broken-symmetry mean-field reference (thermal BCS in our case) and then perform the number-projection, $$\vert \Psi \rangle \simeq \mathcal{P}_{N_0} \Omega(\beta) \, \vert 0 (\beta) \rangle; \quad \vert 0 (\beta) \rangle = e^{-\beta H_0 / 2} \vert \mathbb{I} \rangle.$$ The correlation operator $\Omega$ is built out of number non-conserving BCS quasiparticles, [@henderson_quasiparticle_2014; @signoracci_ab_2015] and the un-projected part of the thermal state, $\Omega (\beta) \vert 0 (\beta) \rangle$, looks like a standard single-reference CI wave function, which simplifies the process of correlating the reference. In order to carry out the projection efficiently, we use an integral form for the projection operator, [@peierls_collective_1957; @bayman_derivation_1960; @ring_nuclear_1980] i.e. $$\mathcal{P}_{N_0} = \frac{1}{2 \pi} \int_0^{2 \pi} d \phi e^{i \phi (N_0 - N)}.$$ Computing matrix elements and overlaps in the presence of $\mathcal{P}$ involves the use of transition density matrices and can be complicated (see e.g., Refs. ) For CAP, we use the thermal AGP state in Eq. \[mean-field-eqns\] as the reference and add correlation using a number-conserving wave operator, $$\vert \Psi \rangle \simeq \Lambda(\beta) \vert \Psi_{AGP}(\beta) \rangle
= \Lambda (\beta) \mathcal{P}_{N_0} \vert 0(\beta) \rangle.$$ Contrasting with CAP, the projection problem here is trivial but adding correlation becomes complicated.
Both of these techniques have been explored extensively for ground-state methods. [@degroote_polynomial_2016; @tsuchimochi_communication_2016; @wahlen-strothman_merging_2017; @hermes_combining_2017; @qiu_projected_2017; @qiu_particle-number_2019; @henderson_geminal-based_2019; @henderson_correlating_2020; @dutta_geminal_2020] Here, we discuss an example for each: a finite-temperature generalization of the number-projected CI, along the lines discussed by Tsuchimochi *et. al.* in Ref. , and an imaginary-time perturbation theory based on the thermal AGP as the reference, as explored in Refs.
{width="0.49\linewidth"} {width="0.49\linewidth"}
### Projection after correlation
The number-projected thermal CI state is parameterized as $$\vert \Psi_{c} (\beta) \rangle =
\mathcal{P}_{N_0} e^{t_0} \left ( 1 + T \right ) \vert 0 (\beta) \rangle,$$ where $\vert 0 (\beta) \rangle$ is the thermal BCS state at inverse temperature $\beta$, $t_0$ keeps track of the norm of the state (related to the grand potential) and $T$ creates quasiparticle excitations on the BCS, $$T = \sum_{pq} t_{pq} a_p^\dagger \tilde{a}_q^\dagger + \frac{1}{4} \sum_{pqrs} t_{pqrs} a^\dagger_p a^\dagger_q \tilde{a}^\dagger_s \tilde{a}^\dagger_r + \ldots.$$ The CI amplitudes can be determined in two different ways. One can compute them in the grand-canonical ensemble, as we have done in Ref. , and then perform a one-shot projection. This approach is generally known as *projection after variation* (PAV). Alternatively, the amplitudes can be computed in the presence of the projection operator by solving the imaginary-time evolution equation, referred to as *variation after projection* (VAP). VAP allows for more variational freedom and thus, performs better than PAV. Accordingly, we focus our attention on VAP hereafter.
Substituting this CI ansatz into Eq. \[imag-time-evol-eq\] and evaluating overlaps of the resulting equation against the ground and excited BCS states, we get $$\begin{gathered}
\int_0^{2\pi} d\phi \, \langle 0 (\beta) \vert \nu e^{i \phi (N_0 - N)} \left(
(1 + T) \frac{d t_0}{d \beta} + \frac{d T}{d \beta}
\right) \vert 0 (\beta) \rangle
\\
= \int_0^{2\pi} d \phi \, \langle 0(\beta) \vert \nu e^{i \phi (N_0 - N)} \bar{H} \vert 0(\beta) \rangle,
\label{ci-imaginary-time}\end{gathered}$$ where $\bar{H}$ is the effective Hamiltonian, $$\bar{H} =
-\frac{1}{2} \Big(H (1 + T) - (1 + T) H_0 \Big),$$ and $\nu$ takes values from $\{1,~ a_p^\dagger \tilde{a}_q^\dagger,~ a_p^\dagger a_q^\dagger \tilde{a}^\dagger_s \tilde{a}^\dagger_r, ~\ldots \}$ to construct ground and excited BCS states for the bra. Both the amplitudes as well as the quasiparticle operators are functions of temperature, therefore the $\beta$-derivative can be broken down into the derivative of the amplitudes and that of the operator parts, $$\frac{d T}{d \beta} = \frac{d_{amp} T}{d \beta} + \frac{d_{op} T}{d \beta}.$$ We can rewrite Eq. \[ci-imaginary-time\] as a system of first-order ODEs that govern the evolution of the CI-amplitudes, $$\sum_{\mu} A_{\nu \mu} \cdot \frac{\partial t_\mu}{\partial \beta} = B_\nu,
\label{cisd-evolution-equation}$$ where $A$ is the overlap matrix,
$$\begin{aligned}
A_{\nu \mu} &=
\int_0^{2\pi} d\phi \, \langle \nu (\beta) \vert e^{-i \phi (N - N_0)} \mathcal{L}_\mu \vert 0(\beta) \rangle,
\\
\mathrm{with} \quad
\mathcal{L}_\mu &= \begin{cases}
1 + T, & \mu = 1\\
\mu, & \mu \in \{a_p^\dagger \tilde{a}_q^\dagger, \, a_p^\dagger a_q^\dagger \tilde{a}_s^\dagger \tilde{a}_r^\dagger \}
\end{cases}.
\end{aligned}$$
The right hand side vector $B_\nu$ is given by
$$\begin{aligned}
B_\nu &=
\int_0^{2\pi} d \phi \, \langle \nu (\beta) \vert e^{-i \phi (N - N_0)} \mathcal{R} \vert 0 (\beta) \rangle,
\\
\mathcal{R} &=
\bar{H} - \frac{d_{op} T}{\partial \beta}.
\end{aligned}$$
Here, we have used $\nu, \mu$ as a composite notation for the ground and excited quasiparticle states. Equation \[cisd-evolution-equation\] can be integrated starting from $\beta = 0$, where $T = 0$ is the exact initial condition. However, exact evolution requires inversion of the overlap matrix $A$, which is computationally expensive. Moreover, $A$ may also have zero (or near-zero) eigen modes. To avoid these issues, at each $\beta$-grid point, we solve for the derivative vector iteratively using MinresQLP, [@choi_minres-qlp_2011; @choi_algorithm_2014] a robust algorithm for singular linear systems, and then use a fourth order Runge-Kutta method to perform the integration. We further observe that the partial traces of higher rank terms in the CI operator are proportional to the lower-rank terms, e.g. CI with single and double excitations is equivalent to CI with just the double excitations. To avoid linear dependencies in the overlap matrix, we keep only the highest rank terms in our truncated CI theory. Finally, we also note that the projection integral converges rapidly as the number of grid points becomes greater than the number of spin-orbitals.
### Correlation after projection
For correlation after projection method, a numerical integration to perform the projection is not required as it uses a strictly number conserving state, the thermal AGP, as the reference. As an example for this approach, we consider the perturbation theory (PT), where we partition the Hamiltonian as $H = H_0 + \lambda V$, where $H_0$ is the mean-field contribution and $V$ acts as a perturbation. The canonical thermal state can be expanded as a series in $\lambda$,
$$\begin{aligned}
\vert \Psi(\beta) \rangle
&= \vert \Psi_0 \rangle + \lambda \vert \Psi_1 \rangle + \lambda^2 \vert \Psi_2 \rangle + \ldots,
\\
&= e^{-\beta H_0 / 2} \left( \vert \phi_0 \rangle + \lambda \vert \phi_1 \rangle + \lambda^2 \vert \phi_3 \rangle + \ldots \right).
\label{pt-series-psi}
\end{aligned}$$
Substituting this form for $\vert \Psi \rangle$ in Eq. \[imag-time-evol-eq\] and collecting terms at various orders in $\lambda$ gives $\partial \vert \phi_0 \rangle / \partial \tau = 0$, or equivalently $\vert \Psi_0 \rangle = \vert \Psi_{AGP}(\beta) \rangle$ for terms at $\mathcal{O}(\lambda^0)$, and $$\frac{\partial}{\partial \tau} \vert \phi_n \rangle
=
-\frac{1}{2} e^{\tau H_0 / 2} V e^{-\tau H_0 / 2} \vert \phi_{n-1} \rangle
\label{pt-eqns}$$ for $\mathcal{O}(\lambda^n), \,n \geq 1$. Integrating Eq. \[pt-eqns\] yields perturbative corrections identical to those in a time-dependent interaction picture theory. We work in a basis where $H_0$ is diagonal. This allows us to integrate the equations analytically. Detailed notes on both the projected CI and the AGP based perturbation theory are available in the Supplemental Information.
Results
=======
We apply the projected CI with double excitations (CID), as well as the second order perturbation theory (PT2) to small molecular and model systems to highlight the performance of these finite-temperature canonical ensemble methods against exact benchmark results. Figure \[fig1:h2-sto3g\] (left) shows error in internal energy for a Hydrogen molecule in the minimal STO-3G basis and at a bond length of $0.74 {\textup{\AA}}$. The results compare the performance of optimized and unoptimized projected thermal BCS (or AGP), PAV and VAP projected thermal CID, and PT2. To make comparison with the grand-canonical results worked out in Refs. , we also plot the internal energy errors in the grand-canonical quasiparticle CI with singles and doubles (CISD) theory, computed with respect to the exact grand-canonical results. In the low temperature limit, where the canonical ensemble is of special interest, both the optimized and the unoptimized thermal AGPs perform comparably. We use the ground-state spin-restricted Fock operator as our unoptimized $H_0$. It is apparent that the mean-field approach misses out a lot of correlation, a part of which is recovered by CID and PT2. In fact, the VAP CID, like its ground-state analogue and unlike the grand-canonical CISD, is exact for a two-electron system like the $H_2$ molecule, and expectedly outperforms the PAV approach. The second order perturbation theory, though not exact, also improves upon the mean-field results. All the CI and PT results approach their appropriate ground-state counterparts in the zero temperature limit.
![Comparison of specific heat trends for the six-site Hubbard model with $U/t= 6$ as a function of temperature. The exact results highlight the difference between canonical and the grand-canonical ensemble.[]{data-label="fig3:hubbard-cv"}](fig2){width="0.90\linewidth"}
The second plot in Figure \[fig1:h2-sto3g\] shows similar results for the six-site Hubbard model with $U/t = 2$ and $6$ at half-filling. For $U/t=2(6)$, we use the spin-restricted (unrestricted) Fock operator as our $H_0$. Having noted already that the optimized and the unoptimized thermal AGP perform comparably, we just plot the latter. Besides, it is the unoptimized $H_0$ that defines the reference state upon which we build both the CI and PT2. Similarly, we only consider the VAP approach for the projected CID results. Once again, we observe that the projected CID recovers a fair amount of correlation over the mean-field, more so in the weakly correlated regime. The trends for the grand-canonical and the canonical CI errors are alike and also analogous to the ground-state theory. The AGP based PT2 results perform well for $U/t = 2$ but do not introduce any significant improvement over the mean-field results for $U/t=6$. Ground-state PT2 shows similar trends. This validates our observation that the performance of thermal methods generally follows their ground-state counterparts.
To further highlight the merits of the projected CI theory over mean-field, as well as the distinction between canonical and grand-canonical ensemble properties, we plot the specific heat curves for the half-filled six-site Hubbard model with $U/t=6$ in Figure \[fig3:hubbard-cv\]. We compare the thermal AGP and projected CID results against the full CI (FCI) or exact results, and also plot the grand-canonical CISD and FCI data. The two different peaks in the FCI curves (shown in black) correspond to the spin and charge excitation energy scales. We find that the mean-field entirely fails to account for the spin excitations whereas the projected CID performs better both qualitatively and quantitatively. The difference between the exact results in the canonical and the grand-canonical ensembles also corroborates the importance of the canonical ensemble for accurate description of finite many-body systems.
Conclusion
==========
We have presented a theory to generalize correlated ground-state wave function theories, namely Hartree-Fock, perturbation theory, and CI, to study canonical ensemble thermal properties in many-electron systems. In the low temperature regime, where the canonical ensemble is most applicable, these methods perform as well as their ground-state counterparts for the benchmark problems studied. The ability to build both canonical and grand-canonical methods also signifies the robustness of thermofield theory for finite-temperature wave function methods. At zero temperature, one is generally required to go to much higher orders in CI or PT to obtain highly accurate results and better alternatives, such as the coupled cluster theory and multi-reference methods, are generally preferred. Our work is a first step towards achieving finite-temperature analogues of such sophisticated techniques.
This work was supported by the U.S. Department of Energy, Office of Basic Energy Sciences, Computational and Theoretical Chemistry Program under Award No. DE-FG02-09ER16053. G.E.S. acknowledges support as a Welch Foundation Chair (No. C-0036).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
For an interconnection network $G$, the [*$\omega$-wide diameter*]{} $d_\omega(G)$ is the least $\ell$ such that any two vertices are joined by $\omega$ internally-disjoint paths of length at most $\ell$, and the [*$(\omega-1)$-fault diameter*]{} $D_{\omega}(G)$ is the maximum diameter of a subgraph obtained by deleting fewer than $\omega$ vertices of $G$.
The enhanced hypercube $Q_{n,k}$ is a variant of the well-known hypercube. Yang, Chang, Pai, and Chan gave an upper bound for $d_{n+1}(Q_{n,k})$ and $D_{n+1}(Q_{n,k})$ and posed the problem of finding the wide diameter and fault diameter of $Q_{n,k}$. By constructing internally disjoint paths between any two vertices in the enhanced hypercube, for $n\ge3$ and $2\le k\le n$ we prove that $D_\omega(Q_{n,k})=d_\omega(Q_{n,k})=d(Q_{n,k})$ for $1 \leq \omega
< n-\lfloor\frac{k}{2}\rfloor$; $D_\omega(Q_{n,k})=d_\omega(Q_{n,k})=d(Q_{n,k})+1$ for $n-\lfloor\frac{k}{2}\rfloor \leq \omega \leq n+1$, where $d(Q_{n,k})$ is the diameter of $Q_{n,k}$. These results mean that interconnection networks modelled by enhanced hypercubes are extremely robust.
author:
- |
[Meijie Ma$^{a,b,}$[^1] Douglas B. West$^{b,c}$ Jun-Ming Xu$^{d}$ ]{}\
[$^a$ School of Management Science and Engineering]{}\
[Shandong Institute of Business and Technology, Yantai 264005, China]{}\
[$^b$Department of Mathematics]{}\
[Zhejiang Normal University, Jinhua 321004, China]{}\
[$^c$Department of Mathematics]{}\
[University of Illinois, Urbana, IL 61801, USA]{}\
[$^d$School of Mathematical Sciences]{}\
[University of Science and Technology of China, Hefei 230026, China]{}\
title: The vulnerability of the diameter of enhanced hypercubes
---
\#1\#2\#3[\#1\_[\#2]{},…,\#1\_[\#3]{}]{}
: interconnection network; enhanced hypercube; wide diameter; fault diameter.
Introduction
============
An interconnection network is conveniently represented by an undirected graph. The vertices (or edges) of the graph represent the nodes (or links) of the network. Throughout this paper, vertex and node, edge and link, graph and network are used interchangeably. Reliability and efficiency are important criteria in the design of interconnection networks. In the study of fault-tolerance and transmission delay of networks, wide diameter and fault diameter are important parameters that have been studied by many researchers. They combine connectivity with diameter to measure simultaneously the fault-tolerance and efficiency of parallel processing computer networks. These parameters were studied by several authors for some Cartesian product graphs [@ez2013; @x04; @xxh05] and for the hypercube and its variants [@c10; @de97; @kllht09; @l1993; @sl12; @sy1997; @tct2014].
Let $u$ and $v$ be two vertices in a network $G$. A [*$u,v$-path*]{} is a path with endpoints $u$ and $v$. The [*distance*]{} between $u$ and $v$, denoted by $d(u,v)$, is the minimum length (number of edges) of a $u,v$-path. The [*diameter*]{} of $G$, denoted by $d(G)$, is the maximum distance between vertices. The [*connectivity*]{} $\kappa(G)$ is the minimum number of vertices whose removal results in a disconnected or $1$-vertex network. We say that $G$ is [*$k$-connected*]{} when $0 < k \leq \kappa(G)$. By Menger’s Theorem [@Menger], in a $k$-connected network there exist $k$ internally disjoint paths joining any two vertices ([*internally disjoint*]{} means that the only shared vertices are the endpoints).
Given a $k$-connected graph $G$, fix $\omega$ with $1\le\omega\le k$. The [*$\omega$-wide diameter*]{} of $G$, denoted by $d_\omega(G)$, is the least $\ell$ such that for any $u,v\in V(G)$ there exist $\omega$ internally disjoint $u,v$-paths of length at most $\ell$. Throughout this paper, we abuse terminology by writing “disjoint paths” to mean “internally disjoint paths”. Note that $d_1(G)$ is just the diameter $d(G)$ of $G$. From the definition, $$d(G)=d_1(G)\leq d_2(G)\leq \cdots \leq d_{k-1}(G)\leq d_k(G).$$
Failures are inevitable when a network is put in use. Therefore, it is important to consider faulty networks. The [*$(\omega-1)$-fault diameter*]{} of a graph $G$, denoted by $D_\omega(G)$, is the maximum diameter among subgraphs obtained from $G$ by deleting fewer than $\omega$ vertices; it measures the worst-case effect on the diameter when vertex faults occur. Note that $D_\omega(G)$ is well-defined if and only if $G$ is $\omega$-connected, moreover, $$d(G)=D_1(G)\leq D_2(G)\leq \cdots \leq D_{k-1}(G)\leq D_k(G).$$
From the definitions, it follows that $D_\omega(G)\leq d_\omega(G)$ when $G$ is $k$-connected and $1\le\omega\le k$ [@lcch98]. Equality holds for some well-known networks [@dhl96; @lc99].
As a topology for an interconnection network of a multiprocessor system, the hypercube is a widely used and well-known model, since it possesses many attractive properties such as regularity, symmetry, logarithmic diameter, high connectivity, recursive construction, ease of bisection, and relatively low link complexity [@l92; @ss88; @x01]. We study an important variant of the hypercube $Q_n$, the enhanced hypercube $Q_{n,k}$ proposed by Tzeng and Wei [@tw1991]; its properties have been studied in [@clthh2009; @l08; @w1994; @w1999; @ycpc2014]. We give the definition and basic properties of $Q_{n,k}$ in Section 2.
It was shown by Liu [@l08] that $\kappa (Q_{n,k})=n+1$. Thus, the wide diameter $d_\omega(Q_{n,k})$ and the fault diameter $D_\omega(Q_{n,k})$ are well-defined when $\omega \leq n+1$. Yang, Chang, Pai, and Chan [@ycpc2014] gave an upper bound for $d_{n+1}(Q_{n,k})$ and $D_{n+1}(Q_{n,k})$, and they posed the problem of finding the wide diameter and fault diameter of $Q_{n,k}$. In this paper, for $n\ge3$ and $2\le k\le n$, we prove $$D_\omega(Q_{n,k})=d_\omega(Q_{n,k})=\begin{cases}
d(Q_{n,k}) & \textrm{for $1 \leq \omega < n-\lfloor\frac{k}{2}\rfloor$;}\\
d(Q_{n,k})+1 & \textrm{for $n-\lfloor\frac{k}{2}\rfloor \leq \omega \leq n+1$.}
\end{cases}$$ The special case $k=n$ (folded hypercube) was obtain earlier by Simó and Yebra [@sy1997], along with the same values for edge deletions. For enhanced hypercubes also, our arguments yield the same values for edge deletions as for vertex deletions.
Properties of $Q_{n,k}$
=======================
Let $x_n\cdots x_1$ be an $n$-bit binary string. We call the rightmost bit the first bit and the leftmost bit the $n$th bit. For simplicity we use $a^i$ to mean that the bit $a$ is repeated $i$ times; for example, $01^30^2 = 011100$. The [*Hamming distance*]{} between strings $u$ and $v$, denoted by $H(u,v)$, is the number of positions where the two strings differ.
The [*$n$-dimensional hypercube*]{} $Q_n$ is the graph whose vertices are the $n$-bit binary strings and whose edges are the pairs of vertices differing in exactly one position. An edge of $Q_n$ is a [*$j$-dimensional edge*]{} if the two endpoints differ in the $j$th position. For $1\le j\le n$, let $E_j$ denote the set of $j$-dimensional edges in $Q_n$.
As a variant of the hypercube, the [*$n$-dimensional folded hypercube*]{} $FQ_n$, proposed first by El-Amawy and Latifi [@as91], is obtained from the hypercube $Q_n$ by making each vertex $u$ adjacent to its complementary vertex, denoted $\bar u$ and obtained from $u$ by subtracting each bit from $1$. Such an edge is often called a [*complementary edge*]{}.
For $2\leq k\leq n$, the [*$n$-dimensional enhanced hypercube*]{} $Q_{n,k}$ is obtained from the hypercube $Q_n$ by adding the edge $uv$ whenever $u$ and $v$ are related by $u=x_n\cdots x_1$ and $v=x_n\cdots x_{k+1}\bar{x}_k\bar{x}_{k-1}\cdots \bar{x}_1$; that is, the first $k$ bits are complemented. Such an edge is called a [*$k$-complementary edge*]{}. For convenience, we use $E_{0}$ to denote the set of $k$-complementary edges. Thus $E(Q_{n,k})=E(Q_n)\cup E_{0}$. When $k=n$, we have $Q_{n,n}=FQ_n$; hence the enhanced hypercube is a generalization of the folded hypercube. The graphs shown in Fig. \[f1\] are $Q_{3,3}$ and $Q_{4,3}$, where the hypercube edges and $3$-complementary edges are represented by solid lines and dashed lines, respectively.
\[fig:EQ1\]
(-0.6,-.0)(3,4)
(1,1)[000]{}(.6,0.75)[000]{} (3.5,1)[010]{}(3.8,.75)[010]{} (1,3.5)[100]{}(.5,3.5)[100]{} (3.5,3.5)[110]{}(3.4,3.8)[110]{}
(2.3,1.8)[001]{}(2.5,1.6)[001]{} (4.8,1.8)[011]{}(5.25,1.98)[011]{} (2.3,4.3)[101]{}(2.3,4.6)[101]{} (4.8,4.3)[111]{}(4.8,4.6)[111]{}
(2.3,0.1)[(a) $Q_{3,3}$]{}
(-1.6,-.0)(6,4)
(1,1)[0000]{}(.6,0.75)[0000]{} (3.5,1)[0010]{}(3.8,.75)[0010]{} (1,3.5)[0100]{}(.5,3.5)[0100]{} (3.5,3.5)[0110]{}(3.36,3.75)[0110]{}
(2.3,1.8)[0001]{}(2.5,1.6)[0001]{} (4.8,1.8)[0011]{}(5.25,1.98)[0011]{} (2.3,4.3)[0101]{}(2.3,4.6)[0101]{} (4.8,4.3)[0111]{}(4.8,4.6)[0111]{}
(6,1)[1010]{}(5.6,0.75)[1010]{} (8.5,1)[1000]{}(8.8,.75)[1000]{} (6,3.5)[1110]{}(6.7,3.65)[1110]{} (8.5,3.5)[1100]{}(8.4,3.8)[1100]{}
(7.3,1.8)[1011]{}(7.5,1.6)[1011]{} (9.8,1.8)[1001]{}(10.25,1.8)[1001]{} (7.3,4.3)[1111]{}(7.3,4.6)[1111]{} (9.8,4.3)[1101]{}(9.8,4.6)[1101]{}
(4.8,0.1)[(b) $Q_{4,3}$]{}
A graph $G$ is [*vertex-transitive*]{} if for any $u,v\in V(G)$ there is some $\sigma \in Aut(G)$, the automorphism group of $G$, such that $\sigma(u)=v$; it is [*edge-transitive*]{} if for any $xy,
uv \in E(G)$ there is some $\sigma \in Aut(G)$ such that $\{\sigma(x), \sigma(y)\} = \{u, v\}$. The hypercube $Q_n$ and folded hypercube $FQ_{n}$ are vertex-transitive and edge-transitive, and the enhanced hypercube $Q_{n,k}$ is vertex-transitive but not edge-transitive when $k<n$ [@mx2010; @x01; @ycpc2014].
The [*Cartesian product*]{} $G \Box H$ of graphs $G$ and $H$ is the graph with vertex set $V(G)\times V(H)$, in which vertices $(u,v)$ and $(u',v')$ are adjacent whenever $uu'\in E(G)$ and $v=v'$, or $u=u'$ and $vv'\in E(H)$. By the definition of $Q_{n,k}$, we have $Q_{n,k}=Q_{n-k}\Box FQ_{k}$. Although $Q_{n,k}$ is not edge-transitive when $2\leq k< n$, we have the following property.
\[prop1\] Permuting the first $k$ positions and/or permuting the last $n-k$ positions in the names of the vertices of $Q_{n,k}$ does not change the graph.
Exchanging the $i$th bit and $j$th bit among the first $k$ or among the last $n-k$ preserves the adjacency relation, since the number of coordinates in which two vertices differ is not changed by exchanging such coordinates. An arbitrary permutation is obtained by a succession of such exchanges.
\[prop2\] Given $u,v\in V(Q_{n,k})$, let $r$ and $s$ be the numbers of positions in which $u$ and $v$ differ among the first $k$ and last $n-k$ positions, respectively. The distance between $u$ and $v$ is computed by $d(u,v)=s+\min\{r,k-r+1\}$.
The distance is the minimum number of steps to change $u$ into $v$. All steps change one bit, except that a $k$-complementary edge changes the first $k$ bits. If no $k$-complementary edge is used, then the number of steps is at least the Hamming distance, and this suffices. For this reason, a shortest path uses at most one $k$-complementary edge. If a $k$-complementary edge is used, then the $k-r$ positions in which $u$ and $v$ agree among the first $k$ must be changed individually.
Proposition \[prop2\] immediately yields $d(Q_{n,k})=(n-k)+\lceil\frac {k}{2}\rceil$, which equals $n-\lfloor \frac {k}{2}\rfloor$. This was observed by Tzeng and Wei [@tw1991], along with an algorithm for finding shortest paths joining vertices. Note that if $u$ and $v$ differ in more than $\lceil \frac {k}{2}\rceil$ positions among the first $k$, then every shortest $u,v$-path contains exactly one $k$-complementary edge.
Construction of Paths
=====================
Many properties of interconnection networks were investigated by different construction methods of paths [@chf14; @l15]. In this section we will prove our main results by constructing disjoint paths of bounded length joining any two vertices in $Q_{n,k}$.
Let $P$ be a path $u_0\rightarrow u_1\rightarrow \cdots \rightarrow
u_{\ell-1}\rightarrow u_{\ell}$ from the vertex $u_0$ to a vertex $u_\ell$ in $Q_{n,k}$. The path $P$ traverses the edges $u_0u_1,u_1u_2,\ldots,u_{\ell-1}u_\ell$. Since the edge $u_{j-1}u_j$ is in $E_{d_j}$ for some $d_j\in\{0,1,\ldots,n\}$, and every vertex is incident with exactly one edge in $E_{d_j}$, we can represent the path $P$ from $u_0$ by the list $(d_1,\ldots, d_\ell)$, where $d_j$ indicates the type of the edge joining $u_{j-1}$ and $u_j$. For example, in $Q_{5,3}$ the path originating from $00000$ determined by $(2,0,5)$ is $00000\xrightarrow{2} 000{\bf 1}0 \xrightarrow{0}
00{\bf 101} \xrightarrow{5} {\bf 1}0101$. Note that the length of a path determined by a list $I$ is the number of elements in $I$.
We use the following two lemmas to construct disjoint paths. A [*proper segment*]{} of a list is a string of consecutive elements that is not the full list.
\[lem4\] No two cyclic permutations of a list of distinct elements have a common proper initial segment.
Let $I_1=(d_1,d_2, \ldots, d_\ell)$ and $I_j=(d_j,\ldots,d_\ell,
d_1,\ldots, d_{j-1})$ for $2\leq j \leq \ell$.
Consider $j$ and $j'$ with $1\leq j'< j \leq \ell$. A proper initial segment of $I_j$ for $j\geq 2$ contains $d_j$ and not $d_{j-1}$, but this is not true for any initial segment of $I_{j'}$, since when $d_j$ is not the first element, $d_{j-1}$ also occurs in any initial segment containing $d_j$. Therefore, $I_{j}$ and $I_{j'}$ have no common proper initial segment.
\[distinct\] Let $S$ be a set of $\ell$ distinct elements of $\{0,1,\dots,n\}$. Let $\VEC I1m$ be orderings of $S$ such that no two have a common proper initial segment. If $\{0,1,\ldots,k\}$ are not all in $S$, then for any vertex $u$ in $Q_{n,k}$, the lists $I_1, \ldots, I_m$ determine disjoint paths to a single vertex.
Let $S=\{d_1,\ldots, d_\ell\}$ and $[k]=\{1,\ldots,k\}$. If $0\notin S$, then each path reaches the vertex $v$ that differs from $u$ in the positions of $S$. If $0\in S$, then each path reaches the vertex $v$ that differs from $u$ in the positions of $([k]-S)\cup (S-\{0\}-[k])$.
If two subsets of $S$ both contain $0$ or both omit $0$, then they produce paths to the same vertex only if they are the same set. If $0\in T\subseteq S$ and $0\notin T'\subseteq S$, then $T$ and $T'$ produce paths to the same vertex only if they agree outside $[k]$ and intersect $[k]$ in complementary subsets. By the hypothesis that $0,1,\dots,k$ are not all present in $S$, this cannot occur.
Therefore, the paths from $u$ determined by $I_j$ and $I_{j'}$ have a common internal vertex if and only if they have a common proper initial segment.
\[lem5\] Let $I=(d_1,\ldots,d_\ell)$ with all $d_i$ distinct and in $\{0,1,\ldots,n\}$. If $\{0,1,\ldots,k\}$ are not all in $I$, then for any vertex $u$ in $Q_{n,k}$, the $\ell$ cyclic permutations of $I$ determine disjoint paths to a single vertex.
The conclusion follows immediately from Lemmas \[lem4\] and \[distinct\].
When $G$ is the complete graph $K_n$ with $n\ge3$, we have $D_\omega(G)=1$ but $d_\omega(G)=2$ for $2\le\omega\le n-1$. Since $Q_{2,2}=K_4$, we consider $Q_{n,k}$ with $n\ge 3$ and $2\leq k\leq n$.
\[lem1\] For any two distinct vertices $u$ and $v$ in $Q_{n,k}$ with $n\ge3$ and $2\leq k\leq n$, there exist $n+1$ disjoint $u,v$-paths of length at most $d(Q_{n,k})+1$, such that at least $n-\lfloor\frac{k}{2}\rfloor-1$ of the paths have length at most $d(Q_{n,k})$.
The vertex transitivity of $Q_{n,k}$ and Proposition \[prop1\] allow us to assume $u=0^{n}$ and $v=0^{n-k-j}1^{j}0^{k-i}1^{i}$, where $0\leq i \leq k$ and $0\leq j \leq n-k$. We will construct the desired paths. Let $r=\min\{i+j,k-i+j+1\}$. Recall that $d(Q_{n,k})=n-\lfloor\frac k2\rfloor$. We consider two cases according to the relationship between $r$ and $d(Q_{n,k})$.
[*Case 1: $r< d(Q_{n,k})$.*]{} We first specify a list $I$ of length $r$, in two cases (Table\[tab1\]). Note that $r=i+j$ when $i\le
k-i$ and $r=k-i+j+1$ when $i>k-i$.
Case Condition $I$
------ ------------ ---------------------------------
A $i\le k-i$ $(1,\dots,i,k+1,\dots,k+j)$
B $i> k-i$ $(0,i+1,\dots,k,k+1,\dots,k+j)$
: Two cases for $r< d(Q_{n,k})$.[]{data-label="tab1"}
Not all of $0,1,\dots,k$ appear in $I$, since having $0$ requires $i>k/2$ (Case B), and then also having $1$ requires $i=0$, a contradiction. Hence Lemma \[lem5\] applies, so the $r$ cyclic permutations of $I$ determine $r$ disjoint $u,v$-paths of length $r$.
The remaining $n-r+1$ paths, with length $r+2$, are specified by adding one of $\{0,1,\dots,n\}-I$ at both the beginning and the end of $I$. Let $I'$ be the list obtained by adding $h$, and let $P$ be the path from $u$ determined by $I'$.
If $h>k$ or if $h\in[k]$ and $0\notin I$, then $u$ and $v$ agree in position $h$, and $P$ is the only path in the constructed set containing vertices that differ from them in position $h$. Furthermore, all internal vertices of $P$ differ from $u$ and $v$ in position $h$.
If $h=0$, then $i\le k-i$ (Case A). All internal vertices of $P$ differ from $u$ and $v$ in positions $k$ and $k-1$, and no other path has any such vertices.
If $h\in[k]$ and $0\in I$, then $1\le h\le i$ (Case B). The first vertex of $P$ after $u$ differs from $u$ only in position $h$. The next vertex disagrees with $u$ on all of positions $1,\dots,i$ except $h$, and this remains true of all other internal vertices of $P$, because $I$ contains no element of $\{1,\dots,i\}$. All the other paths in the construction have no vertices satisfying either of these conditions.
Since $r<d(Q_{n,k})$, all the paths have length at most $d(Q_{n,k})+1$; in fact, all have length at most $d(Q_{n,k})$ unless $r=d(Q_{n,k})-1$. In this case there are $r$ paths of length $r$, which suffices since $r=d(Q_{n,k})-1=n-\fktwo-1$.
[*Case 2: $r\ge d(Q_{n,k})$.*]{} Let $s=d(Q_{n,k})$. Since $r=\min\{i+j,k-i+j+1\}$, we have $i+j\ge s$ and $k-i+j+1\ge s$, so $k+2j+1\ge 2s$. If $j\le n-k-1$, then $2n-k-1\ge 2s=2n-2\fktwo$, which is impossible. Hence $j=n-k$. With $j=n-k$, we have $i\ge
\cktwo$ and $k+1-i\ge\cktwo$, so $\cktwo\le i\le \fktwo+1$. When $i=\cktwo$, we have $r=i+j$; when $i=\fktwo+1$, we have $r=k-i+j+1$. Both cases apply when $k$ is odd. In either case, $r=n-\fktwo=d(Q_{n,k})$, and we define three lists(Table \[tab2\]).
Case $i$ $I$ $J$ $I'$
------ ------------ ------------------------------- ------------------ -------------------
A $\cktwo$ $(1,\dots,i,k+1,\dots,n)$ $(k+1,\ldots,n)$ $(0,i+1,\dots,k)$
B $\fktwo+1$ $(0,i+1,\dots,k,k+1,\dots,n)$ $(k+1,\ldots,n)$ $(1,\dots,i)$
: Two cases for $j=n-k$ and $\cktwo\le i\le \fktwo+1$.[]{data-label="tab2"}
Since $j=n-k$, in each case $I$ has length $r$. By Lemma \[lem5\], the cyclic permutations of $I$ yield $r$ disjoint $u,v$-paths of length $r$. Since $r=n-\fktwo=d(Q_{n,k})$, this yields enough paths of length at most $d(Q_{n,k})$. Let $T=\{i+1,\dots,k\}$ in Case A, $T=\{1,\dots,i\}$ in Case B. Every vertex in each of these paths is constant in the positions of $T$ (all-$0$ or all-$1$), and in fact all-$0$ in Case A.
Since $r=n-\fktwo$, we only need to find $\fktwo+1$ more paths of length at most $d(Q_{n,k})+1$. Note that $I'$ has length $\fktwo+1$. Form $\fktwo+1$ lists by inserting $J$ after the first element of each cyclic permutation of $I'$. The first and last lists are $(0,J,i+1,\dots,k)$ and $(k,J,0,i+1,\dots,k-1)$ in Case A, $(1,J,2,\dots,i)$ and $(i,J,1,\dots,i-1)$ in Case B. Each of these lists has length $n-\cktwo+1$, which is at most $d(Q_{n,k})+1$.
Each of these lists is an ordering of a single set of elements. By an argument like that of Lemma \[lem4\], they have no common proper initial segments. Since they also do not contain all of $\{0,1,\dots,k\}$, by Lemma \[distinct\] these paths are disjoint.
In Case A, each internal vertex on each of these paths is not all $0$ in the positions of $T$. In Case B, each internal vertex on each of these paths has between $1$ and $|T|-1$ nonzero positions in $T$. Hence these paths are disjoint from the earlier paths.
Consequences
============
From Theorem \[lem1\] and the definition of wide diameter, we immediately obtain an upper bound on $d_{\omega}(Q_{n,k})$.
\[cor1\] If $n\ge3$ and $2\leq k \leq n$, then $$d_{\omega}(Q_{n,k}) \le \left\{\begin{array}{ll}
d(Q_{n,k})\ & {\rm for}\ 1 \leq \omega < n-\fktwo,\\
d(Q_{n,k})+1\ & {\rm for}\ n-\lfloor\frac{k}{2}\rfloor \leq \omega \leq n+1.
\end{array}\right.$$
When $\omega<n-\fktwo$, Theorem \[lem1\] provides at least $\omega$ disjoint paths with length at most $d(Q_{n,k})$ joining any two vertices in $Q_{n,k}$. When $\omega\le n+1$, it provides at least $\omega$ such paths with length at most $d(Q_{n,k})+1$.
We next give a lower bound on the fault diameter $D_\omega(Q_{n,k})$.
\[lem2\] Fix $n\ge3$. If $2\leq k \leq n$ and $n-\lfloor\frac{k}{2}\rfloor \leq \omega \leq n+1$, then $$D_{\omega}(Q_{n,k}) \geq d(Q_{n,k})+1.$$
Since $D_\omega(G)$ is nondecreasing in $\omega$, proving $D_{n-\lfloor\frac{k}{2}\rfloor}(Q_{n,k}) \geq d(Q_{n,k})+1$ is sufficient. Let $u=0^{n}$ and $v=1^{n-k}0^{k-i}1^{i}$, where $i=\cktwo-1$. Note that $v$ has $1$s in $n-\fktwo-1$ positions. Let $W$ be the set of neighbors of $u$ whose single $1$ occurs in a position where $v$ has a $1$, so $|W|=n-\fktwo-1$. On any $u,v$-path in $Q_{n,k}-W$, the neighbor of $u$ has a single $1$ in a position among $i+1,\dots,k$ or is $0^{n-k}1^k$.
By Proposition \[prop2\], the distance between $v$ and a neighbor $u'$ of $u$ not in $W$ is $n-k+\min\{i+1,k-i\}$. Since $i=\cktwo-1$, the distance is $n-\fktwo$, which equals $d(Q_{n,k})$. Hence every $u,v$-path in $Q_{n,k}-W$ has length at least $d(Q_{n,k})+1$.
\[T3\] For $3\leq n$ and $2\leq k \leq n$, $$D_\omega(Q_{n,k})=d_\omega(Q_{n,k})=\begin{cases}
d(Q_{n,k}) & \textrm{for $1 \leq \omega < n-\lfloor\frac{k}{2}\rfloor$;} \\
d(Q_{n,k})+1 & \textrm{for $n-\lfloor\frac{k}{2}\rfloor \leq \omega \leq n+1$.}
\end{cases}$$
Since $d(Q_{n,k}) \leq D_\omega(Q_{n,k})\leq d_\omega(Q_{n,k})$ for $1\leq \omega \leq n+1$, Corollary \[cor1\] yields $D_\omega(Q_{n,k})=d_\omega(Q_{n,k})=d(Q_{n,k})$ for $1 \leq \omega <n-\lfloor\frac{k}{2}\rfloor$.
For $n-\lfloor\frac{k}{2}\rfloor \leq \omega \leq n+1$, Corollary \[cor1\] and Lemma \[lem2\] yield $D_\omega(Q_{n,k})=d_\omega(Q_{n,k})=d(Q_{n,k})+1$.
Theorem \[T3\] shows that the fault diameter $D_\omega(Q_{n,k})$ equals the wide diameter $d_\omega(Q_{n,k})$ for the enhanced hypercubes $Q_{n,k}$. More importantly, they equal the traditional diameter when $1\leq \omega < n-\lfloor\frac{k}{2}\rfloor$, and they exceed it only by $1$ when $n-\lfloor\frac{k}{2}\rfloor \leq \omega \leq n+1$. Thus, the resilience of enhanced hypercubes is similar to that of hypercubes, which increases the appeal of enhanced hypercubes.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported partially by NSFC (Nos. 11101378, 11571044). Research of D.B. West was supported by 1000 Talent Plan, State Administration of Foreign Experts Affairs, China.
[s2]{} A. El-Amawy and S. Latifi, Properties and performance of folded hypercubes, IEEE Trans. Parallel Distrib. Syst. 2 (1991) 31–42.
C.-H. Chang, C.-K. Lin, J.J.M. Tan, H.-M. Huang, L.-H. Hsu, The super spanning connectivity and super spanning laceability of the enhanced hypercubes, J. Supercomput. 48 (2009) 66–87.
X.-B. Chen, Edge-fault-tolerant diameter and bipanconnectivity of hypercubes, Inform. Proc. Lett. 110 (2010) 1088–1092.
D. Cheng, R.-X. Hao, Y.-Q. Feng, Two node-disjoint paths in balanced hypercubes, Appl. Math. Comput. 242 (2014) 127–142.
K. Day, A. E. Al-Ayyoub, Fault diameter of $k$-ary $n$-cube networks, IEEE Trans. Parallel Distrib. Syst. 8 (1997) 903–907.
D.Z. Du, D.F. Hsu, Y.D. Lyuu, On the diameter vulnerability of Kautz digraphs, Discrete Math. 151 (1996) 81–85.
R. Erveš, J. Žerovnik, Mixed fault diameter of Cartesian graph bundles, Discrete Appl. Math. 161 (2013) 1726–1733.
T.-L. Kung, C.-K. Lin, T. Liang, L.-Y. Hsu, J.J.M. Tan, Fault diameter of hypercubes with hybrid node and link faults, J. Inter. Networks 10 (3) (2009) 233–242.
P.-L. Lai, Paths and cycles identifying vertices in twisted cubes, Appl. Math. Comput. 259 (2015) 620–627.
S. Latifi, Combinatorial analysis of the fault-diameter of the $n$-cube, IEEE Trans. Comput. 42 (1993) 27–33.
F.T. Leighton, Introduction to Parallel Algorithms and Architectures: Arrays, Trees, and Hypercubes, Morgan Kaufmann, 1992.
S.C. Liaw, G.J. Chang, Wide diameters of butterfly networks, Taiwan. J. Math. 3 (1) (1999) 83–88.
S.C. Liaw, G.J. Chang, F. Cao, D.F. Hsu, Fault-tolerant routing in circulant networks and cycle prefix networks, Ann. Combin. 2 (1998) 165–172.
H. Liu, Properties of enhanced hypercube networks, J. Syst. Sci. Inform. 6 (2008) 251–216.
M. Ma, J.-M. Xu, Algebraic properties and panconnectivity of folded hypercubes, Ars Combin. 95 (2010) 179–186.
K. Menger, Zur allgemeinen Kurventheorie, Fund. Math. 10 (1927) 95–115.
Y. Saad and M.H. Schultz, Topological properties of hypercubes, IEEE Trans. Comput. 37 (1988) 867–872.
T. Shi, M. Lu, Fault-tolerant diameter for three family interconnection networks, J. Combin. Optim. 23 (2012) 471–482.
E. Sinó, J.L.A. Yebra, The vulnerability of the diameter of folded $n$-cubes, Discrete Math. 174 (1997) 317–322.
T.-H. Tsai, Y.-C. Chen, J.J.M. Tan, Topological properties on the wide and fault diameters of exchanged hypercubes, IEEE Trans. Parallel Distrib. Syst. 25 (2014) 3317–3327.
N.-F. Tzeng, S. Wei, Enhanced hypercubes, IEEE Trans. Comput. 40 (1991) 284–294.
D. Wang, Diagnosability of enhanced hypercubes, IEEE Trans. Comput. 43 (1994) 1054–1061.
D. Wang, Diagnosability of hypercubes and enhanced hypercubes under the comparison diagnosis model, IEEE Trans. Comput. 48 (1999) 1369–1374.
J.-M. Xu, Topological Structure and Analysis of Interconnection Networks, Kluwer Academic Publishers, Dordrecht/Boston/London, 2001.
J.-M. Xu, Wide-diameter of Cartesian product graphs and digraphs, J. Combin. Optim. 8 (2) (2004), 171–181.
M. Xu, J.-M. Xu, X.-M. Hou, Fault diameter of Cartesian product graphs, Inform. Proc. Lett. 93 (2005) 245–248.
J.-S. Yang, J.-M. Chang, K.-J. Pai, H.-C. Chan, Parallel construction of independent spanning trees on enhanced hypercubes, IEEE Trans. Parallel Distrib. Syst. 26 (2015) 3090–3098.
[^1]: Corresponding author: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Any model of incomplete information situations has to consider the players’ hierarchies of beliefs, which can make the modeling very cumbersome. Harsányi [@Harsanyi1967-68] suggested that the hierarchies of beliefs can be replaced by types, i.e., a type space can substitute for the hierarchies of beliefs (henceforth Harsányi program). In the purely measurable framework Heifetz and Samet [@HS3] formalized the concept of type space, and proved that there is universal type space, i.e., the most general type space exists. Later Meier [@Mei1] showed that the universal type space is complete, in other words, the universal type space is a consistent object. After these results, only one step is missing to prove that the Harsányi program works, that every hierarchy of beliefs is in the complete universal type space, put it differently, every hierarchy of beliefs can be replaced by type. In this paper we also work in the purely measurable framework, and show that the types can substitute for all hierarchies of beliefs, i.e., the Harsányi program works.'
author:
- |
Miklós Pintér\
Corvinus University of Budapest[^1]
title: 'Every hierarchy of beliefs is type[^2]'
---
Introduction
============
It is recommended that the models of incomplete information situations to be able to consider the players’ hierarchies of beliefs, e.g. player $1$’s beliefs about the parameters of the game, player $1$’s beliefs about player $2$’s beliefs about the parameters of the game, player $1$’s beliefs about player $2$’s beliefs about player $1$’s beliefs about the parameters of the game, and so on. However the explicit use of hierarchies of beliefs[^3] makes the analysis very cumbersome, hence it is desirable to evade that they appear explicitly in the models.
In order to make the models of incomplete information situations more handy, Harsányi [@Harsanyi1967-68] suggested that the hierarchies of beliefs could be replaced by types. He wrote[^4] “It seems to me that the basic reason why the theory of games with incomplete information has made so little progress so far lies in the fact that these games give rise, or at least appear to rise, to infinite regress in reciprocal expectations on the part of the players. $\ldots$ The purpose of this paper is to suggest an alternative approach to the analysis of games with incomplete information. $\ldots$ As we have seen, if we use the Bayesian approach, then the sequential-expectations model for any given $I$-game $G$ will have to be analyzed in terms of infinite sequences of higher and higher-order subjective probability distributions, i.e., subjective probability distributions over subjective probability distributions. In contrast, under own model, it will be possible to analyze any given $I$-game $G$ in terms of one *unique* probability distribution $R^\ast$ (as well as certain conditional probability distributions derived from $R^\ast$). $\ldots$ Instead of assuming that certain important *attributes* of the players are determined by some hypothetical random events at the beginning of the game, we may rather assume that the players *themselves* are drawn at random from a certain hypothetical population containing the mixture of different “types", characterized by different attribute vectors (i.e., by different combinations of relevant attributes). $\ldots$ Our analysis of $I$-games will be based on the assumption that, in dealing with incomplete information, every player $i$ will use Bayesian approach. That is, he will assign a *subjective* joint probability $P_i$ to all variables unknown to him $-$ or at least to all unknown *independent* variables, i.e., to all variables no depending on the players’ own strategy choices.”
In other words, Harsányi’s main concept was that the types can substitute for the hierarchies of beliefs, and all types can be collected into an object on which the probability measures are for the players’ (subjective) beliefs. Henceforth, we call this method of modeling Harsányi program.
However, at least two questions come up in connection with the Harsányi program: (1) is the concept of type itself appropriate for the proposes under consideration? (2) can every hierarchy of beliefs be a type?
Question (1) consists of two subquestions. First, can all types be collected into one object? The concept of universal type space formalizes this requirement: the universal type space in a certain category of type spaces is a type space (a) which is in the given category, and (b) into which, every type space of the given category can be mapped in a unique way. In other words, the universal type space is the most general type space, it contains all type spaces (all types). In the purely measurable framework Heifetz and Samet [@HS3] introduced the idea of (universal) type space, and proved that the universal type space exists.
Second, can every probability measure on the object of the collected types (type space) be a (subjective) belief? Brandenburger [@Brand] introduced the notion of complete type space: a type space is complete, if the type functions in it are surjective (onto). Put it differently, a type space is complete, if all probability measures on the object consisting of the types of the model are corresponded to types. Quite recently Meier [@Mei1] showed that the purely measurable universal type space is complete. Summing up the above discussion, the answer for question (1) is affirmative, i.e., in the purely measurable framework the complete universal type space exists.
Question (2) is on that whether or not the universal type space contains every hierarchy of beliefs. Mathematically the problem is the following: every hierarchy of beliefs defines an inverse system of measure spaces, and the question is that: do these inverse systems of measure spaces have inverse limits? Kolmogorov Extension Theorem is on this problem, however it calls for topological concepts, e.g. for inner compact regular probability measures. Therefore up to now, all papers on this problem (e.g. Mertens and Zamir [@MZ1], Brandenburger and Dekel [@BD2], Heifetz [@Ha2], Mertens et al. [@MSZ1], Pintér [@Pinter2005] among others) used topological type spaces instead of purely measurable ones. Although these papers give positive answer for question (2) (i.e. their type spaces contain all “considered” hierarchies of beliefs), very recently Pintér [@pmp2] showed that there is no universal topological type space (there is no such a topological type space that contains every topological type space), therefore the answer for question (1) is negative in this case i.e., in the topological framework the Harsányi program breaks down.
In the above mentioned papers the authors answer question (2) (affirmatively) by constructing an object consisting all considered hierarchies of beliefs, called beliefs space, and show that the constructed beliefs space defines (is equivalent to) a topological type space.
In this paper we work with the category of type spaces introduced by Heifetz and Samet, i.e., in the purely measurable framework. It is our main result that (in the purely measurable framework) every hierarchy of beliefs is type, put it differently, the Harsányi program works. The strategy of the proof is the same as in the above papers, i.e., we construct such an object that contains every hierarchy of beliefs (see definition \[velemenyter\].) and generates a type space. More exactly, it is showed that the (purely measurable) beliefs space is equivalent to the complete universal type space.
As we have already mentioned the above strategy depends on the Kolmogorov Extension Theorem. Since we work in the purely measurable framework, therefore we avoid the direct use of topological concepts and use a non-topological variant of the Kolmogorov Extension Theorem. Mathematically, we use a new result of Pintér [@pmp3] to show that the inverse systems of measure spaces under consideration have inverse limits.
One important remark, our result does not contradict with Heifetz and Samet’s [@HS4] counterexample, since their hierarchy of beliefs is not in the purely measurable beliefs space (for the the details see section \[comparison\].).
The paper is organized as follows: in the first section we introduce an example illustrating our main result. Section \[beliefspace\]. presents the technical setup and some basic results of the field. Our main result (theorem \[fotetel\].) comes up in section \[bs\]. Section \[beliefspace\]. is on the proof of theorem \[fotetel\]. Section \[comparison\]. is for a detailed discussion of the connection between our result and two other papers Heifetz and Samet [@HS4], and Pintér [@pmp2]. The last section briefly concludes.
An example {#example}
==========
In this section we introduce an example for illustrating the importance of the hierarchies of beliefs.
Consider a $2 \times 2$ game in strategic form, two players: Player $1$ and Player $2$, both have two actions $U$, $D$ and $L$, $R$ respectively, there are two states of the nature in the model: $s_1$ and $s_2$ ($S = \{s_1,s_2\}$) with the payoffs in tables \[abra1\] and \[abra2\].
$$\begin{array}{cc|cc}
& & \multicolumn{2}{c}{\text{Player 2}} \\
& & L & R \\ \hline
\text{Player 1} &
\begin{tabular}{c}
\textit{U} \\
\textit{D}
\end{tabular}
&
\begin{tabular}{c}
(2,3) \\
(3,4)
\end{tabular}
&
\begin{tabular}{c}
(4,2) \\
(5,5)
\end{tabular}
\end{array}$$
$$\begin{array}{cc|cc}
& & \multicolumn{2}{c}{\text{Player 2}} \\
& & L & R \\ \hline
\text{Player 1} &
\begin{tabular}{c}
\textit{U} \\
\textit{D}
\end{tabular}
&
\begin{tabular}{c}
(4,5) \\
(5,3)
\end{tabular}
&
\begin{tabular}{c}
(3,4) \\
(2,2)
\end{tabular}
\end{array}$$
In this example that an action is rationalizable for a certain player means that there is such a state of the world that the common belief of rationality implies that the player under consideration plays the given action. Although the usual (see e.g. Osborne and Rubinstein’s textbook [@OR]) and the above rationalizability concepts differ, both catch the same intuition of rationalizability, and the only reason for introducing a new terminology is that this “different” concept of rationalizability reflects the main massage of the example more and makes us possible to keep the discussion quite simple.
It is easy to verify that in this example for both players both actions are rationalizable. If player $1$ believes with probability $1$ that the state of the nature is $s_1$ then her rationality implies that she plays action $D$. Furthermore, if player $1$ believes with probability $1$ that the state of the nature is $s_2$ and that player $2$ believes with probability $1$ that the state of the nature is $s_1$ and that she (player $1$) also believes with probability $1$ that the state of the nature is $s_1$, then that she believes that player $2$ is rational, and that player $2$ believes that she is rational imply that she believes with probability $1$ that player $2$ believes with probability $1$ that she plays action $D$, hence she believes with probability $1$ that player $2$ plays action $R$, therefore she plays action $U$.
If player $2$ believes with probability $1$ that the state of the nature is $s_2$ then her rationality implies that she plays action $L$. If it is mutually believed with probability $1$ that the state of the nature is $s_1$ then that player $2$ believes that player $1$ is rational implies that she believes with probability $1$ that player $1$ plays action $D$, hence she plays action $R$.
Summing up the above discussion, an adequate type space must reflect the fact that for both players both actions are rationalizable. If this does not happen then the given type space is inappropriate for modeling the incomplete information situation under consideration.
In the following we look into the question of what kind of type spaces can be appropriate for the modeling proposes under discussion.
Case 1: The type space is neither complete nor universal.
Consider the type space (see definition \[tipuster\].)
$$\label{tipuster1}
(S,(\Omega,\mathcal{M}_i)_{i = 0,1,2},g,\{f_i\}_{i=1,2}) \ ,$$
where $\Omega \circeq S \times \{ t_1 \} \times \{ t_2 \}$, $f_1 \circeq \delta_{(s_2,t_2)}$, $f_2 \circeq \delta_{(s_2,t_1)}$ (the Dirac measures concentrated at point $(s_2,)$ and $(s_2,t_1)$ respectively), $g : \Omega \rightarrow S$ is the coordinate projection, $\mathcal{M}_i \circeq \mathcal{P} (\Omega)$ (the class of all subsets of $\Omega$) $i = 0,1,2$.
It is easy to verify that in model at every state of the world both players believe that they play the game at state of the nature $s_2$, hence e.g. for player $2$ action $R$ is not rationalizable.
Case 2: The type space is complete but not universal.
Consider the type space
$$\label{tipuster2}
(\{s_2\},(\Omega,\mathcal{M}_i)_{i = 0,1,2},g,\{f_i\}_{i=1,2}) \ ,$$
where $\Omega \circeq \{ s_2 \} \times \{ t_1 \} \times \{ t_2 \}$, $g : \Omega \rightarrow S$ is $g \circeq s_2$ (the natural embedding of $\Omega$ into $S$), $f_1 \circeq \delta_{(s_2,t_2)}$, $f_2 \circeq \delta_{(s_2,t_1)}$, and $\mathcal{M}_i \circeq \{\emptyset,\Omega\}$, $i=0,1,2$.
It is easy to verify that this type space is complete (see definition \[teljesseg\].), and at every state of the world (there is only one in this model) it is commonly believed (with probability $1$) that the state of the nature is $s_2$, hence e.g. for player $1$ action $U$ is not rationalizable.
Case 3: Complete universal type space.
From Heifetz and Samet [@HS3], and Meier [@Mei1]: the complete universal type space (see definitions \[tt3\]. and \[teljesseg\].) exists. Therefore in this example it contains the type space
$$\label{ts1}
(S,(\Omega,\mathcal{M}_i)_{i = 0,1,2},g,\{f_i\}_{i=1,2}) \ ,$$
where $T_1 \circeq T_2 \circeq [0,1]$, $\Omega \circeq S \times T_1 \times T_2$, $g :\Omega \rightarrow S$, $i=1,2$: $pr_i : \Omega \to T_i$ are coordinate projections, $\forall x \in [0,1]$: $\mu (x) \in \Delta (S)$ is such that $\mu (\{s_1\}) = x$, $\mathcal{M}_0 \circeq \sigma(\mathcal{P} (S) \otimes \{T_1\} \otimes \{T_2\})$ ($\sigma$-field generated by the sets $\mathcal{P} (S) \otimes \{T_1\} \otimes \{T_2\}$), $\mathcal{M}_1 \circeq \sigma(\{S\} \otimes B(T_1) \otimes \{T_2\})$ ($B(T_1)$ is for the Borel $\sigma$-field of $T_1$), $\mathcal{M}_2 \circeq \sigma(\{S\} \otimes \{T_1\} \otimes B(T_2))$, and last $f_1 (\omega) \circeq \mu(pr_1 (\omega)) \times \delta_{1}$ (the product measure of the measures $\mu(pr_1 (\omega))$ and $\delta_{1}$), $f_2 (\omega) \circeq \mu(pr_2 (\omega)) \times \delta_{1}$.
In this model at every state of the world every player believes that the other player believes that the state of the nature is $s_1$ and that the given player believes that the state of the nature is $s_1$, hence, as we have already discussed, for both players both actions are rationalizable.
Therefore, in this example the complete universal type space reflects the main intuitions of the modeled situation.
Case 4: Complete universal type space that does not contain every hierarchy of beliefs.
Only one question has remained, whether or not the complete universal type space contains every hierarchy of beliefs. Although in this example the universality implies that the model reflects the main intuitions of the situation we considered, in general[^5], if the complete universal type space misses some hierarchies of beliefs then it is possible to construct a game in which the missing hierarchy(ies) of beliefs is(are) important, i.e., there is a game for which the complete universal type space is not appropriate (as in e.g. Case 1).
Therefore if the above mentioned failure happens then the Harsányi program breaks down, since the complete universal type space cannot reflect all important details of incomplete information situations.
The main result of this paper (theorem \[fotetel\].) argues that in the purely measurable framework Case 4 cannot happen, i.e., the complete universal type space contains every hierarchy of beliefs, in other words, the Harsányi program works.
Our result heavily depends on that we work in the purely measurable framework, i.e. with the measurable structure introduced in definition \[merhetoseg\]. However, doing so is not restrictive at all, in contrary the richer structures bring only irrelevant details into the model, hence they are useless and more, as Pintér’s result [@pmp2] shows, they can be harmful.
Type space {#typespace}
==========
First some notations. Let $N$ be the set of the players, w.l.o.g. we can assume that $0 \notin N$, and let $N_0 \circeq N \cup \{0\}$, where $0$ is for the nature as an extra player.
Let $A$ be arbitrary set, then $\# A$ is for the cardinality of set $A$. For any $A \subseteq \mathcal{P} (X)$: $\sigma (A)$ is the coarsest $\sigma$-field which contains $A$. Let $(X,\mathcal{M})$ and $(Y,\mathcal{N})$ be arbitrary measurable spaces. Then $(X \times Y,\mathcal{M} \otimes \mathcal{N})$ or briefly $X \otimes Y$ is the measurable space on the set $X \times Y$ equipped by the $\sigma$-field $\sigma (\{A \times B \mid A \in \mathcal{M}, \ B \in \mathcal{N}\})$.
The measurable spaces $(X,\mathcal{M})$ and $(Y,\mathcal{N})$ are measurable isomorphic if there is such a bijection $f$ between them that both $f$ and $f^{-1}$ are measurable.
Let the measurable space $(X,\mathcal{M})$ and $x \in X$ be arbitrarily fixed. Then $\delta_x$ is for the Dirac measure on $(X,\mathcal{M})$ concentrated at point $x$.
In the following, practically, we use the terminologies that were introduced by Heifetz and Samet [@HS3].
\[merhetoseg\] Let $(X,\mathcal{M})$ be arbitrarily fixed measurable space, and denote $\Delta(X,\mathcal{M})$ the set of the probability measures on it. Then the $\sigma$-field $\mathcal{A}^\ast$ on $\Delta(X,\mathcal{M})$ is defined as follows:
$$\mathcal{A}^\ast \circeq \sigma(\{\{ \mu \in \Delta(X,\mathcal{M}) \mid \mu (A) \geq p\}, \ A \in \mathcal{M}, \ p \in [0,1]\}) \ .$$
In other words, $\mathcal{A}^\ast$ is the smallest $\sigma$-field among the $\sigma$-fields which contain the sets $\{ \mu \in \Delta(X,\mathcal{M}) \mid \mu (A) \geq p\}$, where $A \in \mathcal{M}$ and $p \in [0,1]$ are arbitrarily chosen.
In incomplete information situations it is necessary to consider the events like player $i$ believes with probability at least $p$ that an event occurs (beliefs operator see e.g. Aumann [@Aumann1999b]). For this reason $\{ \mu \in \Delta(X,\mathcal{M}) \mid \mu (A) \geq p\}$ must be an event (measurable set). To keep the class of events as small (coarse) as possible, we use the $\mathcal{A}^\ast$ $\sigma$-field[^6].
Notice that $\mathcal{A}^\ast$ is not a fixed $\sigma$-field, it depends on the measurable space on which the probability measures are defined. Therefore $\mathcal{A}^\ast$ is similar to the $weak^\ast$ topology, which depends on the topology of the base (primal) space.
Let $(S,\mathcal{A})$ be a fixed parameter space.
Henceforth we assume that $(S,\mathcal{A})$ is the fixed parameter space that contains all states of the nature. For instance in the example of section \[example\]. $S$ has two elements: $S = \{s_1,s_2\}$, and $\mathcal{A} = \mathcal{P} (S)$.
\[vilagallapot\] Let $\Omega$ be the space of the states of world, and $\forall i \in N_0$: $\mathcal{M}_i$ be a $\sigma$-field on $\Omega$. The $\sigma$-field $\mathcal{M}_i$ represents player $i$’s information, $\mathcal{M}_0$ is for the information available for the nature, hence it is the representative of $\mathcal{A}$, the $\sigma$-field of the parameter space $S$. Let $\mathcal{M} \circeq \sigma (\bigcup \limits_{i \in N_0} \mathcal{M}_i)$, the smallest $\sigma$-field which contains all $\sigma$-fields $\mathcal{M}_i$.
Every point in $\Omega$ provides a complete description of the actual state of world. It includes both the state of the nature and the players’ states of mind. The different $\sigma$-fields are for modeling the informedness of the players, they have the same role as the partitions in e.g. Aumann’s [@Aumann] paper have. Therefore, if $\omega,\omega' \in \Omega$ are not distinguishable [^7] in the $\sigma$-field $\mathcal{M}_i$ then player $i$ is not able to discern difference between them, i.e., she believes the same things, and behaves in the same way at the two states $\omega$ and $\omega'$. $\mathcal{M}$ represents all information available in the model, it is the $\sigma$-field got by pooling the information of the players and the nature.
For the sake of brevity, henceforth - if it does not make confusion - we do not indicate the $\sigma$-fields. E.g. instead of $(S,\mathcal{A})$ we write $S$, or $\Delta(S)$ instead of $(\Delta(S,\mathcal{A}),\mathcal{A}^\ast)$. However, in some cases we refer to the non-written $\sigma$-field: e.g. $A \in \Delta(X,\mathcal{M})$ is a measurable set in $\mathcal{A}^\ast$, i.e., in the measurable space $(\Delta(X,\mathcal{M}),\mathcal{A}^\ast)$, but $A \subseteq \Delta(X,\mathcal{M})$ keeps its original meaning: $A$ is a subset of $\Delta(X,\mathcal{M})$.
\[tipuster\] Let $(\Omega,\mathcal{M})$ be the space of the states of world (see definition \[vilagallapot\].). The type space based on the parameter space $S$ is a tuple $(S,\{(\Omega,\mathcal{M}_i)\}_{i \in N_0},$ $g,\{f_i\}_{i \in N})$, where
1. $g : \Omega \rightarrow S$ is $\mathcal{M}_0$-measurable,
2. $\forall i \in N$: $f_i : \Omega \to \Delta(\Omega,\mathcal{M}_{-i})$ is $\mathcal{M}_i$-measurable,
where $\mathcal{M}_{-i} \circeq \sigma(\bigcup \limits_{j \in N \setminus \{i\}} \mathcal{M}_j)$.
Put definition \[tipuster\]. differently, $S$ is the parameter space, it contains the “types” of the nature. $\mathcal{M}_i$ represents the information available for player $i$, hence it corresponds to the concept of type (Harsányi [@Harsanyi1967-68]). $f_i$ is the type function of player $i$, it maps the player $i$’s types to her (subjective) beliefs.
The above definition of type space differs from Heifetz and Samet’s concept, but it is similar to Meier’s [@Mei1], [@Mei3] type space. We do not use Cartesian product space, but refer only to the $\sigma$-fields. By following strictly Heifetz and Samet’s paper, if one takes the Cartesian product of the parameter space and the type sets, and defines the $\sigma$-fields as the $\sigma$-fields induced by the coordinate projections (e.g. $\mathcal{M}_0$ is induced by the coordinate projection $pr_0 : S \times \times_{i \in N} T_i \rightarrow S$, for the notations see their paper) then she gets at our concept. However, if the Cartesian product is not used directly then it is necessary to connect the parameter space into the type space in some way. For this we use $g$ (Mertens and Zamir [@MZ1] and Meier [@Mei3] use a similar formalism), hence $g$ and $pr_0$ have the same role in the two formalizations, in this and in Heifetz and Samet’s paper respectively.
A further difference between the two formalizations lies in the role of the parameter space. While in Heifetz and Samet the entire parameter space must appear in the type space, in our approach this is not required (see Case $2$ in the example of section \[example\].). We emphasize that this difference is not relevant.
\[tt2\] The type morphism between the type spaces
$$(S,\{(\Omega,\mathcal{M}_i)\}_{i \in N_0},g,\{f_i\}_{i \in N}) \mspace{20mu} \text{and} \mspace{20mu} (S,\{(\Omega',\mathcal{M}'_i)\}_{i \in N_0},g',\{f'_i\}_{i \in N})$$
$\varphi : \Omega \rightarrow \Omega'$ is such an $\mathcal{M}$-measurable function that
1. diagram is commutative (i.e. $\forall A \in S$: $g^{-1} (A) = (g' \circ \varphi)^{-1} (A)$)
$$\label{diag1}
\begin{diagram}
\Omega & & \\
\dTo^\varphi & \rdTo^{g} \\
\Omega' & \rTo^{g'} & S \\
\end{diagram}$$
2. $\forall i \in N$: diagram is commutative (i.e. $\forall \omega \in \Omega$, $\forall A \in \mathcal{M}'_{-i}$: $f'_i \circ \varphi(\omega) (A)$ $= f_i (\omega) (\varphi^{-1} (A))$)
$$\label{morphism}
\begin{diagram}
\Omega & \rTo^{f_i} & \Delta (\Omega,\mathcal{M}_{-i}) \\
\dTo^{\varphi} & \mspace{200mu} & \dTo^{\varphi} \\
\Omega' & \rTo^{f'_i} & \Delta (\Omega',\mathcal{M}'_{-i}) \\
\end{diagram}$$
$\varphi $ type morphism is a type isomorphism, if $\varphi $ is a bijection and $\varphi^{-1}$ is also a type morphism.
The above definition is practically the same as Heifetz and Samet’s, hence all intuitions, they discussed, remain valid, i.e., the type morphism maps type profiles from a type space to type profiles in an other type space in a way that the corresponded types induce equivalent beliefs. In other words, the type morphism reserves the players’ beliefs.
The type spaces that are based on the parameter space $S$ as objects and the type morphisms form a category. Let $\mathcal{C}^S$ denote this category of type spaces.
It is a direct corollary of definitions \[tipuster\]. and \[tt2\].
Heifetz and Samet introduced the concept of universal type space.
\[tt3\] The type space $(S,\{(\Omega,\mathcal{M}_i)\}_{i \in N_0},g,\{f_i\}_{i \in N})$ is universal, if for any type space $(S,\{(\Omega',\mathcal{M}'_i)\}_{i \in N_0},g',\{f'_i\}_{i \in N})$ there is a unique type morphism $\varphi$
$$\text{from} \mspace{20mu} (S,\{(\Omega',\mathcal{M}'_i)\}_{i \in N_0},g',\{f'_i\}_{i \in N}) \mspace{20mu} \text{to} \mspace{20mu} (S,\{(\Omega,\mathcal{M}_i)\}_{i \in N_0},g,\{f_i\}_{i \in N}) \ .$$
In other words, the universal type space is the most general, the broadest type space among the type spaces. It contains all types that appear in the type spaces of the given category.
The universal type space is terminal (final) object in $\mathcal{C}^S$.
It comes directly from definition \[tt3\].
From the viewpoint of category theory the uniqueness of universal type space is really straightforward.
\[egyetlen\] The universal type space is unique up to type isomorphism.
Every terminal object is unique up to isomorphism.
The only question is the existence of universal type space.
\[egyetemes\] There is universal type space, in other words, there is terminal object in $\mathcal{C}^S$.
See Heifetz and Samet Theorem 3.4.
As we have already mentioned, Heifetz and Samet’s formalization of type space is a little bit different from ours. However the difference between the two approaches is quite slight, and we prove a stronger result in theorem \[fotetel\]., hence we have omitted the formal proof of the above proposition.
Next, we turn our attention to an other property of type spaces, the completeness.
\[teljesseg\] The type space $(S,\{(\Omega,\mathcal{M}_i)\}_{i \in N_0},g,\{f_i\}_{i \in N})$ is complete if $\forall i \in N$: $f_i$ is surjective (onto).
The above concept was introduced by Brandenburger [@Brand]. Completeness recommends that for any player $i$, every probability measure on $(\Omega,\mathcal{M}_{-i})$ be in the range of the given player’s type function. In other words, for any player $i$: all measures on $(\Omega,\mathcal{M}_{-i})$ must belong to types of the given player.
\[fontosallitas\] The universal type space is complete.
See Meier [@Mei1] Theorem 4.
We can say again that Meier’s type space is a little bit different from ours, however the difference is really slight, and we prove a stronger result in theorem \[fotetel\]., hence we have omitted the formal proof of the above proposition.
Beliefs space {#bs}
=============
In the following we formalize the intuition of hierarchy of beliefs, i.e., the “infinite regress in reciprocal expectations.” First we give a rough description (see Mertens and Zamir’s [@MZ1]), take player $i$, and examine the situation from her viewpoint:
$$\label{alap}
\begin{array}{lcl}
T_0 & \circeq & S \\
T_1 & \circeq & T_0 \otimes \Delta(T_0)^{N \setminus \{i\}} \\
T_2 & \circeq & T_1 \otimes \Delta(T_1)^{N \setminus \{i\}} = T_0 \otimes \Delta(T_0)^{N \setminus \{i\}} \otimes \Delta(T_0 \otimes \Delta(T_0)^{N \setminus \{i\}})^{N \setminus \{i\}} \\
& \vdots & \\
T_n & \circeq & T_{n-1} \otimes \Delta(T_{n-1})^{N \setminus \{i\}} = T_0 \otimes \bigotimes \limits_{m=0}^{n-1} \Delta(T_m)^{N \setminus \{i\}} \\
& = & T_0 \otimes \bigotimes \limits_{m=0}^{n-2} \Delta(T_m)^{N \setminus \{i\}} \otimes \Delta(T_0 \otimes \bigotimes \limits_{m=0}^{n-2} \Delta(T_m)^{N \setminus \{i\}})^{N \setminus \{i\}} \\
& \vdots &
\end{array}$$
The above formalism can be interpreted as follows. $T_0$ describes the basic uncertainty of the modeled situation, it consists of the states of the nature. $T_1$ is for $T_0$ and the first order beliefs of the other players (not $i$) $\Delta (T_0)^{N \setminus \{i\}}$ (${N \setminus \{i\}}$ is the players’ set except player $i$), i.e., what the other players believe about the states of the nature. In general, $T_n$ describes $T_{n-1}$ and the $n$th order beliefs of the other payers $\Delta (T_{n-1})^{N \setminus \{i\}}$, i.e., what the other players believe about $T_{n-1}$.
However, there is some redundancy[^8] in the above description. E.g. $\Delta(T_0 \otimes \Delta(T_0)^{N \setminus \{i\}})^{N \setminus \{i\}}$ determines $\Delta (T_0)^{N \setminus \{i\}}$ and so does $\Delta(T_{n-1})^{N \setminus \{i\}}$ $\forall (0 \leq m \leq n-2)$: $\Delta(T_m)^{N \setminus \{i\}}$, therefore we can rewrite the above formalisms into the following form:
$$\label{alap2}
\begin{array}{lcl}
T_0 & \circeq & S \\
T_1 & \circeq & T_0 \otimes \Delta(T_0)^{N \setminus \{i\}} \\
T_2 & \circeq & T_0 \otimes \Delta(T_0 \otimes \Delta(T_0)^{N \setminus \{i\}})^{N \setminus \{i\}} \\
T_3 & \circeq & T_0 \otimes \Delta(T_0 \otimes \Delta(T_0)^{N \setminus \{i\}} \otimes \Delta(T_1)^{N \setminus \{i\}})^{N \setminus \{i\}} \\
& \vdots & \\
T_n & \circeq & T_0 \otimes \Delta(T_0 \otimes \bigotimes \limits_{m=0}^{n-2} \Delta(T_m)^{N \setminus \{i\}})^{N \setminus \{i\}} \\
& \vdots &
\end{array}$$
Let $\# \Theta_{-1}^i = 1$, and $\forall n \in \mathbb{N}$: let $\Theta_n^i \circeq \Delta (T_n)$, $q_{-10}^i: \Theta_0^i \rightarrow \Theta_{-1}^i$. Moreover, $\forall n \in \mathbb{N}$, and $\forall \mu \in \Theta_{n+1}^i \circeq \Delta (T_0 \otimes \Delta(T_0 \otimes \bigotimes \limits_{m=0}^{n-1} \Delta (T_m)^{N \setminus \{i\}}))$: $\mu$ can be naturally defined on $T_0 \otimes \Delta ( T_0 \otimes \bigotimes \limits_{m=0}^{n-2} \Delta (T_m)^{N \setminus \{i\}})$ as a restriction of $\mu$, i.e. let $q_{nn+1}^i : \Theta_{n+1}^i \rightarrow \Theta_{n}^i$ be as follows $\forall \mu \in \Theta_{n+1}^i$:
$$q_{nn+1}^i (\mu) \circeq \mu|_{T_0 \otimes \Delta (T_0 \otimes \bigotimes \limits_{m=0}^{n-1} \Delta (T_m)^{N \setminus \{i\}})} \ .$$
Then $\forall n$: $q_{nn+1}^i$ is a measurable mapping.
The next definition, which is from Mertens et al.’s [@MSZ1], summarizes the above discussions and formalizes the “infinite regress in reciprocal expectations.”
\[velemenyter\] In the diagram
$$\label{limit2}
\begin{diagram}
\Theta^i & \mspace{100mu} & \Delta (S \otimes \Theta^{N \setminus \{i\}}) \\
\dTo^{p^i_{n+1}} & & \dTo^{id_S} \mspace{35mu} \dTo_{p^{N \setminus \{i\}}_n} \\
\Theta_{n+1}^i & \circeq & \Delta (S \otimes \Theta_{n}^{N \setminus \{i\}}) \\
\dTo^{q_{nn+1}^i} & & \dTo^{id_S} \mspace{35mu} \dTo_{q_{n-1n}^{N \setminus \{i\}}} \\
\Theta_{n}^i & \circeq & \Delta (S \otimes \Theta_{n-1}^{N \setminus \{i\}}) \\
\end{diagram}$$
- $i \in N$ is an arbitrarily fixed player,
- $n \in \mathbb{N}$,
- $S$ is the fixed parameter space,
moreover $\forall i \in N$:
- $\# \Theta_{-1}^i = 1$,
- $\forall n \in \mathbb{N} \cup \{-1\}$: $\Theta_n^{N \setminus \{i\}} \circeq \bigotimes \limits_{j \in {N \setminus \{i\}}} \Theta^{j}_n$,
- $q_{-10}^i : \Theta_0^i \rightarrow \Theta_{-1}^i$,
- $\forall m,n \in \mathbb {N}$, $m \leq n$, $\forall \mu \in \Theta_n^i$:
$$q_{mn}^i (\mu) \circeq \mu|_{S \otimes \Theta_{m-1}^{N \setminus \{i\}}} \ ,$$
therefore $q_{mn}^i$ is a measurable mapping.
- $\Theta^i \circeq \varprojlim (\Theta_n^i,\mathbb{N} \cup \{-1\},q_{mn}^i)$,
- $\forall n \in \mathbb{N} \cup \{-1\}$: $p_n^i : \Theta^i \rightarrow \Theta_{n}^i$ is canonical projection,
- $\forall m,n \in \mathbb{N} \cup \{-1\}$, $m \leq n$: $q_{mn}^{N \setminus \{i\}}$ is the product of the mappings $q_{mn}^j$, $j \in {N \setminus \{i\}}$, and so is $p_n^{N \setminus \{i\}}$ of $p_n^j$, $j \in {N \setminus \{i\}}$, therefore both mappings are measurable,
- $\Theta^{N \setminus \{i\}} \circeq \bigotimes \limits_{j \in {N \setminus \{i\}}} \Theta^{j}$.
Then $T \circeq S \otimes \Theta^N$ is called beliefs space.
The interpretation of the beliefs space is the following. For any $\theta^i \in \Theta^i$: $\theta^i = (\mu_{1}^i,\mu_{2}^i,\ldots)$, where $\mu_{n}^i \in \Theta_{n-1}^i$ is the $n$th order belief of player $i$. Therefore every point $\Theta^i$ defines an inverse system of measure spaces
$$\label{hb}
((S \otimes \Theta_n^{N \setminus \{i\}},p_{n+1}^i (\theta^i)),\mathbb{N} \cup \{-1\},(id_S,q_{mn}^{N \setminus \{i\}})) \ .$$
We call inverse system of measure spaces like player $i$’s *hierarchy of beliefs*[^9].
To sum up, $T$ consists of all states of world: all states of the nature, the points in $S$, and all states of mind, the points in the set $\Theta^N$, therefore $T$ contains all players’ all hierarchies of beliefs.
Our main result:
\[fotetel\] The complete universal type space contains all players’ all hierarchies of beliefs.
All next section is devoted to the proof of the above theorem.
The proof of theorem \[fotetel\]. {#beliefspace}
=================================
The strategy of the proof is to show that the belief space (see definition \[velemenyter\].) generates (is equivalent to) the complete universal type space (in category $\mathcal{C}^S$). The key point of the proof is to demonstrate that in the diagram $\forall i \in N$:
$$\Theta^i = \Delta (S \otimes \Theta^{N \setminus \{i\}}) \ ,$$
i.e. they are measurable isomorphic. Therefore in the following we focus on this point.
\[indukalttipuster\] The belief space $T$ generates a type space in the category $\mathcal{C}^S$.
Let $\forall i \in N$: $pr_i : T \rightarrow \Theta^{i}$, $pr_0 : T \rightarrow S$ be coordinate projections, and $\forall i \in N \cup \{0\}$: the $\sigma$-fields $\mathcal{M}_i^\ast$ be induced by $pr_i$. Pintér’s [@pmp3] result implies that $\forall i \in N$:
$$\label{fontos}
\Theta^i = \Delta (S \otimes \Theta^{N \setminus \{i\}}) \ ,$$
i.e., they are measurable isomorphic.
Furthermore, let $g^\ast \circeq pr_0$, and $\forall t \in T$: $f_i^\ast (t) \circeq pr_i (t)$. Then
$$(S,\{(T,\mathcal{M}_i^\ast)\}_{i \in N},g^\ast,\{f_i^\ast\}_{i \in N})$$
is a type space in category $\mathcal{C}^S$.
\[onetoone\] In the type space $(S,\{(T,\mathcal{M}_i^\ast)\}_{i \in N},g^\ast,\{f_i^\ast\}_{i \in N})$ $\forall i \in N$: if $pr_i (t) \neq pr_i (t')$ then $f_i^\ast (t) \neq f_i^\ast (t')$.
It is the direct corollary of that different inverse systems of measure spaces have different inverse limits, and $\Theta^i$ consists of different inverse systems of measure spaces (hierarchies of beliefs).
\[teljeskell2\] The type space $(S,\{(T,\mathcal{M}_i^\ast)\}_{i \in N},g^\ast,\{f_i^\ast\}_{i \in N})$ is complete.
It is the direct corollary of .
\[atetel2\] The type space $(S,\{(T,\mathcal{M}_i^\ast)\}_{i \in N},g^\ast,\{f_i^\ast\}_{i \in N})$ is a universal type space.
Let $(S,\{(\Omega,\mathcal{M}_i)\}_{i \in N},g,\{f_i\}_{i \in N})$ be an arbitrarily fixed type space (an object in $\mathcal{C}^S$), and $i \in N$ and $\omega \in \Omega$ be also arbitrarily fixed.
The first order belief of player $i$ at state of world $\omega$ $v_i^1 (\omega)$ is the measure defined as follows $\forall A \in S$:
$$v_1^i (\omega) (A) \circeq f_i (\omega) (g^{-1} (A)) \ .$$
$f_i$ is $\mathcal{M}_i$-measurable, hence $v_i^1$ is also $\mathcal{M}_i$-measurable.
The second order belief of player $i$ at state of world $\omega$ $v_2^i (\omega)$ is the measure defined as follows $\forall A \in S \otimes \Theta_0^{N \setminus \{i\}}$:
$$v_2^i (\omega) (A) \circeq f_i (\omega) ((g^{-1},(v_1^{N \setminus \{i\}})^{-1})) (A) \ ,$$
where $v_1^{N \setminus \{i\}}$ is the product of the mappings $v_1^j$, $j \in N \setminus \{i\}$. Since $f_i$ is $\mathcal{M}_i$-measurable, hence $v_i^2$ is also $\mathcal{M}_i$-measurable.
Let $n > 1$ be arbitrarily fixed, then the $n$th order belief of player $i$ at state of world $\omega$ $v_n^i (\omega)$ is the measure defined as follows $\forall A \in S \otimes \Theta_{n-2}^{N \setminus \{i\}}$:
$$v_n^i (\omega) (A) \circeq f_i (\omega) ((g^{-1},(v_{n-1}^{N \setminus \{i\}})^{-1})) (A) \ .$$
Since $f_i$ is $\mathcal{M}_i$-measurable, hence $v_n^i$ is also $\mathcal{M}_i$-measurable.
To sum up, there is a well defined mapping $\phi : \Omega \rightarrow S \otimes T$ as follows $\forall \omega \in \Omega$:
$$\label{psidef}
\phi (\omega) \circeq (g(\omega),(v^i_1(\omega),v^i_2 (\omega),\ldots)_{i \in N}) \ .$$
Then it is easy to verify that
\(1) $\phi$ is $\mathcal{M}$-measurable.
\(2) The above construction implies that $\forall \omega \in \Omega$, $\forall i \in N$, $\forall A \in T$:
$$f^\ast_i \circ \phi (\omega) (A) = f_i (\omega) (\phi^{-1} (A)) \ ,$$
i.e., $\phi$ is a type morphism.
\(3) From corollary \[onetoone\]. $\phi$ is the unique type morphism from the type space $(S,\{(\Omega,\mathcal{M}_i)\}_{i \in N},g,\{f_i\}_{i \in N})$ to $(S,\{(T,\mathcal{M}_i^\ast)\}_{i \in N},g^\ast,\{f_i^\ast\}_{i \in N})$.
It is worth noticing that $\phi$ in the above proof is not injective (one to one). If there are duplicate types in a type space, i.e. such types that generate the same hierarchy of beliefs (see e.g. Ely and Peski’s [@ElyPeski] example), then the $\phi$ image of this duplication is one point in the universal type space. Therefore, it is not surprising at all that there are no duplicates in the universal type space, i.e., it can be complete.
From proposition \[atetel2\].
$$\label{univ}
(S,\{(T,\mathcal{M}_i)\}_{i \in N},g^\ast,\{f_i^\ast\}_{i \in N})$$
is a universal type space.
Then corollary \[egyetlen\]. implies that Heifetz and Samet’s [@HS3] universal type space and coincide (they are type isomorphic).
From proposition \[teljeskell2\]. is complete (Meier [@Mei1] also proved this).
Finally, from definition \[velemenyter\]. contains all hierarchies of beliefs.
Related papers {#comparison}
==============
In this section theorem \[fotetel\]. is compared to the results of Heifetz and Samet [@HS4], and Pintér [@pmp2]. These papers seem to contradict our main result, however in the following we show that it is not the case at all.
Heifetz and Samet in their paper “Coherent beliefs are not always types,” as the title indicates, give an example, a hierarchy of beliefs, that can not be type in any type space. Mathematically, their counterexample is based on an exercise of Halmos’s book [@Halmos1], an example for an inverse system of measure spaces without inverse limit. First, we summarize their example.
\[plHS\] Some notations: $l^\ast$ and $l_\ast$ are respectively the outer and inner measures induced by the Lebesgue measure. Let $\{A_n\}_n$ be the Vitali sets from Halmos’ [@Halmos1] example, so it is true that $\forall n$: $A_{n+1} \subseteq A_n \subseteq [0,1]$, $l^\ast(A_n) = 1$, $l_\ast (A_n) = 0$, and $\bigcap \limits_n A_n = \emptyset$. Moreover, $\forall n$: let $\mu_n$ be the probability measures on $B(\prod \limits_{k=0}^n A_k)$[^10] also from Halmos’ example.
Look at the following inverse system of measure spaces:
$$\label{kell}
((\prod \limits_{k=0}^n A_k,B(\prod \limits_{k=0}^n A_k),\mu_n),\mathbb{N},pr_{nn+1}) \ ,$$
where $pr_{nn+1} : \prod \limits_{k=0}^{n+1} A_k \rightarrow \prod \limits_{k=0}^n A_k$ is coordinate projection.
Furthermore, if $X \circeq \prod \limits_{k=0}^n X_k$ is a product space, and $\delta_x$ is the Dirac measure concentrated at $x \circeq (x_0,x_1,\ldots,x_n)$, then $\delta_x = \prod \limits_{k=0}^n \delta_{x_k}$, where $\delta_{x_k}$ is the Dirac measure concentrated at $x_k$.
Interpretation: There are two players, we chose one of them. Let $A_0$ be the parameter space (the set of the states of the nature). $A_1 \subseteq A_0$, and $\forall x \in A_1$ let $x$ be $\delta_{x}$, i.e. $A_1$ is[^11] the set of some first order beliefs of the given player. Moreover, $A_2 \subseteq A_1$ and $\forall x \in A_2$ let $x$ be $\delta^2_{x}$, where $\delta^2_x \circeq \delta_{\delta_{x}}$, i.e. $A_1 \times A_2$ is the set of some second order beliefs of the given player. In general, $\forall n \geq 3$, $\forall x \in A_n \subseteq A_{n-1}$: let $x$ be $\delta^n_x$, where $\delta^n_{x} \circeq \delta_{\delta^{n-1}_{x}}$, i.e. $\prod \limits_{k=1}^n A_k$ is the set of some $n$-order beliefs of the given player.
To sum up, $\forall n$: $(a_0,a_1,\ldots,a_n) \in \prod \limits_{k=0}^n A_k$ is $(a_0,\delta_{a_1},\delta^2_{a_2},\ldots,\delta^n_{a_n}) = a_0 \times \delta_{(a_1,\ldots,a_n)}$.
Put it differently, $\prod \limits_{k=0}^\infty A_k$ is the space of some coherent hierarchies of beliefs, therefore is a hierarchy of beliefs on coherent hierarchies of beliefs. However, from Halmos’ example this hierarchy of beliefs is not type ( has no limit).
Next we show that Heifetz and Samet’s hierarchy of beliefs is not in the beliefs space $T$.
is not in $T$.
It is enough to show that the diagonal of $A_0 \times A_1$ is not a measurable subset of $A_0 \otimes \Delta (A_0)$. The strategy is the following: if the diagonal of $A_0 \times A_1$ is measurable subset of $A_0 \otimes \Delta (A_0)$ then $\forall B \subseteq A_0 \times \Delta (A_0)$: the intersection of the diagonal of $A_0 \otimes A_1$ and $B$ is a measurable subset in subspace $B$.
Examine $A_0 \otimes A_0$. Then from example \[plHS\]. $A_0 \otimes A_0$ is measurable isomorphic (actually more, it is homeomorphic) to $A_0 \otimes \Delta_D (A_0)$, where $\Delta_D (A_0)$ is for the Dirac measures on $A_0$. However, form the definitions of $\{A_n\}_n$ the diagonal of $A_0 \times A_1$ is not a measurable subset of (the diagonal of ) $A_0 \times A_0$, hence $\mu_1 \notin \Delta (A_0 \otimes \Delta (A_0))$, where $\mu_1$ is from example \[plHS\].
To sum up, Heifetz and Samet’s counterexample is such a hierarchy of beliefs that is not among the purely measurable hierarchies of beliefs, i.e. that is not in the purely measurable beliefs space. Therefore, Heifetz and Samet’s result does not contradict our result.
Quite recently, Pintér provided a negative result, he argues that there is no universal topological type space in the category of topological type spaces. Actually, this non-existence is got by topological argument, hence his negative result does not contradict this paper’s positive one.
On the other hand, Pintér’s result clearly shows that irrelevant details, brought in the model by topological concepts, can make real difficulties, which culminate in that the goal proving that the Harsányi program works is unreachable in the topological approach.
Conclusion
==========
The main result of this paper is theorem \[fotetel\]. concludes that in the purely measurable framework the Harsányi program works, i.e., the incomplete information situations can be modeled by type spaces. In this sense, this paper ends the sequence of papers focusing on the Harsányi program: Heifetz and Samet [@HS3], and Meier [@Mei1] among others.
Theorem \[fotetel\]. with Pintér’s [@pmp2] result raise the problem that although in the literature mostly the topological models are popular, the purely measurable and not the topological framework is appropriate for modeling incomplete information situations. Can every result in the topological framework be translated into the purely measurable one? For this question future research can answer.
[10]{} Aumann, R.J.: “Interactive epistemology I., Knowledge" *International Journal of Game Theory* **28**, 263–300. (1999)
Aumann, R.J.: “Interactive epistemology II., Probability" *International Journal of Game Theory* **28**, 301–314. (1999)
Battigalli, P., M. Siniscalchi: “Hierarchies of Conditional Beliefs and Interactive Epistemology in Dynamic Games” *Journal of Economic Theory* **88**, 188–230. (1999)
“On solutions of bayesian games" *International Journal of Game Theory* **8**, 193–215. (1979)
“On the Existence of a ’Complete’ Possibility Structure" *Cognitive Processes and Economic Behavior* edited by Marcello Basili, Nicola Dimitri, and Itzhak Gilboa, Routledge 30–34. (2003)
“The power of paradox: some recent developments in interactive epistemology" *International Journal of Game Theory* **35**, 465–492. (2007)
“Hierarchies of beliefs and common knowledge" *Journal of Economic Theory* **59**, 189–198. (1993)
“An Impossibility Theorem on Beliefs in Games" *Studia Logica* **84**, 211–240. (2006)
“Rationality and knowledge in game theory" *Advances in Economics and Econometrics: Theory and Applications (Seventh World Congress of Econometric Society Vol. 1.)* 87–171, (1997)
“Hierarchies of belief and interim rationalizability" *Theoretical Economics* **1**, 19–65. (2006)
*Measure Theory*, Springer-Verlag (1974)
“Games with incomplete information played by bayesian players part I., [II]{}., [III]{}." *Management Science* **14**, 159–182., 320–334., 486–502. (1967-1968)
“The bayesian formulation of incomplete information - the non-compact case" *International Journal of Game Theory* **21**, 329–338. (1993)
“Knowledge Spaces with Arbitrarily High Rank" *Games and Economic Behavior* **22**, 260–273. (1998)
“Topology-free typology of beliefs" *Journal of Economic Theory* **82**, 324–341. (1998)
“Coherent beliefs are not always types" *Journal of Mathematical Economics* **32**, 475–488. (1999)
“An infinitary probability logic for type spaces" *CORE Discussion paper No. 0161* (2001)
“Finitely additive beliefs and universal type spaces" *The Annals of Probability* **34**, 386–422. (2006)
Meier, M.: “Universal knowledge–belief structures" *Games and Economic Behavior* **62**, 53-–66. (2008)
“Formulations of bayesian analysis for games with incomplete informations" *International Journal of Game Theory* **14**, 1–29. (1985)
“Repeated games part A" *CORE Discussion Paper No. 9420* (1994)
Osborne, M. J., A. Rubinstein: *A course in game theory*, The MIT Press, Cambridge, Mass. (1994)
“Type space on a purely measurable parameter space" *Economic Theory* **26**, 1239–139. (2005)
“The non-existence of universal topological type space" *working paper*, (2007)
“The existence of inverse limit of inverse systems of measure spaces – a special case" *working paper*, (2008)
[^1]: Department of Mathematics, Corvinus University of Budapest, 1093 Hungary, Budapest, Fővám tér 13-15., [email protected]
[^2]: Thanks. This work was supported by the János Bolyai Research Scholarship of the Hungarian Academy of Sciences and by grant OTKA 72856.
[^3]: In this paper we use the terminology hierarchy of beliefs instead of the longer coherent hierarchy of beliefs.
[^4]: [@Harsanyi1967-68] pp.163–167.
[^5]: Brandenburger and Dekel’s [@BD2] result implies that in this very simple case every hierarchy of beliefs is in the complete universal type space. The general case is that when the parameter space $S$ is arbitrary measurable space.
[^6]: For a more detailed argument see e.g. Meier [@Mei3] p. 56. “Why should the knowledge operators of the players just operate on measurable sets and not on all subsets of the space? The justification for this is that we think of events as those sets of states that the players can describe, and only those can be the objects of their reasoning. In view of this interpretation a statement saying “player i knows that the actual state of the world is in E,” where E is an entity of states he cannot represent in his mind, is meaningless. Of course, it might well be that in some knowledge–belief spaces all subsets of the space of states of the world can be described by the players (for example in the finite knowledge–belief spaces), but we do not want to assume this in general.”
[^7]: Let $(X,\mathcal{T})$ be arbitrarily fixed measurable space, and $x,y \in X$ be also arbitrarily fixed. $x$ and $y$ are measurably indistinguishable if $\forall A \in \mathcal{T}$: $(x \in A) \Leftrightarrow (y \in A)$.
[^8]: This redundancy is called *coherency* and *consistency* in the literature of game theory and mathematics respectively.
[^9]: In the literature this system is usually called coherent hierarchy of beliefs. Since it does not make confusion, in this paper we omit the adjective coherent.
[^10]: $B(\cdot)$ is for the Borel $\sigma$-field.
[^11]: Henceforth in context like this “is” means the two spaces are homeomorphic
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Zijie Qu,$^{1\ast}$ Fatma Zeynep Temel,$^{1,2}$ Rene Henderikx$^{1,3}$, Kenneth S. Breuer$^{1}$\
\
\
\
\
bibliography:
- 'scibib.bib'
title: Changes in the flagellar bundling time account for variations in swimming behavior of flagellated bacteria in viscous media
---
Although the motility of the flagellated bacteria, *Escherichia coli*, has been widely studied, the effect of viscosity on swimming speed remains controversial. The swimming mode of wild-type *E.coli* is often idealized as a “run-and-tumble” sequence in which periods of swimming at a constant speed are randomly interrupted by a sudden change of direction at a very low speed. Using a tracking microscope, we follow cells for extended periods of time in Newtonian liquids of varying viscosity, and find that the swimming behavior of a single cell can exhibit a variety of behaviors including run-and-tumble and “slow-random-walk” in which the cells move at relatively low speed. Although the characteristic swimming speed varies between individuals and in different polymer solutions, we find that the skewness of the speed distribution is solely a function of viscosity and can be used, in concert with the measured average swimming speed, to determine the effective running speed of each cell. We hypothesize that differences in the swimming behavior observed in solutions of different viscosity are due to changes in the flagellar bundling time, which increases as the viscosity rises, due to the lower rotation rate of the flagellar motor. A numerical simulation and the use of Resistive Force theory provide support for this hypothesis.
The survival of motile bacteria depends in part on the ability to navigate their environment, swimming towards attractants (e.g. food) and away from repellents (e.g. toxins). In order to move in a low Reynolds number environment and to avoid the time-reversibility of Stokesian dynamics [@purcell1977life], flagellated bacteria such as *Escherichia coli* exhibit a non-reciprocal swimming behavior first described by Berg and Brown [@berg1972chemotaxis]. The “run-and-tumble” behavior is characterized by extended linear movements (“runs”) punctuated by sudden changes in direction (“tumbles”). The tumbling event is initiated by the clockwise (CW) rotation of one or more of the flagellar motors [@berg2008coli; @turner2000real] ($T_o$ in Fig. 1A). This precipitates the unravelling of the flagellar bundle which causes the cell to immediately stall and re-orient ($T_o - T_1$). As the motor returns to counter-clockwise (CCW) rotation ($T_2$), the flagellar bundle re-forms ($T_2 - T_3$) [@turner2000real; @kim2003macroscopic] and the cell accelerates back to the characteristic run speed, $U_o$. Note that the value of $U_o$ can vary, and depends on the cell metabolism, the number, length and spatial distribution of flagella and the conditions of the surrounding fluid (temperature, presence or absence of specific nutrients, etc).
This mode of cell motility has been studied extensively over the past decades (e.g. [@lighthill1976flagellar; @hotani1982micro; @berg1993torque; @berg1995cells; @fraser1999swarming; @powers2002role; @dalton2011vivo; @martinez2014flagellated; @patteson2015running]) but while it remains a compelling idealized model for multi-flagellated motion, there remain questions. For example, Molaei *et al.* analyzed thousands of individual cell motion histories [@molaei2014failed] and reported that only $70$% of the *E.coli* cells exhibited run-and-tumble style of motion while the rest of the cells, moved in a different mode, termed “slow-random-walk” and characterized by a slower average speed and absent clearly defined tumbling events. More recently, a close examination of cell motility and flagellar motion [@turner2016visualizing] revealed intermediate states, such as partial unbundling, which also contributed to a wider variety of swimming modalities than the binary “run” and “tumble” states.
Bacteria live in varied fluid environments that can exhibit viscous and/or viscoelastic properties [@kimsey1990motility], measurements and calculations of cell motility in these complex fluids have yielded numerous seemingly contradictory results [@martinez2014flagellated; @patteson2015running; @berg1979movement; @lauga2007propulsion; @fu2007theory; @leshansky2009enhanced; @liu2011force; @spagnolie2013locomotion; @dasgupta2013speed]. Even for cells swimming in (assumed to be) Newtonian polymer solutions of varying viscosity, the picture is unclear. One of the earliest experimental studies in polymeric solutions shows that the tumbling of *E.coli* cells is suppressed and swimming speed is increased even when the polymer concentration is low [@berg1979movement]. The authors explain this phenomenon by appealing to the properties of the loose and quasi-rigid polymer network and its interactions with the nanoscale flagellar propulsors. Magariyama and Kudo proposed a simple model based on Resistive Force Theory (RFT) [@purcell1977life; @magariyama2002mathematical], but modified by the introduction of two apparent viscosities that depend on the length, morphology, and the interaction between polymer molecules [@magariyama2002mathematical]. A further complication arises from the observation that the level of biological activity appears to change with the addition of the thickening polymer [@martinez2014flagellated], probably due to the metabolism of small polymer fragments by the bacteria.
In order to fully understand the different swimming modes, cells must be observed for relatively long time periods and in different fluid environments. Two methodologies are commonly described. In most studies, cells are tracked under a stationary microscope platform (e.g. [@martinez2014flagellated; @molaei2014failed]) which, though effective and straightforward, only permits tracking for short times as the cells quickly pass through the microscope’s field of view and focal plane. Alternatively, one can track individual cells in three dimensions by physically moving the objective and the microscope stage in real time [@berg1972chemotaxis; @liu2014helical; @turner2016visualizing]. Although the tracking microscope is inefficient, in terms of the number of observed individuals, the extended tracking time permits detailed observation of similarities and differences in the swimming behavior for both a single cell and between individual cells in an identical genetic population. In this manuscript, in an attempt to understand the different swimming modalities and the role of viscosity on cell motility, we report on the use of tracking microscopy to measure the detailed behavior of wild-type *E.coli* swimming in Newtonian fluids of varying viscosity. Different solutions of polymers using two molecular weights were prepared, and cell trajectories in both native and dialyzed polymer solutions were recorded.
A typical time history of speed and angular change (Fig. 1B, C) shows good qualitative and quantitative agreement with the classic results of Berg and Brown [@berg1972chemotaxis]. Using their definition of the run and tumble phases (Fig. 1A), we find that the run time and tumbling frequency are not affected by the fluid properties (Table S1). However, close inspection of the time-traces indicates that a single cell exhibits both classical run-and-tumble events as well as periods of extended low-speed swimming or “slow-random-walk” [@molaei2014failed]. This is quantitatively reflected by the probability density function (PDF) of swimming speed (Fig. 1D) which shows two peaks; one at high speed, which we associate with the observed run behavior, and a second peak at a lower speed corresponding to the “slow-random-walk” behavior.
From these results, we assert that the “slow-random-walk” mode of motility is not the result of different cells illustrating different swimming modalities. Rather, over an extended period of time, a single cell can exhibit multiple modes of motility. Indeed, more complex combinations of speed and orientation changes are observed, (e.g. Fig. 1B, $t \approx 0$ s $-$ $4$ s) which might be due to partial unbundling [@turner2016visualizing].
The shape of the speed ($U$) distribution, in particular the skewness, $K = \overline{(U-\overline{U}})^{3}/\sigma^{3}$, where bar denotes the mean value and $\sigma$ is the sample standard deviation, proves to be a valuable means to quantify differences between swimming behaviors. For the example shown (Fig. 1) a bi-modal PDF ($K = 0.22$) illustrates a co-existence between run-and-tumble and “slow-random-walk” behaviors. One can imagine that a swimmer exhibiting a pure run-and-tumble behavior would have a PDF characterized by a sharp peak at the run speed with a broad low-speed tail. Such a speed distribution would have a negative skewness ($K < 0$). Similarly, a cell that spends more time in a tumbling state, with only short runs would have a low mean speed and a positively-skewed PDF ($K > 0$). A cell that tumbles continuously, would have a zero mean speed and zero skewness. Note that the skewness is independent of the absolute swimming speed.
Every cell trajectory that we have measured exhibits this characteristic twin-peaked speed distribution. However, even though the average run times and tumble frequencies are relatively constant (Table S1), there is considerable cell-to-cell variation in absolute swimming speed (Fig. 2A), most likely due to natural variations in the length and number of flagella. In addition, we observe that there is a marked difference between the swimming speed in dialyzed and native polymer solutions despite the fact that these solutions have the same bulk viscosity (Fig. S1). Although the average swimming speed does decrease as viscosity rises, there does not appear to be a uniform scaling; in addition the swimming speed in native solutions increases initially, before decaying, a phenomenon that has been previously observed [@berg1979movement; @magariyama2002mathematical; @martinez2014flagellated] and variously attributed to non-Newtonian interactions with the polymer network [@magariyama2002mathematical] and the effect of small polymer fragments on the metabolic activity of the cell [@martinez2014flagellated].
Characterizing motility purely by the average swimming speed thus appears to be too blunt a tool; however, looking at the skewness of the speed distributions we see that, as the viscosity increases, the speed distribution changes smoothly reflecting a shift from a predominantly run-and-tumble style, characterized by a negative skewness, to a predominantly slow-random-walk style of swimming, characterized by a skewness close to zero, or even slightly positive (Fig. 2B). The same behavior is observed in all four polymer solutions (two different molecular weights, dialyzed and native solutions) suggesting that the speed distribution is consistent with the solution viscosity and independent of the absolute swimming speed.
What might be the cause of this change in the swimming speed distribution? Assuming that the geometry of the cell body and flagellar filaments do not depend on the fluid viscosity, the hydrodynamics of the run scale linearly with viscosity [@purcell1977life; @lauga2009hydrodynamics]. Furthermore, the tumbling frequency, is independent of viscosity (Table S1). However, the time for the flagellar bundle to unravel and reform during the tumble *does* change with viscosity. Kim *et al.* [@kim2003macroscopic] showed that the flagellar bundling of elastic helices depends on a non-dimensional parameter, $M = \mu \omega L^4/ E I$, where $\mu$ is the fluid viscosity, $\omega$ is the rotation rate, $L$ the filament length, $E$ the elastic modulus and $I$ the moment of inertia. $M$ represents the balance between the viscous and elastic stresses in the filament and Kim *et al.* demonstrated that flagellar bundling occurs after about $15$ rotations for values of $M$ greater than about $100$. For a fixed torque motor [@chen2000torque; @darnton2007torque], the flagella rotation rate will decrease as the fluid viscosity increases indicating that the bundling process, which requires a fixed number of rotations [@kim2003macroscopic] will take longer at higher viscosity. In addition, Turner *et al.* observed that the swimming speed of the cell remains depressed after tumble due to the rebundling process [@turner2000real]. Thus it seems plausible that, as the viscosity rises, the cell spends less time running at full speed, and more time at lower speed recovering from tumbles. This hypothesis is consistent with our observation that the speed distribution skewness approaches or passes zero as the viscosity rises (Fig. 2B). The effect is obscured in the speed vs. viscosity data (Fig. 2A) by the confounding factors of individual variations in morphology and metabolic activity as well as the effects of polymers on cell activity level.
A numerical simulation confirms the relationship between the bundling time, the scaled average speed, and the skewness of the speed distribution. We model the swimming as a combination of a run at a given “characteristic run speed”, $U_o$, punctuated by tumbles that occur randomly according to a Poisson distribution. The acceleration from the tumble back to $U_o$ is changed by the effect of varying viscosity on the bundling dynamics. Using this idealized simulation, we generate synthetic trajectories, speed histories and speed distributions associated with different bundling times (Fig. 3A, B) that are both qualitatively and quantitatively similar to the experimentally-measured distributions (e.g. Fig. 1D). Plotting the distribution skewness against the ratio of the average speed to the characteristic run speed, $\overline{U}/U_o$, we find that the data exhibits a linear trend: $\overline{U}/U_o = -0.185K + 0.627$. More importantly, the simulation results allow us to use measurements of the speed distribution skewness, $K$, and the average swimming speed, $\overline{U}$, to estimate the characteristic run speed, $U_o$ - a parameter that varies from cell to cell and is not directly measurable. With the estimate for $U_o$, and typical values for the geometry of the cell and flagella [@magariyama2002mathematical], we use RFT to calculate the motor torque, $T$, as well as the cell and flagellar rotation rate, $\omega_c$ and $\omega_f$ respectively.
Although there is scatter in the data, the motor torque is estimated to lie between $0.25$ and $0.75$ $\times 10^{-18}$ Nm (Fig. 4A), which agrees well with the measurement of Darnton *et al.* [@darnton2007torque] who used a similar technique, but is lower than the measurement of Reid *et al.* [@reid2006maximum]. It is worthwhile to note that the motor torque in the native polymer solutions (circles in Fig. 4A) is higher than the torque in the dialyzed solutions (squares in Fig. 4A), consistent with the observations both here, and by Martinez *et al.* [@martinez2014flagellated] that the cell activity is generally higher in the native polymer solutions.
Using the motor torque and flagellar rotation rate obtained from RFT, we calculate the bundling time, $T_b$ (Fig. 4B), assuming that $20$ rotations are required for complete bundling. The results confirm the hypothesis that the bundling time is a function of viscosity, rising from approximately $0.1$ seconds in pure motility buffer to about $0.2$ seconds in the most viscous medium.
A second, independent, estimate of the bundling time can be found from the measured speed vs. time history of each cell. To accomplish this, we first use the skewness of the measured speed distribution to determine the characteristic run speed, $U_o$ (Fig. 3C). Using Berg & Brown’s definition of a change in angular orientation greater than $35^\circ/0.08$ s we identify the start of each tumble ($T_o$ in Fig. 1A) and mark the completion of the re-bundling process as the time at which the swimming speed first reaches the characteristic run speed ($T_3$ in Fig. 1A). The bundling time, $T_b$, is then defined as $T_3 - T_o - 0.32$, where $0.32$s is used as the duration of the CW rotation ($T_2 - T_o$) (Fig. 1A) [@berg2008coli; @turner2000real; @darnton2007torque].
The estimate of the flagellar bundling time obtained using this method (Fig. 4C) agrees well with the results obtained using RFT (Fig. 4B), demonstrating that the bundling time increases with viscosity, rising from about $0.08$ to $0.3$ seconds over the five-fold increase in viscosity. The scatter in the data likely results from our inability to accurately estimate the exact duration of the CW rotation ($T_2 - T_o$) and the variability associated with the determination of $T_3$.
In summary, we confirm that the motility of a wild-type *E. coli* cell is quite nuanced, exhibiting both run-and-tumble and slow-random-walk during its natural swimming behavior. Furthermore, we believe that we have clarified the confusion surrounding cell motility in viscous media by demonstrating that the swimming behavior in Newtonian fluids of different viscosities can be explained using classical Resistive Force Theory coupled with the recognition that the flagellar bundling process takes longer at higher viscosity due to the slower rotation of the flagellar motor. Lastly, we note that the skewness of the swimming speed is a more useful metric than the average swimming speed, and can be used in the analysis of cell swimming trajectories to control for differences in the characteristic run speed that arise due to cell-to-cell variations and for differences related to the uncontrolled presence of biological stimulants in the surrounding medium.
{width="1\linewidth"}
{width="1\linewidth"}
{width="1\linewidth"}
{width="1\linewidth"}
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to Coli Genetic Stock Center (Yale University) for strains and advice. We thank A.Tripathi for the help in measuring the sample viscosity. The work was supported by National Science Foundation (CBET\#1336638).
Supplementary materials {#supplementary-materials .unnumbered}
=======================
Materials and Methods\
Fig. S1\
Table S1\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we use radio polarimetric observations of the jet of the nearby bright quasar 3C345 to estimate the fluid velocity on kiloparsec scales. The jet is highly polarized, and surprisingly, the electric vector position angles in the jet are “twisted” with respect to the jet axis. Simple models of magnetized jets are investigated in order to study various possible origins of the electric vector distribution. In a cylindrically-symmetric transparent jet a helical magnetic field will appear either transverse or longitudinal due to partial cancellations of Stokes parameters between the front and back of the jet. Synchrotron opacity can break the symmetry, but it leads to fractional polarization less than that observed, and to strong frequency dependence that is not seen. Modeling shows that differential Doppler boosting in a diverging jet can break the symmetry, allowing a helical magnetic field to produce a twisted electric vector pattern. Constraints on the jet inclination, magnetic field properties, intrinsic opening angle, and fluid velocities are obtained, and show that highly relativistic speeds ($\beta \ga 0.95$) are required. This is consistent with the observed jet opening angle, with the absence of a counter-jet, with the polarization of the knots at the end of the jet, and with some inverse-Compton models for the X-ray emission from the 3C345 jet. This model can also apply on parsec scales and may help explain those sources where the electric vector position angles in the jet are neither parallel nor transverse to the jet axis.'
author:
- 'David H. Roberts[^1] & John F. C. Wardle'
title: |
Evidence for Highly Relativistic Velocities\
in the Kiloparsec-Scale Jet of the Quasar 3C345
---
INTRODUCTION
============
An important question about extragalactic radio jets is whether or not the highly relativistic velocities often present on parsec scales continue to kiloparsec scales. Here we discuss a novel piece of information suggesting that this is indeed the case in at least one quasar jet.
3C345 is a superluminal quasar with inferred Lorentz factor $\Gamma$ up to at least 20 on parsec scales [@Lister]. In a companion paper [@RWM], hereafter RWM, we present detailed VLA observations of the kiloparsec-scale jet of 3C345. Its most surprising feature is a “twist” to the electric vector distribution, suggesting a helical magnetic field. This was first seen by @KWR, and is unlike the jets in other Fanaroff-Riley Type II radio sources, where the electric vectors are typically transverse to the jet axis, suggesting a longitudinal magnetic field [@BridlePerley]. In this paper we examine possible origins of this unusual electric vector distribution. Section \[s:data\] summarizes the relevant data on the 3C345 jet, §\[s:models\] describes simple models for the polarization of the jet, and §\[s:concl\] presents our conclusions.
KEY DATA ON THE KILOPARSEC-SCALE JET OF 3C345 {#s:data}
=============================================
Here we summarize the key features of the 3C345 jet as determined by RWM. (1) The fractional polarization in the jet is high, ranging from $0.2$ to $0.5$, and is systematically greater at the edges. (2) Unusually, the inferred magnetic field direction in the main body of the jet is neither longitudinal nor transverse, but makes an apparent helix (Figure \[fig:Poln\]a). At the center of the jet the twist is about $35^\circ$. (3) The “twist” in the apparent magnetic field is not due to Faraday rotation, which is a few degrees or less (Figure \[fig:Poln\]b). (4) The jet diverges slightly with an apparent semi-opening angle of about $\phi_a = 9.4^\circ$. (5) The mean spectral index of the jet is 0.85 ($I_\nu \propto \nu^{-\alpha}$). (6) There is no counter-jet to a limit of 5% of the brightness of the main jet.
In Figure \[fig:slices\] we show cross-sectional slices of the Stokes parameters $I$, $Q$, & $U$ for six position in the jet. The purpose of this paper is to use these data to constrain fluid velocities on kiloparsec scales in 3C345.
![Linear polarization of the kiloparsec-scale jet of 3C345 at 5 GHz. (a) Contours of linearly polarized intensity with ticks showing the orientation of the electric vectors. Note the unusual twist of the electric vectors in the main body of the jet. (b) Contours of linearly polarized intensity with ticks displaying the [*differences*]{} between the electric vector orientations at 5 and 8 GHz as position angle. Vertical lines indicate no Faraday rotation. These figures have been rotated by $-50^\circ$. From RWM.\[fig:Poln\]\
](Figure1a "fig:"){width="1.0\columnwidth"} ![Linear polarization of the kiloparsec-scale jet of 3C345 at 5 GHz. (a) Contours of linearly polarized intensity with ticks showing the orientation of the electric vectors. Note the unusual twist of the electric vectors in the main body of the jet. (b) Contours of linearly polarized intensity with ticks displaying the [*differences*]{} between the electric vector orientations at 5 and 8 GHz as position angle. Vertical lines indicate no Faraday rotation. These figures have been rotated by $-50^\circ$. From RWM.\[fig:Poln\]\
](Figure1b "fig:"){width="1.0\columnwidth"}
\
\
MODELS FOR THE TWISTED POLARIZATION\
OF THE 3C345 JET {#s:models}
====================================
Is the magnetic field in the 3C345 helical, as it appears to be? There are few similar examples in the literature, the best observed one being M87. There the well-resolved jet shows a twisted filamentary structure in the radio [@Owen] that has been ascribed to a Kelvin-Helmholtz instability [@Hardee], and the magnetic field structure seems to follow the filaments. Optically it looks quite similar to the radio image [@Fraix]. In 3C345 we see no filamentary structure in the jet, and suggest instead that the twisted electric field pattern is the result of relativistic effects. A truly helical field in a transparent homogeneous cylindrical jet would appear either transverse or longitudinal, not helical, due to cancellations between the back and the front (in a horizontal jet this means that $U$ is identically zero). Two possibilities suggest themselves; either the symmetry through the jet is broken by opacity or it is broken by relativistic effects. In the latter case, the gentle divergence in the profile of the jet demonstrated by RWM suggests a similarly diverging velocity field. In such a field, differential Doppler boosting due to differing line-of-sight velocity components in the front and back of the jet will break the symmetry, permitting there to be non-zero $U$, and thus creating a twist to the observed electric vectors if the underlying magnetic field is helical. Here we show that this is sufficient to produce the twist seen in 3C345, and use it to determine important physical parameters of the jet.
Synchrotron Opacity
-------------------
We have made simple radiative transfer calculations using the equations in @P70 that show that adding some synchrotron opacity can produce an apparently-helical field, but that the fractional polarization is reduced well below the $\sim 25\%$ typical of the 3C345 jet, and that the polarization structure would be a strong function of frequency, which is not the case here. Thus we reject this explanation.
Differential Doppler Boosting
-----------------------------
In order to assess the effects of differential Doppler boosting in a slightly conical jet, we have investigated the following model. The jet is taken to be a uniform-density axisymmetric cylinder filled with relativistic electrons with energy index $1+2\alpha$ containing a helical magnetic field made up of two parts. The first is a toroidal component generated by a uniform current density, and the second component is a uniform longitudinal field; we parameterize the helicity of the field by the ratio $b$ of longitudinal component to the toroidal component at the surface. The effect of Doppler boosting is included in an [*ad hoc*]{} manner by superimposing a diverging velocity field whose outer profile is constrained by the observed opening angle of the jet. The angle of the velocity vectors to the jet axis varies with normalized radius $r$ as $$\eta(r) = \phi_i r^\epsilon$$ where the radial parameter satisfies $0 \leq r \leq 1$, $\phi_i$ is the intrinsic half-opening angle of the jet, and $\epsilon$ is an adjustable parameter that we take to be unity. The radiative transfer was done in the observer frame using the equations in [@P70], incorporating the relationship between the source magnetic field and radiation electric vectors in a relativistically-moving medium [@LPB].
If the fluid speed is $\beta c$ and the inclination of the jet axis to the line of sight is $i$, in the fluid frame the photon paths make an angle $\theta^\prime$ with the jet axis, where $$\cos{\theta^\prime} = \frac{\cos{i}-\beta}{1-\beta \cos{i}}.$$ The computations show that the symmetry of the profiles of $I$, $Q$, and $U$ across the jet constrain the angle $\theta^\prime$ to be very close to $90^\circ$. This means that $\sin i \simeq 1/\Gamma$, so $\beta$ and $i$ cannot be chosen independently. The intrinsic opening angle of the jet $\phi_i$ and the apparent opening angle $\phi_a$ are related by geometry according to $$\tan \phi_i = \tan \phi_a \sin i,$$ which means that increasing the speed of a model jet reduces the intrinsic opening angle of the jet for a given observed opening angle, partially counteracting the increased differential Doppler effect. One upshot of this is that the maximum $U/I$ that can be generated is sensitive to $\beta$ only linearly, and is not explicitly a function of $\Gamma$.
We searched for solutions in the $b$–$\beta$ plane using contours of constant $Q/I$ and $U/I$ at the center of the jet, where the values should be approximately $0.05$ and $0.22$, respectively, prior to convolution with the beam (see Figure \[fig:ContourPlot\]). The nominal model without convolution with the observing beam is shown in Figure \[fig:ModelUnconv\]; its parameters are $\beta=0.97$, $i=14^\circ$, $\phi_i = 2.3^\circ$, and $b = -0.19$. Figure \[fig:models\] shows the average observed profiles of $I$, $Q$, and $U$, each normalized by the peak of $I$, as functions of distance across the jet versus the predictions of this model when it is convolved with the observing beam of RWM. We find the following. (1) It is possible to choose parameters $\beta$ and $b$ for the jet that produce Stokes parameters that closely match those of the 3C345 jet. (2) The ratio of surface longitudinal to toroidal magnetic field strength in the observer frame must satisfy $-0.3 \lesssim b \lesssim -0.1$. (3) The inclination of the jet to the line of sight is $8^\circ \lesssim i \lesssim 16^\circ$. (4) The fluid speed is $\beta \ga 0.95$.
We also did numerical and analytic calculations in the plasma frame that were consistent with these results when transformed to the observer frame (see @CH). For example, in the special case of a spectral index of $\alpha = 1$ and $\theta^\prime = 90^\circ$, it can be shown that the normalized profiles of $I$, $Q$, and $U$ are $$\frac{I(x)}{I(0)} \simeq \frac{3b^{\prime 2}+1-x^2}{1+3b^{\prime 2}} (1-x^2)^{1/2} + O((\phi_a \beta)^2),$$ $$\frac{Q(x)}{I(0)} \simeq \frac{3}{4} \frac{3b^{\prime 2}-1+x^2}{1+3b^{\prime 2}} (1-x^2)^{1/2} + O((\phi_a \beta)^2),$$ $$\frac{U(x)}{I(0)} \simeq - \frac{9 b^\prime \beta \phi_a}{2(1+3b^{\prime 2})} (1-x^2)^{3/2} + O((\phi_a \beta)^3).$$ Here $-1 \le x \le 1$ is normalized position across the jet and $b^\prime = \Gamma b $ is the pitch angle parameter in the frame of the fluid. These profiles agree very well with numerical integrations for the same parameters.
For such a simple model the success is gratifying. All of the essential features of the linear polarization of the jet are reproduced within the observational uncertainties. Models without highly relativisitic bulk flows are unable to reproduce the twist of the electric vectors, and we regard the fit of data and model to be strong evidence for such flows on kiloparsec scales in 3C345. All of this assumes that the apparent semi-opening angle of the jet is about $9^\circ$ (RWM). Since $U/I$ at the center of the jet is proportional to $b \beta \phi_a$ and $b$ is constrained by the observed values of $Q/I$, even a 20% reduction in $\phi_a$ renders the models untenable.
If 3C345 is typical, and helical magnetic fields are present is all kiloparsec-scale jets, we must explain why we don’t see twists in other highly-polarized FR II jets. One possibility is that 3C345 is core-dominated and thus is inclined near the line of sight, while those observed by, e.g., @Bridle94, are lobe-dominated and their jets are inclined at much larger angles to the line of sight. When we apply our model for 3C345 to such jets the result is apparently longitudinal magnetic fields.
The symmetry of $I$, $Q$, & $U$ across the jet means that it is seen at an angle $\theta^\prime \simeq 90^\circ$ in the frame of the fluid, and that the inclination to the line of sight is $i \simeq 1/\Gamma$. Since this is the condition for maximal superluminal motion, it suggests that deep polarimetric observations of the kiloparsec-scale jets in other core-dominated and/or superluminal sources be undertaken to search for further examples of twisted electric vector configurations.
CONCLUSIONS {#s:concl}
===========
Possible origins of the twisted electric vector orientations in the kiloparsec-scale jet of 3C345 are investigated, and it is found that differential Doppler boosting in a conical jet is adequate to permit a helical magnetic field to produce a twisted electric field pattern. Constraints on the jet speed and inclination to the line of sight are derived, and are found to be consistent with the presence of superluminal motion on parsec scales and with the absence of a counter-jet on kiloparsec scales. Acceptable models have speed $\beta \ga 0.95$, intrinsic jet opening angle $\phi_i \lesssim 2.6^\circ$, and inclination to the line of sight $i \lesssim 16^\circ$. Models with $\beta \lesssim 0.95$ are unsuccessful, so we conclude that the fluid speeds in the kiloparsec-scale jet of 3C345 are highly relativistic. This is in agreement with the conclusions drawn from inverse-Compton models of the X-ray emission seen from 3C345 by Chandra [@CHANDRA]. This model can also apply on parsec scales, and may help explain those sources where the electric vector position angles in the parsec scale jet are neither parallel nor transverse to the direction of the jet ([@CWRG]; @GPC; @PTZ; @LH).
ACKNOWLEDGMENTS
===============
The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. D. H. R. gratefully acknowledges the support of the William R. Kenan, Jr. Charitable Trust and of the National Radio Astronomy Observatory. J. F. C. W. is supported by NSF grant AST-1009262. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. We thank Jean Eilek for helpful conversations and a careful reading of the manuscript.
Bridle, A. H., Hough, D. H., Lonsdale, C. L., Burns, J. O., & Laing, R. A. 1994, , 108, 766 Bridle, A. H., & Perley, R. A. 1984, , 22,319 Cawthorne, T. V., Wardle, J. F. C., Roberts, D. H. & Gabuzda, D. C. 1993, , 416, 519 Cocke, W. J., & and Holm, D. A. 1972, Nature, 240, 161 Fraix-Burnet, D., Le Borgne, J.-F. & Nieto, J.-L. 1989, A&A, 224, 17. Hardee, P. E. & Eilek, J. A. 2011, , 735, 61 Gabuzda, D. C., Pushkarev, A. B. and Cawthorne, T. V. 2000, , 319, 1109 Kharb, P., Lister, M. L., Marshall, H. L., & Hogan, B. S. 2012, , 748, 81 Kollgaard, R. I., Wardle, J. F. C., & Roberts, D. H. 1989, , 97, 1550 Laing, R. A. 1980, , 193, 439 Lister, M. L., & Homan, D. C. 2005, , 130, 1389 Lister, M. L. et al. 2009, , 138, 1874 Lyutikov, M., Pariev, V. I., & Blandford, R. D. 2003, , 597, 998 Owen, F. N., Hardee, P. E. & Cornwell, T. J. 1989, , 340, 698 Pacholczyk, A. G. 1970, [*Radio Astrophysics*]{}, Freeman (San Francisco) Pollack, L. K., Taylor, G. B., & Zavala, R. T. 2003, , 589, 733 Roberts, D. H., Wardle, J. F. C., & Marchenko, V. V. 2012, , submitted (arXiv:???) (RWM)
[^1]: Visiting Astronomer, National Radio Astronomy Observatory
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We derive a new sufficient condition for the existence of $\aleph_0$-categorical universal structures in classes of relational structures with constraints, augmenting results by Cherlin, Shelah, Chi [@CheSheChi99], and Hubička and Nešetřil [@HubNes09b].
Using this result we show that the hom-equivalence class of any countable weakly oligomorphic structure has up to isomorphism a unique model-complete smallest and greatest element, both of which are $\aleph_0$-categorical.
As the main tool we introduce the category of constraint structures, show the existence of universal homogeneous objects, and study their automorphism groups.
All constructions rest on a category-theoretic version of [Fraïssé]{}’s Theorem due to Droste and Göbel. We derive sufficient conditions for a comma category to contain a universal homogeneous object.
This research is motivated by the observation that all countable models of the theory of a weakly oligomorphic structure are hom-equivalent—a result akin to (part of) the Ryll-Nardzewski Theorem.
address:
- |
Christian Pech\
Lößnitzgrundstraße 47\
01445 Radebeul\
Germany
- |
Maja Pech\
Department of Mathematics and Informatics\
University of Novi Sad\
Trg Dositeja Obradovića 4\
21000 Novi Sad\
Serbia
author:
- Christian Pech
- Maja Pech
bibliography:
- 'PP11arxiv.bib'
title: 'Universal homogeneous constraint structures and the hom-equivalence classes of weakly oligomorphic structures'
---
Introduction {#introduction .unnumbered}
============
Weak oligomorphy is a natural weakening of the notion of oligomorphy. The latter concept was coined by Peter Cameron in the 1970$^{\text{th}}$ and it turned out to be fundamental in several fields of mathematics [@Cam90]. Not only many combinatorial enumeration problems have a natural encoding by oligomorphic permutation groups, but the Engeler-Ryll-Nardzewski-Svenonius Theorem links oligomorphic permutation groups to $\aleph_0$-categorical structures and hence to model theory [@Hod97]. Both notions, oligomorphy and $\aleph_0$-categoricity, are in turn closely related to the concept of (ultra-) homogeneity and thus with the theory of [Fraïssé]{}-limits [@Fra53; @Mac11].
In their seminal paper [@CamNes06], Peter Cameron and Jaroslav Nešetřil introduced several variations to the concept of homogeneity, one of them being homomorphism-homogeneity—saying that every homomorphism between finitely generated substructures of a given structure extends to an endomorphism of that structure. The relevance of this notion in the theory of transformation monoids on countable sets was realized quickly [@Dol12; @Dol12b; @MPPhD; @AUpaper; @Pon05]. Also a classification theory for homomorphism-homogeneous structures emerged quickly [@CamLoc10; @DolMas11; @IliMasRaj08; @JunMas12; @Mas07; @Mas12; @MasNenSko11] and classes of high complexity of finite homomorphism-homogeneous structures were discovered [@Mas12b; @RusSch10].
Weak oligomorphy is a phenomenon that arrises naturally in the context of homomorphism-homogeneity. A countable relational structure is weakly oligomorphic if its endomorphism monoid is oligomorphic, i.e., it has of every arity only finitely many invariant relations. It is not hard to see that every homomorphism-homogeneous relational structure over a finite signature is weakly oligomorphic. It is less obvious that every weakly oligomorphic relational structure has a positive existential expansion that is homomorphism-homogeneous (cf. [@MasPec11; @MPPhD; @AUpaper]). Clearly, every oligomorphic structure is weakly oligomorphic, but the reverse does not hold, in general.
It turns out that weak oligomorphy is the key for the development of a model-theory of homomorphism-homogeneous structures running in parallel to the model theory of homogeneous structures [@MasPec11; @AUpaper]. However, we think that in order to be sustainable, the theory of weak oligomorphy and homomorphism-homogeneity has not only to exist for its own sake but it has to augment the branches of mathematics from whom it was inspired. First steps into this direction were undertaken in [@AgeHHnew], where it was shown that every weakly oligomorphic structure is homomorphism-equivalent to an oligomorphic substructure. This, e.g., has the consequence that every constraint satisfaction problem with a weakly oligomorphic template is equivalent to one with an $\aleph_0$-categorical template. Another consequence is that a set of positive existential propositions is the complete positive existential theory of a weakly oligomorphic structure if and only if it is the positive existential part of an $\aleph_0$-categorical theory.
In this paper we are going further into the direction of creating cross-links between the theory of weakly oligomorphic structures and the theory of oligomorphic structures.
Our foremost interest belongs to the hom-equivalence classes of weakly oligomorphic structures. In the Section \[s2\] we give some motivations, why we are interested in the structure of hom-equivalence classes of weakly oligomorphic structures. While the Ryll-Nardzewski Theorem says that any countable oligomorphic structure is up to isomorphism determined by its first order theory, we will show that countable weakly oligomorphic structures are up to homomorphism equivalence determined by their first order theory. This property is called weak $\aleph_0$-categoricity.
In Section \[s3\], we describe the extremal elements of hom-equivalence classes of weakly oligomorphic structures with respect to inclusion. We show that every such class has a canonical largest and smallest element both of which are oligomorphic and model-complete (i.e. finite or $\aleph_0$-categorical). In fact we prove a more general result about the existence of universal elements in given classes of countable relational structures that augments results from [@CheSheChi99; @Hub10; @HubNes09].
In Section \[s4\] we introduce and study constraint structures. These are basically solutions to instances of constraint satisfaction problems. More precisely, for a template ${\mathbf{T}}$ a ${\mathbf{T}}$-colored structure is a pair $({\mathbf{A}},a)$ such that $a$ is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{T}}$. We introduce strong and weak homomorphisms between constraint structures and show the existence of universal homogeneous ${\mathbf{T}}$-colored structures. This is done using a categorical variant of [Fraïssé]{}’s theorem due to Droste and Göbel [@DroGoe92; @DroGoe93]. Universal homogeneous constraint structures are used for proving the existence result of universal elements from Section \[s3\] and thus, ultimately they give rise to universal elements in hom-equivalence classes of countable relational structures. We give conditions under which the strong or weak automorphism group of a universal homogeneous constraint structure is oligomorphic.
Universal homogeneous constraint structures are closely related to monotone free amalgamation classes (indeed, if the universe in which they live is monotone, then the age of a universal homogeneous constraint structure is monotone and free). For relational structures these two properties have strong consequences. From one hand, given that the relational signature is finite, it means that the automorphism group of the [Fraïssé]{}-limit has the small index property [@Her98; @Mac11], and hence the [Fraïssé]{}-limit of the class can be reconstructed up to first order interpretation from its automorphism group (considered as abstract group). From the other hand, for any monotone free [Fraïssé]{}-class ${\mathcal{C}}$, the class of linear ordered structures $({\mathcal{C}},\prec)$ is a Ramsey-class [@HubNes09; @NesRoe77]. In Section \[s5\] we give sufficient conditions for the (strong) automorphism group of a universal homogeneous constraint structure to have the small index property, the Bergman-property, and uncountable cofinality (cf. [@Ber06; @DroGoe05; @KecRos07]).
Finally, in Section \[s6\], we develop the necessary categorical framework for the construction of universal homogeneous constraint structures. In particular, we give sufficient conditions under which a comma-category contains a universal homogeneous object. This is the technical backbone of our approach to universal structures. Its high level of abstraction makes it perhaps seem like a tool too big for the task at hand. However, due to its categorical nature it has potentially a much wider field of application. This, in our opinion, makes it worthwhile to be included into this paper.
Preliminaries
=============
This paper deals (mainly) with relational structures. Here, under a relational structure we understand a model-theoretic structure without operations and constants. Accordingly, a relational signature is a model-theoretic signature without operational- or constant-symbols. Whenever we do not state otherwise, a relational signature can contain any number of relational symbols. Relational structures will be denoted like ${\mathbf{A}},{\mathbf{B}},{\mathbf{C}},\dots$. There carriers are denoted like $A,B,C,\dots$. Tuples are denoted like ${\bar{a}},{\bar{b}},{\bar{c}}$, and usually $a_i$ will denote the $i$-th coordinate of ${\bar{a}}$, etc.
As usual, a homomorphism between relational structures is a function between the carriers that preserves all relations.
If $f:{\mathbf{A}}\to{\mathbf{B}}$, then we call ${\mathbf{A}}$ the *domain* of $f$ and ${\mathbf{B}}$, the *codomain*. Moreover, the structure induced by $f(A)$ is called the *image* of $f$.
*Epimorphisms* are surjective homomorphisms and *monomorphisms* are injective homomorphisms. Isomorphisms are bijective homomorphisms whose inverse is a homomorphism, too. *Embeddings* are monomorphisms that not only preserve relations but also reflect them. That is, a monomorphism is an embedding if and only if it is an isomorphism to its image.
For classes ${\mathcal{A}}$, ${\mathcal{B}}$ of relational structures we write ${\mathcal{A}}\to{\mathcal{B}}$ if for every ${\mathbf{A}}\in{\mathcal{A}}$ there exists a ${\mathbf{B}}\in{\mathcal{B}}$ and a homomorphism $f:{\mathbf{A}}\to{\mathbf{B}}$. Instead of $\{{\mathbf{A}}\}\to{\mathcal{B}}$ we write ${\mathbf{A}}\to{\mathcal{B}}$ and instead of ${\mathcal{A}}\to \{{\mathbf{B}}\}$ we write ${\mathcal{A}}\to{\mathbf{B}}$. An important special case is the notion ${\mathbf{A}}\to{\mathbf{B}}$ which means that there is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$. If ${\mathbf{A}}\to{\mathbf{B}}$ and ${\mathbf{B}}\to{\mathbf{A}}$, then we call ${\mathbf{A}}$ and ${\mathbf{B}}$ *homomorphism-equivalent*.
If ${\mathcal{A}}$ is a class of relational structures over a signature $R$, then by $({\mathcal{A}},\to)$ and $({\mathcal{A}},{\hookrightarrow})$ we will denote the categories of objects from ${\mathcal{A}}$ with homomorphisms or embeddings as morphisms, respectively.
On some occasions we will need the notion of the *Gaifman-graph* of a relational structure ${\mathbf{A}}$. This is a simple graph whose vertex set is $A$, such that two vertices are joint by an edge whenever they occur together in a tuple from one of the basic relations of ${\mathbf{A}}$. We will denote the Gaifman-graph of ${\mathbf{A}}$ by $\Gamma_{\mathbf{A}}$. We call ${\mathbf{A}}$ *connected* if $\Gamma_{\mathbf{A}}$ is connected and we call ${\mathbf{A}}$ tight if its Gaifman-graph is complete.
Ages of relational structures. {#ages-of-relational-structures. .unnumbered}
------------------------------
The *age* of a relational structure ${\mathbf{A}}$ is the class of all finite relational structures that embed into ${\mathbf{A}}$ (it is denoted by ${\operatorname{Age}}({\mathbf{A}})$).
Let ${\mathcal{C}}$ be a class of finite relational structures over the same signature. We say that ${\mathcal{C}}$ has the *Joint embedding property (JEP)* if whenever ${\mathbf{A}},{\mathbf{B}}\in {\mathcal{C}}$, then there exists a ${\mathbf{C}}\in{\mathcal{C}}$ such that both ${\mathbf{A}}$ and ${\mathbf{B}}$ are embeddable in ${\mathbf{C}}$. Moreover, ${\mathcal{C}}$ has the *Hereditary property (HP)* if whenever ${\mathbf{A}}\in {\mathcal{C}}$, and ${\mathbf{B}}< {\mathbf{A}}$, then ${\mathbf{B}}$ is isomorphic to some ${\mathbf{C}}\in {\mathcal{C}}$. We say that ${\mathcal{C}}$ has the *amalgamation property (AP)* if whenever ${\mathbf{A}}, {\mathbf{B}}_1,{\mathbf{B}}_2 \in{\mathcal{C}}$, and $f_1:{\mathbf{A}}\to {\mathbf{B}}_1$ and $f_2:{\mathbf{A}}\to{\mathbf{B}}_2$ are embeddings, then there are ${\mathbf{C}}\in {\mathcal{C}}$, and embeddings $g_1:{\mathbf{B}}_1\to {\mathbf{C}}$ and $g_2:{\mathbf{B}}_2\to{\mathcal{C}}$ such that $g_1\circ f_1=g_2\circ f_2$. Finally, ${\mathcal{C}}$ has the *homo-amalgamation property (HAP)* if whenever ${\mathbf{A}}, {\mathbf{B}}_1,{\mathbf{B}}_2 \in {\mathcal{C}}$, $f_1:{\mathbf{A}}\to {\mathbf{B}}_1$ is a homomorphism, and $f_2:{\mathbf{A}}\to {\mathbf{B}}_2$ is an embedding, then there are ${\mathbf{C}}\in {\mathcal{C}}$, an embedding $g_1:{\mathbf{B}}_1\to {\mathbf{C}}$ , and a homomorphism $g_2:{\mathbf{B}}_2\to{\mathcal{C}}$ such that $g_1\circ f_1=g_2\circ f_2$.
A basic theorem by Roland [Fraïssé]{} states that a class of finite structures of the same type is the age of a countable structure if and only if
1. it has only countably many isomorphism types,
2. it is isomorphism-closed,
3. it has the (HP) and the (JEP).
A class of finite structures with these properties will be called an *age*. If ${\mathcal{C}}$ is an age, then by $\overline{{\mathcal{C}}}$ we denote the class of all countable structures whose age is contained in ${\mathcal{C}}$.
Constraint satisfaction problems. {#constraint-satisfaction-problems. .unnumbered}
---------------------------------
For a relational structure ${\mathbf{T}}$, with ${\operatorname{CSP}}({\mathbf{T}})$ we denote the class of all those finite relational structures ${\mathbf{A}}$ for which ${\mathbf{A}}\to{\mathbf{T}}$. The notion ${\operatorname{CSP}}$ comes from theoretical computer science and reads *constraint satisfaction problem*. It refers to the problem to decide whether for a given finite relational structure ${\mathbf{A}}$ there exists a homomorphism into ${\mathbf{T}}$. If ${\mathbf{T}}$ is finite over a finite signature, then ${\operatorname{CSP}}({\mathbf{T}})$ is called a finite constraint satisfaction problem. Clearly, finite CSPs are in NP. It is conjectured that every finite constraint satisfaction problem is either in P or it is NP-complete (originally, this was stated by Feder and Vardi in [@FedVar99]). Such dichotomy-conjectures are very intensively studied in theoretical computer science and in universal algebra. It is impossible to give a reasonable bibliographic overview of this topic but a few important step-stones are [@BarKoz09; @BarKozNiv08; @BerIdzMarMcKVal10; @Bod08; @Bod12; @BodNes06; @BodPin11; @BulKroJea05; @Bul06; @HelNes90; @Sch78]. In this paper we are not concerned with complexity aspects of CSPs. Rather, CSPs appear in the context of our research as analogues of ages. Yet, our structural results may be of some interest also for the research on the complexity -classification of constraint satisfaction problems. E.g., a consequence of Theorem \[univ-struc\] is that every CSP over a finite or $\aleph_0$-categorical template is equivalent to a substructure-problem and to a weak substructure problem of an $\aleph_0$-categorical template.
We will use the operator ${\operatorname{CSP}}$ for all kinds of relational structures over arbitrary relational signatures. It was observed by Feder (cf. also [@Bod08 Lemma 1]) that a class of relational structures is the constraint satisfaction problem of a structure if and only if it is closed with respect disjoint union and inverse homomorphisms (a class ${\mathcal{C}}$ of finite relational structures is closed with respect to inverse homomorphisms if whenever ${\mathbf{B}}\in{\mathcal{C}}$ and ${\mathbf{A}}$ is finite such that ${\mathbf{A}}\to{\mathbf{B}}$, then also ${\mathbf{A}}\in{\mathcal{C}}$).
Classes of finite relational structures that are closed with respect to disjoint union and inverse homomorphism will just be called *constraint satisfaction problems* or CSPs, for short. Note that every CSP over a countable signature is an age, too.
Weakly oligomorphic structures. {#weakly-oligomorphic-structures. .unnumbered}
-------------------------------
In [@Cam90], Peter Cameron defined the notion of oligomorphic permutation groups. Recall that a permutation group is called *oligomorphic* if it has just finitely many orbits in its action on $n$-tuples for every $n$. A structure ${\mathbf{A}}$ is called oligomorphic if its automorphism group is oligomorphic.
Before coming to the definition of weakly oligomorphic structures, we have to recall some model theoretic notions: Let $\Sigma$ be a relational signature, and let $L(\Sigma)$ be the language of first order logics with respect to $\Sigma$. Let ${\mathbf{A}}$ be a $\Sigma$-structure. For a formula $\varphi({\bar{x}})$ (where ${\bar{x}}=(x_1,\dots,x_n)$), we define $\varphi^{\mathbf{A}}\subseteq A^n$ as the set of all $n$-tuples ${\bar{a}}$ over $A$ such that ${\mathbf{A}}\models\varphi({\bar{a}})$. More generally, for a set $\Phi$ of formulae from $L(\Sigma)$ with free variables from $\{x_1,\dots,x_n\}$, we define $\Phi^{\mathbf{A}}$ as the intersection of all relations $\varphi^{\mathbf{A}}$ where $\varphi$ ranges through $\Phi$. We call $\Phi$ a *type*, and $\Phi^{\mathbf{A}}$ the relation defined by $\Phi$ in ${\mathbf{A}}$.
If $\Phi^{\mathbf{A}}\neq\emptyset$, then we say that ${\mathbf{A}}$ realizes $\Phi$. We call $\Phi$ positive existential if it consists just of positive existential formulae.
For a relation $\varrho\subseteq A^n$ by ${\operatorname{Tp}}_{\mathbf{A}}(\varrho)$ we denote the set of all formulae $\varphi({\bar{x}})$ such that $\varrho\subseteq \varphi^{\mathbf{A}}$. This is the type defined by $\varrho$ with respect to ${\mathbf{A}}$. Analogously, the positive existential type ${\operatorname{Tp}^{\exists_1^+}}_{\mathbf{A}}(\varrho)$ is defined. By ${\operatorname{Th}}({\mathbf{A}})$ we will denote the full first order theory of ${\mathbf{A}}$ while with ${\operatorname{Th}^{\exists_1^+}}({\mathbf{A}})$ we will denote the positive existential part of ${\operatorname{Th}}({\mathbf{A}})$.
Let us come now to the definition of the structures under consideration in this paper.
A relational structure ${\mathbf{A}}$ is called *weakly oligomorphic* if for every arity there are just finitely many relations that can be defined by positive existential types.
One can argue that it would be more appropriate to define a structure ${\mathbf{A}}$ to be weakly oligomorphic if its endomorphism monoid is *oligomorphic* (i.e. there are just finitely many invariant relations of ${\operatorname{End}}({\mathbf{A}})$ of any arity). However, there is no need to worry, since, at least for countable structures, these two definitions are equivalent:
\[prop:weakly:olig\] A countable structure ${\mathbf{A}}$ is weakly oligomorphic if and only if ${\operatorname{End}}({\mathbf{A}})$ is oligomorphic.
Clearly, if a structure is oligomorphic, then it is also weakly oligomorphic. Moreover, weakly oligomorphic structures are closed under retracts. Hence, e.g., retracts of $\aleph_0$-categorical structures are weakly oligomorphic.
First structural results for weakly oligomorphic structures were obtained in [@MasPec11; @AgeHHnew; @AUpaper].
A motivating result {#s2}
===================
In this section we will motivate in greater detail, why we are interested in studying the hom-equivalence classes of weakly oligomorphic structures. The first reason comes from the theory of homogeneous relational structures. Let us recall [Fraïssé]{}’s theorem:
A class ${\mathcal{C}}$ of finite relational structures is the age of some countable homogeneous relational structure if and only if
(i) it is closed under isomorphism,
(ii) it has only countably many non-isomorphic members,
(iii) it has the hereditary property and the amalgamation property.
Moreover, any two countable homogeneous relational structures with the same age are isomorphic.
In [@CamNes06], Peter Cameron and Jaroslav Nešetřil introduced the notion of homomorphism-homogeneous structures. *A local homomorphism* of a structure ${\mathbf{A}}$ is a homomorphism from a finite substructure of ${\mathbf{A}}$ to ${\mathbf{A}}$. A structure ${\mathbf{A}}$ is called *homomorphism-homogeneous* if every local homomorphism of ${\mathbf{A}}$ can be extended to an endomorphism of ${\mathbf{A}}$. For homomorphism-homogeneous relational structures a [Fraïssé]{}-type characterization was given in [@AgeHHnew]:
A class ${\mathcal{C}}$ of finite relational structures is the age of some countable homomorphism-homogeneous relational structure if and only if
(i) it is closed under isomorphism,
(ii) it has only countably many non-isomorphic members,
(iii) it has the hereditary property and the homo-amalgamation property.
Moreover, any two countable homomorphism-homogeneous relational structures with the same age are homomorphism-equivalent.
If we add to this the easy observation that all homomorphism-homogeneous structures over a finite signature are weakly oligomorphic (cf. [@AgeHHnew Cor.6.7]), then this gives the first reason to be interested in the hom-equivalence classes of weakly oligomorphic structures.
Our second motivation comes from the theory of $\aleph_0$-categorical structures. Let us recall the classical result that links $\aleph_0$-categoricity with oligomorphy (cf. [@Cam90 (2.10)]):
\[thm:ryll-nardzewsky\] Let ${\mathbf{A}}$ be a countably infinite relational structure. Then ${\operatorname{Th}}({\mathbf{A}})$ is $\aleph_0$-categorical if and only if ${\operatorname{Aut}}({\mathbf{A}})$ is oligomorphic.
A consequence of this characterization is that two countable oligomorphic structures have the same first order theory if and only if they are isomorphic. For countable weakly oligomorphic relational structures we will show the following related result:
\[woRN\] Let ${\mathbf{B}}$ be a weakly oligomorphic structure, and let ${\mathbf{A}}$ be a countable relational structure. Then the following are equivalent:
1. ${\mathbf{A}}\to{\mathbf{B}}$,
2. ${\operatorname{Th}^{\exists_1^+}}({\mathbf{A}})\subseteq{\operatorname{Th}^{\exists_1^+}}({\mathbf{B}})$,
3. ${\operatorname{Age}}({\mathbf{A}})\to{\operatorname{Age}}({\mathbf{B}})$,
4. ${\operatorname{CSP}}({\mathbf{A}})\subseteq{\operatorname{CSP}}({\mathbf{B}})$.
Before coming to the proof of Theorem \[woRN\], we need to recall a result from [@AgeHHnew], and to prove some additional auxiliary results:
\[prop:wo-hom\] Let ${\mathbf{A}}$, ${\mathbf{B}}$ be a relational structures over the same signature such that ${\operatorname{Age}}({\mathbf{A}})\to{\operatorname{Age}}({\mathbf{B}})$, and suppose that ${\mathbf{A}}$ is countable, and that ${\mathbf{B}}$ is weakly oligomorphic. Then ${\mathbf{A}}\to{\mathbf{B}}$.
\[lem\_wo\_real\] Let ${\mathbf{A}}$ be a weakly oligomorphic structure over the signature $R$, and let $\Psi$ be a positive existential type in the language of $R$-structures. If every finite subset of $\Psi$ is realized in ${\mathbf{A}}$, then $\Psi$ is realized in ${\mathbf{A}}$.
Suppose that every finite subset of $\Psi$ is realized in ${\mathbf{A}}$, but $\Psi$ is not. Following we will define a sequence $(\varphi_i)$ of formulae from $\Psi$, and a sequence $({\bar{d}}_i)$ such that ${\mathbf{A}}\models\varphi_j({\bar{d}}_i)$ for $1\le j \le i$, but ${\mathbf{A}}\nvDash \varphi_{i+1}({\bar{d}}_i)$.
Let $\varphi_0\in\Psi$. Then there exists a ${\bar{d}}_0\in A^m$ such that ${\mathbf{A}}\models\varphi_0({\bar{d}}_0)$. Suppose that $\varphi_i$, and ${\bar{d}}_i$ are defined already. By assumption, ${\bar{d}}_i$ does not realize $\Psi$. Let $\varphi_{i+1}\in\Psi$ such that ${\mathbf{A}}\nvDash\varphi_{i+1}({\bar{d}}_i)$. Again, by assumption, the set $\{\varphi_0,\dots,\varphi_{i+1}\}$ is realized in ${\mathbf{A}}$. Let ${\bar{d}}_{i+1}\in A^m$ a tuple that realizes $\{\varphi_0,\dots,\varphi_{i+1}\}$.
By construction, the sets $\Psi_i=\{\varphi_1,\dots,\varphi_i\}$ define an infinite decreasing chain of distinct non-empty relations in ${\mathbf{A}}$. However, this gives a contradiction with the assumption that ${\mathbf{A}}$ is weakly oligomorphic.
We conclude that $\Psi$ is realizable.
\[lem:wo\_age\] Let ${\mathbf{A}}$, and ${\mathbf{B}}$ be relational structures over the same signature. If ${\operatorname{Th}^{\exists_1^+}}({\mathbf{A}})\subseteq {\operatorname{Th}^{\exists_1^+}}({\mathbf{B}})$, and if ${\mathbf{B}}$ is weakly oligomorphic, then ${\operatorname{Age}}({\mathbf{A}})\to{\operatorname{Age}}({\mathbf{B}})$.
Let ${\mathbf{C}}$ be a finite substructure of ${\mathbf{A}}$, and let $C=\{c_1,\dots,c_n\}$ be its carrier. Define ${\bar{c}}:=(c_1,\dots,c_n)$. Then every finite subset of ${\operatorname{Tp}^{\exists_1^+}}_{\mathbf{A}}({\bar{c}})$ is realized in ${\mathbf{B}}$. Since ${\mathbf{B}}$ is weakly oligomorphic, by Lemma \[lem\_wo\_real\], we get that there is a tuple ${\bar{d}}\in B^n$ that realizes ${\operatorname{Tp}^{\exists_1^+}}_{\mathbf{A}}({\bar{c}})$. Let $D=\{d_1,\dots,d_n\}$, and let ${\mathbf{D}}$ be the substructure of ${\mathbf{B}}$ induced by $D$. Then the mapping $f:{\mathbf{C}}\to{\mathbf{D}}$ given by $c_i\mapsto d_i$ is a homomorphism. This shows that ${\operatorname{Age}}({\mathbf{A}})\to{\operatorname{Age}}({\mathbf{B}})$.
($1\Rightarrow 4$) Clear.
($4\Rightarrow 2$) It is well known (and easy to see) that for every positive primitive proposition $\varphi$ there exists a finite relational structure ${\mathbf{A}}_\varphi$ such that for any relational structure ${\mathbf{C}}$ of the given type, we have ${\mathbf{C}}\models\varphi$ if and only if ${\mathbf{A}}_\varphi\to{\mathbf{C}}$. Thus from ${\operatorname{CSP}}({\mathbf{A}})\subseteq{\operatorname{CSP}}({\mathbf{B}})$ it follows that the positive primitive theory of ${\mathbf{A}}$ is contained in the positive primitive theory of ${\mathbf{B}}$. However, this is the case if and only if ${\operatorname{Th}^{\exists_1^+}}({\mathbf{A}})\subseteq{\operatorname{Th}^{\exists_1^+}}({\mathbf{B}})$.
($2\Rightarrow 3$) This is a direct consequence of Lemma \[lem:wo\_age\].
($3\Rightarrow 1$) This follows from Lemma \[prop:wo-hom\].
Let us recall now a result from [@MasPec11]:
Let ${\mathbf{A}}$ be a countable weakly oligomorphic structure, and let ${\mathbf{B}}$ be a countable model of ${\operatorname{Th}}({\mathbf{A}})$. Then ${\mathbf{B}}$ is weakly oligomorphic, too.
We can combine this with Theorem \[woRN\] to obtain:
Let $T$ be the complete first order theory of a weakly oligomorphic structure. Then all countable models of $T$ are homomorphism-equivalent.
A first order theory, for which all countable models are homomorphism-equivalent can rightfully be called *weakly $\aleph_0$-categorical*. Hence we just proved that the first order theory of each weakly oligomorphic structure is weakly $\aleph_0$-categorical. Models of weakly $\aleph_0$-categorical theory are another motivation to study hom-equivalence classes of weakly oligomorphic structures.
Hom-equivalence classes of weakly oligomorphic structures {#s3}
=========================================================
We are going to give structure to hom-equivalence classes by equipping them with a quasi-order. There are several natural and interesting ways to do so—e.g. we may say that ${\mathbf{A}}$ is below ${\mathbf{B}}$ if ${\mathbf{A}}$ is a retract of ${\mathbf{B}}$, or if ${\mathbf{A}}$ is a homomorphic image of ${\mathbf{B}}$. In this paper we choose the embedding quasi-order (with the hope to come back in the future to the retraction-quasi-order). We define the relation of embeddability on relational structure of a given type. In particular, we write ${\mathbf{A}}{\hookrightarrow}{\mathbf{B}}$ if there exists an embedding from ${\mathbf{A}}$ into ${\mathbf{B}}$. This relation, clearly defines a quasi-order on the relational structures of a given type. We will restrict this quasi-order to hom-equivalence classes of weakly oligomorphic structures. Abusing the terminology, we will further restrict our attention solely to the finite and countably infinite structures. So for a relational structure, the hom-equivalence class ${{\mathcal{E}}}({\mathbf{A}})$ is defined to be the class of all countable relational structures ${\mathbf{B}}$ for which ${\mathbf{A}}\leftrightarrow{\mathbf{B}}$.
We are far away from completely understanding the structure embeddability-quasiorder on hom-equivalence classes of weakly oligomorphic structures. Initially, we are interested in extremal elements—i.e. such elements that embed into all other elements and such elements into which all other elements can be embedded. This promises to be most rewardable as we can expect some exceptional properties of such structures. Of course, being a quasi-order, it can happen that there are several smallest and greatest elements with respect to embeddability, and that not all such elements are of the same beauty. So we aim also to find among all smallest and greatest elements the most natural, distinguished candidates.
Smallest elements
-----------------
For the existence of smallest elements in hom-equivalence classes of weakly oligomorphic structures we only need to collect a couple of results from the literature. Recall that a structure ${\mathbf{A}}$ is called a *core* if all its endomorphisms are embeddings. Clearly, if ${\mathbf{A}}$ is a smallest element in ${{\mathcal{E}}}({\mathbf{T}})$, then ${\mathbf{A}}$ must be a core. And moreover, every core in ${{\mathcal{E}}}({\mathbf{T}})$ is a smallest element. Indeed, if ${\mathbf{B}}\in{{\mathcal{E}}}({\mathbf{T}})$ and and if $f:{\mathbf{A}}\to{\mathbf{B}}$, $g:{\mathbf{B}}\to{\mathbf{A}}$, then $g\circ f$ is an embedding. Hence $f$ is an embedding, too.
The most simple case is when ${\mathbf{T}}$ is finite. In this case up to isomorphism ${\mathbf{B}}$ has a unique retract that is a core. This is up to isomorphism the unique smallest element in ${{\mathcal{E}}}({\mathbf{T}})$.
Let us now treat the case that ${\mathbf{T}}$ is homomorphism-homogeneous and weakly oligomorphic. Then we have:
\[wo-olig\] Every countable, weakly oligomorphic, homomorphism-homogeneous structure ${\mathbf{T}}$ contains, up to isomorphism, a unique homomorphism-equivalent homomorphism-homogeneous core ${\mathbf{F}}$. Moreover, ${\mathbf{F}}$ is oligomorphic and homogeneous.
So in this case we have the existence of a smallest element as well as the existence of a distinguished smallest element (namely, it is the unique one that is homomorphism-homogeneous).
The next case is that ${\mathbf{T}}$ is a countably infinite structure with an oligomorphic automorphism group. By the Ryll-Nardzewski theorem, ${\mathbf{B}}$ is $\aleph_0$-categorical. Now we can use a result by Bodirsky:
\[bod\] Every $\aleph_0$-categorical relational structure ${\mathbf{T}}$ is homomorphism-equivalent to a model-complete core ${\mathbf{C}}$, which is unique up to isomorphism, and $\aleph_0$-categorical or finite. For all $k\ge 1$, the orbits of $k$-tuples in ${\mathbf{C}}$ are primitive positive definable.
So in this case we have the existence of a distinguished smallest element in ${{\mathcal{E}}}({\mathbf{T}})$—the unique model-complete core.
Finally, let us consider the most general case, that ${\mathbf{T}}$ is countably infinite and weakly oligomorphic. In this case we can combine Theorem \[bod\] by Bodirsky with the following result:
Let ${\mathbf{T}}$ be a countable weakly oligomorphic relational structure. Then ${\mathbf{T}}$, is homomorphism-equivalent to a finite or $\aleph_0$-categorical structure ${\mathbf{C}}$. Moreover, ${\mathbf{C}}$ embeds into ${\mathbf{A}}$.
That is, ${\mathbf{T}}$ is homomorphism-equivalent to a model complete core, which is unique up to isomorphism.
Greatest elements {#s32}
-----------------
A largest element in the hom-equivalence class of a relational structure is also known under the name of a universal object for this class. It is a long standing open problem, to characterize such classes of countable structures, that have a universal structure. Much work has been done on classes of structures with forbidden substructures. Striking results in this direction are, e.g., [@CheSheChi99; @HubNes09] (see in particular [@HubNes09] for further references).
Let ${\mathbf{T}}$ be a relational structure. Then the greatest possible age of a structure on ${{\mathcal{E}}}({\mathbf{T}})$ is ${\operatorname{CSP}}({\mathbf{T}})$. If ${\mathbf{T}}$ is weakly oligomorphic, and if ${\mathbf{A}}$ is any countable structure that is universal for $\overline{{\operatorname{CSP}}({\mathbf{T}})}$ then, by Lemma \[prop:wo-hom\], it follows that ${\mathbf{A}}$ is a greatest element of ${{\mathcal{E}}}({\mathbf{T}})$. For finite relational signatures, the existence of such universal structures follows already from the following result by Hubička and Nešetřil. In this theorem, ${{\mathit{Forb}}}_h({{\mathcal{F}}})$ denotes the class of all countable $R$-structures into which no element from ${{\mathcal{F}}}$ maps homomorphically. Moreover, a structure is called connected if its Gaifman-graph is connected.
\[Hub\] Let ${{\mathcal{F}}}$ be a countable set of finite connected relational structures over a finite signature $R$. Then ${{\mathit{Forb}}}_h({{\mathcal{F}}})$ contains a universal structure ${\mathbf{U}}_{{\mathcal{F}}}$.
Indeed, for a structure ${\mathbf{T}}$ over a finite relational signature $R$, it is easy to see that $\overline{{\operatorname{CSP}}({\mathbf{T}})}={{\mathit{Forb}}}_h({{\mathcal{F}}})$ where ${{\mathcal{F}}}$ is equal to the class of all finite $R$-structures not in ${\operatorname{CSP}}({\mathbf{T}})$. Moreover, we have that whenever a structure in ${{\mathcal{F}}}$ is disconnected, then at least one of its connected components is in ${{\mathcal{F}}}$, too, for otherwise ${{\mathit{Forb}}}_h({{\mathcal{F}}})$ would not be closed with respect to direct sums—a contradiction. It follows that if ${{\mathcal{F}}}_c$ denotes the connected structures in ${{\mathcal{F}}}$, then ${{\mathit{Forb}}}_h({{\mathcal{F}}})={{\mathit{Forb}}}_h({{\mathcal{F}}}_c)$. On the other hand, every class of the shape ${{\mathit{Forb}}}_h({{\mathcal{F}}})$ for a class of finite connected structures ${{\mathcal{F}}}$, is of the shape $\overline{{\operatorname{CSP}}({\mathbf{T}})}$ for some countable structure ${\mathbf{T}}$. Finally, since $R$ is finite, any class of finite $R$ structures contains up to isomorphism just countably many elements.
Unfortunately, in general we do not know much about the symmetries of the universal structures given by Theorem \[Hub\]. However, in the special case that ${{\mathcal{F}}}$ is finite we know of the existence of an $\aleph_0$-categorical universal structure in ${{\mathit{Forb}}}_h({{\mathcal{F}}})$ due to a result by Hubička and Nešetřil [@HubNes09 Thm.1.3] which in turn generalizes the analogous result about graphs due to Cherlin, Shelah, and Chi (cf. [@CheSheChi99 Thm.4]).
Though, we know now that the hom-equivalence class of any weakly oligomorphic structure has a largest element, in general we do not know of the existence of a “distinguished” largest element. Here, by distinguished we, ideally, understand $\aleph_0$-categorical because of the following old result by Saracino:
\[saracino\] Let $T$ be an $\aleph_0$-categorical theory with no finite models. Then $T$ has a model-companion $T'$. Moreover, $T'$ is $\aleph_0$-categorical, too.
An immediate consequence is, that for every countably infinite, $\aleph_0$-categorical structure ${\mathbf{U}}$ there exists a countably infinite $\aleph_0$-categorical structure ${\mathbf{U}}'$, such that ${\operatorname{Age}}({\mathbf{U}})={\operatorname{Age}}({\mathbf{U}}')$, and such that ${\mathbf{U}}'$ is model-complete. Moreover ${\mathbf{U}}'$ is uniquely determined, up to isomorphism.
Consequently, whenever we have an $\aleph_0$-categorical universal element in a hom-equivalence class, then we also have a distinguished $\aleph_0$-categorical universal element—namely, the model-complete one.
In order to be able to state our contribution to the existence problem of universal structures, we have to introduce the notion of strict amalgamation classes due to Dolinka (cf. [@Dol12 Sec.2.1]). Let ${\mathcal{C}}$ be a [Fraïssé]{}-class (not necessarily of relational structures). Then we say that ${\mathcal{C}}$ has the *strict amalgamation property* if every pair of morphisms in $({\mathcal{C}},{\hookrightarrow})$ with the same domain has a pushout in $(\overline{{\mathcal{C}}},\rightarrow)$. Note that these pushouts will always be amalgams. Thus the strict amalgamation property postulates canonical amalgams.
Similarly, we say that ${\mathcal{C}}$ has the *strict joint embedding property* if every pair of structures from ${\mathcal{C}}$ has a coproduct in $(\overline{{\mathcal{C}}},\rightarrow)$. If $(\overline{\mathcal{C}},{\hookrightarrow})$ has a finite initial object, then the strict joint embedding property follows from the strict amalgamation property. This is always the case for classes of relational structures.
Finally, a *strict [Fraïssé]{}-class* is a [Fraïssé]{}-class that enjoys the strict joint embedding property and the strict amalgamation property.
If ${\mathcal{U}}$ is a strict [Fraïssé]{}-class, then a [Fraïssé]{}-class ${\mathcal{C}}$ that is a subclass of ${\mathcal{U}}$, will be called *free in ${\mathcal{C}}$* if it is closed with respect to finite coproducts and with respect to canonical amalgams in ${\mathcal{U}}$.
Note that every free amalgamation class of relational structures over the signature $R$, in our terminology, is a free [Fraïssé]{}-class in the class of all finite relational $R$-structures. Moreover, every free amalgamation class is also a strict [Fraïssé]{}-class. However, there are strict [Fraïssé]{}-classes that are not free amalgamation classes. The class of finite partial orders is an example.
The following is an instantiation into the model-theoretic world of a more general category-theoretic result, that will be presented in Section \[xs6\]:
\[univ-struc\] Let ${\mathcal{U}}$ be a strict [Fraïssé]{}-class of relational structures, and let ${\mathcal{C}}$ be a [Fraïssé]{}-class that is free in ${\mathcal{U}}$. Let ${\mathbf{T}}\in\overline{\mathcal{U}}$. Then
1. $\overline{{\mathcal{C}}}\cap\overline{{\operatorname{CSP}}({\mathbf{T}})}$ has a universal element ${\mathbf{U}}_{{\mathcal{C}},{\mathbf{T}}}$,
2. if the [Fraïssé]{}-limit of ${\mathcal{C}}$ and ${\mathbf{T}}$ each have an oligomorphic automorphism group (i.e. each is finite or $\aleph_0$-categorical), then $\overline{{\mathcal{C}}}\cap\overline{{\operatorname{CSP}}({\mathbf{T}})}$ has a universal element ${\mathbf{U}}_{{\mathcal{C}},{\mathbf{T}}}$ that is finite or $\aleph_0$-categorical.
If ${\mathbf{T}}\in\overline{{\mathcal{C}}}$, then ${\mathbf{U}}_{{\mathcal{C}},{\mathbf{T}}}$ can be chosen as a co-retract of ${\mathbf{T}}$.
We will postpone the proof of Theorem \[univ-struc\] to Section \[xs3\], as we need to to construct some tools first. It is our goal to construct ${\mathbf{U}}_{{\mathcal{C}},{\mathbf{T}}}$ as the reduct of a [Fraïssé]{}-limit of a class of expanded structures (a usual method). However, our way of expanding structures will leave the domain of model theory. In the next section we will elaborate on this by introducing and studying constraint relational structures.
Before starting, let us give a direct consequence of Theorem \[univ-struc\] that is relevant to hom-equivalence classes.
Let $R$ be a countable relational signature, and let ${\mathbf{T}}$ be a countable $R$-structure. Then ${{\mathcal{E}}}({\mathbf{T}})$ has a largest element. Moreover, if $R$ is finite and ${\mathbf{T}}$ is weakly oligomorphic, then ${{\mathcal{E}}}({\mathbf{T}})$ has, up to isomorphism, a unique $\aleph_0$-categorical, model complete largest element.
Note that the class of all finite $R$-structures is a free amalgamation class, and hence a strict [Fraïssé]{}-class. Denote this class by ${\mathcal{U}}$. Let ${\mathcal{C}}:={\mathcal{U}}$. Then, by Theorem \[univ-struc\], there exists a universal element ${\mathbf{U}}_{\mathbf{T}}$ in $\overline{{\operatorname{CSP}}({\mathbf{T}})}$. Moreover, ${\mathbf{T}}$ is a retract of ${\mathbf{U}}_{\mathbf{T}}$. In particular, ${\mathbf{U}}_{\mathbf{T}}$ is in ${{\mathcal{E}}}({\mathbf{T}})$.
If $R$ is finite, then ${\mathcal{C}}$ has an $\aleph_0$-categorical [Fraïssé]{}-limit. By Proposition \[wo-olig\], it follows that ${\mathbf{T}}$ is homomorphism-equivalent to a finite or $\aleph_0$-categorical structure. So we can without loss of generality assume that the automorphism group of ${\mathbf{T}}$ is oligomorphic. However, then Theorem \[univ-struc\] gives us additionally, that ${\mathbf{U}}_{\mathbf{T}}$ can be chosen to be finite or $\aleph_0$-categorical. Since ${\operatorname{CSP}}({\mathbf{T}})$ contains structures of arbitrary size, we conclude that ${\mathbf{U}}_{\mathbf{T}}$ is $\aleph_0$-categorical. Now the existence and uniqueness of an $\aleph_0$-categorical, model-complete universal structure in ${{\mathcal{E}}}({\mathbf{T}})$ follows from Theorem \[saracino\].
Constraint structures {#s4}
=====================
Throughout this section we fix a relational signature $R$. Also we fix a universe ${\mathcal{U}}$ of countable $R$-structures which we require to be a strict [Fraïssé]{}-class.
Let ${\mathbf{T}}\in\overline{{\mathcal{U}}}$, and let ${\mathcal{C}}$ be a [Fraïssé]{}-class that is free in ${\mathcal{U}}$. A *${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$* is a pair $({\mathbf{A}},a)$ such that ${\mathbf{A}}\in\overline{{\mathcal{C}}}$ and $a:{\mathbf{A}}\to{\mathbf{T}}$ is a homomorphism. If ${\mathcal{C}}$ is known from the context, we will usually leave away “in $\overline{{\mathcal{C}}}$”.
A (strong) homomorphism $f$ between ${\mathbf{T}}$-colored structures $({\mathbf{A}},a)$ and $({\mathbf{B}},b)$ is a homomorphism from ${\mathbf{A}}$ to ${\mathbf{B}}$ such that $b\circ f=a$. A homomorphism between ${\mathbf{T}}$-colored structures is called *embedding* if it is an embedding between the corresponding $R$-structures.
With ${\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ we will denote the class of all countable ${\mathbf{T}}$-colored structures in $\overline{{\mathcal{C}}}$. As usual, by $({\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}}),\to)$ and $({\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}}),{\hookrightarrow})$ we denote the corresponding categories with homomorphisms and embeddings, respectively. Endomorphisms are homomorphisms of a ${\mathbf{T}}$-colored structure to itself and automorphisms are bijective endomorphisms whose inverse is an endomorphism, too. The group of automorphisms of $({\mathbf{A}},a)$ will be denoted by ${\operatorname{sAut}}({\mathbf{A}},a)$ (we use the notion ${\operatorname{sAut}}$, standing for “strong” automorphisms, instead of ${\operatorname{Aut}}$, because later on we will consider also weak automorphisms of ${\mathbf{T}}$-colored structures).
We have the following [Fraïssé]{}-type result for ${\mathbf{T}}$-colored structures:
\[hom-constraints\] With the notions from above, there exists a countable $({\mathbf{U}},u)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ such that
1. for every countable ${\mathbf{T}}$-colored structure $({\mathbf{A}},a)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ there exists an embedding $\iota:({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ (universality),
2. for every finite ${\mathbf{T}}$-colored structure $({\mathbf{A}},a)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$, and for all embeddings $\iota_1,\iota_2:({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ there exists an automorphism $f$ of $({\mathbf{U}},u)$ such that $f\circ\iota_1=\iota_2$ (homogeneity).
Moreover, all countable universal homogeneous ${\mathbf{T}}$-colored structures are mutually isomorphic.
For proving this theorem we are standing at a cross-road and see three ways to proceed. Either we translate the ${\mathbf{T}}$-colored structures to models and use [Fraïssé]{}’s theorem, or we prove yet another [Fraïssé]{}-type theorem especially for ${\mathbf{T}}$-colored structures, or we use an already available [Fraïssé]{}-type theorem for categories, such as the one due to Droste and Göbel (cf. [@DroGoe92]). All three possibilities require considerable technical work. We decided for the third way, as it seems to us the cleanest and it promises to be the most useful in future research. The necessary techniques are quite independent of the rest of the paper. Therefore, in order not to disturb the flow of presentation, we postpone the category theoretical part to Section \[s6\], and the actual proof of Theorem \[hom-constraints\] to Section \[xs4\].
Let us have a look onto the symmetries of universal homogeneous ${\mathbf{T}}$-colored structures:
\[hom-const-saut\] Let $({\mathbf{U}},u)$ be a universal homogeneous ${\mathbf{T}}$-colored structure in $\overline{\mathcal{C}}$. If the [Fraïssé]{}-limit of ${\mathcal{C}}$ is finite or $\aleph_0$-categorical, and if ${\mathbf{T}}$ is finite, then ${\operatorname{sAut}}({\mathbf{U}},u)$ is oligomorphic.
Let ${\bar{a}}=(a_1,\dots,a_n)$, and ${\bar{b}}=(b_1,\dots,b_n)$ be tuples of elements from $U$, such that the mapping $a_i\mapsto b_i$ is a local isomorphism of $R$-structures, and suppose that $u({\bar{a}})=u({\bar{b}})$. Let ${\mathbf{A}}$ be the substructure of ${\mathbf{U}}$ that is induced by $\{a_1,\dots,a_n\}$, $a:{\mathbf{A}}\to{\mathbf{T}}$ be the restriction of $u$ to ${\mathbf{A}}$. Let $\iota_1:({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ be the identical embedding, and define $\iota_2 : ({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ by $a_i\mapsto b_i$. Then, by the assumptions on ${\bar{a}}$ and ${\bar{b}}$, $\iota_2$ is an embedding, and by the homogeneity of $({\mathbf{U}},u)$, there exists an automorphism $f$ of $({\mathbf{U}},u)$ such that $f\circ\iota_1=\iota_2$. In particular, $f({\bar{a}})={\bar{b}}$. Since there are only finitely many isomorphism types of $n$-tuples in ${\mathbf{U}}$, and since there are only finitely many $n$-tuples in ${\mathbf{T}}$, it follows that ${\operatorname{sAut}}({\mathbf{U}},u)$ has only finitely many $n$-orbits. Hence it is oligomorphic.
So far, what we know about the symmetries of ${\mathbf{U}}$ is not enough to prove Theorem \[univ-struc\]. For an $\aleph_0$-categorical structure ${\mathbf{T}}$ the strong automorphism group of $({\mathbf{U}},u)$ does not yield enough information about ${\mathbf{U}}$. For this reason we will add morphisms to the category of ${\mathbf{T}}$-colored structures:
A *weak homomorphism* from $({\mathbf{A}},a)$ to $({\mathbf{B}},b)$ is a pair $(f,g)$ such that $f:{\mathbf{A}}\to{\mathbf{B}}$ is a homomorphism, $g$ is is an automorphism of ${\mathbf{T}}$, such that $b\circ f = g \circ a$. Weak embeddings are weak homomorphisms whose first component is an embedding. Accordingly, the group of weak automorphisms of $({\mathbf{A}},a)$ is defined like $${\operatorname{wAut}}({\mathbf{A}},a):=\{(f,g)\mid f\in{\operatorname{Aut}}({\mathbf{A}}),\,a\circ f=g\circ a\}.$$ We define $(f_1,g_1)\circ(f_2,g_2):=(f_1\circ f_2,g_1\circ g_2)$. With $({\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}}),\to_w)$ and $({\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}}),{\hookrightarrow}_w)$ we will denote the categories of ${\mathbf{T}}$-colored structures in $\overline{{\mathcal{C}}}$ with weak homomorphisms and weak embeddings, respectively.
The group of strong automorphisms embeds naturally into the group of strong automorphisms through $f\mapsto (f,1_{\mathbf{T}})$.
Clearly, the two projections $\pi_1:(f,g)\mapsto f$, $\pi_2:(f,g)\mapsto g$ are group homomorphisms from ${\operatorname{wAut}}({\mathbf{A}},a)$ to ${\operatorname{Aut}}({\mathbf{A}})$ and ${\operatorname{Aut}}({\mathbf{T}})$, respectively. The image of $\pi_1$ will be called the *color automorphism group* of $({\mathbf{A}},a)$, and it will be denoted by ${\operatorname{cAut}}({\mathbf{A}},a)$. The kernel of $\pi_2$ is isomorphic to ${\operatorname{sAut}}({\mathbf{A}},a)$.
Let us call a ${\mathbf{T}}$-colored structure $({\mathbf{U}},u)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ *w-homogeneous* if for every finite ${\mathbf{T}}$-colored structure $({\mathbf{A}},a)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$, and for all weak embeddings $(f_1,g_2),(f_2,g_2):({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ there exists a weak automorphism $(f,g)$ of $({\mathbf{U}},u)$ such that $(f,g)\circ(f_1,g_1)=(f_2,g_2)$.
Then we have the following [Fraïssé]{}-type result whose proof will be postponed till Section \[xs4\], when all technical prerequisites are provided:
\[hom-constraints2\] Let $({\mathbf{U}},u)$ be a universal homogeneous ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$. Then $({\mathbf{U}},u)$ is w-homogeneous. Moreover, all countable ${\mathbf{T}}$-colored structures in $\overline{{\mathcal{C}}}$ that are universal and $w$-homogeneous are mutually isomorphic.
Our next observation is:
Let $({\mathbf{A}},a)$ be a ${\mathbf{T}}$-colored structure in $\overline{\mathcal{C}}$, such that $a$ is an epimorphism. Then ${\operatorname{wAut}}({\mathbf{A}},a)\cong{\operatorname{cAut}}({\mathbf{A}},a)$.
It is enough to show that the projection $\pi_1$ is injective. Let $(f,g_1)$, $(f,g_2)$ be weak automorphisms of $({\mathbf{A}},a)$. Then, by definition $a\circ f=g_1\circ a= g_2\circ a$. Since $a$ is an epimorphism, it follows that $g_1=g_2$.
Let us now reexamine, when universal, homogeneous ${\mathbf{T}}$-colored structures give rise to $\aleph_0$-categorical structures:
\[hom-const-waut\] Let $({\mathbf{U}},u)$ be a universal, homogeneous ${\mathbf{T}}$-colored structure in $\overline{\mathcal{C}}$. If the [Fraïssé]{}-limit of ${\mathcal{C}}$ is finite or $\aleph_0$-categorical, and if ${\mathbf{T}}$ has an oligomorphic automorphism group, then ${\operatorname{cAut}}({\mathbf{U}},u)$ is oligomorphic.
Let ${\bar{a}}=(a_1,\dots,a_n)$, and ${\bar{b}}=(b_1,\dots,b_n)$ be tuples of elements from $U$, such that the mapping $a_i\mapsto b_i$ is a local isomorphism of $R$-structures, and suppose that $u({\bar{a}})$ and $u({\bar{b}})$ are in the same $n$-orbit of ${\operatorname{Aut}}({\mathbf{T}})$.
Let ${\mathbf{A}}$ be the substructure of ${\mathbf{U}}$ that is induced by the entries of ${\bar{a}}$. Let $\iota$ be the identical embedding of ${\mathbf{A}}$ into ${\mathbf{U}}$, and let $f:{\mathbf{A}}\to{\mathbf{U}}$ given by $f(a_i)=b_i$. Clearly, both $\iota$ and $f$ are homomorphisms (of $R$-structures). Let $a:=\iota\circ h$. Then $({\mathbf{A}},\iota\circ u)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$. Clearly, $(\iota,1_{\mathbf{C}}):({\mathbf{A}},\iota\circ u){\hookrightarrow}({\mathbf{U}},u)$. Since $u({\bar{a}})$ and $u({\bar{b}})$ are in the same $n$-orbit of ${\operatorname{Aut}}({\mathbf{T}})$, there exists a $g\in {\operatorname{Aut}}({\mathbf{T}})$ that maps $u({\bar{a}})$ to $u({\bar{b}})$. It is now easy to see that $(f,g):({\mathbf{A}},\iota\circ u){\hookrightarrow}({\mathbf{U}},u)$ is a weak embedding. Hence, by w-homogeneity of $({\mathbf{U}},u)$, there exists a weak automorphism $(v,w)$ of $({\mathbf{U}},u)$ such that $(v,w)\circ(\iota,1_{\mathbf{C}})=(f,g)$. In particular, $v$ is a color automorphism of $({\mathbf{U}},u)$ that maps ${\bar{a}}$ to ${\bar{b}}$.
Since there are only finitely many isomorphism types of $n$-tuples in ${\mathbf{U}}$, and since ${\operatorname{Aut}}({\mathbf{T}})$ has only finitely many $n$-orbits, it follows that ${\operatorname{cAut}}({\mathbf{U}}, u)$ has only finitely many $n$-orbits. Hence it is oligomorphic.
\[retract\] Suppose that ${\mathbf{T}}\in\overline{{\mathcal{C}}}$. Let $({\mathbf{U}},u)$ be a universal homogeneous ${\mathbf{T}}$-colored structure in $\overline{\mathcal{C}}$. Then $u:{\mathbf{U}}\to{\mathbf{T}}$ is a retraction.
Since ${\mathbf{T}}\in\overline{{\mathcal{C}}}$, we have that $({\mathbf{T}},1_{\mathbf{T}})\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$. Hence, there exists an embedding $\iota:({\mathbf{T}},1_{\mathbf{T}}){\hookrightarrow}({\mathbf{U}},u)$. That is, we have $u\circ\iota=1_{\mathbf{T}}$. Thus, $\iota$ is a right-inverse of $u$, and $u$ is a retraction.
An immediate consequence is:
Suppose that ${\mathbf{T}}\in\overline{{\mathcal{C}}}$. Let $({\mathbf{U}},u)$ be a universal homogeneous ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$. Then ${\operatorname{cAut}}({\mathbf{U}},u)\cong{\operatorname{wAut}}({\mathbf{U}},u)$.
In the following it will be our goal to represent the color automorphism group of a universal homogeneous ${\mathbf{T}}$-colored structure as the automorphism group of a relational structure. First, let us expand the signature $R$ of ${\mathbf{T}}$ by another binary relational symbol $\kappa$ and denote the new signature by $\tilde{R}$. For a ${\mathbf{T}}$-colored structure $({\mathbf{A}},a)$ define $\widetilde{{\mathbf{A}}}$ to be the $\tilde{R}$ structure whose $R$-reduct is equal to ${\mathbf{A}}$ and in which the relational symbol $\kappa$ is interpreted as the kernel of $a$.
\[locally\_closed\] Suppose that ${\mathbf{T}}\in\overline{{\mathcal{C}}}$. Let $({\mathbf{U}},u)$ be a universal homogeneous ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$. Then ${\operatorname{cAut}}({\mathbf{U}},u)={\operatorname{Aut}}(\widetilde{{\mathbf{U}}})$ (in particular, ${\operatorname{cAut}}({\mathbf{U}},u)$ is closed in $S_\infty$).
Let $\kappa$ be the kernel of $u$. For every $\varrho\in R$ we introduce a new symbol $\hat{\varrho}$ and define $\hat{\varrho}_{\mathbf{U}}:=\{{\bar{a}}\mid u({\bar{a}})\in\varrho_{\mathbf{T}}\}$. Let $\widehat{{\mathbf{U}}}$ be the structure obtained from ${\mathbf{U}}$ by expansion through $\kappa$ and through all $\hat{\varrho}_{\mathbf{U}}$. We claim that ${\operatorname{Aut}}(\widehat{{\mathbf{U}}})={\operatorname{cAut}}({\mathbf{U}},u)$.
Let $f\in{\operatorname{cAut}}({\mathbf{U}},u)$. Then there exists a $g\in{\operatorname{Aut}}({\mathbf{T}})$ such that $g\circ u= u\circ f$. If $(a,b)\in\kappa$ then $(a,b)\in\ker(g\circ u)$ hence $(a,b)\in\ker(u\circ f)$ whence $(f(a),f(b))\in\kappa$. Let ${\bar{a}}\in\hat{\varrho}_{\mathbf{U}}$. Then $g(u({\bar{a}}))\in \varrho_{\mathbf{T}}$. Hence $u(f({\bar{a}}))\in\varrho_{\mathbf{T}}$, whence $f({\bar{a}})\in\hat{\varrho}_{\mathbf{A}}$. We conclude that $f\in{\operatorname{Aut}}(\widehat{{\mathbf{U}}})$.
Let now $f\in{\operatorname{Aut}}(\widehat{{\mathbf{U}}})$. Let us define a binary relation $\gamma$ on $U$ by $\gamma:=\{(u(a),u(f(a)))\mid a\in U\}$. We claim that $\gamma$ is the graph of an automorphism of ${\mathbf{T}}$. If $(x,y_1),(x,y_2)\in\gamma$ then there exists $a_1,a_2\in U$ such that $u(a_1)=u(a_2)=x$ and $u(f(a_1))=y_1$, $u(f(a_2))=y_2$. But then $(a_1,a_2)\in \kappa$ and since $f$ preserves $\kappa$, we get $y_1=y_2$. Since $u$ is surjective (recall that it is a retraction of ${\mathbf{U}}$ onto ${\mathbf{T}}$), it follows that $\gamma$ is the graph of a function. Suppose now that $(x_1,y),(x_2,y)\in\gamma$. Then there are $a_1,a_2\in U$ such that $u(a_1)=x_1$, $u(a_2)=x_2$, $u(f(a_1))=u(f(a_2))=y$. Hence $(f(a_1),f(a_2))\in\kappa$. Since $f^{-1}$ is an automorphism of $\widehat{{\mathbf{U}}}$, we have $(a_1,a_2)\in\kappa$ whence $x_1=x_2$. It follows that $\gamma$ is the graph of an injective function. Since $f$ is a bijection, it follows that $\gamma$ is the graph of a bijective function. Let $g$ be this function. Clearly, by construction we have $g\circ u=u\circ f$.
It remains to show that $g$ is an automorphism of ${\mathbf{T}}$. Let $\varrho\in R^{(n)}$ and ${\bar{c}}\in\varrho_{\mathbf{T}}$. Let ${\bar{a}}\in U^n$ such that $u({\bar{a}})={\bar{c}}$. Then ${\bar{a}}\in\hat{\varrho}_{\mathbf{U}}$. Hence $f({\bar{a}})\in\hat{\varrho}_{\mathbf{U}}$ and $u(f({\bar{a}}))\in\varrho_{\mathbf{T}}$. By construction of $g$ we have that $u(f({\bar{a}}))=g({\bar{c}})$. Hence $g\in{\operatorname{Aut}}({\mathbf{T}})$. We conclude that $f\in{\operatorname{cAut}}({\mathbf{U}},u)$.
Finally, we have to show that the $\tilde{R}$-reduct $\widetilde{{\mathbf{U}}}$ of $\widehat{{\mathbf{U}}}$ has the same automorphism group like $\widehat{{\mathbf{U}}}$. It is enough to show that the relations $\hat{\varrho}_{\mathbf{U}}$ are first order definable in $\widetilde{{\mathbf{U}}}$. For $\varrho\in R^{(n)}$ consider the formula $$\varphi_\varrho(x_1,\dots,x_n) \equiv \exists x_{n+1}\dots\exists x_{2n} \bigwedge_{i=1}^n \kappa(x_i,x_{i+n}) \land \varrho(x_{n+1},\dots x_{2n}).$$ Let $\varphi_{\varrho,\widetilde{{\mathbf{U}}}}$ be the interpretation of $\varphi_\varrho$ in $\widetilde{{\mathbf{U}}}$. We claim that $\varphi_{\varrho,\widetilde{{\mathbf{U}}}}=\hat\varrho_{\mathbf{U}}$. Clearly, $\varphi_{\varrho,\widetilde{{\mathbf{U}}}}\subseteq\hat\varrho_{\mathbf{U}}$. For the other inclusion we use that $u$ is a retraction onto ${\mathbf{T}}$. Let $\iota:{\mathbf{T}}{\hookrightarrow}{\mathbf{U}}$ be a co-retraction of $u$. Let ${\bar{a}}\in \hat{\varrho}_{\mathbf{U}}$. Let ${\bar{b}}:=\iota(u({\bar{a}}))$. Then ${\bar{b}}\in\varrho_{\mathbf{U}}$, and $u({\bar{b}})=u({\bar{a}})$. Hence for all $1\le i\le n$ we have that $(a_i,b_i)\in\kappa_{\mathbf{U}}$. Hence ${\bar{a}}\in \varphi_{\varrho,\widetilde{{\mathbf{U}}}}$.
A class of ${\mathbf{T}}$-colored structures with the small index property {#s5}
==========================================================================
Following [@DolMas12; @KecRos07], we say that a [Fraïssé]{}-class ${\mathcal{C}}$ has the *Hrushovski property* if for every ${\mathbf{A}}\in{\mathcal{C}}$ there exists a ${\mathbf{B}}\in{\mathcal{C}}$ such that ${\mathbf{A}}\le{\mathbf{B}}$ and such that every isomorphism between substructures of ${\mathbf{A}}$ extends to an automorphism of ${\mathbf{B}}$. We say that a homogeneous structure ${\mathbf{F}}$ has the Hrushovski property if ${\operatorname{Age}}({\mathbf{F}})$ does.
As the name suggests, the Hrushovski property was first proved for the class of finite simple graphs by Hrushovski in [@Hru92].
Hodges, Hodkinson, Lascar, and Shelah showed how the Hrushovski property is related to the small index property: Let $G$ be a permutation group acting on a countable set. Then $G$ is said to have the *small index property* if every subgroup of index less than $2^{\aleph_0}$ contains the stabilizer of a finite tuple. A structure ${\mathbf{A}}$ is said to have the small index property if its automorphism group does. Herwig, in [@Her98], generalizing Hrushovski’s ideas, gave a sufficient condition for relational structures over finite signatures to have the Hrushovski property and the small index property. Let us now reproduce some important notions from [@Her98].
Let $R$ be a relational signature. A finite $R$-structure ${\mathbf{A}}$ is called a *link-structure*, if either $|A|=1$ or there exist $a_1,\dots,a_n\in A$ such that $A=\{a_1,\dots,a_n\}$ and for some $\varrho\in R^{(n)}$ we have $(a_1,\dots,a_n)\in\varrho_{{\mathbf{A}}}$ (note that the $a_i$ do not need to be distinct).
If ${\mathcal{L}}$ is a set of link-structures, then we say that a structure ${\mathbf{A}}$ has *link type* ${\mathcal{L}}$ if every substructure of ${\mathbf{A}}$ that is a link structure, is isomorphic to some structure from ${\mathcal{L}}$.
Let ${{\mathcal{F}}}$ be a set of finite $R$-structures. Then an $R$-structure ${\mathbf{A}}$ is called *${{\mathcal{F}}}$-free*, if no member of ${{\mathcal{F}}}$ homomorphically maps to ${\mathbf{A}}$.
If ${\mathcal{L}}$ is a set of link-structures and ${{\mathcal{F}}}$ is a set of finite $R$-structures, then by ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ denotes the class of all finite ${{\mathcal{F}}}$-free $R$-structures of link-type ${\mathcal{L}}$.
A finite $R$ structure ${\mathbf{A}}$ is called *packed* if any two distinct elements of $A$ lie in a tuple of some basic relation of ${\mathbf{A}}$ (in other words, the Gaifman-graph of ${\mathbf{A}}$ is the complete graph).
Now we are ready to formulate Herwig’s criterion:
\[herwig\] Let $R$ be a finite relational signature. Let ${{\mathcal{F}}}$ be a set of finite packed $R$-structures. Let ${\mathcal{L}}$ be a set of link-structures. Then ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ is a free amalgamation class that has the Hrushovski property. Moreover, the automorphism group of the [Fraïssé]{}-limit of ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ has ample generics and the small index property.
We are not going to define the notion of ample generics but only remark that this property (originating from ideas from [@Las91; @Tru92], cf. also [@Tru07]) is central to showing the small index property of a permutation group.
Note that the classes of the shape ${\mathcal{K}}_{\emptyset{{\mathcal{F}}}}$ are precisely the so called monotone free amalgamation classes (cf. [@Her98 Lem.14]). Another way to put this is, that a CSP has the amalgamation property if and only if it can be represented like ${\mathcal{K}}_{\emptyset{{\mathcal{F}}}}$ for some set ${{\mathcal{F}}}$ of finite packed structures.
We are not happy about the requirement that the relational signature has to be finite. In order to overcome this problem, we have to restrict the possible link structures. Let $R$ be an arbitrary relational signature. A finite $R$-structure ${\mathbf{A}}$ is called *sparse* if all but finitely many relational symbols from $R$ are interpreted by $\emptyset$ in ${\mathbf{A}}$. An arbitrary $R$-structure is called sparse if all its finite substructures are sparse. For an $R$-structure ${\mathbf{A}}$, by $R({\mathbf{A}})$ we will denote the set of all relational symbols in $R$ that have a non-empty interpretation in ${\mathbf{A}}$.
Now we can formulate our criterion:
\[extherwig\] Let $R$ be a relational signature. Let ${{\mathcal{F}}}$ be a set of finite packed $R$-structures. Let ${\mathcal{L}}$ be a set of sparse link-structures. Then ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ has the Hrushovski property. If ${\mathcal{L}}$ is countable, then ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ is a free amalgamation class whose [Fraïssé]{}-limit has the small index property.
Let ${\mathbf{A}}\in{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$, and let $p_1,\dots,p_n$ be isomorphisms between substructures of ${\mathbf{A}}$. Let $R':=R({\mathbf{A}})$. Since ${\mathbf{A}}$ is sparse, it follows that $R'$ is finite. Let ${\mathcal{L}}'$ be the subclass of of all link-structures ${\mathbf{L}}\in{\mathcal{L}}$ for which $R({\mathbf{L}})\subseteq R'$. Similarly, let ${{\mathcal{F}}}'$ be the subclass of all structure ${\mathbf{K}}\in{{\mathcal{F}}}$ for which $R({\mathbf{K}})\subseteq R'$. Then ${\mathcal{K}}_{{\mathcal{L}}'{{\mathcal{F}}}'}\subseteq {{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ and ${\mathbf{A}}\in{\mathcal{K}}_{{\mathcal{L}}'{{\mathcal{F}}}'}$. By Theorem \[herwig\], it follows that ${\mathcal{K}}_{{\mathcal{L}}'{{\mathcal{F}}}'}$ has the Hrushovski property. Hence, there is a finite superstructure ${\mathbf{B}}$ of ${\mathbf{A}}$ in ${\mathcal{K}}_{{\mathcal{L}}'{{\mathcal{F}}}'}$ such that $p_1,\dots,p_n$ extend to automorphisms of ${\mathbf{B}}$. But we have that ${\mathbf{B}}$ is also an element of ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$. Hence ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ has the Hrushovski property, too.
Clearly, ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ has the hereditary property and it has free amalgams. However, in order to form a [Fraïssé]{}-limit, we need in addition, that it contains up to isomorphism just countably many structures. However, this is assured by the requirement, that ${\mathcal{L}}$ should be countable. In this case, from the Hrushovski property of ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ and from the fact that ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ is a free amalgamation class, it follows that the automorphism group of the [Fraïssé]{}-limit of ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$ has ample generics and the small index property (cf. [@Her98 Prop.8,Thm.11]).
The Hrushovski property can be straight forwardly defined also for [Fraïssé]{}-classes of ${\mathbf{T}}$-colored structures with respect to strong embeddings: We say that ${\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ the *Hrushovski property* if for every finite $({\mathbf{A}},a)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ there exists a finite $({\mathbf{B}},b)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ such that $({\mathbf{A}},a)\le({\mathbf{B}},b)$ and such that every strong isomorphism between substructures of $({\mathbf{A}},a)$ extends to a strong automorphism of $({\mathbf{B}},b)$. Our main result in this section will be:
\[SIP\] Let $R$ be a relational signature. Let ${{\mathcal{F}}}$ be a set of finite packed $R$-structures and let ${\mathcal{L}}$ be a countable set of sparse link-structures. Let ${\mathbf{T}}$ be any countable $R$-structure. Then ${\operatorname{Col}}_{{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}}({\mathbf{T}})$ has the Hrushovski property. If $({\mathbf{U}},u)$ is a universal homogeneous object in ${\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$, then ${\operatorname{sAut}}({\mathbf{U}},u)$ has the small index property.
The first step will be to encode ${\mathbf{T}}$-colored structures as relational structures over an extended signature $\hat{R}$: For every $t\in T$ we add a new unary relation-symbol $M_t$ to $R$, obtaining a new signature $\hat{R}$. Now, to every ${\mathbf{T}}$-colored structure $({\mathbf{A}},a)$ we associate an $\hat{R}$-structure $S({\mathbf{A}},a)$ by setting $M_{t,S({\mathbf{A}},a)}:=a^{-1}(t)$. Clearly this process is one-to-one. That is, we can reconstruct $({\mathbf{A}},a)$ from $S({\mathbf{A}},a)$. Moreover, $f:({\mathbf{A}},a){\hookrightarrow}({\mathbf{B}},b)$ if and only if $f:S({\mathbf{A}},a){\hookrightarrow}S({\mathbf{B}},b)$.
Of course, not all $\hat{R}$-structures can be obtained in this way. Let ${\mathcal{L}}':=\{({\mathbf{L}},l)\mid {\mathbf{L}}\in{\mathcal{L}},\,l:{\mathbf{L}}\to{\mathbf{T}}\}$ and let $\widehat{\mathcal{L}}:=\{S({\mathbf{L}},l)\mid ({\mathbf{L}},l)\in{\mathcal{L}}'\}$. We can consider ${{\mathcal{F}}}$ as a set of $\hat{R}$ structures in a natural way, by interpreting all additional relational symbols with the empty set. We will show now that $S$ induces a concrete isomorphism between $({\operatorname{Col}}_{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}({\mathbf{T}}),{\hookrightarrow})$ and $(\overline{{\mathcal{K}}_{\widehat{\mathcal{L}}{{\mathcal{F}}}}},{\hookrightarrow})$. Let $({\mathbf{A}},a)\in{\operatorname{Col}}_{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}({\mathbf{T}})$. Then every link-structure of ${\mathbf{A}}$ induces a ${\mathbf{T}}$-colored substructure that is in ${\mathcal{L}}'$. Hence it induces a structure from $\widehat{\mathcal{L}}$ in $S({\mathbf{A}},a)$. Moreover, if ${\mathbf{F}}\in{{\mathcal{F}}}$, and if $h:{\mathbf{F}}\to S({\mathbf{A}},a)$, then $h:{\mathbf{F}}\to {\mathbf{A}}$—a contradiction. Hence $S({\mathbf{A}},a)\in \overline{{\mathcal{K}}_{\widehat{\mathcal{L}}{{\mathcal{F}}}}}$. It remains to show that $S$ is surjective. Let $\widehat{\mathbf{B}}\in\overline{{\mathcal{K}}_{\widehat{\mathcal{L}}{{\mathcal{F}}}}}$, and let ${\mathbf{B}}$ be its $R$-reduct. Clearly, ${\mathbf{B}}\in\overline{{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}}$. Let $u\in \widehat{\mathbf{B}}$. Then $u$ induces a link-structure from $\widehat{\mathcal{L}}$. Hence there exists a unique $t_u\in T$ such that $u\in M_{t_u,\widehat{\mathbf{B}}}$. Define $b:B\to T$ by $u\mapsto t_u$. We claim that $b:{\mathbf{B}}\to{\mathbf{T}}$. Suppose this was not the case. Then there exists a relational symbol $\varrho\in R^{(n)}$ and a tuple ${\bar{u}}\in B^n$ such that ${\bar{u}}\in\varrho_{\mathbf{B}}$ but $b({\bar{u}})\notin\varrho_{\mathbf{T}}$. However, then the elements of ${\bar{u}}$ induce a link-structure in $\widehat{{\mathbf{B}}}$ that is not in $\widehat{\mathcal{L}}$—contradiction. Thus $({\mathbf{B}},b)$ is a ${\mathbf{T}}$-colored structure in $\overline{{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}}$. By construction we have $S({\mathbf{B}},b)=\widehat{{\mathbf{B}}}$. Moreover, all elements of $\widehat{\mathcal{L}}$ are sparse. Also, since ${\mathbf{T}}$ is countable, so is $\widehat{\mathcal{L}}$.
From Proposition \[extherwig\], it follows that ${\mathcal{K}}_{\widehat{\mathcal{L}}{{\mathcal{F}}}}$ has the Hrushovski property. Hence ${\operatorname{Col}}_{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}({\mathbf{T}})$ has the Hrushovski property. Moreover, the [Fraïssé]{}-limit of ${\mathcal{K}}_{\widehat{\mathcal{L}}{{\mathcal{F}}}}$ has the small index property. Let $({\mathbf{U}},u)$ be a universal homogeneous ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$. Then $S({\mathbf{U}},u)$ is a universal homogeneous structure in $\overline{{\mathcal{K}}_{\widehat{\mathcal{L}}{{\mathcal{F}}}}}$. By the construction of $S$ we have ${\operatorname{sAut}}({\mathbf{U}},u)={\operatorname{Aut}}(S({\mathbf{U}},u))$. Hence ${\operatorname{sAut}}({\mathbf{U}},u)$ has the small index property.
If we assume that ${\mathbf{T}}$ is an element of ${{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$, then we can make much stronger statements. From one hand we observe that in this case there exists a finite tuple ${\bar{a}}$ over $U$, such that the stabilizer of ${\bar{a}}$ in ${\operatorname{cAut}}({\mathbf{U}},u)$ is contained in ${\operatorname{sAut}}({\mathbf{U}},u)$. Hence ${\operatorname{cAut}}({\mathbf{U}},u)$ has the small index property, too.
On the other hand, if ${\mathbf{T}}\in{{\mathcal{K}_{\mathcal{L}\mathcal{F}}}}$, then ${\operatorname{sAut}}({\mathbf{U}},u)$ is oligomorphic. Moreover, as it was shown above, it is closed in $S_\infty$ and has ample generics (we refer to [@KecRos07] for a definition of this notion). From this it follows that ${\operatorname{sAut}}({\mathbf{U}},u)$ is not equal to the union of any $\omega$-chain of proper subgroups (i.e., it has uncountable cofinality, cf. the remark after Theorem 6.12 of [@KecRos07]). Moreover, from [@KecRos07 Thm.6.19] it follows that in this case ${\operatorname{sAut}}({\mathbf{U}},u)$ is $21$-Bergman. That is, whenever $W_0\subseteq W_1 \subseteq W_2\subseteq\dots\subseteq {\operatorname{sAut}}({\mathbf{U}},u)$ is an exhaustive sequence of of subsets of ${\operatorname{sAut}}({\mathbf{U}},u)$, then there is an $n$, such that $W_n^{21}={\operatorname{sAut}}({\mathbf{U}},u)$. In particular, ${\operatorname{sAut}}({\mathbf{U}},u)$ has the Bergman-property (cf. [@Ber06; @DroGoe05]).
Universal homogeneous objects in comma categories {#s6}
=================================================
In this Section we will provide the main tool needed for proving Theorems \[hom-constraints\] and \[hom-constraints2\]. The exposition is more or less independent from the rest of the paper. The main result will be a theorem that gives sufficient conditions that a comma-category contains a universal homogeneous object. This result relies on an earlier categorical version of [Fraïssé]{}’s theorem due to Droste and Göbel [@DroGoe92].
For the convenience of the reader, this part is kept relatively self-contained. For basic notions from category theory we refer to [@HCA1].
Comma-categories
----------------
Recall the definition of the comma categories:
Let ${\mathfrak{A}}$,${\mathfrak{B}}$,${\mathfrak{C}}$ be categories, let $F:{\mathfrak{A}}\to{\mathfrak{C}}$, $G:{\mathfrak{B}}\to{\mathfrak{C}}$ be functors. The arrow category $(F \downarrow G)$ has as objects triples $(A,f,B)$ where $A\in{\mathfrak{A}}$, $B\in{\mathfrak{B}}$, $f:FA\to GB$. The morphisms from $(A,f,B)$ to $(A',f',B')$ are pairs $(a,b)$ and such that $a:A\to A'$, $b:B\to B'$ such that the following diagram commutes: $$\begin{psmatrix}
FA & GB\\
FA' & GB'
\ncline{->}{1,1}{1,2}^{f}
\ncline{->}{2,1}{2,2}^{f'}
\ncline{->}{1,1}{2,1}<{Fa}
\ncline{->}{1,2}{2,2}<{Gb}
\end{psmatrix}$$
There are two projection functors $U:(F \downarrow G)\to {\mathfrak{A}}$ and $V:(F \downarrow G)\to {\mathfrak{B}}$ defined by $U:(A,f,B)\mapsto A, (a,b)\mapsto a$ and $V:(A,f,B)\mapsto B, (a,b)\mapsto b$. Moreover there is a canonical natural transformation $\alpha: F\circ U\to G\circ V$ defined by $\alpha_{(A,f,B)}=f$ (cf. [@HCA1 Prop.1.6.2]). Finally, the comma-category has the following universal property:
With the notions from above, let ${\mathfrak{D}}$ be another category, let $U':{\mathfrak{D}}\to{\mathfrak{A}}$, $V':{\mathfrak{D}}\to{\mathfrak{B}}$ be functors, and let $\alpha':FU'\Rightarrow GV'$ be a natural transformation. Then there exists a unique functor $W:{\mathfrak{D}}\to(F \downarrow G)$ such that $UW=U'$, $VW=V'$, $\alpha*1_W=\alpha'$ (where $\alpha*1_W$ denotes the horizontal composition of $\alpha$ and $1_W$, i.e. $\alpha'_D=\alpha_{WD}$ for all $D\in{\mathfrak{D}}$).
The following lemmata show how arrow categories behave with respect to (weak) colimits. Nothing here is really new. But since it is not easy to find these facts in literature, we give them here with proof.
\[weak\] With the notions from above, let ${\mathfrak{D}}$ be a small category, and let $H:{\mathfrak{D}}\to (F \downarrow G)$. Suppose that
1. $U\circ H$ has a weak colimit $(L,(p_D)_{D\in{\mathfrak{D}}})$,
2. $V\circ H$ has a compatible cocone $(M,(q_D)_{D\in{\mathfrak{D}}})$,
3. $(FL,(Fp_D)_{D\in{\mathfrak{D}}})$ is a weak colimit of $FUH$.
Then there is a morphism $h: FL\to GM$ such that $((L,h,M), (p_D,q_D)_{D\in{\mathfrak{D}}})$ is a compatible cocone for $H$. In case that $(FL,(Fp_D)_{D\in{\mathfrak{D}}})$ is a colimit of $FUH$, $h$ is unique.
Recall $\alpha:FU\Rightarrow GV$ with $\alpha_{(A,f,B)}=f$ is a natural transformation. Consider $\alpha':=\alpha * 1_H:FUH\Rightarrow GVH$ given by $\alpha'_D=\alpha_{HD}$. Then $(GM,(Gq_D\circ\alpha'_D)_{D\in {\mathfrak{D}}})$ is a compatible cocone for $FUH$. Hence there is a morphism $h: FL\to GM$ such that for every $D\in {\mathfrak{D}}$ we have $Gq_D\circ\alpha'_D=h\circ Fp_D$. So indeed, $(p_D,q_D):HD\to (L,h,M)$. In case that $(FL,(Fp_D)_{D\in{\mathfrak{D}}})$ is a colimit, there is a unique such $h$ such that $(p_D,q_D):HD\to(L,h,M)$ is a morphism.
Let $d: D\to D'$ be a morphism of ${\mathfrak{D}}$. Then $HD=(UHD, \alpha'_D, VHD)$ and $HD'=(UHD',\alpha'_{D'},VHD')$. Moreover, $Hd=(UHd,VHd)$. So we have that the following diagram commutes: $$\begin{psmatrix}[rowsep=1cm]
&[name=FUHD]FUHD & [name=GVHD]GVHD\\
[name=FL]FL & & & [name=GM]GM\\
&[name=FUHD']FUHD' & [name=GVHD']GVHD'\\[-1cm]
\ncline{->}{FUHD}{GVHD}^{\alpha'_D}
\ncline{->}{FUHD'}{GVHD'}^{\alpha'_{D'}}
\ncline{->}{FUHD}{FUHD'}<{FUHd}
\ncline{->}{GVHD}{GVHD'}<{GVHd}
\ncline{->}{FUHD}{FL}^{Fp_D}
\ncline{->}{FUHD'}{FL}_{Fp_{D'}}
\ncline{->}{GVHD}{GM}^{Gq_D}
\ncline{->}{GVHD'}{GM}_{Gq_{D'}}
\ncarc[arcangle=-90]{->}{FL}{GM}^h
\end{psmatrix}$$ It follows that $(p_D,q_D)=(p_{D'},q_{D'})\circ Hd$.
\[colimits\] With the notions from above, let ${\mathfrak{D}}$ be a small category, and let $H:{\mathfrak{D}}\to (F \downarrow G)$. Suppose that
1. $U\circ H$ has a colimit $(L,(p_D)_{D\in{\mathfrak{D}}})$,
2. $V\circ H$ has a colimit $(M,(q_D)_{D\in{\mathfrak{D}}})$,
3. $(FL,(Fp_D)_{D\in{\mathfrak{D}}})$ is a colimit of $FUH$,
Then there is a unique morphism $h:FL\to GM$ such that $((L,h,M),(p_D,q_D)_{d\in{\mathfrak{D}}})$ is a compatible cocone of $H$. Moreover, this cocone is a colimit of $H$.
From Lemma \[weak\] it follows that there is a morphism $h:FL\to GM$ such that $((L,h,M),(p_D,q_D)_{D\in{\mathfrak{D}}})$ is a compatible cocone for $H$. The uniqueness of $h$ follows from Lemma \[weak\], too. So it remains to show that $((L,h,M),(p_D,q_D)_{D\in{\mathfrak{D}}})$ is a colimit.
Let $((L',h',M'), (p'_D,q'_D)_{D\in{\mathfrak{D}}})$ be another compatible cocone for $H$. Then $(L',(p'_D)_{D\in{\mathfrak{D}}})$ is a compatible cocone for $UH$, and $(M',(q'_D)_{D\in{\mathfrak{D}}})$ is a compatible cocone for $VH$. Hence, there are unique morphisms $r:L\to L'$ and $s:M\to M'$ such that $r\circ p_D=p'_D$ and $s\circ q_D=q'_D$, for all $D\in{\mathfrak{D}}$. We will show that $(r,s):(L,h,M)\to (L',h',M')$ is the unique mediating morphism. First we need to show that it is a morphism at all: For this, we use that $(FL,(Fp_D)_{D\in{\mathfrak{D}}})$ is a colimit of $FUH$. Consider the following diagram: $$\begin{psmatrix}
& [name=FL] FL & [name=GM] GM\\
& [name=FL'] FL' & [name=GM'] GM'\\
[name=FUHD]FUHD & & & [name=GVHD] GVHD
\ncline{->}{FL}{GM}^h
\ncline{->}{FL'}{GM'}^{h'}
\ncline{->}{FUHD}{GVHD}^{\alpha_D}
\ncline{->}{FL}{FL'}<{Fr}
\ncline{->}{GM}{GM'}>{Gs}
\ncline{->}{FUHD}{FL'}^{Fp'_D}
\ncline{->}{GVHD}{GM'}^{Gq'_D}
\ncarc[arcangle=30]{->}{FUHD}{FL}<{Fp_D}
\ncarc[arcangle=-30]{->}{GVHD}{GM}>{Gq_D}
\end{psmatrix}$$ The lower quadrangle commutes, because $(p'_D,q'_D)$ is a morphism. We already saw that the two triangles commute. Note that $(GM', (Gs\circ h\circ Fp_D)_{D\in{\mathfrak{D}}})$ is a compatible cocone for $FUH$ with the mediating morphism $Gs\circ h$. Now we compute $$\begin{aligned}
h'\circ Fr\circ Fp_D &= h'\circ Fp'_D\\
&= Gq'_D\circ\alpha_D \\
&= Gs\circ Gq_D\circ\alpha_D\\
&= Gs\circ h \circ Fp_D.
\end{aligned}$$ Hence, $h'\circ Fr$ is another mediating morphism and we conclude that $h'\circ Fr= Gs\circ h$ and hence $(r,s)$ is a morphism. We already noted, that the two triangles of the above given diagram commute. However, this means that $(r,s)$ is mediating. Let us show the uniqueness of $(r,s)$:
Suppose that $(r',s')$ is another mediating morphism. Then $U(r',s')=r'$ is a mediating morphism between $(L, (p_D)_{D\in{\mathfrak{D}}})$ and $(L', (p'_D)_{D\in{\mathfrak{D}}})$, and $V(r',s')=s'$ is a mediating morphism between $(M, (q_D)_{D\in{\mathfrak{D}}})$ and $(M', (q'_D)_{D\in{\mathfrak{D}}})$. Hence $r=r'$ and $s=s'$.
Algebroidal categories
----------------------
The notion of algebroidal categories goes back to Banaschewski and Herrlich [@BanHer76]. We need this concept in order to be able to make use of the category-theoretic version of [Fraïssé]{}’s theorem due to Droste and Göbel [@DroGoe92]. We closely follow the exposition from [@DroGoe92].
Let $\lambda$ be a regular cardinal. Let us consider $\lambda$ as a category. A *$\lambda$-chain* in ${\mathfrak{C}}$ is a functor from $\lambda$ to ${\mathfrak{C}}$. An object $A$ of ${\mathfrak{C}}$ is called *$\lambda$-small* if whenever $(S,(f_i)_{i\in\lambda})$ is the colimit of a $\lambda$-chain $F$ in ${\mathfrak{C}}$, and $h:A\to S$, then there exists a $j\in\lambda$, and a morphism $g:A\to F(j)$ such that $h=f_j\circ g$. With ${\mathfrak{C}}_{<\lambda}$ we will denote the full subcategory of ${\mathfrak{C}}$ whose objects are all $\lambda$-small objects of ${\mathfrak{C}}$. The category ${\mathfrak{C}}$ will be called *semi-$\lambda$-algebroidal* if all $\mu$-chains in ${\mathfrak{C}}_{<\lambda}$ have a colimit in ${\mathfrak{C}}$, and if every object of ${\mathfrak{C}}$ is the colimit of a $\lambda$-chain in ${\mathfrak{C}}_{<\lambda}$. Moreover, ${\mathfrak{C}}$ will be called *$\lambda$-algebroidal* if
1. it is semi-$\lambda$-algebroidal,
2. ${\mathfrak{C}}_{<\lambda}$ contains at most $\lambda$ isomorphism classes of objects, and
3. between any two objects from ${\mathfrak{C}}_{<\lambda}$ there are at most $\lambda$ morphisms.
Let us now have a look onto $\lambda$-small objects in comma-categories.
\[smallthings\] Let ${\mathfrak{A}}$, ${\mathfrak{B}}$, ${\mathfrak{C}}$ be categories, such that ${\mathfrak{A}}$ and ${\mathfrak{B}}$ have colimits of $\lambda$-chains and such that all morphisms of ${\mathfrak{B}}$ are monomorphisms. Let $F:{\mathfrak{A}}\to{\mathfrak{C}}$ be $\lambda$-continuous, and let $G:{\mathfrak{B}}\to{\mathfrak{C}}$ any functor that preserves monomorphisms.
Let $(A,f,B)$ be an object of $(F \downarrow G)$ such that $A$ is $\lambda$-small in ${\mathfrak{A}}$ and $B$ is $\lambda$-small in ${\mathfrak{B}}$. Then $(A,f,B)$ is $\lambda$-small in $(F \downarrow G)$.
Let $H:\lambda\to (F \downarrow G)$ be a $\lambda$-chain, and let $((L,h,M),(a_i,b_i)_{i\in\lambda})$ be a colimit of $H$. Let $(a,b):(A,f,B)\to(L,h,M)$ be a morphism. Since ${\mathfrak{A}}$ and ${\mathfrak{B}}$ have colimits of $\lambda$-chains, it follows that $U\circ H$, and $V\circ H$ have colimits, and since $F$ is $\lambda$-continuous, by Lemma \[colimits\], we have that $(L,(a_i)_{i\in\lambda})$ is a colimit of $U\circ H$, and $(M,(b_i)_{i\in\lambda})$ is a colimit of $V\circ H$. Since $A$ is $\lambda$-small in ${\mathfrak{A}}$ and $a:A\to L$, there exists an $i\in\lambda$, and a morphism $\hat{a}:A\to UHi$ such that $a_i\circ\hat{a}=a$. Also, since $B$ is $\lambda$-small in ${\mathfrak{B}}$, and since $b:B\to M$, it follows that there exists some $j\in\lambda$ and a morphism $\hat{b}:B\to VHi$ such that $b_i\circ\hat{b}=b$. Without loss of generality, $i=j$. Consider the following diagram: $$\begin{psmatrix}
[name=FL]FL & & & [name=GM]GM\\
& [name=FA]FA & [name=GB]GB\\
[name=FUHi]FUHi & & & [name=GVHi]GVHi
\ncline{->}{FL}{GM}^h
\ncline{->}{FA}{GB}^f
\ncline{->}{FUHi}{GVHi}^{h_i}
\ncline{->}{FA}{FUHi}^{F\hat{a}}
\ncline{>->}{GB}{GVHi}^{G\hat{b}}
\ncline{->}{FA}{FL}^{Fa}
\ncline{>->}{GB}{GM}^{Gb}
\ncline{->}{FUHi}{FL}<{Fa_i}
\ncline{>->}{GVHi}{GM}<{Gb_i}
\end{psmatrix}$$ where $Hi=(UHi,h_i,VHi)$. By the assumptions, the upper quadrangle and the two triangles of this diagram commute. We compute $$\begin{aligned}
Gb_i\circ h_i\circ F\hat{a} &= h\circ Fa_i\circ F\hat{a}\\
&=h\circ Fa\\
&=Gb\circ f\\
&=Gb_i\circ G\hat{b}\circ f
\end{aligned}$$ Since $Gb_i$ is a monomorphism, we conclude that $G\hat{b}\circ f=h_i\circ F\hat{a}$ whence the whole diagram commutes. Hence $(\hat{a},\hat{b}):(A,f,B)\to Hi$ and $(a_i,b_i)\circ(\hat{a},\hat{b})=(a,b)$.
Such $\lambda$-small objects $(A,f,B)$ in $(F \downarrow G)$ for which $A$ and $B$ are $\lambda$-small in ${\mathfrak{A}}$ and ${\mathfrak{B}}$, respectively, will be called *inherited $\lambda$-small objects*. In principle, there may be non-inherited $\lambda$-small objects in $(F \downarrow G)$.
\[inherited\] With the notions from above, if in $(F \downarrow G)$ every object is the colimit of a $\lambda$-chain of inherited $\lambda$-small objects, then every $\lambda$-small object of $(F \downarrow G)$ is inherited.
Let $(A,f,B)$ be $\lambda$-small in $(F \downarrow G)$, let $H:\lambda\to (F \downarrow G)$ such that $Hi=(A_i,f_i,B_i)$ is inherited $\lambda$-small for all $i\in\lambda$, and such that $((A,f,B),(p_i,q_i)_{i\in\lambda})$ is a colimit of $H$. Consider the identity morphism $(1_A,1_B)$ of $(A,f,B)$. Since $(A,f,B)$ is $\lambda$-small, there is some $i\in\lambda$ and some $(a,b):(A,f,B)\to(A_i,f_i,B_i)$ such that $(1_A,1_B)=(p_i,q_i)\circ(a,b)$. In other words, $(A,f,B)$ is a retract of $(A_i,f_i,B_i)$. It follows that $A$ is a retract of $A_i$ and $B$ is a retract of $B_i$. Now it is easy to see that retracts of $\lambda$-small objects are $\lambda$-small. Hence $A$ is $\lambda$-small in ${\mathfrak{A}}$, and $B$ is $\lambda$-small in ${\mathfrak{B}}$. By Lemma \[smallthings\], it follows that $(A,f,B)$ is inherited.
\[semialgebroidal\] Let ${\mathfrak{A}},{\mathfrak{B}},{\mathfrak{C}}$ categories such that ${\mathfrak{A}}$ and ${\mathfrak{B}}$ are semi-$\lambda$-algebroidal, and such that all morphisms of ${\mathfrak{B}}$ are monomorphisms. Let $F:{\mathfrak{A}}\to{\mathfrak{C}}$, $G:{\mathfrak{B}}\to{\mathfrak{C}}$ be $\lambda$-continuous functors such that $F$ preserves $\lambda$-smallness and $G$ preserves monos. Then $(F \downarrow G)$ is semi-$\lambda$-algebroidal.
Let us first show, that every object of $(F \downarrow G)$ is the colimit of a $\lambda$-chain of inherited $\lambda$-small objects: Let $(A,f,B)\in (F \downarrow G)$. Since ${\mathfrak{A}}$ is semi-$\lambda$-algebroidal, there is a $\lambda$-chain $H$ of $\lambda$-small objects and morphisms $a_i:Hi\to A$ (for all $i\in\lambda$) such that $(A,(a_i)_{i\in\lambda})$ is a colimit of $H$. Similarly, since ${\mathfrak{B}}$ is semi-$\lambda$-algebroidal, there is a $\lambda$-chain $K$ in of $\lambda$-small objects in ${\mathfrak{B}}$ and a family of morphisms $b_i: Ki\to B$ ($i\in\lambda$), such that $(B,(b_i)_{i\in\lambda})$ is a colimit of $K$. Since $G$ is $\lambda$-continuous, we have that $(GB, (Gb_i)_{i\in\lambda})$ is a colimit of $GK$. Since $F$ preserves $\lambda$-smallness, we have that $FHi$ is $\lambda$-small. Hence, there exists a $j=j(i)$ and $b_i: FHi\to GKj(i)$ such that the following diagram commutes: $$\begin{psmatrix}
[name=FA]FA & [name=GB]GB\\
[name=FHi]FHi & [name=GKji]GKj(i)
\ncline{->}{FA}{GB}^f \ncline{->}{FHi}{GKji}_{h_i} \ncline{->}{FHi}{FA}<{Fa_i}
\ncline{>->}{GKji}{GB}>{Gb_{j(i)}} \end{psmatrix}$$ Whenever a factoring morphism of $f\circ Fa_i$ exists through $GKj$, then it exists also through $GKj'$ for all $j'>j$. Hence the function $J:\lambda\to\lambda: i\mapsto j(i)$ can be chosen to be increasing in a way that the sequence $(j(i))_{i\in\lambda}$ is cofinal in $\lambda$. By taking $K':= KJ$, we have that $\chi:=(h_i)_{i\in\lambda}$ is a natural transformation from $FH$ to $GJK$. Moreover, by cofinality, we have that $(B, (b_{j(i)})_{i\in\lambda})$ is a colimit of $JK$. By the universal property of the comma-categories, there exists a unique functor $W:\lambda\to (F \downarrow G)$ such that $UW=H$, $VW=K'$, $\alpha*W=\chi$. It follows that $(A,f,B)$ is a colimit of $W$ and it follows from Lemma \[smallthings\] that $Wi$ is $\lambda$-small for all $i\in\lambda$.
It remains to show that $(F \downarrow G)$ has colimits of all $\mu$-chains for $\mu<\lambda$. However, this is a direct consequence of Lemma \[colimits\].
In the proof of Proposition \[semialgebroidal\] we showed that every object of $(F \downarrow G)$ is the colimit of a $\lambda$-chain of inherited $\lambda$-small objects. From Lemma \[inherited\] it follows that under the assumptions of Proposition \[semialgebroidal\], all $\lambda$-small objects of $(F \downarrow G)$ are inherited. This enables us, to formulate the following result:
\[algebroidal\] Let ${\mathfrak{A}},{\mathfrak{B}},{\mathfrak{C}}$ categories such that ${\mathfrak{A}}$ and ${\mathfrak{B}}$ are $\lambda$-algebroidal, and such that all morphisms of ${\mathfrak{B}}$ are monomorphisms. Let $F:{\mathfrak{A}}\to{\mathfrak{C}}$, $G:{\mathfrak{B}}\to{\mathfrak{C}}$ be $\lambda$-continuous functors such that $F$ preserves $\lambda$-smallness and $G$ preserves monos. Additionally, suppose that for all $\lambda$-small objects $A\in{\mathfrak{A}}_{<\lambda}$, $B\in{\mathfrak{B}}_{<\lambda}$ there are at most $\lambda$ morphisms between $FA$ and $GB$. Then $(F \downarrow G)$ is $\lambda$-algebroidal.
Proposition \[semialgebroidal\] we have that $(F \downarrow G)$ is semi-$\lambda$-algebroidal. We already noted, that all $\lambda$-small objects of $(F \downarrow G)$ are inherited. By this reason, the number of $\lambda$-small objects in $(F \downarrow G)$ is at most $\lambda^3=\lambda$. Also, the number of morphisms between $\lambda$-small objects of $(F \downarrow G)$ is at most $\lambda^2=\lambda$. Hence, $(F \downarrow G)$ is $\lambda$-algebroidal.
The Droste-Göbel-machine
------------------------
In [@DroGoe92; @DroGoe93], Droste and Göbel developed a categorical version of a classical model theoretic theorem by [Fraïssé]{} that characterizes universal homogeneous countable structures. This generalization is staged in $\lambda$-algebroidal categories, and we need to introduce a few more notions in order to be able to state it.
In the following, let ${\mathfrak{C}}$ be a category in which all morphisms are monomorphisms. Let ${\mathfrak{C}}^*$ be a full subcategory of ${\mathfrak{C}}$.
Let $U\in{\mathfrak{C}}$. Then we say that
$U$ is ${\mathfrak{C}}^*$-universal
: if for every $A\in{\mathfrak{C}}^*$ there is a morphism $f:A\to U$,
$U$ is ${\mathfrak{C}}^*$-homogeneous
: if for every $A\in{\mathfrak{C}}^*$ and for all morphisms $f,g:A\to U$ there exists an automorphism $h$ of $U$ such that $h\circ f=g$,
We say that
${\mathfrak{C}}^*$ has the joint embedding property
: if for all $A,B\in{\mathfrak{C}}^*$ there exists a $C\in{\mathfrak{C}}^*$ and morphisms $f:A\to C$ and $g:B\to C$,
${\mathfrak{C}}^*$ has the amalgamation property
: if for all $A$, $B$, $C$ from ${\mathfrak{C}}^*$ and $f:A\to B$, $g:A\to C$, there exists $D\in{\mathfrak{C}}^*$ and $\hat{f}:C\to D$, $\hat{g}:B\to D$ such that the following diagram commutes: $$\begin{psmatrix}
[name=A]A & [name=B]B\\
[name=C]C & [name=D]D.
\ncline{->}{A}{B}^f \ncline{->}{A}{C}<g \ncline{->}{C}{D}^{\hat{f}} \ncline{->}{B}{D}<{\hat{g}} \end{psmatrix}$$
We call a category ${\mathfrak{C}}$ a *$\lambda$-amalgamation category* if
1. all morphisms of ${\mathfrak{C}}$ are monomorphisms,
2. ${\mathfrak{C}}$ is $\lambda$-algebroidal,
3. ${\mathfrak{C}}_{<\lambda}$ has the joint embedding property,
4. ${\mathfrak{C}}_{<\lambda}$ has the amalgamation property.
Let us state now the result by Droste and Göbel:
\[DroGoe\] Let $\lambda$ be a regular cardinal, and let ${\mathfrak{C}}$ be a $\lambda$-algebroidal category in which all morphisms are monomorphisms. Then there exists a ${\mathfrak{C}}$-universal, ${\mathfrak{C}}_{<\lambda}$-homogeneous object in ${\mathfrak{C}}$ if and only if ${\mathfrak{C}}$ is a $\lambda$-amalgamation category. Moreover, any two ${\mathfrak{C}}$-universal, ${\mathfrak{C}}_{<\lambda}$-homogeneous objects in ${\mathfrak{C}}$ are isomorphic.
A [Fraïssé]{}-type theorem for comma-categories {#xs6}
-----------------------------------------------
Before we can come to the formulation of a sufficient condition that the comma-category of two functors has a universal homogeneous object, we need to introduce some more notions.
Let $\widehat{\mathfrak{A}}$ be a category and let ${\mathfrak{A}}\le\widehat{\mathfrak{A}}$ be a subcategory. We say that ${\mathfrak{A}}$ is *isomorphism closed* in $\widehat{\mathfrak{A}}$ if for all $A\in{\mathfrak{A}}$ and for every isomorphism $f\in\widehat{\mathfrak{A}}(A\to B)$ we have that $B\in{\mathfrak{A}}$ and $f\in{\mathfrak{A}}(A\to B)$.
We say that ${\mathfrak{A}}$ has the *free joint embedding property in $\widehat{\mathfrak{A}}$* if for all $A,B\in{\mathfrak{A}}$ there exist $C\in{\mathfrak{A}}$, and $f\in{\mathfrak{A}}(A\to C)$, and $g\in{\mathfrak{A}}(B\to C)$ such that $(C,f,g)$ is a weak coproduct in $\widehat{\mathfrak{A}}$.
We say that ${\mathfrak{A}}$ has the *free amalgamation property in $\widehat{\mathfrak{A}}$* if for all $A,B,C\in{\mathfrak{A}}$ and for all $f\in{\mathfrak{A}}(A\to B)$, $g\in{\mathfrak{A}}(A\to C)$ there exists a $D\in{\mathfrak{A}}$ and $\hat f\in{\mathfrak{A}}(C\to D)$, $\hat g\in{\mathfrak{A}}(B\to D)$ such that the following diagram is a weak pushout-square in $\widehat{\mathfrak{A}}$: $$\begin{psmatrix}
B & D \\
A & C
\ncline{->}{1,1}{1,2}^{\hat g}
\ncline{->}{2,1}{2,2}^{g}
\ncline{<-}{1,1}{2,1}<{f}\ncline{<-}{1,2}{2,2}<{\hat f}\end{psmatrix}$$
A pair of categories $({\mathfrak{A}},\widehat{\mathfrak{A}})$ is called a $\lambda$-amalgamation pair if
1. ${\mathfrak{A}}\le\widehat{\mathfrak{A}}$ is isomorphism closed,
2. all morphisms of ${\mathfrak{A}}$ are monomorphisms,
3. ${\mathfrak{A}}$ is $\lambda$-algebroidal,
4. ${\mathfrak{A}}_{<\lambda}$ has the free joint embedding property in $\widehat{\mathfrak{A}}$, and
5. ${\mathfrak{A}}_{<\lambda}$ has the free amalgamation property in $\widehat{\mathfrak{A}}$.
$\lambda$-amalgamation pairs capture the idea of free amalgamation classes and of strict [Fraïssé]{}-classes, that we talked about in Section \[s32\].
Now we are ready to link up our previous observations in the following result:
\[mainconstruction\] Let $(\widehat{\mathfrak{A}},{\mathfrak{A}})$ be a $\lambda$-amalgamation pair, ${\mathfrak{B}}$ be a $\lambda$-amalgamation category, and let ${\mathfrak{C}}$ be a category. Let $\hat F:\widehat{\mathfrak{A}}\to{\mathfrak{C}}$, $G:{\mathfrak{B}}\to{\mathfrak{C}}$ and let $F$ be the restriction of $\hat{F}$ to ${\mathfrak{A}}$. Further suppose that
1. $\hat F$ preserves weak coproducts and weak pushouts in ${\mathfrak{A}}_{<\lambda}$,
2. $F$ and $G$ are $\lambda$-continuous,
3. $F$ preserves $\lambda$-smallness,
4. $G$ preserves monomorphisms,
5. for every $A\in{\mathfrak{A}}_{<\lambda}$ and for every $B\in {\mathfrak{B}}_{<\lambda}$ there are at most $\lambda$ morphisms in ${\mathfrak{C}}(FA\to GB)$.
Then $(F \downarrow G)$ has a $(F \downarrow G)$-universal, $(F \downarrow G)_{<\lambda}$-homogeneous object. Moreover, up to isomorphism there is just one such object in $(F \downarrow G)$.
By construction, all morphisms of $(F \downarrow G)$ are monomorphisms. From Proposition \[algebroidal\], it follows that $(F \downarrow G)$ is $\lambda$-algebroidal.
Let $(A_1,f_1,B_1),(A_2,f_2,B_2)\in(F \downarrow G)_{<\lambda}$. Then $A_1,A_2\in F_{<\lambda}$, and $B_1,B_2\in{\mathfrak{B}}_{<\lambda}$. Since ${\mathfrak{A}}_{<\lambda}$ has the free joint embedding property in $\widehat{\mathfrak{A}}$, it follows that there exists a $C\in{\mathfrak{A}}_{<\lambda}$, $p_{A_1}\in{\mathfrak{A}}_{<\lambda}(A_1\to C)$, $p_{A_2}\in{\mathfrak{A}}_{<\lambda}(A_2\to C)$, such that $(C,p_{A_1},p_{A_2})$ is a weak coproduct of $A_1$ and $A_2$ in $\widehat{\mathfrak{A}}$. By assumption we have $(\hat FC,\hat Fp_{A_1},\hat Fp_{A_2})$ is a weak coproduct of $\hat FA_1$ and $\hat FA_2$ in ${\mathfrak{C}}$. On the other hand, from the amalgamation property of ${\mathfrak{B}}_{<\lambda}$ it follows that there exists $M\in{\mathfrak{B}}_{<\lambda}$ and morphisms $q_{B_1}\in{\mathfrak{B}}(B_1\to M)$ and $q_{B_2}\in{\mathfrak{B}}(B_2\to M)$. From Lemma \[weak\], it follows that there exists an $h:\hat FC\to GM$ such that $(p_{A_1},q_{B_1}):(A_1,f_1,B_1)\to(C,h,M)$ and $(p_{A_2},q_{B_2}):(A_1,f_1,B_1)\to(C,h,M)$. In other words, $(F \downarrow G)_{<\lambda}$ has the joint embedding property.
Analogously, it can be shown that $(F \downarrow G)_{<\lambda}$ has the amalgamation property. Now the existence and uniqueness of an $(F \downarrow G)$-universal, and $(F \downarrow G)_{<\lambda}$-homogeneous object in $(F \downarrow G)$ follows from Theorem \[DroGoe\].
Tying up loose ends
===================
Missing proofs from Section \[s4\] {#xs4}
----------------------------------
Let ${\mathfrak{C}}:=(\overline{{\mathcal{U}}},\to)$, ${\mathfrak{A}}:=(\overline{{\mathcal{C}}},{\hookrightarrow})$, $\widehat{\mathfrak{A}}:=(\overline{{\mathcal{C}}},\to)$ and let ${\mathfrak{B}}$ be the category that has just one object $I$ and one morphism $1_I$. Let $\hat F:\widehat{\mathfrak{A}}\to{\mathfrak{C}}$ be the identical embedding, $F$ be the restriction of $\hat F$ to ${\mathfrak{A}}$, and let $G:{\mathfrak{B}}\to{\mathfrak{C}}$ be the unique functor that maps $I$ to ${\mathbf{T}}$. Then a routine check shows that the conditions of Theorem \[mainconstruction\] are fulfilled (with $\lambda=\aleph_0$). It remains to note that $({\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}}),{\hookrightarrow})$ is isomorphic to $(F \downarrow G)$.
For the proof of Theorem \[hom-constraints2\] we need some more preparations:
Let $G$ be a subgroup of ${\operatorname{Aut}}({\mathbf{T}})$. Let $({\mathbf{A}},a)$ and $({\mathbf{B}},b)$ be ${\mathbf{T}}$-colored structures in $\overline{{\mathcal{C}}}$. A $G$-embedding is a weak embedding $(f,g)$ such that $g\in G$. A countable ${\mathbf{T}}$-colored structure $({\mathbf{U}},u)$ in $\overline{{\mathcal{C}}}$ is called *$G$-universal* if for every $({\mathbf{A}},a)\in{\operatorname{Col}}_{\mathbf{C}}({\mathbf{T}})$ there exists a $G$-embedding from $({\mathbf{A}},a)$ to $({\mathbf{U}},u)$ Moreover, we call $({\mathbf{U}},h)$ *$G$-homogeneous* if for all finite $({\mathbf{A}},a)\in{\operatorname{Col}}_{\mathcal{C}}({\mathbf{T}})$ and all $G$-embeddings $(f_1,g_1),(f_2,g_2):({\mathbf{A}},a)\to({\mathbf{U}},u)$ there exists a $G$-automorphism $(f_3,g_3)$ of $({\mathbf{U}},u)$ such that $(f_3,g_3)\circ(f_1,g_1)=(f_2,g_2)$.
\[guniversal\] With the notions from above, let $G$ be a countable subgroup of ${\operatorname{Aut}}({\mathbf{T}})$ . Then there exists a unique (up to isomorphism) $G$-universal and $G$-homogeneous structure $({\mathbf{U}},u)$ in $\overline{{\mathcal{C}}}$.
Let ${\mathfrak{C}}:=(\overline{{\mathcal{U}}},\to)$, $\widehat{\mathfrak{A}}:=(\overline{{\mathcal{C}}},\to)$, ${\mathfrak{A}}:=(\overline{{\mathcal{C}}},{\hookrightarrow})$, and let ${\mathfrak{B}}$ be the category that has just one object ${\mathbf{T}}$ and whose morphisms are the elements of $G$. Now let $\hat F:\widehat{\mathfrak{A}}\to{\mathfrak{C}}$, $G:{\mathfrak{B}}\to{\mathfrak{C}}$ be identical embeddings, respectively. Let $F$ be the restriction of $\hat F$ to ${\mathfrak{A}}$, and let ${\mathfrak{D}}:=(F \downarrow G)$.
A routine check shows that $(\widehat{\mathfrak{A}},{\mathfrak{A}})$, ${\mathfrak{B}}$, ${\mathfrak{C}}$, $\hat F$, $G$ fulfill the assumptions of Theorem \[mainconstruction\] (with $\lambda=\aleph_0$). Hence, ${\mathfrak{D}}$ has (up to isomorphism) a unique ${\mathfrak{D}}$-universal and ${\mathfrak{D}}_{<\aleph_0}$-homogeneous object $({\mathbf{U}},u,{\mathbf{T}})$.
Clearly, $({\mathbf{U}},u)$ is $G$-universal and $G$-homogeneous.
\[equal-univ\] Let $({\mathbf{U}},u)$ be a $G$-universal, $G$-homogeneous ${\mathbf{T}}$-colored structure in $\overline{\mathcal{C}}$ (here $G$ is an arbitrary subgroup of ${\operatorname{Aut}}({\mathbf{T}})$). Then $({\mathbf{U}},u)$ is also a universal, homogeneous ${\mathbf{T}}$-colored structure.
Let $(\widehat{{\mathbf{U}}},\hat{u})$ be a countable universal, homogeneous ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$. Then, there exists a $G$-embedding $(f,g):(\widehat{{\mathbf{U}}},\hat{u}){\hookrightarrow}({\mathbf{U}},u)$. Now let $({\mathbf{A}},a)$ be any countable ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$. Then there exists an embedding $\iota:({\mathbf{A}},g^{-1}\circ a)$ into $(\widehat{{\mathbf{U}}},\hat{u})$. But then $f\circ\iota:({\mathbf{A}},g^{-1}\circ a){\hookrightarrow}({\mathbf{U}},g^{-1}\circ u)$ is an embedding. Hence also $f\circ\iota:({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ is an embedding. Consequently, $({\mathbf{U}}, u)$ is universal.
Let now $({\mathbf{A}},a)$ be any finite ${\mathbf{T}}$-colored structure in $\overline{\mathcal{C}}$ and let $f_1,f_2:({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ be embeddings. Then $(f_1,1_{\mathbf{T}})$ and $(f_2,1_{\mathbf{T}})$ are $G$-embeddings. Since $({\mathbf{U}},u)$ is $G$-homogeneous, there is a $G$-automorphism $(h,g)$ of $({\mathbf{U}},u)$ such that $(h,g)\circ(f_1,1_{\mathbf{T}})=(f_2,1_{\mathbf{T}})$. That is, $h\circ f_1=f_2$, and $g\circ 1_{\mathbf{T}}=1_{\mathbf{T}}$. In other words, $g=1_{\mathbf{T}}$ and $h$ is an automorphism of $({\mathbf{U}},u)$. Hence $({\mathbf{U}},u)$ is homogeneous.
A consequence of Proposition \[equal-univ\] is, that the construction of the universal homogeneous ${\mathbf{T}}$-colored structure is essentially equivalent to the construction of $G$-universal, $G$-homogeneous ${\mathbf{T}}$-colored structure in $\overline{\mathcal{C}}$. However, the latter construction uncovers more symmetries.
Now we are ready to prove Theorem \[hom-constraints2\]:
Let $({\mathbf{U}},u)$ be a universal homogeneous ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$.
Let $({\mathbf{A}},a)$ be a finite ${\mathbf{T}}$-colored structure in $\overline{{\mathcal{C}}}$, and let $(f_1,g_1), (f_2,g_2):({\mathbf{A}},a){\hookrightarrow}({\mathbf{U}},u)$ be weak embeddings. Let $G$ be the subgroup of ${\operatorname{Aut}}({\mathbf{T}})$ that is generated by $g_1$ and $g_2$. Since $G$ is countable, we have, by Proposition \[guniversal\] that $({\mathbf{U}},u)$ is $G$-homogeneous. It follows that there is a $G$-automorphism $(f,g)$ of $({\mathbf{U}},u)$ such that $(f,g)\circ(f_1,g_1)=(f_2,g_2)$.
Thus, $({\mathbf{U}},u)$ is w-homogeneous.
Any two countable universal w-homogeneous ${\mathbf{T}}$-colored structures in $\overline{{\mathcal{C}}}$ are homogeneous, by Proposition \[equal-univ\]. Hence, by Theorem \[hom-constraints\], they are isomorphic.
Missing proofs from Section \[s3\] {#xs3}
----------------------------------
About 1
: From Theorem \[hom-constraints\], it follows that there exists a universal ${\mathbf{T}}$-colored structure $({\mathbf{U}},u)$ in $\overline{{\mathcal{C}}}$. So we can choose ${\mathbf{U}}_{{\mathcal{C}},{\mathbf{T}}}:={\mathbf{U}}$.
About 2
: From Proposition \[hom-const-waut\], it follows that in this case ${\operatorname{cAut}}({\mathbf{U}},u)$ is oligomorphic. Since ${\operatorname{cAut}}({\mathbf{U}},u)\le{\operatorname{Aut}}({\mathbf{U}})$, it follows that ${\operatorname{Aut}}({\mathbf{U}})$ is oligomorphic, too.
Finally, if ${\mathbf{T}}\in\overline{{\mathcal{C}}}$, then from Proposition \[retract\], it follows that $u:{\mathbf{U}}\to{\mathbf{T}}$ is a retraction.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'IP Geolocation databases are widely used in online services to map end user IP addresses to their geographical locations. However, they use proprietary geolocation methods and in some cases they have poor accuracy. We propose a systematic approach to use publicly accessible reverse DNS hostnames for geolocating IP addresses. Our method is designed to be combined with other geolocation data sources. We cast the task as a machine learning problem where for a given hostname, we generate and rank a list of potential location candidates. We evaluate our approach against three state of the art academic baselines and two state of the art commercial IP geolocation databases. We show that our work significantly outperforms the academic baselines, and is complementary and competitive with commercial databases. To aid reproducibility, we open source our entire approach.'
author:
- Ovidiu Dan
- Vaibhav Parikh
- 'Brian D. Davison'
bibliography:
- 'sigproc.bib'
title: IP Geolocation through Reverse DNS
---
<ccs2012> <concept> <concept> <concept\_id>10002951.10003227.10003236.10003101</concept\_id> <concept\_desc>Information systems Location based services</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10003033.10003099.10003101</concept\_id> <concept\_desc>Networks Location based services</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept\_id>10003456.10010927.10003618</concept\_id> <concept\_desc>Social and professional topics Geographic characteristics</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10003033.10003079.10011704</concept\_id> <concept\_desc>Networks Network measurement</concept\_desc> <concept\_significance>300</concept\_significance> </concept> <concept> <concept\_id>10003033.10003106.10010924</concept\_id> <concept\_desc>Networks Public Internet</concept\_desc> <concept\_significance>100</concept\_significance> </concept> <concept> </ccs2012>
Introduction {#Sec:Introduction}
============
IP Geolocation databases map IP addresses to their corresponding geographical locations. They are often used to find the approximate location of an IP address at the city level. Records in these databases typically contain IP ranges along with their physical location. These databases are vital to a variety of online services when the exact location of a user is not available. Table \[table:Example of rows from IP Geolocation DB\] lists a few examples of such records. For instance, the second example in the table maps a /24 subnet (256 IPs) to Hengyang, a city in China. While some users opt-in to share their exact coordinates to online services through devices with global positioning sensors or Wi-Fi based geolocation, others decline or use devices without such features. IP geolocation is therefore a valuable source of information on user location.
A practical application of IP geolocation is **personalized local search results** in the context of search engines. Figure \[fig:Effect of missing location on search results\] demonstrates the striking difference in results for the query “restaurants” when the user location is unknown, compared to when it is known. The generic nationwide results require the user to requery for more specific restaurants in their area, while the personalized results directly list restaurants tailored to a specific location. Previous work has shown that personalizing results to a user’s location leads to increased user satisfaction and conversely that missing location information leads to user dissatisfaction [@dan2016improving; @bennett2011inferring; @kliman2015location]. IP geolocation databases are also used in many other applications, including: **content personalization and online advertising** to serve content local to the user [@bennett2011inferring; @hannak2013measuring; @kolmel2002location], **content delivery networks** to direct users to the closest datacenter [@huang2011public], **law enforcement** to fight cybercrime [@Shue2013], **geographic content licensing** to restrict content delivery to licensed geographic regions [@macvittie2010geolocation], and **e-commerce** to display variable pricing based on local taxes and shipping [@svantesson2007commerce].
![Example of information that can be extracted from reverse DNS hostnames, including location information such as city name, state, country, as well as physical connection characteristics.[]{data-label="fig:ReverseDNSExample1"}](Figures/ReverseDNSExample1.pdf){width="0.8\linewidth"}
[.45]{} \[fig:sub1\]
[.55]{} \[fig:sub2\]
[ L L L L L ]{} [ ]{} **StartIP** & **EndIP** & **Country** & **Region** & **City**\
1.0.16.0 & 1.0.16.255 & JP & Tokyo & Tokyo\
124.228.150.0 & 124.228.150.255 & CN & Hunan & Hengyang\
131.107.147.0 & 131.107.147.255 & US & Washington & Redmond\
Commercial IP geolocation databases such as MaxMind [@maxmind2018], Neustar IP Intelligence [@neustar2018], and IP2Location [@ip2location2018] are considered state of the art. They combine multiple IP location sources, including WHOIS lookups, network latency, network topology, reverse DNS, as well as direct contracts with Internet Service Providers [@muir2009internet]. However, the exact methods they use are proprietary. Some of these approaches have been studied in academia to some extent, as described in Section \[sec:Related Work\]. Related work has shown that while they have high coverage, commercial databases are sometimes inaccurate or are missing location information for some IP ranges [@shavitt2011geolocation; @poese2011ip; @gharaibeh2017look].
**Our work focuses on extracting location information from reverse DNS hostnames assigned to IP addresses**, which has many potential advantages including high coverage and accuracy. These hostnames can be periodically collected in a short amount of time by performing a reverse DNS lookup for every address in the IP space. Figure \[fig:ReverseDNSExample1\] exemplifies the information that can be parsed from reverse DNS hostnames. Here we can derive both the location and connection characteristics for the hostname of an IP address. A person reading the name of the hostname can reasonably determine that it references *Wallingford*, a town in Connecticut, USA. **Reverse DNS is the opposite of Forward DNS**. Forward DNS starts from a domain or subdomain such as `www.bing.com` and resolves to zero, one, or more IP addresses [@rfc1034]. Note that multiple subdomains can map to the same IP. Conversely, reverse DNS lookups start from an IP address and typically return zero or one hostnames [@rfc2317]. The reverse DNS hostname does not need to be the same as the Forward DNS hostname. While Forward DNS lookups are used by Internet users to get to websites, Reverse DNS hostnames are typically used to name and describe the underlying physical infrastructure that makes up the Internet.
Figure \[fig:ForwardReverseDNS\] contains examples of both forward and reverse DNS resolution.
[ ![Example of the difference between Forward DNS (top) and Reverse DNS (bottom)[]{data-label="fig:ForwardReverseDNS"}](Figures/ForwardDNS.pdf "fig:"){width="40.00000%"} ]{}
[ ![Example of the difference between Forward DNS (top) and Reverse DNS (bottom)[]{data-label="fig:ForwardReverseDNS"}](Figures/ReverseDNS.pdf "fig:"){width="40.00000%"} ]{}
**Given a reverse DNS hostname, our task is to determine its location at the city level**. This task poses multiple challenges. First, the naming schemes of Internet Service Providers are often ad-hoc and do not always contain the full names or common abbreviations of cities. For example, the `drr01.cral.id.frontiernet.net` hostname is located in *Coeur D’Alene, Idaho*. Determining that the `cral` substring maps to this location is difficult even for a human. Second, many cities around the world have ambiguous names. Take for instance *Vancouver, Canada* and *Vancouver, USA*. A hostname which only contains the substring *vancouver* is not specific enough to determine a single location correctly. Even unambiguous city names can become ambiguous when abbreviations are used instead of their full names. Does `nwmd` refer to ***N**e**w** Rich**m**on**d**, WI* or to ***N**e**w** **M**arylan**d**, NB*, or to neither of them? Third, sometimes hostnames contain multiple or conflicting locations. For example, it is difficult to determine if `sur01.tacoma.wa.seattle.comcast.net` is located in *Seattle, WA*, *Tacoma, WA*, or maybe even ***Su**mne**r**, WA*.
We propose a systematic approach for using reverse DNS hostnames to geolocate IP addresses. Our contributions are:
1. **In our preliminary investigation** we determine reverse DNS coverage in the entire IPv4 address space. We also find an upper bound of exact city and airport code matches.
2. **We present a machine learning approach for extracting locations from hostnames**. We cast the task as a machine learning problem where for a given hostname, we split the hostname into its constituent terms, we generate a list of potential location candidates, and then we classify each hostname and candidate pair using a binary classifier to determine which candidates are plausible. Finally, we rank the remaining candidates by confidence, and we break the ties by location popularity.
3. **We evaluate our approach against state-of-the-art baselines**. Using a large ground truth set, we evaluate our approach against three academic baselines and two commercial IP geolocation databases. We show that our method significantly outperforms academic baselines. We also show that the academic baselines contain incorrect rules which impact their performance. Finally, we demonstrate that our approach is both competitive and complementary to commercial geolocation baselines, which shows that our method can help improve their accuracy.
4. **We release our approach as open source**. To help the academic community reproduce our results, we release our reverse DNS geolocation software as open source.
Related Work {#sec:Related Work}
============
We divide IP geolocation research in two broad categories, based on the methods they use: **network delay and topology** approaches use ping, traceroute, and BGP network structure information; **web mining** approaches use diverse information mined from the web, including WHOIS databases, social graphs, and reverse DNS.
The majority of IP geolocation research relies on active **network delay measurements** to locate addresses. These approaches issue pings or traceroutes from multiple geographically distributed landmark servers to a target IP, then triangulate the location [@padmanabhan2001investigation; @jayant2004toward; @gueye2006constraint; @katz2006towards; @youn2009statistical; @laki2010model; @eriksson2010learning; @li2012ip; @ciavarrini2018geolocation]. Multiple systems such as Octant [@wong2007octant], Alidade [@chandrasekaran2015alidade], or HLOC [@scheitle2017hloc] combine delay measurement methods with other data sources such as reverse DNS and WHOIS information.
Network delay and topology methods have significant limitations. First, they have scalability issues as they require access to landmark servers distributed around the world and each target IP needs separate measurements. Second, not all networks allow ping and traceroute. Third, routes on the Internet do not necessarily map to geographic distances. These problems often lead to lackluster results, with error distance in the order of hundreds of kilometers [@padmanabhan2001investigation; @gueye2006constraint]. Some of the earlier research is also plagued by extremely small ground truth datasets, often focusing on a handful of IP addresses in a few US universities [@padmanabhan2001investigation; @katz2006towards; @gueye2006constraint; @youn2009statistical].
Our work addresses several of these limitations. First, using reverse DNS hostnames for geolocation does not require any network delay measurements. Reverse DNS hostnames can be obtained much faster than performing active delay measurements, by querying DNS servers. Second, our ground truth dataset is several orders of magnitude larger than the ones used in previous work and it spans the entire planet. Third, our approach of extracting locations from hostnames can be performed offline. Fourth, our results have lower median error distance than most previous work.
**Web mining approaches** use diverse information mined from the web. Guo et al. [@guo2009mining] extract locations mentioned in web pages and assign the locations to the IPs which host the content. Although they report city level agreement for 87% of the IPs, they use an IP geolocation database with unknown accuracy as ground truth. Endo and Sadok [@endo2010whois] propose using *whois* information. Unfortunately, the evaluation section lacks a comparison against ground truth. Wang et al. [@wang2011towards] combine a network delay approach with extracting the location of web servers from the web pages that they host. They obtain a median error of 0.69 kilometers for the best data set, which contains only 88 target IP addresses. Backstrom et al. [@backstrom2010find] propose deriving the location of target users based on the locations of friends. However, this approach requires access to users’ social graphs. It achieves a median error distance of 590 km on 2,830 IPs.
**In this work we propose extracting IP locations from their reverse DNS hostnames**. *GeoTrack* [@padmanabhan2001investigation], proposed by Padmanabhan and Subramanian, is one of the earliest reverse DNS geolocation approaches. It uses manual rules to determine locations of hostnames in the United States using city names, airport codes, and country codes. It then combines this approach with traceroutes to estimate the location of a target IP. Our machine learning approach does not require manual rules and it achieves a median error of tens of kilometers using a test set of millions of IP addresses. *Undns* is the most well-known and widely used reverse DNS geolocation approach [@undns2002]. Similarly to *GeoTrack*, it consists of manual rules which are expressed as regular expressions at the domain level. For example, the rule `([A-Z]3,4)[0-9]?.verizon-gni.net` matches the hostname `PHIL.verizon-gni.net`. A domain specific location dictionary is then used to match the extracted slot `PHIL` to *Philadelphia, PA*. The obvious disadvantage of this approach is that each domain requires manually generated and potentially error prone rules. In comparison, our approach is more scalable since it does not require human generated rules. It also handles unique situations better, since it considers the terms of each hostname individually, without requiring domain-specific training. In Section \[sec:Academic baselines\] we show that our approach significantly outperforms *undns*. Multiple geolocation and network topology papers use *undns* as-is to draw conclusions or perform experiments [@freedman2005geographic; @2018nurgeography]. Unfortunately, we demonstrate that *undns* results suffer due to catch-all rules.
*DRoP*, another state of the art reverse DNS based approach, aims to geolocate hostnames using automatically generated rules generated by finding patterns across all the hostname terms of a domain [@drop2014]. For example, it may find that for the domain `cogentco.com`, the second term from the right often contains airport codes. These rules are then validated using network delay information. *DRoP* places 99% of IPs in 5 test domains within 40 km of their actual location. However, it uses network delay measurements, its method of splitting hostnames is rudimentary, and as we show in Section \[sec:Academic baselines\], it performs poorly on worldwide ISP domains.
Finally *DDec* [@ddec2018] combines *undns* and *DRoP* rules by giving precedence to *undns* and using *DRoP* as fallback.
Datasets {#sec:Datasets}
========
This section contains descriptions of the datasets we use throughout this paper for experiments, training and testing.
**Our ground truth set** contains 67 million IP addresses with known geographic location. To the best of our knowledge, it is the largest and most diverse set used in geolocation literature. We compiled the ground truth set in March 2018 by randomly sampling the query logs of a large-scale commercial search engine. We describe the characteristics of this dataset in more detail in Section \[sec:Ground Truth\].
**GeoNames** is a free database with geographical information [@geonames2018]. The March 2018 snapshot we used contains information on 11.5 million geographic features from all countries in the world. From Geonames we used multiple subsets available separately for download. *Cities 1000* consists of information on all cities in the world with a population of at least 1,000, including coordinates, original names, ASCII names, alternate names, and codes of administrative divisions. *Alternate Names* contains more alternate names for some cities such as abbreviations, colloquial names, and historic names. More importantly, it also contains *airport codes* issued by *IATA*, *ICAO*, and *FAAC*, which are travel organizations. *Admin 1 Codes* is comprised of the codes and names of first-level administrative regions. *Country Info* contains general information about countries, including their Internet top-level domain (TLD).
**CLLI** is an abbreviation for Common Language Location Identifier. These codes are used by the North American telecommunications industry to designate names of locations and functions of telecommunications equipment. While historically only used by the Bell Telephone companies, they were more recently adopted by other companies as well. Multiple codes can map to the same location. For example, all the following codes map to *Chicago, Illinois*: `chcgil`, `chchil`, `chciil`, `chcjil`, and `chclil`. Note that the codes cannot necessarily be derived from the name of the city. This database is available from multiple sources. We acquired a May 2017 snapshot from TelcoData [@telcodata2018] for a token amount.
**UN/LOCODE**, which stands for United Nations Code for Trade and Transport Locations, is a worldwide geographic coding scheme developed and maintained by the UN. It assigns codes to locations used in trade and transport, such as rail yards, sea ports, and airports. The code assigned to Paris, France is `FRPAR` and the functions listed for this location are: `port`, `rail`, `road`, and `postal`. This dataset is updated twice a year and it is available for free on the United Nations Economic Commission for Europe website [@unlocode2018]. We used the December 2017 release, which was the latest available version.
**Public Suffix List**, maintained by the Mozilla Foundation, is a free list of domain suffixes under which Internet users can directly register names [@publicsuffix2018]. Examples include `cloudapp.net` and `gov.uk`.
**Rapid7 Reverse DNS** consists of reverse DNS hostnames of the entire IPv4 address space. The dataset is available for free and it is updated weekly. The archive contains snapshots going back to 2013 [@rapid7rdns2018]. We discuss this dataset in detail in the next section.
Reverse DNS {#sec:Reverse DNS}
===========
Forward DNS lookups convert hostnames such as `www.bing.com` into IP addresses, while reverse DNS lookups work in the other direction; they start from an IP address and find a hostname. **Since the forward and reverse DNS lookups are set in different DNS records, they do not need to have the same hostname.** Reverse hostnames are more likely to be used to name the underlying networking infrastructure, while forward hostnames are used to name websites or other online services [@rfc1034].
IPv6 addresses also have reverse DNS hostnames. The only difference is that the records are under the `ip6.arpa` domain. **While we evaluate our approach on IPv4, all methods described in this paper can be equally applied to IPv6 addresses as well.**
To determine the viability of using reverse DNS hostnames for geolocation, we studied the *Rapid7 reverse DNS dataset* [@rapid7rdns2018], which covers the entire IPv4 address space. Rapid7 compiles it by performing IPv4 PTR lookups over the entire address space as described above, except for ranges that are blacklisted or private. The archive contains snapshots going back to 2013 [@rapid7rdns2013]. This preliminary investigation is based on a snapshot taken in January 2017, while in the rest of the paper we use a more recent dataset from March 2018.
The IPv4 address space consists of all 32-bit numbers. This limits the possible address space to 2^32^ (4.3 billion) addresses. The number of usable IPs is actually only 3.7 billion, since some IP ranges are designated as special-use or private [@ianaspprp2017]. Since not all IP addresses have a reverse DNS hostname, we parsed the Rapid7 dataset to find the actual coverage. We found that 1.25 billion addresses have a reverse DNS hostname. This finding shows that although they have significant coverage, these hostnames need to be augmented with other data to obtain a complete geolocation database.
We then quantified how many of the hostnames are valid, since the DNS records are unrestricted strings. We parsed each hostname and rejected the ones that did not respect Internet host naming rules [@rfc1123]. We also rejected hostnames that did not have a valid suffix as defined by the *Public Suffix List*, which is a list of valid domain suffixes from the Mozilla Foundation previously described in Section \[sec:Datasets\]. This left us with 1.24 billion hostnames, of which 1.15 billion were distinct. Our findings are summarized in Table \[table:Usage of Reverse DNS hostnames\], which shows that 33.4% of usable IP addresses have a valid reverse DNS hostname, and 31.1% are distinct. Considering that not all IPv4 addresses are yet allocated, the actual percentage is likely higher.
Next, we set out to determine if reverse DNS hostnames are a valuable source of geolocation information. We searched for exact city names and airport codes in the hostnames, using the *Cities 1000* and the *Alternate Names* dataset, respectively. We find that 163.7 million hostnames could contain exact city names, and 272.9 million hostnames could contain airport codes. This approach represents an upper-bound of the number of hostnames that could contain exact city names or airport codes. The results contain true positives such as `sur01.seattle.wa.seattle.comcast.net` in *Seattle, Washington* and `inovea5.gs.par.ivision.fr` in *Paris, France*. However, this naive approach also matches false negatives such as `node-j.pool-1-0.dynamic.totbb.net` which is not in *Pool, UK* and `mobile.bigredgroup.net.au` which is not in *Mobile, Alabama*. Nevertheless, the results summarized in Table \[table:Usage of Reverse DNS hostnames\] show that there are potentially hundreds of millions of hostnames that could contain geographic information, using just these two features alone. We conclude that while the results are promising, a more sophisticated approach could achieve higher coverage and accuracy.
[ @ l c c c ]{} **Set name** & **Size** & **% of usable** & **% of distinct**\
**Total IPv4 space** & **4.3 B** & &\
(0,0.7ex) – (0,0) – (0.75em,0); (0.55em,0.2em) – (0.75em,0) – (0.55em,-0.2em);
Reserved IP addresses & 0.6 B & &\
(0,0.7ex) – (0,0) – (0.75em,0); (0.55em,0.2em) – (0.75em,0) – (0.55em,-0.2em);
Usable IP addresses & 3.7 B & &\
(0,0.7ex) – (0,0) – (0.75em,0); (0.55em,0.2em) – (0.75em,0) – (0.55em,-0.2em);
IPs with Reverse DNS hostnames & 1.25 B & 33.7% &\
(0,0.7ex) – (0,0) – (0.75em,0); (0.55em,0.2em) – (0.75em,0) – (0.55em,-0.2em);
Valid Reverse DNS hostnames & 1.24 B & 33.4% &\
(0,0.7ex) – (0,0) – (0.75em,0); (0.55em,0.2em) – (0.75em,0) – (0.55em,-0.2em);
Distinct DNS hostnames & 1.15 B & 31.1% &\
(0,0.7ex) – (0,0) – (0.75em,0); (0.55em,0.2em) – (0.75em,0) – (0.55em,-0.2em);
Exact city match (naive) & 0.16 B & 4.4% & 14.1%\
(0,0.7ex) – (0,0) – (0.75em,0); (0.55em,0.2em) – (0.75em,0) – (0.55em,-0.2em);
Airport code match (naive) & 0.27 B & 7.4% & 23.5%\
[ llcX ]{} **Domain** & **Count** & **% of valid** & **Top hostname components sorted in descending order by how often they appear in all hostnames of this domain**\
comcast.net & 50.0M & 4.0% & c, hsd, hsd1, m, , , , , , , a, , , d, f, , b, e, , , , , , , , , , , npls, dd, , , ,\
bbtec.net & 37.2M & 3.0% & softbank, biz\
rr.com & 31.1M & 2.5% & res, cpe, mta, , biz, rrcs, , neo, , , kya, , , , , , , , , ,\
myvzw.com & 29.6M & 2.4% & sub, qarestr\
sbcglobal.net & 28.4M & 2.3% & lightspeed, adsl, dsl, , , , , , , , , , , , , ,\
t-ipconnect.de & 24.5M & 2.0% & dip, dip0, p, b, e, a, f, d, c, pd, fc, fd, fe, ff, de, dd, dc, df, ee, bb, bd, bc, ae, ac, aa, ab, af, ad, ba, bf, ea, eb, be, ec, fa, ed, fb, ef, db, da, ca, cf\
telecomitalia.it & 19.4M & 1.6% & host, static, business, b, r, retail, dynamic, host156, host15, host94, host61, host127, host232, host112, host95, host72, host107, host220\
ge.com & 16.7M & 1.4% & static, n, n003, n003-000-000-000, n129, n144, n144-220-000-000, n129-201-000-000, n129-202-000-000, n165-156-000-000m n165, n192\
ocn.ne.jp & 16.2M & 1.3% & p, ipngn, , , ipbf, , ipbfp, omed, omed01, , , , , ,\
spcsdns.net & 16.0M & 1.3% & pools, static\
To further familiarize ourselves with hostname naming conventions, we extracted the top hostname components of the largest 10 domains in the Rapid7 dataset. We divided each subdomain on the dotted terms, and then we further split the components on dashes and on the transitions between numbers and letters. For example, we split `soc-l.wht2.ocn.ne.jp` into `soc`, `l`, `wht`. We then manually labeled the components that we found to reasonably correspond to geographic locations. We also cross-checked our findings with commercial geolocation databases. Table \[table:Top hostname n-grams\] shows a sumary of the results. We observe that only 4 out of the top 10 domains contain indicators of geographic location. However, those that use geographic encodings do so extensively. We found that service providers use various naming conventions across different networks and within a single network. For instance, the hostnames under the *sbcglobal.net* domain owned by *AT&T* make use of abbreviations such as `pltn` to refer to *Pleasanton, CA*. But they also use combinations of city abbreviations with State names such as `chcgil` to refer to *Chicago, Illinois*. Our findings are in line with previous work by Chabarek and Barford [@chabarek2013s]. They found that all 8 of the providers they studied used multiple naming schemes. They also found that 20 out of 22 North American providers they surveyed use geographic encodings in their hostnames.
We also studied the distribution of top-level domains (TLDs) such as `.com` and `.fr` in the *Rapid7* dataset to determine if country-specific domains can be used as location hints. We observed that most hostnames contain a `.net` domain at 33.2%, followed by `.com` with only 17.2%. This is the opposite of forward DNS, where `.com` is more popular. The difference is due to Internet Service Providers preferring to use `.net` domains for hostnames that describe the underlying physical architecture of their *network*. After removing the `.com`, `.net`, `.edu`, and `.mil` domains which together make up 51.6% of valid hostnames, we are left with approximately 600 million hostnames, the vast majority of which are country-specific. We found very few novelty TLDs used in reverse DNS hostnames. We conclude that the corresponding country of a reverse DNS domain could potentially be a useful hint in geolocation.
Finally, we compared snapshots of the dataset, each collected in the month of January of years 2014 to 2017, inclusive. Our goal was to determine how the characteristics of the hostnames change in time. For each IP in the snapshot we compared the hostname values in consecutive years. Table \[table:Reverse DNS hostname changes across 4 years\] shows a summary of the results. We found that a maximum of 14.7% of hostnames changed year over year and 63.7% of them remain the same across all four years. These numbers include the cases where one side of the comparison had a hostname but the other side was empty due to the DNS query returning an empty hostname, or due to the request failing because of network failures during data collection. We then performed a similar comparison, this time counting only the cases where both sides of the comparison contained non-empty hostnames. Here we found that a maximum of 2.2% hostnames change over the years, if both the values are present. To understand why there is such a large discrepancy between these two findings we also determined the number of hosts that were gained or list between the years. By hostnames gained we mean that in the older year a hostname was missing, while in the subsequent year it was present, and by hostnames lost we mean the opposite. The results show that yearly more hostnames are gained than lost, at about a ratio of 2:1.
In summary, we determined that 1.24 billion IP addresses have valid reverse DNS hostnames with 1.15 billion distinct values, many of which contain exact city or airport code matches.
[ @ l c c c ]{} **Change / Year Pair** & & &\
Hostnames changed (incl. empty) & & &\
Hostnames changed (non-empty) & & &\
Hostnames gained & & &\
Hostnames lost & & &\
Approach {#sec:Approach}
========
We cast the problem of extracting locations from reverse DNS hostnames as a machine learning problem. We train a binary classifier on a dataset where each training sample is a hostname and location candidate pair, along with a binary label which signifies if the hostname is *likely* or *unlikely* to be in the candidate location. Given a new hostname, our proposed approach splits the hostname into components, finds a preliminary list of location candidates, generates primary and secondary features for each candidate, then classifies each potential location using the classifier, also assigning each candidate a confidence score. For instance, for the hostname `ce-salmor0w03w.cpe.or.portland.bigisp.net` our approach considers tens of potential location candidates, including *Portland, UK* and *Salmoral, Spain*. In the end however, it ranks *Salem, Oregon* and *Portland, Oregon* as the most likely candidates.
Splitting hostnames {#sec:Splitting hostnames}
-------------------
Drawing from our preliminary analysis in Section \[sec:Reverse DNS\], as well as further manual analysis, we implemented multiple heuristics for splitting hostnames into their constituent components.
First, we apply the *ToUnicode* algorithm described in RFC3490 [@rfc3490] to convert International Domain Names (IDN) to Unicode. The reason we perform this translation is that international hostnames are stored as ASCII strings using Punycode transcription. For example, the hostname `xn–0rsod70av79j.xn–j6w193g` gets converted to
[UTF8]{}[gbsn]{}夏威夷舞.香港
. This allows us to perform location lookups using the original language of the hostname. Second, we separate the subdomain from the domain and the public suffix, using the list provided by the Mozilla Foundation previously described in Section \[sec:Datasets\]. These suffixes are a superset of normal TLDs because they also contain entire domains under which users can create subdomains. For example, the list contains the pseudo-TLD `azurewebsites.net` since users of Azure cloud services can register their own subdomains under this name. At this point we also extract the native TLD. For instance, for `dps8099.denver.k12.co.us` we extract `denver.k12.co.us` as the domain because `k12.co.us` is a public suffix, we extract `dps8099` as the subdomain, and finally we extract `.us` as the TLD.
Third, we split the extracted subdomain at three levels of aggregation: on the dotted elements, on hyphens within the dotted elements, and on the transitions between letters and numbers within the hyphenated elements, saving the results at each level. Figure \[fig:HostnameSplitter\] contains a specific example represented intuitively as a tree structure. The bottom three levels of the tree correspond to the three levels of aggregation. As a last step, we trim the leaf nodes. We remove any leaf node consisting solely of numbers. We also remove common terms terms related to connection characteristics, such as `dsl`, `fiber`, and `nas`. We obtained them by counting the top extracted leaf nodes in the training set and manually selecting the ones which are unrelated to geolocation but clearly related to the underlying network infrastructure. The list is available in the source code we are publishing along with this paper.
![Hostname Splitter example with pruning[]{data-label="fig:HostnameSplitter"}](Figures/HostnameSplitter.pdf){width="0.7\columnwidth"}
Features {#sec:Features}
--------
Starting from the results of the hostname splitter, we find the location candidates along with their [*primary*]{} and [*secondary*]{} features, as defined below. **The list of preliminary location candidates is defined by the union of locations which match any of the primary features of the hostname.** Figure \[fig:FeatureGeneration\] shows a concrete example. Primary features can be derived directly from a hostname. These features are matched using a single contiguous string which indicate a location at city level granularity. Primary feature generation and candidate selection happen at the same time. Secondary features are generated in the context of a hostname and location candidate pair. These features require the context of a primary candidate to match. In our example two location candidates and their primary features are first selected based on the term `roch` in the hostname. Then we compute secondary features separately for each candidate. In the context of *Rochester, Minnesota*, we match the `mn` term as a secondary feature that captures the administrative region for this candidate.
{width="1\columnwidth"}
\[fig:FeatureGeneration\]
**Primary features** are based on the *GeoNames*, *UN/LOCODE*, and *CLLI* datasets described in Section \[sec:Datasets\]. From *GeoNames* we use the *Cities 1000*, *Alternate Names*, and *Admin 1 Codes* subsets. The primary feature categories are listed in Table \[table:Primary Features Categories\]. Each of these categories is represented by three specific features: *IsMatch*, *Population*, and *MatchedLettersCount*. The *IsMatch* feature is a boolean which indicates if the feature matched the current hostname and current location candidate. The *Population* feature contains the population of the current location, if *IsMatch* is *true*. We use population as a proxy for the importance of a city candidate. Finally, *MatchedLettersCount* contains the number of characters which matched. As the number of characters in common between a hostname and a location increases, it could mean a higher confidence match. For instance, if the hostname contains the letters `seattle` and the current location candidate is *Seattle, Washington*, then the *CityName-MatchedLettersCount* Feature would have a value of seven.
\[table:Primary Features Categories\]
[ lll ]{} **Category** & **Example** & **Location**\
City Name & p907072-li-mobac01..ocn.ne.jp & Osaka, JP\
Alternate names & 178235248188..vectranet.pl & Warsaw, PL\
Abbreviations & cpe-68-173-83-248..res.rr.com & New York City\
City + Admin1 & -rta-1.inhouse.compuserve.com & Toronto, ON\
City + Country & er1-ge-7-1.5.savvis.net & London, UK\
No Vowels Name & static-50-47-60-130..wa.frontiernet.net & Seattle, WA\
First Letters & 97-90-205-107.dhcp..ca.charter.com & Los Angeles\
Airport Code & 62.80.122.50..de.eunx.net & Frankfurt, DE\
CLLI Code & 99-166-111-251..sbcglobal.net & Tucker, GA\
UN/LOCODE & 16.151.88.129,19d.kor.hp.com & Korea, Seoul\
Host Patterns & -651-1-29-109.abo.wanadoo.fr & Toulon, FR\
\[table:Secondary Features Categories\]
[ lll ]{} **Category** & **Candidate** & **Match Example**\
Admin1 & Johnstown, PA & 138-207-246-119...atlanticbb.net\
First Letters Admin 1 & Ft. Huachuca, AZ & -bw-noc..aisco.ngb.army.mil\
Country & Paris, FR & ci77.12eme..psi.net\
Country TLD & Barcelona, ES & .fib.upc.\
While most feature categories in Table \[table:Primary Features Categories\] are self-explanatory, we describe them here briefly. The *City Name* category matches entire names of cities. *Alternate names* matches translations and colloquial names of locations. *Abbreviations* are based on the first letters of cities with longer names, such as `sf` for *San Francisco*. The *City + Admin1* category consists of concatenations of city and administrative regions, such as `seattlewa`. Similarly, *City + Country* matches combinations of city and country names.
The intent of the *No Vowel Name* feature is to match city names without vowels. It allows partial matches using the first 3 or more letters of the names. For example, this allows matching `gnvl` to *Greenville, SC* and `rvrs` to *Riverside, CA*. Furthermore, based on our observations we extended this feature with more complex variations. We select the first and last letters of each word in the name, even if the letters are vowels. We then generate combinations of letters from this list, in order. Examples matched by this variation include `oxfr` for *Oxford, MA*, and `ftmy` for *Fort Meyers, FL*.
The *First Letters* features use the first consecutive letters of locations. The *Airport Code* category spans airport codes from travel organizations. *CLLI* and *UN/LOCODE* codes match telecommunications and transportation codes of locations, respectively.
Finally, *Host Patterns* attempts to capture rules not encompassed by the other features. Using training data we extract frequently co-occurring hostname term permutations of one or two terms. We then aggregate the training data per domain and within a domain on the term permutations. If at least 40% of the training locations for a permutation are located within a 20 kilometer radius, we convert the term permutation into a rule. We determined the support ratio and the distance radius using a small validation set. For example, this feature determines that whenever a hostname in the `frontiernet.net` domain contains the term `or` in the rightmost position and the term `mmvl` in the second rightmost position, then the hostname is most likely located in `McMinnville, Oregon`. This feature would then match for the hostname `static-50-126-80-6.mmvl.or.frontiernet.net`.
**Secondary features** are determined in the context of a hostname and location candidate pair. As shown in Figure \[fig:FeatureGeneration\], we first determine all candidates before we can compute the secondary features. An example of secondary features for the *Rochester, MN* candidate is *Admin1 Match*, which is *true* only if the administrative region of the candidate location can be found in a different term of the hostname. Since the hostname contains the term `mn`, which is an abbreviation of Minnesota, then this secondary feature is *true* for the first candidate. However, it is *false* for the second candidate, because *Rocha* is in an administrative region also called *Rocha*, and it cannot be found in the hostname. *First Letters Admin 1* is similar, but it matches at least 3 first consecutive letters of administrative names. *Country* and *Country TLD* both try to match the country of the current candidate by searching for a country code in the hostname terms or in the domain *TLD*, respectively.
Classifier {#sec:Classifier}
----------
For a given hostname, our reverse DNS geolocation can extract and evaluate tens of potential location candidates. For example, if one of the terms of the hostname is `york`, the initial list of candidates will contain all locations named *York* in the world. We run a binary classifier on each of the initial candidates. The classifier uses the primary and secondary features to evaluate if it is plausible for the hostname to be located in a candidate location. All the candidates where the classifier returns `false` are discarded. The remaining plausible candidates are sorted by confidence and returned in a list.
Although determining the optimal type of binary classifier is outside of the scope of this work, we tested four variations of the classifier: logistic regression, C4.5 decision trees, random forest, and SVM. Logistic regression had the best performance on a small validation set. Consequently, we performed all experiments in Section \[sec:Evaluation\] using this classifier.
Sampling Strategy {#sec:Sampling Strategy}
-----------------
We propose sampling the training set to account for data bias, to improve generalization, and to reduce the amount of required training data. First, the entire set of reverse DNS hostnames is naturally skewed towards the largest Internet Service Providers, which own the most addresses. Second, some feature categories such as *City Name* occur much more often than others such as *Abbreviations*. This can lead the classifier to ignore less frequent features categories. Third, during training multiple location candidates can be generated for each hostname, out of which at most one can be correct. Since the classifier is trained on hostname and candidate pairs, this also introduces another type of bias where the number of negative samples significantly outweighs the number of positive ones. Therefore, we sample data to account for some of this bias and to improve generalization through increased training data diversity.
We perform stratified sampling on the domain of the hostname, keeping at most $\mathcal{X}$ samples per domain. This approach ensures that naming schemes of large organizations do not significantly skew the training data. We further increase feature diversity by keeping a ratio of $\mathcal{Y}:1$ between the number of samples that contain the most commonly occurring feature and the ones that contain the least occurring feature. Finally, we also enforce a ratio of $\mathcal{Z}:1$ between the number of negative and positive examples. We evaluate our data sampling strategy and its three parameters in Section \[sec:Preliminary Evaluation\].
Evaluation {#sec:Evaluation}
==========
We evaluate our approach against three state-of-the-art academic baselines and two commercial geolocation databases. We show that our method significantly outperforms academic baselines and is complementary and competitive to commercial databases.
Ground Truth {#sec:Ground Truth}
------------
Our ground truth dataset contains 67 million IP addresses with known IP location, of which we used 40 million for training and 27 million for testing. We compiled the dataset in March 2018 from a subset of the query log of a major search engine. Each IP address has a corresponding location obtained from users that opted in to provide their location through devices connected to cellular networks or home Wi-Fi networks. We discarded any IP address that was present in multiple cities over the course of a month. The locations were aggregated at IP and city level by an automated pipeline. We did not have access to the locations of individual users.
Preliminary Evaluation {#sec:Preliminary Evaluation}
----------------------
We conducted two experiments to evaluate the binary classifier in isolation. In the first experiment, we randomly selected 100,000 IP addresses from the training set and performed ten-fold cross validation. We did not further sample the data in any other way. For each hostname, we extracted location candidates, then ran the binary classifier on all the pairs between the target hostname and each of its candidates. Since **our approach can return multiple plausible locations for a given hostname**, we choose the candidate with the highest classifier confidence. We break ties by selecting the location with the highest population, as a proxy of popularity. We obtained an overall accuracy of 99%, mostly because the vast majority of results were true negatives. However, the true positive rate was only 67.6%, precision was 80.9%, and recall was 67.6%.
In the second experiment we introduced training data sampling as described in Section \[sec:Sampling Strategy\]. We set the $\mathcal{X}$, $\mathcal{Y}$, and $\mathcal{Z}$ parameters to *200*, *10*, and *3*, respectively. We again performed ten-fold cross validation. Although accuracy decreased to 92.9%, we obtained better results for true positive rate, precision, and recall, at 78.8%, 88.5%, and 78.8%, respectively. We varied the values of the $\mathcal{X}$, $\mathcal{Y}$, and $\mathcal{Z}$ parameters using exhaustive search but this did not alter the results significantly. In conclusion, our sampling strategy helps the classifier generalize and it significantly improves results.
Academic baselines {#sec:Academic baselines}
------------------
\[table:Academic Evaluation\]
[ lr|rrrr|rrrr|rrrr|rrrr ]{} & & & &\
**Domain** & \# & **undns** & **DRoP** & **DDec** & **RDNS** & **undns** & **DRoP** & **DDec** & **RDNS** & **undns** & **DRoP** & **DDec** & **RDNS** & **undns** & **DRoP** & **DDec** & **RDNS**\
163data.com.cn & 166K & 1,517.5 & N/A & 1,517.5 & & 1,495 & N/A & 1,495 & & [100%]{} & N/A & [100%]{} & 94.5% & 0.67 & N/A & 0.67 &\
bell.ca & 200K & N/A & 5,875.2 & 5,875.2 & & N/A & 5,807 & 5,807 & & N/A & 2.3% & 2.3% & [95.7%]{} & N/A & 0.00 & 0.00 &\
brasiltelecom.net.br & 32K & 808.7 & 5,628.7 & 808.7 & & 889 & 5,620 & 889 & & [100%]{} & 69.7% & [100%]{} & 73.9% & 1.12 & 0.12 & 1.12 &\
charter.com & 580K & 60.8 & N/A & 60.8 & & & N/A & & 484 & 78.0% & N/A & 78.0% & [89.0%]{} & 1.63 & N/A & 1.63 &\
frontiernet.net & 67K & 36.5 & 6,247.6 & 36.5 & & 785 & 6,101 & 785 & & 3.6% & 0.8% & 3.6% & [99.4%]{} & 0.05 & 0.00 & 0.05 &\
nttpc.ne.jp & 0.9K & 9.5 & 9,259.9 & 16.2 & & & 9,161 & 4,976 & 3,694 & 12.0% & 16.2% & 16.2% & [57.6%]{} & 0.06 & 0.02 & 0.03 &\
optusnet.com.au & 100K & 704.4 & 16,134.6 & 704.4 & & 1,175 & 16,374 & 1,175 & & [100%]{} & 49.8% & [100%]{} & 98.9% & 0.85 & 0.03 & 0.85 &\
qwest.net & 408K & 3,426.6 & 8,038.7 & 8,038.7 & & 6,856 & 7,361 & 7,361 & & 0.0% & 4.1% & 4.1% & [94.0%]{} & 0.00 & 0.01 & 0.01 &\
Overall & 1.6M & 163.9 & 13,974.2 & 177.9 & & 924.0 & 12,640.4 & 1,497.5 & & 48.3% & 6.1% & 49.7% & [92.3%]{} & 0.52 & 0.00 & 0.33 &\
\[table:Incorrect DRoP Examples\]
[ llll ]{} **Hostname** & **Location extracted incorrectly by *DRoP*** & **Correct location** & ***DRoP* Rule**\
`2259w-lp140-03-50-100-186-228.dsl.bell.ca` & **malton** → Malton, North Yorkshire, England & **malton**, **.ca** → Malton, Canada & `%<<pop>>([^L]+L+D*){3}.bell.ca`\
`200-96-182-198.cbace700..brasiltelecom.net.br` & **dsl** → Daru, Sierra Leone & **cbace**, **.br** → Cuiabá, Brazil & `%<<iata>>.brasiltelecom.net.br`\
`70-100-143-28.dsl2-.roch.ny.frontiernet.net` & **pixley** → Pixley, California, USA & **roch**, **ny** → Rochester, New York, USA & `%<<pop>>([^L]+L+D*){2}.frontiernet.net`\
`st0120.931.m-hiroshima.nttpc.ne.jp` & **nas** → Nassau, Bahamas & **hiroshima**, **.jp** → Hiroshima, Japan & `%<<iata>>([^L]+L+D*){2}.nttpc.ne.jp`\
`d49-194-53-51.meb1..optusnet.com.au` & **vic** → Vicenza, Italy & **meb**, **vic**, **.au** → Melbourne, Victoria & `%<<iata>>.optusnet.com.au`\
`71-209-14-48..qwest.net` & **bois** → ’s-Hertogenbosch, The Netherlands & **bois** → Boise, Idaho, USA & `%<<pop>>.qwest.net`\
We next evaluate against three state-of-the-art academic baselines. Like our approach, they receive a hostname as input and attempt to extract its location. The *undns* baseline from University of Washington consists of manually generated rules that map hostname patterns to locations [@undns2002]. The *DRoP* baseline from CAIDA at University of California-San Diego relies on automatically generated rules derived from hostname patterns and validated by active measurement data (traceroutes) [@drop2014]. Finally, the *DDec* baseline also from CAIDA combines the results from *undns* and *DRoP* [@ddec2018].
Since all three baselines are accessed from a public web endpoint [@ddec2018], we had to restrict the number of requests we made to a manageable size, out of politeness. For testing we initially selected multiple service providers of different sizes, spanning various countries around the world. However, the baselines were missing **any** rules for several of these providers, including *airtelbroadband.in* from India, *bigpond.net.au* from Australia, and *megared.net.mx* in Mexico. Although the baselines have good rule coverage in North America, they are at least partially lacking in international coverage. In the interest of fairness, we selected a list of eight providers, each of which are covered by at least two of the baselines.
To train the classifier, our sampling strategy only considered approximately 60,000 data points out of the 40 million hostnames in our training set. From our test set of 27 million IP addresses, we selected all of the ground truth data points which intersected the eight target providers, which yielded a testing subset of 1.6 million hostnames. We issued these requests to the CAIDA web endpoint and parsed the responses from each of the baselines.
Table \[table:Academic Evaluation\] lists the results for each of the eight domains, as well as the overall results across the entire testing subset. Our approach is labeled *RDNS* in the table. We define the *error distance* in kilometers to be the distance between where a model places the location of a hostname, and the actual location of the IP address behind that hostname. The first block of results shows median error distance in kilometers. We observe that **our model significantly outperforms the baselines** and its results are generally more stable across all domains. We also observe that the median error distance for several domains is abnormally high for the *DRoP* baseline, and sometimes for the other baselines as well. To further investigate this surprising finding we manually verified a small sample of results. Table \[table:Incorrect DRoP Examples\] lists examples of locations extracted incorrectly by the *DRoP* baseline. In the last column of the table we list the rule that caused the incorrect extraction. For example, *DRoP* incorrectly determines that the hostname `d49-194-53-51.meb1..optusnet.com.au` is in Vicenza, Italy, using the rule `%<<iata>>.optusnet.com.au`. Although the *IATA* airport code *vic* is indeed located in Vicenza, the correct location is Melbourne, Victoria. We could not find any `optusnet.com.au` hostname where the rule was correct. In conclusion, the *DRoP* baseline contains incorrect rules for some domains. The results for the *undns* baseline also indicate high error distance for multiple test domains. After investigating the results, we found that *undns* sometimes maps entire TLDs to a single city. For example, the locations for all `163.data.com.cn` hostnames are extracted as *Beijing, China*. Lastly, since *DDec* is a combination of *undns* and *DRoP*, it is also affected by incorrect rules.
The advantage of using median as a metric is that it is impervious to outliers, which can favor our model that can place false positives far from the actual location, generating larger outliers. To fairly characterize the results, we also computed *RMSE*, a metric at the other extreme of the spectrum. *RMSE*, which stands for root mean squared error, easily gets swayed by large outliers. This poses a disadvantage for our model. We compute it using the error distance in kilometers for each hostname. The *RMSE* results in Table \[table:Academic Evaluation\] show that generally our approach still outperforms the baselines in 6 out of 8 domains. In the two cases where our model has higher *RMSE* than the models, the coverage of our model is higher.
In 3 out of 8 cases the *undns* baseline has 100% coverage. We define coverage as the total number of hostnames where a model made a decision, over the total number of hostnames in the test set. *undns* having high coverage is a side effect of it using catch-all rules that map entire TLDs to a single city. In all three cases this leads to poor results for both median error and *RMSE*.
We define the combined score as the inverse of *RMSE* multiplied by coverage. As error distance improves (gets smaller), the combined score increases, and vice versa. Similarly, higher coverage also improves the combined score, and vice versa. **Our approach significantly outperforms all academic baselines** when considering the combination of error distance and coverage.
Finally, Figure \[fig:Academic Evaluation\] displays the cumulative error distance in kilometers. The X axis represents the maximum distance between the real location and the predicted location. The Y axis shows how many hostnames and their IP addresses fall within the error distance. For instance, the *<20 km* column shows that our method, labeled *RDNS*, places approximately 54% of hostnames in the ground truth set within 20 kilometers of their actual location. Our method outperforms the baselines by a large margin. The *DRoP* baseline yields the worst results, significantly underperforming the other methods.
![Academic Evaluation Error Distance[]{data-label="fig:Academic Evaluation"}](Figures/Academic8Evaluation.pdf){width="1\columnwidth"}
Commercial baselines {#sec:Commercial baselines}
--------------------
In this work we focus on improving reverse DNS geolocation, which is only one source of geolocation information. Table \[table:Usage of Reverse DNS hostnames\] reveals that about a third of IP addresses have reverse DNS hostnames. A further subset of these hostnames contain location hints. While this can result in hundreds of millions of hostnames with location information, this is insufficient to completely cover the *IPv4* space.
Commercial geolocation databases combine and conflate multiple geolocation data sources. Information from reverse DNS hostnames is required but not sufficient to compile a full geolocation database. Our approach, which can output multiple potentially valid location candidates for a given hostname, lends itself to being combined with other data source to form a more complete database.
Although reverse DNS geolocation on its own cannot match commercial databases, we evaluate our approach to show that our approach can complement and potentially improve existing databases. We trained our classifier as described in Section \[sec:Academic baselines\]. We then obtained two state of the art commercial IP geolocation databases. We tested our approach against the two commercial database providers *A* and *B* using our entire test dataset of 27 million hostnames. The first four graphs in Figure \[fig:Commercial Evaluation\] show that on certain domains our approach outperforms, and thus can be used to improve, commercial databases. However, as expected, the fifth graph shows that overall the commercial databases still outperform our method. Results show that median error is 43.7, 16.7, 11.1 kilometers, and *RMSE* is 4649, 545.3, 545.9 for *RDNS*, *Provider A*, and *Provider B*, respectively.
![Commercial Evaluation Error Distance[]{data-label="fig:Commercial Evaluation"}](Figures/CommercialEvaluation.pdf){width="1\columnwidth"}
Privacy {#sec:Privacy}
=======
Online privacy is becoming increasingly important. For example, Pew research has found in 2016 that while many Americans are willing to share personal information in exchange for accessing online services, they are often cautious about disclosing their information and frequently unhappy about what happens to that information once companies have collected it [@rainie2016privacy]. We have designed both our approach and our evaluation with this sensitive subject in mind.
Our proposed geolocation method relies on reverse DNS hostnames shared publicly by Internet Service Providers. These hostnames provide only coarse city-level or region level location information. Therefore our approach may be more privacy conscious than the widespread industry practice of requesting exact GPS coordinates through mobile apps or the HTML5 Geolocation API.
Furthermore, the ground truth set was anonymized before we had access to it by modifying raw locations in a random direction by 200 meters, aggregating all locations reported for an IP address, and reducing location accuracy to city-level.
Reproducing Results {#sec:Reproducing Results}
===================
To aid in reproducing and extending our results, we are open sourcing all the major components of our approach, including the hostname splitter and the terms blacklist, our sampling strategy, the primary and secondary feature generators, as well as the classifier itself. For feature generation we have purposely used mainly freely available datasets as described in Section \[sec:Datasets\]. While we cannot include our ground truth set because it is proprietary, we will make available a binary version of our model. We will also publish instructions on creating a ground truth set using public datasets by using our sampling strategy to minimize any manual labeling.
Conclusions and Future Work {#sec:Conclusions and Future Work}
===========================
We presented a machine learning approach to geolocating reverse DNS hostnames. Our method significantly outperforms several state of the art academic baselines and it is competitive and complementary with commercial baselines. Our method outputs multiple plausible locations in case of ambiguity. It thus lends itself to being combined with other data sources to form a more complete geolocation database. Our future work will focus on combining reverse DNS hostname information with WHOIS databases and network delay to form a geolocation database across the entire IP space.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Explosive nucleosynthesis is a combination of the nuclear physics of thermonuclear reactions, and the hydrodynamics of the plasma in which the reactions occur. It depends upon the initial conditions—the stellar evolution up to the explosive instability, and the nature of the explosion mechanism.
Some key issues for explosive nucleosynthesis are the interaction of burning with hydrodynamics, the degree of microscopic mixing in convective zones, and the breaking of spherical symmetry by convection and rotation. Recent experiments on high intensity lasers provides new opportunities for laboratory testing of astrophysical hydrodynamic codes. Implications of supernovae 1987A and 1998bw (GRB980425?), and $\eta$ Carina are discussed, as well as the formation of black holes or neutron stars.
address: 'Steward Observatory, University of Arizona, Tucson AZ 85721, USA'
author:
- David Arnett
title: 'Explosive Nucleosynthesis: Prospects'
---
,
Nucleosynthesis ,Hydrodynamics ,Supernovae ,Lasers ,Black Holes ,GRB’s ,Rotation ,Convection ,Jets
Introduction
============
This paper is presented to honor David Norman Schramm. It will be a personal view of where this field is going, not a review of work already done, and in a sense it is a formal continuation of many previous discussions Dave and I had on this topic.
There is a growing consensus that the synthesis of the elements and their isotopes may be divided into three major components:
- Cosmological synthesis of the light elements,
- Hydrostatic synthesis in stars, and
- Hydrodynamic synthesis in stellar explosions.
The last item, “explosive nucleosynthesis,” will be the focus here. Although thermonuclear burning has a morphological equivalent in terrestrial burning by chemical reactions, this discussion will be couched in the language of physics, and named dimensionless numbers, whose meaning may not be obvious to scientists outside the combustion and/or the fluid dynamics community, will be used sparingly. These numbers can be constructed from ratios of the time scales and of length scales discussed below.
Any discussion of nucleosynthesis requires consideration of the empirical basis of the nuclear reaction rates. The historical context and references are collected in [@da96]. The recent appearance of an excellent review [@ktw98] allows us the freedom here to concentrate instead upon topics less discussed but of not less importance: the hydrodynamic context of the burning. However, the brevity of this mention should not be interpreted as an indication that the construction of an empirical basis for nucleosynthesis theory is a finished topic. To use nucleosynthesis products as a probe of stellar environments, it is necessary to insure that abundance features observed and simulated are due to the history of physical conditions, not poorly known reaction rates. Nuclear physics is the foundation upon which we build.
Explosive nucleosynthesis is a combination of the nuclear physics of thermonuclear reactions, and the hydrodynamics of the plasma in which the reactions occur. It depends upon the initial conditions—the stellar evolution up to the explosive instability, and the nature of the explosion mechanism.
Relevant Scales of Length and Time
==================================
There is a vast difference between macroscopic and microscopic lengths in stars (see [@da96], Chapter 11). Burning is a nuclear process which occurs in a context of enormous dimension. The radius of the a typical star, the Sun, is $7 \times 10^{10}\rm cm$, while the internuclear spacing is roughly $1.2 \times 10^{-8}{\rm\ cm} (A/\rho)^{1/3}$, where $\rho$ is the density in cgs units and $A$ the mass number of the most abundant nucleus. The mean density of the Sun is 1.4 g/cc (but in a presupernova densities increase to above $10^{9}$ g/cc). Suppose that collision cross sections have a scale of order $10^{-16}\rm\ cm^2$, corresponding to a collisional mean free path of $$\lambda_{col} \approx 1.6 \times 10^{-8} A/\rho \rm\ cm.$$ A characteristic scale for electron-photon interactions is the Thomson cross section, corresponding to a mean free path of roughly $$\lambda_{rad} \approx 1.6 A/\rho\rm \ cm.$$ The precise determination of cross sections, properly averaged over the relevant distributions is complex, but the qualitative result is the same as given by these simple estimates. This suggests that thermal energy moves much more easily than composition, and that our terrestrial intuition with flames may be colored by the ease with which heat moves on the length scales with which we are familiar.
Consider a homogeneous sphere of radius $R$; its diffusion time is $$\tau_{dif} = {3 \over \pi^2} R^2/\lambda v = {3 \over \pi^2} R^2 \rho N_A
\sigma/ v,$$ where $v$ is the velocity of the diffusing entities, $\sigma$ is the cross section, and $N_A$ is Avagadro’s number. Diffusion is slow for large objects like stars. However, time scales for binary interactions are $$\tau_{reac} \approx 1/(\rho N_A \sigma v).$$ This dramatically different density dependence makes reactions faster and diffusion slower at higher densities, that is, at advanced burning stages.
Simulations of stellar hydrodynamics and evolution presume that the stellar plasma is homogeneous on scales below the size of the computational zone. However, this is unlikely for the late stages of stellar life, which occur on time scales of seconds, not billions of years. This scaling with density insures that the problem of incomplete mixing gets worse as the star evolves. It is exacerbated in massive stars because they evolve most rapidly; they are primary sites of explosive nucleosynthesis.
Types of Nuclear Burning in Stars
=================================
The simplest sort of stellar burning is that of a “radiative zone,” in which heat can diffuse in and out, but composition is only changed by the conversion of fuel to ashes. This is thought to be the case in the center of the sun, for example, and the driving change which causes such stars to become red giants.
Even quasistatic nuclear burning can give locally intense heating, so that steep temperature gradients are formed which may drive convection. Convective currents are likely to be turbulent in stars because their large size and small viscosity give a large Reynolds number, that is, chaotic flow. This turbulent mixing is more effective than diffusion. As we shall see, the simulation of this represents a major challenge. Although such burning often is not explosive, it sets the stage for the explosive events, and seems to be their direct predecessor.
A violent but relatively simple process is detonation [@llfm]. The process is explosive, and supersonic. It is local in the sense that shock compression heats the fuel to the flash point, so that it incinerates. The motion of the shock is purely hydrodynamic except that the shock is enhanced by the energy released by the burning. The difficulty lies in determining how the detonation begins; this is an old and complex problem [@fd79].
Another violent type of burning is deflagration [@llfm], in which new fuel is ignited by heat flow from regions already burned. Unlike detonation, this depends upon the nature of the heat flow (conduction or radiative diffusion). Deflagration is subsonic, and in that sense is milder than detonation, into which it may develop. Deflagration is often unstable in stars (see Fig. 11.1 in [@da96]), which adds to the complexity of an already complex situation.
These themes, sufficiently complicated in their own right, are the basis for explosive nucleosynthesis. This level of complexity, and the need for quantitative predictions, makes computer simulations a necessary tool.
Laboratory Supernova
====================
Before trusting hydrodynamic simulations, detailed quantitative testing of the computational framework is necessary. High intensity lasers have been successfully used to study the Rayleigh-Taylor (RT) and Richtmeyer-Meshkov (RM) instabilities well into the nonlinear regime [@br94]. The RT instability occurs when gravity tries to pull a heavier fluid through an underlying lighter one. The RM instability is similar except that the role of gravity is replaced by the inertial acceleration from the passage of a shock wave. Core collapse supernovae are driven by a poweful shock, and such shocks are the breeding ground of hydrodynamic instabilities. Observations of SN 1987A strongly suggested the occurrence of mixing of radioactive material outward, a phenomena not seen in one dimensional (1D) simulations but predicted in 2D simulations [@afm89].
It is prudent to compare astrophysical codes to those mature computer codes used by the inertial confinement fusion (ICF) community on problems for which both should be applicable. This has been successfully done [@jk97]; both CALE (ICF) and PROMETHEUS (astrophysics) codes were used to simulate an experiment on the NOVA laser at Lawrence Livermore National Laboratory (LLNL). The observed instabilities (RT and RM) were well simulated by both codes; the bubble and spike positions were reproduced, even into the strongly nonlinear regime.
A theoretical look at the relation between the hydrodynamics occurring in SN 1987A and in the laser experiments shows that a rigorous mapping exists [@ry99]. Consider the He-H interface in the SN at 2000 s, and the Cu-CH interface in the laser experiment at 20 ns. In both, the ratio of inertial to viscous forces (the Reynolds number) and the ratio of the convective to conductive heat transport (the Peclet number) are large. Therefore the viscosity and thermal diffusivity are negligible, and the dynamics of the interface well described by Euler’s equations. These equations are invariant under a scale transformation, which maps lengths of $10^{11}\rm\ cm$ into $50 \rm\ \mu m$, densities of $8 \times 10^{-3}\rm\ g\ cm^{-3}$ into $4 \rm\ g\ cm^{-3}$, and pressures of $40 \rm\ Mbar$ into $0.6 \rm\ Mbar$, for example, at a star time of 2000 s and a laser time of 20 nanoseconds. Thus, in a very real sense, the experiment reconstructed a part of the supernova event.
Such experiments are also crucial for another problem: multidimensional geometry. The experiments are inherently 3D, but can be configured to give primarily 2D behavior. The computational load for 3D scales from 2D as the number of zones in the new dimension. At present, a single workstation can easily produce 2D simulatons with good resolution, but 3D requires parallel processing. For the next few years we will need to explore with 2D while we develop the capability of doing reliable and resolved 3D simulations with ease. The laser experiments can help us discover the qualitative and quantitative limitations of 2D in real world situations.
Laboratory experiments can have other impacts on explosive nucleosynthesis. Experiments modelling turbulent mixing, combustion, and flame propogation are needed. However, because the scales of the systems are so different, care must be taken in mapping the experiments into the astrophysical domain.
Thermonuclear Supernovae
========================
Supernovae of Type Ia are thought to be produced in a white dwarf star by a runaway thermonuclear reaction. They produce radioactive , and its decay to and thence to give the characteristic light curve [@da96]. SNIa’s are a major source of and other iron-group nuclei. Because of the empirical relationship between their brightness and the width of the luminosity peak (the Phillips relation), they are the best distance indicators now known, and of fundamental importance for cosmology [@mp93; @rpk95; @sp97].
Supernovae of Type Ia present several outstanding puzzles.
- What are their progenitors?
- How do they evolve to ignition?
- How do they evolve from ignition to explosion?
- Why does the Phillips relation between brightness and peak width work?
At present there is no unique and satisfactory scenario for their evolution up to explosion. For example, SNIa’s may result from one member of a binary pair growing in mass from accretion by its companion (there are many possibilities for the nature of that companion and of the matter accreted), or by the merger of a pair of white dwarfs (which is an inherently 3D problem, and not yet computable for secular time scales and good resolution).
If the accretion model is the correct one, how does it ignite? The favorite notion is that it ignites at the core of the white dwarf. To avoid collapse or excessive production of neutron rich isotopes, the ignition must not occur at densities which are too high ($\rho >> 2 \times 10^9 \rm\ g/cc$). From ignition to thermal runaway there is a period of about $10^3$ years, during which a convective region with Urca cooling probably evolves ( see references and discussion in [@da96; @mo96]), but such a process has only been simulated in 1D. What occurs in this “lost millenium” remains puzzling. See also [@rgi99] for simulations of the closely related problem of the core evolution of a 11 $M_\odot$ star.
If the burning does proceed to thermal runaway, the following evolution remains a subject of debate [@nh95; @sew94; @kh91]. It may take a new generation of simulations to resolve the issue.
The merger scenario also has uncertainty regarding the cause of explosion. Many discussions ( e.g., [@ii99]) assume that given a combined mass above the chandrasekhar limit, explosion must ensue. This is untrue. Collapse or benign ignition are also possible, maybe even more likely [@ml90; @sn98]. Pioneering efforts to simulate the merger process probably need better resolution and longer evolutionary times to get at this problem, or the related one of mergers involving other combinations of constituents (white dwarfs, neutron stars, black holes) [@dbh94; @scm97; @fr94; @mr97; @rj99].
Presumably from all this confusion will emerge a natural and convincing reason for the Phillips relation.
Almost Explosive Burning — Setting the Stage
============================================
Consider an evolved massive star, nearing core collapse. Its oxygen burning shell is an important region for explosive nucleosynthesis: this layer is the site of explosive oxygen and silicon burning as it is ejected by the supernova shock. Its formation and development set the stage for the collapse of the burned core to form a neutron star or black hole. Any discussion of core collapse, explosion mechanisms, or continued collapse to a black hole must presume characteristics of this formation and development, an issue we will return to below.
Almost all simulations of the stages prior to core collapse have been one dimensional (1D), assuming spherical symmetry and instantaneous mixing of radial layers over all angles. In 2D it is possible to treat convection as a real hydrodynamic process, although the vortices are pegged to the grid by assumption. Such simulations [@ba98; @aa99] are qualitatively different from the previous 1D ones.
This should not be surprising. The time scales are very short for such a large object. The duration of shell oxygen burning is only $10^4$ seconds or so, the connective turnover time about $10^2$ seconds, and the sound travel time about 1 second. The sound speed is about 0.01 of lightspeed. There is little time for the convection or the burning to settle into a steady state, or to make the zone well mixed in composition (or even in heat), contrary to the assumptions built into the stellar evolutionary codes. Strong downdrafts develop, and the convection is nonlinear and nonsymmetric with regard to up and down flows. The convective mach numbers approach tenths, and the perturbations in pressure and density are of the order of tens of percent at the flame zone and at the interface at the top of the formally convective zone. The burning is sporadic and flashy. Perhaps the most erroneous aspect of the 1D codes is their treatment of the boundary conditions on convection. Material moves across formally stable regions; in the 2D simulations was entrained from across the outer interface of the oxygen convective zone, and brought down into the flame zone where it flashed vigorously. This occurred after about 400 seconds in the Bazan-Arnett computation [@ba98].
These results have now been confirmed and extended by a completely independent hydrodynamic code and method [@aa99]. At 900 seconds there seems to be a new state developing, strongly dynamic but roughly steady on average. The nuclear luminosity increased more than a factor of 30 above the value obtained in the 1D simulations. The burning resembles a series of “mini-explosions”, in which the burning happens at higher temperatures in flashes, separated by relatively quiet intervals.
The most obvious conclusion is that the 1D simulations of the final stages of massive stars are unrealistic, and their degree of relevance is in question, at least regarding details of presupernovae and pre-explosion nucleosynthesis. We can already see that the shell luminosities are incorrect, as are the mixing algorithms. This brings into question the pre-collapse states hitherto used for core collapse simulations. These simulations show a dependence upon the neutronization (that is, upon $ Y_e$), the mass of the burned core, and its entropy. All these features may change.
While these new simulations do provide a first hydrodynamic description of convection for this evolutionary stage, they must be improved. First, they should be pushed all the way to core collapse, in order to determine the quantitative extent of the changes. Second, their 2D nature may be suspect. Rotation and magnetic fields, unavoidable is this stellar plasma, may reduce the geometric complexity, tending back toward 2D. On the other hand, vortex wandering in 3D may reduce the effects seen in 2D. This challenge is becoming tractable with progress in computer hardware and software.
It may be that the notion, that the extent of the burning shell is determined by the local adiabatic gradient, is flawed. The convective velocity field is certainly NOT local. Rather, the depth to which a blob sinks depends upon which fuels it has to flash and reverse its descent, upon how low its entropy drops due to previous neutrino cooling, and upon its history of electron capture. The actual compositional structure may be better thought of as an ensemble average of blobs being subjected to these effects. This view would imply revisions of hydrostatic as well and explosive nucleosynthesis yields, at least in detail, and perhaps in general.
In this picture, an important argument against significant rotation is removed. The abundances in a rapidly stirred region would be representative of the flame zone in that region, so that rotational mixing would tend to destroy the compositional layering needed to reproduce the solar system abundance pattern. However, if blobs were self limiting in their motion, depending upon their composition, the layering would represent both the temperature and compositon, and could survive.
Core Collapse Supernovae
========================
The dramatic rotational symmetry of the rings of SN1987A, and of the eruption of $\eta$ Carina, suggest that rotation is important for at least the late stages of evolution of massive stars. Given this hint, let us re-examine how rotation might and might not have consequences for supernovae. It appears that one of the worst problems, the destruction of compositonal layering just discussed, may be moot with the new view of convective burning in the presupernova.
The Neutrino Diffusion Model
----------------------------
In the Colgate model of core collapse [@cw66], the collapsing core was supposed to be both thick and thin to neutrinos. It had to be thick so that neutrinos were copiously produced, but thin so they would stream out, but thick enough again to deposit energy as they escaped. The neutrino transport was not actually calculated in [@cw66], but assumed to work in this fashion. That is, the neutrinos were assumed to diffuse quickly out of the collapsed core, and half of their energy was deposited in the mantle. All the models so calculated gave violent explosions; no black holes were formed.
The first radiation hydrodynamic calculations [@da66; @da67] showed that if such fine tuning were allowed, explosions did result, but also showed that black hole formation with no explosion was also likely. This follows from a simple argument. The diffusion time out of a homogeneous sphere of mass $M$ is $$\tau_{dif} = {3 \over \pi^2c}\kappa ( 3M/4\pi)^{2/3} \rho^{1/3},$$ where $\kappa$ is the neutrino opacity and $\rho$ the density. The collapse becomes supersonic, so that it takes a time which is of the order of and scales with the free fall time, $ \tau_{ff} \propto R/v_{ff}$. Since $v_{ff}^2 = 2 G M / R$, we have $\tau_{c} \propto \rho^{-1/2}$. The degree of neutrino trapping depends upon the ratio of diffusion time to collapse time, $$\tau_{dif} / \tau_{c} \propto \kappa \rho^{5/6} M^{2/3}.$$ Increasing the neutrino opacity, the density, or the core mass $M$ tends to increase the neutrino trapping, and reduce the chance of explosion. Massive cores tended to make black holes. With the advent of the neutral current theory of weak interactions, the effective value of the neutrino opacity increased, increasing trapping. Modifying the inital models or the nuclear equation of state to give higher density at bounce also increased trapping. If the neutrinos are trapped in the core, the collapse continues on to black hole formation. Nor does arbitrary tuning of $\kappa$ fix things. If the neutrino opacity is low, the neutrinos escape but do little heating of the surrounding and infalling mantle.
These parameters are not freely variable. The mass $M$ is constrained by the progenitor core mass, which itself cannot be less than the chandrasekhar mass. The weak interaction determines $\kappa$ fairly precisely. The density of the core has varied in simulations, but more care, realistic evaluation of nuclear equation of state, and inclusion of general relativity, give more tightly constrained values. It seems fair to say that the neutrino diffusion model does not work with realistic physics. See [@mb98] for a recent review of the status of core collapse models.
Energetics
----------
The unusual supernova 1998bw, and its possible identification [@ga98; @iwa98; @wwh98] with the gamma-ray burst GRB980425, suggest that the supernova mechanism should be able to provide large explosion energies ($ E_{exp} \ge 10^{52}\rm\ erg$, or more than 10 foe). The gravitational potential energy is $$\Omega \approx G (4\pi/3)^{1/3} M^{5/3} \rho^{1/3},$$ in the newtonian approximation. The minimum core density is about that of the atomic nucleus, $$\rho_{nuc} = 2 \times 10^{14}\rm\ g/cc.$$ A schwarzschild black hole has an average density of $$\rho_{bh} = 3c^6/32\pi G^3M^2 = 2.85 \times 10^{16} (M_\odot/M)^2 \rm\ g/cc,$$ but bounce densities as low as $10^{15} \rm\ g/cc$, dynamical formation of a black hole can occur. Thus $ 2 \times 10^{14} \le \rho \le 10^{15}$ gives an estimate for the maximum energy available for explosion. In the spherically symmetric case, black hole formation at higher masses will limit the energy available for explosion; just below this boundary the energy supply is at a maximum, e.g. [@vra79].
The Shock Mechanism
-------------------
If diffusion of neutrinos will not give adequate transport of energy, shock propagation might; Bethe and Brown have led the pursuit of this possibility [@be90]. Upon reaching nuclear density, or higher, the pressure is adequate to support the collapsing core. The size of this region is given by equating the pressure gradient force to gravity. The mass, for which this is true, falls as a unit, with velocity proportional to radius, so it is called the “homologous core.” Prior to collapse the central density reaches $\rho \ge 2 \times 10^9 \rm\ g/cc$. Neutrino cooling keeps the entropy low, so that the “iron core mass” approaches the chandrasekhar value, $$M_{ch} / M_\odot \approx 1.45 (2 Y_e)^2,$$ where $Y_e$ is the number of electrons per nucleon. The electron fermi energy is several MeV. Electron capture occurs relatively slowly, but at $Z/A = Y_e \approx 0.42$, the nuclei have a threshold for electron capture of several Mev as well, and the neutronization is almost stopped [@da96]. The smallest iron core is about $1.0 M_\odot$ for $y_e = 0.42$. The largest homologous core would occur if no leptons were lost in subsequent collapse, so their pressure would have the largest possible value. At nuclear density, the difference between proton and neutron chemical potentials will be small compared to the fermi energies of electrons and neutrinos (of order 100 Mev), so at weak-interaction equilibrium, $$\mu_e \approx \mu_\nu.$$ This implies $ N_e/g_e = N_\nu/g_\nu$, where the helicity phase space factors are $g_e = 2$ and $g_\nu = 1$, so $Y_e = 2 Y_\nu$, and with no lepton escape we have for the collapsed core $Y_e = 0.28$ and $Y_\nu = 0.14$. The pressure defict is $ P/P_0 = (2/3)^{4/3} + {1 \over 2}(1/3)^{4/3} \approx 0.700$. The corresponding mass deficit is $ (0.7)^{3/2} = 0.585$, so that only about 60% of the iron core is still in the homologous core when it reaches nuclear density and bounces. The shock must propagate through about 40% of that iron core which is still infalling. For the shock to be strong, it must dissociate the iron, requiring 8 MeV per nucleon, or 6.4 foe ($6.4 \times 10^{51} \rm\ erg$). The energies available in such small collapsing cores is almost always smaller than this, thus making the shock mechanism doubtful.
The Convective Mechanism
------------------------
Although the core collapse releases much more energy than seems to be necessary for the typical supernova, the problem lies in getting it out of the nascent neutron star/black hole. One possible solution is “convective overturn.” The term “overturn” is important because the process is unlikely to resemble a well-developed turbulent cascade, but rather a more violent and transient large scale turnover of the lepton-rich deep regions. Epstein [@rie79] first examined the consequences of this possibility. Early numerical simulations [@bbl79; @lbc80] and arguments [@cp80] based on this idea were shown by Smarr et al. [@swbb81] to be overly enthusiastic, although the overturn of the outer core would be a generic and important feature of the core collapse.
Given the difficulties of the neutrino diffusion and the shock models, it would appear better to allow almost all of the iron core to fall in, then release the neutrinos at a later time. Because there would be less mass to dissociate, this minimizes the dissociation losses. The first simulation which showed such “delayed” behavior is due to Wilson [@bw85]. This was a 1D simulation and therefore had a dubious treatment of convective flow. It did stimulate multidimensional simulations [@hbhfc94; @bhf95; @jm96; @mcbb97] which gave results still being argued.
The First Rotational Mechanism
------------------------------
Fred Hoyle [@fh46; @fh64] proposed that rotation played an important role in supernovae. If there is enough angular momentum in the oxygen shell layer which surrounds the iron core, collapse might induce explosion by oxygen burning. This did not result in the Colgate-White simulations because their collapse generated an excessively strong rarefaction wave which swallowed the oxygen shell. A more careful treatment of the onset of core collapse [@da77] showed a longer initial contraction, in which the oxygen shell did burn violently (but not quite explosively). Following Fowler and Hoyle’s suggestion [@fh64], Bodenheimer and Woosley [@bw83] simulated some simple explosions of this type, with rough estimates of the rotational state of the presupernova. These should be re-examined using more realistic multidimensional precollapse models and better resolution. The combined effects of rotation and hydrodynamics on the core mass, the neutron excess, and the entropy would be interesting in its own right. Even in the 1D case the explosion of SN1987A would have gotten about 10% of its energy from explosive burning of oxygen.
Rotation Revisited
------------------
Rotation could cause significant effects in the core itself [@mm89]. Extreme rotation would cause the collapse to halt due to centrifugal forces at a density less than nuclear density. While this would aid neutrino escape, it would also release less gravitational energy and lower the neutrino energies, making them less able to deposit energy in the outer layers. An unresolved issue is the rate of angular momentum transport, which in this case would determine the secular evolution to the neutron star or black hole state.
Magnetic fields might be important as well [@fh46; @go71; @lbw70]. The plasma is matter, not field, dominated (a high $\beta$ plasma), so that magnetic fields would be subtle, at least initially. With dynamo action the field would be strengthened, and buoyancy would tend to move it to regions in which its effects might be still more important. This is justifiably an old problem because it is inherently 3D and time dependent. The failure of the pulsar models to predict a luminosity below the observed radioactive decay of in SN1987A may indicate that magnetic effects are not a dominant feature; however evidence for a pulsar (or alternatively a black hole) would clarify this point.
A more modest and perhaps realistic (?) role for rotation is to induce mixing, but not centrifugal braking. The rotation and magnetic field would guide the overturn, emphasizing large scale motion (low modes). This may help the “convection” model discussed above.
In any case, rotation is likely to have an important effect in that it breaks the spherical symmetry in a characteristic way. While neutrino diffusion of energy will tend to be spherically symmetric, rotation tends to give the rotation axis a special role. Centrifugal force tends to evacuate these regions, so that they would have a lower density and, if heat transport is effective, a higher entropy. Such polar hot spots might be conducive to the formation of jets, and trigger overturn. Preliminary attempts to examine the consequences [@kh99; @mw98] are promising.
Summary
=======
- Explosive nucleosynthesis is a combination of nuclear reactions with hydrodynamics, and depends upon explosions mechanisms for supernovae.
- Incomplete mixing gets worse for massive stars and explosive conditions.
- Nuclear burning, and yield predictions, is complicated by hydrodynamic convection in presupernovae.
- Laboratory experiments with high intensity lasers has become a good testing ground for astrophysics codes. With careful scaling, such experiments can reproduce supernova phenomena.
- Understanding of SNIa’s is impeded by a lack of progenitor information, and by theoretical problems with approach to ignition, runaway, and stellar merger.
- Simulations of oxygen shell burning using actual (2D) hydrodynamics differ drastically from 1D results. The first simulations have been confirmed by an independent code, and is being carried further.
- For core collapse explosion mechanisms, neutrino diffusion and prompt shock models are dead, and pure convection models may be sick.
- Rotation must be included in progenitor and core collapse evolution.
- SN1998bw shows that the energy problem with core collapse supernovae is worse than previously supposed. There are events having energies much larger than several foe ($10^{51}\rm\ erg$).
- Because newtonian gravity and centrifugal force are scale free, jet formation is likely to occur on scales having effective heat flow. This may connect protostar jets and galactic jets (with heat flow by radiative diffusion and convection), and core collapse supernovae (with heat flow by neutrino diffusion and convection).
The prospects are simply wonderful. Our tools are getting much better, and may finally be up to the task of simulation of explosive nucleosynthesis events in realistic geometry. Computer technology, nuclear reaction rates, and hydrodynamics rates are improving and being verified in new ways. Meteoritic data [@ez98] has presented quantitative challenges to 1D model predictions. With the impending crash of SN1987A into its rings, the renewed activity of $\eta$ Carina, the active supernova searches out to large redshift, and the possibility of a connection between core collapse supernovae and GRB’s, we may expect to learn many new things.
Acknowledgment {#acknowledgment .unnumbered}
==============
This research is supported by DOE grant DE-FG03-98DP00214/A001.
[9]{}
D. Arnett, 1966, Can. J. Phys. 44, 2553
D. Arnett, 1967, Can. J. Phys. 45, 1621
D. Arnett, 1977, ApJ 218, 815
D. Arnett, 1996, [*Supernovae and Nucleosynthesis*]{}, (Princeton University Press: Princeton NJ)
D. Arnett, B. A. Fryxell, & E. Müller, 1989, ApJ 341, L63
S. Ashida, E. Livne, & D. Arnett, 1999, in preparation
G. Bazan & D. Arnett, 1998, ApJ 496, 316
H. A. Bethe, 1990, Rev. Mod. Phys 68, 801
H. A. Bethe & J. R. Wilson, 1985, ApJ 295, 14
P. Bodenheimer, & S. E. Woosley, 1983, ApJ 269, 381
S. W. Bruenn, J. R. Buchler, & M. Livio, 1979, ApJ 234, L183
A. Burrows, J. Hayes, & B. A. Fryxell, 1995, ApJ 450, 830
S. A. Colgate & R. H. White, 1966, ApJ 143, 626
S. A. Colgate & A. G. Petchek, 1980, ApJ 236, L115
M. B. Davies, W. Benz, & J. G. Hills, 1994, ApJ 424, 870
R. I. Epstein, 1979, MNRAS 188, 305.
W. A. Fowler & F. Hoyle, 1964, ApJS 9, 201
W. Fickett & W. C. Davis, 1979, [*Detonation*]{}, (University of California Press: Berkeley CA)
T. J. Galama, et al., 1998, Nature 395, 670
J. E. Gunn, & J. P. Ostriker, 1971, ApJ 160, 979
M. E. Herant, W. Benz, W. R. Hix, C. Fryer, & S. A. Colgate, 1994, ApJ 435, 339
F. Hoyle, 1946, NMRAS 106. 343
I. Iben, Jr., & A. V. Tutukov, 1999, ApJ 511, 324
K. Iwamoto, et al., 1998, Nature 395, 672
H.-Th. Janke, & E. Müller, 1996, AAp 306, 167
J. Kane, [*et al.*]{}, 1997, ApJ 478, L75
F. Käppler, F.-K. Thielemann, & M. Wiescher, 1998, [*Annu. Rev. Nucl. Part. Sci.*]{} [**48**]{}, 175
A. Khokhlov, 1991, AAp 245, 114
A. Khokhlov, et al., 1999, ApJ, submitted
L. D. Landau & E. M. Lifshitz, 1959, [*Fluid Mechanics*]{}, (Addison-Wesley: New York)
J. M. LeBlanc & J. R. Wilson, 1970, ApJ 161, 541
M. Livio, J. R. Buchler, & S. A. Colgate, 1980, ApJ 238, L139
A. MacFadyen & S. E. Woosley, 1999, astro-ph/9810274
A. Mezzacappa & S. W. Bruenn, 1998, in proceedings of [ *Future Directions is Supernova Research*]{}, to appear in [*Memoirs of the Italian Astronomical Society*]{}, held at Assergi, Italy, Sept. 29-Oct. 2, 1998
A. Mezzacappa, et al., 1997, ApJ 495, 911
R. Mönchmeyer & E. Müller, 1989, in [*NATO ASI on Timing Neutron Stars,*]{} ed. H. Ögelman & E. van den Heuvel, (Dircrecht: Kluwer), p. 549
R. Moschkovitch, 1996, AAp 311, 152
R. Moschkovitch, & M. Livio, 1990, AAp 236, 378
J. Niemeyer & W. Hillebrandt, 1995, ApJ 452, 779
S. Perlmutter, [*et al.*]{}, 1997, ApJ 517, 565
M. M. Phillips, 1993, ApJ 413, L105
F. Rasio, 1994, BAAS 184, 41.12
B. A. Remington, [*et al.*]{}, 1994, [*Phys. Plasmas*]{} 4, 1994
A. Riess, W. H. Press, & R. P. Kirshner, 1995, ApJ 438, L17
C. Ritosso, E. Garcia-Berro, & I. Iben, Jr., 1999, ApJ 515, 381
M. Ruffert, 1997, in Proceedings of Fourth Huntsville Gamma-Ray Burst Symposium, 15-20 Sept., 1997
M. Ruffert, & H.-Th. Janke, 1999, AAp 344, 573
D. Ryutov, [*et al.*]{}, ApJ, in press (June, 1999)
H. Saio, & K. Nomoto, 1999, ApJ 500, 388
L. Sergretain, C. Chabrier, & R. Mochkovitch, 1997, ApJ 481, 355
L. Smarr, J. R. Wilson, R. T. Barton, & R. L. Bowers, 1981, ApJ 246, 515
K. A. van Riper, & D. Arnett, 1979, ApJ 225, L129
L. Wang & J. C. Wheeler, 1998, ApJ 504, L87
S. E. Woosley, 1994, in [*Supernovae*]{}, ed. S. A. Bludmann, R. Mochkovitch, J. Zinn-Justin, (Amsterdam: Elsevier Science)
E. Zinner, 1998, Annu. Rev. Earth Planet. Sci. 26, 147
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report results from the Supernova Photometric Classification Challenge (), a publicly released mix of simulated supernovae (SNe), with types (Ia, , and II) selected in proportion to their expected rate. The simulation was realized in the $griz$ filters of the Dark Energy Survey (DES) with realistic observing conditions (sky noise, point-spread function and atmospheric transparency) based on years of recorded conditions at the DES site. Simulations of type SNe are based on confirmed light curves that include [*unpublished*]{} samples donated from the Carnegie Supernova Project (CSP), the Supernova Legacy Survey (SNLS), and the Sloan Digital Sky Survey-II (SDSS–II). A confirmed subset was provided for training. We challenged scientists to run their classification algorithms and report a type and for each SN. Participants from groups contributed entries for the sample that included a host-galaxy for each SN and entries for the sample that had no redshift information. Several different classification strategies resulted in similar performance, and for all entries the performance was significantly better for the training subset than for the unconfirmed sample. For the unconfirmed subset, the entry with the highest average figure of merit for classifying SNe Ia has an efficiency of and an SN Ia purity of . As a public resource for the future development of photometric SN classification and estimators, we have released updated simulations with improvements based on our experience from the , added samples corresponding to the Large Synoptic Survey Telescope (LSST) and the , and provided the answer keys so that developers can evaluate their own analysis.'
bibliography:
- 'SNchallenge\_results.bib'
title: Results from the Supernova Photometric Classification Challenge
---
Motivation {#sec:intro}
==========
To explore the expansion history of the universe, increasingly large samples of high-quality SN Ia light curves are being used to measure luminosity distances as a function of redshift. With rapidly increasing sample sizes, there are not nearly enough resources to confirm each SN. Currently, the largest samples are from the Supernova Legacy Survey (SNLS: [@Astier06]) and the Sloan Digital Sky Survey-II (SDSS-II: @York00 [@Frieman07]), each with more than 1000 SNe Ia, yet less than half of their SNe are confirmed. The numbers of SNe are expected to increase dramatically in the coming decade: thousands for the Dark Energy Survey (DES: @DES-moriond2009) and a few hundred thousand for the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS)[^1] and the Large Synoptic Survey Telescope (LSST: @Ivezic_08 [@LSSTSB09]). Since only a small fraction of these SNe will be confirmed, is crucial to fully exploit these large samples.
In the discovery phase of accelerated cosmological expansion, results were based on tens of high-redshift SNe Ia, and some samples included a significant fraction of events that were not classified from a spectrum [@Riess_1998; @Riess_2004; @Perl_1997; @Tonry_2003]. While human judgment played a significant role in classifying these “photometric” SNe, more formal methods of classification have been developed over the past decade: @Poz2002 [@Poz2007a_classify; @Dahlen2002; @GalYam2004; @Sulli2006; @Johnson2006; @Kuz2007; @BEAMS2007; @Rodney2009]. Some of these techniques have been used to select candidates for observations and rate measurements [@Barris2006; @Neill2006; @Poz2007b_rate; @Kuz2008; @Dilday2008], but these methods have not been used to select a significant SN Ia sample for a Hubble-diagram analysis. In short, cosmological parameter estimates from the much larger recent surveys are based solely on confirmed SNe Ia (SNLS: @Astier06, ESSENCE: @WV07, CSP: @Freedman_2009, SDSS-II: @K09).
The main reason for the current reliance on is that vastly increased resources have been used in these more recent surveys. In spite of these increased resources, however, more than half of the discovered SNe lack , and therefore methods must be used to classify the majority of the SNe. There are two difficulties limiting the application of classification. First is the lack of adequate data for training algorithms. Many classification algorithms were developed using publicly available Nugent templates,[^2] consisting of a single spectral energy distribution (SED) template for each type. The Nugent templates were constructed from averaging and interpolating a limited amount of confirmed data [@Levan2005; @Hamuy2002; @Gill99; @Baron2004; @Cap1997], and therefore the impact of the diversity has not been well studied. The second difficulty is that there is no standard testing procedure, and therefore it is not clear which classification methods work best.
To aid in the transition to using SN classification, we have released a public “SN Photometric Classification Challenge,” hereafter called . The announcement of the challenge and instructions to participants were given in a challenge release note [@SNchallenge2010], and an electronic mail message alert was sent to several dozen SN experts. The consisted of a blinded mix of simulated SNe, with types (Ia, Ib, Ic, II) selected in proportion to their expected rate. From 2010 January 29 through June 1, the public challenge was open for scientists to run their classification algorithms and report a type for each SN. A confirmed subset was provided so that algorithms could be tuned with a realistic training set. The goals of this challenge were to (1) learn the relative strengths and weaknesses of the different classification algorithms, (2) improve the algorithms, (3) understand what confirmed subsets are needed to properly train these algorithms, and (4) improve the simulations.
To address the paucity of data, the CSP, SNLS and SDSS-II generously contributed [*unpublished*]{} confirmed light curves. These data are high-quality multiband light curves, and we are grateful to the donating collaborations. Since these SNe are from surveys focused mainly on collecting type Ia SNe, this sample is brighter than the true population. In spite of this bias toward brighter , we anticipated that this challenge would be a useful step away from the overly simplistic studies that have relied on a handful of templates.
The outline of this article is as follows. In §\[sec:sim\] we present full details of the simulation, including strengths, weaknesses and bugs found during the . In §\[sec:take\_challenge\] we describe the classification methods used by the participating groups. The figure of merit used for evaluation is defined in §\[sec:eval\], and the results for all of the participants are presented in §\[sec:results\]. Updated simulations are described in §\[sec:sim\_update\], and we conclude in §\[sec:end\].
The Simulation {#sec:sim}
==============
Here we present full details of how the simulated samples were generated using the software package[^3] [@SNANA09]. Both the strengths and weaknesses are discussed to motivate improvements in future simulations. The limited information available to participants during the challenge is given in § 2 of the challenge release note [@SNchallenge2010].
Simulation Overview {#subsec:sim_overview}
-------------------
The simulation was realized in the $griz$ filters of the Dark Energy Survey (DES), and distances were calculated assuming a standard $\Lambda$CDM cosmology with $\OM=0.3$, $\OL=0.7$ and $w=-1$. The sky-noise, point-spread function and atmospheric transparency were evaluated in each filter and each epoch using a year long history of actual conditions from the ESSENCE project at the Cerro Tololo Inter-American Observatory (CTIO).[^4] For the five SN fields selected for the DES (3 deg$^2$ per field), the cadence was based on allocating 10% of the DES photometric observing time and most of the nonphotometric time. The cadence used in this publicly available simulation was generated by the Supernova Working Group within the DES collaboration.[^5] Since the DES plans to collect data during five months of the year, incomplete light curves from temporal edge effects are included; i.e., the simulated explosion times extend well before the start of each survey season, and extend well beyond the end of the season.
The included a sample with a host-galaxy photometric redshift () and another sample with no redshift information (). For the former, the estimates were based on simulated galaxies (for DES) analyzed with the methods in @Oy08a [@Oy08b]. The average host-galaxy resolution is 0.03, and the distribution includes non-Gaussian outliers. A challenge with precise redshifts was not given because using accurate redshifts makes little difference on the classifications compared with using a host-galaxy .
Two simple selection criteria were applied. First, each object must have an observation in two or more passbands with a signal-to-noise ratio (S/N) above 5. Second, there must be at least five observations after explosion, and there is no S/N requirement on these observations. These requirements are relatively loose because part of the challenge was to determine the optimal selection criteria. For the five seasons planned for the DES, the total number of generated SNe for all types was $1.01\times 10^5$. The number satisfying the loose selection requirements and included in the was $1.8\times 10^4$.
Type Ia Model {#subsec:simIa}
-------------
Simulated SNe Ia were based on an equal mix of the [@Guy07] and models [@JRK07; @K09]. Since these two models do not agree in the ultraviolet region, we used a special version in which the ultraviolet region was adjusted to match that of the model. The treatment of color variations corresponding to each model was used. For , extinction by dust resulted in reddened SNe Ia. The dust parameter $R_V$ was drawn from an asymmetric Gaussian distribution peaked at $R_V=2.0$ with sigmas of 0.2 and 0.5 for the low and high side, respectively, and the $R_V$ values were constrained to lie between 1.5 and 4.1; this $R_V$ distribution has a mean value of 2.2. For , the color-magnitude adjustment was given by $\beta c$ where $\beta = 2.7$ and $c$ is the color excess, $E(B-V)$. The $c$ parameter was drawn from a Gaussian distribution with $\sigma_c = 0.1$ and the constraint $\vert c \vert < 0.4$.
In addition to the model parameters, we have simulated the anomalous Hubble scatter with random color variations. For each passband $f$, a random shift was drawn from a Gaussian distribution with $\sigma_m = 0.09$ mag, and this magnitude shift was applied coherently to all epochs within the passband. The scatter in each color was therefore $0.09\cdot \sqrt{2}$ mag.
For the it is important to include photometric passbands that correspond to rest-frame wavelengths outside the nominally defined ranges of the SN Ia models: specifically, the $g$ and $r$ bands at higher redshifts that probe the far ultraviolet region. Without an estimate of the redshift, analysis programs cannot initially select observations that correspond to a particular rest-frame wavelength range. Since the spectral surfaces of the SN Ia models are defined over a much larger range than that where the models are defined, it is straightforward to extend the wavelength range in the simulation. For both models, the lower wavelength range[^6] was reduced to 2500 Å. To simulate redder passbands for , the upper range was extended from 7000 [Å]{} to 8700 [Å]{}.
Non-Ia SN Model {#subsec:sim_nonIa}
---------------
Simulated photometry of SNe was based on confirmed type light curves from the CSP, SNLS, and SN surveys. The basic strategy is to smoothly warp a standard SED to match the observed photometry and then use the warped SEDs to simulate SNe at all redshifts. After correcting the light curves for Galactic extinction, the light curve for each passband was smoothed using a general function based on that used in the rate analysis in @Bazin2009, $$f(t) = A_0 [ 1 + a_1(t-t_0) + a_2(t-t_0) ]
\frac{ e^{-(t-t_0)/\Tfall} }
{1 + e^{-(t-t_0)/\Trise} }~.
\label{eq:smooth}$$ The parameters $A_0$, $t_0$, $\Trise$, $\Tfall$ and $a_{1,2}$ are fit separately for each passband. The polynomial parameters $a_{1,2}$ were initially fixed to zero; in cases where the fit was inadequate as determined by visual inspection, the fit was redone with the additional $a_{1,2}$ parameters. Examples of smoothed light curves, also called templates, are shown in Fig. \[fig:smooth\_non1a\] for the SNe that were most commonly misidentified as an SN Ia during the (§\[sec:results\]). To use a template in the , the corresponding light curve was required to have good sampling in all passbands, and this requirement was based on visual examination rather than rigorous cuts. Among the 86 confirmed from the , 34 were selected for the ; for the CSP, 5 of 6 were selected, and for the SNLS, 2 of 9 were selected. A list of the SNe used in the is shown in Table \[tb:nonIa\_list\]; combining the surveys, the numbers of types , and are , , and , respectively (also see Table \[tb:nonIa\_subtypes\]).
While the general fitting function (Eq. \[eq:smooth\]) appears adequate upon visual inspection, we note that the rise-time parametrization is not always accurate. For SN 14475 in Fig. \[fig:smooth\_non1a\], the rise time is well sampled and hence the smoothed template is reliable in this region of the light curve. For CSP-2006ep, however, the $u$-band rise time is not well sampled and therefore the smoothed rise time is dependent on the particular parametrization. Ideally the rise-time shape from well-measured light curves would be used as an additional constraint in the smoothing function, but such constraints were not used in this .
The next step is to create a rest-frame time series of SEDs such that the redshifted synthetic magnitudes match those of the smoothed light-curve template at each epoch. These spectral time sequences are called “ template SEDs.” The starting SED for each subtype is taken from the Nugent template, and it is then warped at each epoch to match the observer-frame photometry. For a simulated type and redshift, the corresponding template SED is used to compute observer-frame $griz$ magnitudes.
In addition to the template SEDs we have also included four Nugent SED templates, each representing a composite average over one of the subtypes shown in Table \[tb:nonIa\_subtypes\]. The magnitudes were drawn from Gaussian distributions as described in @Richardson2002.
The final step is to apply random color variations in the same manor as for the type Ia SNe. While the anomalous scatter in the SN Ia Hubble diagram motivates this step in the SN Ia simulation, the motivation for the simulation is to describe a potentially broader class of objects. In the limit of a large and complete set of templates there would be no need to simulate additional sources of magnitude variation. We have made the assumption, however, that our set of templates is not large enough to describe the population.
[llccc]{}
\[tb:nonIa\_list\]
SN Rates and Template Weights {#subsec:sim_rates}
-----------------------------
Following @Dilday2008, the SN Ia volumetric rate ($r_V$) was parametrized as $r_v = \alpha(1+z)^{\beta}$ with $\alpha_{\rm Ia} = 2.6\times 10^{-5}$ ${\rm Mpc}^{-3}\,h_{70}^3\,{\rm year}^{-1}$, $\beta_{\rm Ia} = 1.5$, and $h_{70} = H_0/(70\,{\rm km}\,s^{-1}\,{\rm Mpc}^{-1})$ where $H_0$ is the present value of the Hubble parameter. Integrating out to a redshift of $z = 1.1$, the total number of generated SN Ia for the DES survey is $\sim \NGENIa$, and the number written for the (i.e., passing the loose cuts in §\[subsec:sim\_overview\]) is $\sim \NCUTIa$.
For the rate, we assumed that the redshift dependence has the same general form as for the SNe Ia. The exponent term $\beta_{\rm nonIa} = 3.6$ was taken to match that of the star formation rate. To estimate $\alpha_{\rm nonIa}$ we use the result of @Bazin2009 which reports an observed non-Ia/Ia rate ratio of $4.5 \pm 1.0$ for $z<0.4$. We then calculate $\alpha_{\rm nonIa} = 6.8\times 10^{-5}$ such that the /Ia rate ratio matches the observed ratio. Since the rate has a much larger at redshifts above 0.4, and to increase the sample of misclassified , the rate was arbitrarily increased by a factor of 1.3 at all redshifts. Integrating out to a redshift of $z = 1.1$, the total number of generated for the DES survey is $\sim \NGENnonIa$, and the number written out for the is $\sim \NCUTnonIa$.
The generated non-Ia/Ia ratio over all redshifts is 12. After applying the loose selection requirements for the sample (§\[subsec:sim\_overview\]), this ratio drops to 2.4. We have likely overestimated the contribution, but this overestimate was intentional in order to increase the statistics of SNe that are misidentified as SN Ia.
The breakdown of the into subtypes (,,, and ) is taken from @Smartt2009, and the subtype fractions are shown in Table \[tb:nonIa\_subtypes\] along with the number of templates used to represent each subtype. Within a subtype class, each template is given equal weight in the generation of simulated samples. For each subtype a composite Nugent template is included, and is given the same generation weight as each template based on an observed light curve.
[lccc]{} & & No. of & No. of\
Non-Ia & & measured & composite\
subtype & Fraction & templates & templates\
& & & 1\
& & & 1\
& & & 1\
& & & 1\
\[tb:nonIa\_subtypes\]
Spectroscopic Subset {#subsec:sim_spec_subset}
--------------------
To allow participants to train their classification algorithms, a confirmed training subset was provided. This subset was based on from a 4 m class telescope with a limiting $r$-band magnitude of $\rspeclimit$, and on from an 8 m class telescope with a limiting $i$-band magnitude of $\ispeclimit$. Using this selection resulted in a subset of $\NSPEC$ objects, or about 7% of the total number of objects in the . This training sample is not a random subset and is, in fact, a highly biased subset, as shown in Fig. \[fig:imag\]; the true SN Ia fraction for the confirmed SNe is 70%, compared with only 26% for the unconfirmed SNe. While a truly random subset would be ideal for training classification algorithms, limited resources in future surveys are much more likely to obtain a biased sample unless there is sufficient motivation to modify the targeting strategy.
If each SN spectrum were taken exactly at the epoch of peak brightness ($t_0$), then the for obtaining a spectrum adequate for classification would depend only on the peak magnitude. However, a spectrum is typically taken slightly before or after $t_0$, when the SN is slightly dimmer than at peak brightness; therefore we have parametrized the efficiency for obtaining a spectrum ($\effspec$) to be $$\effspec = \epsilon_0(1-x^{\ell})~,~~~
x \equiv \frac{\magpeak - \magmin}{\maglim - \magmin}~,$$ where the parameters $\ell$, $\magmin$ and $\maglim$ are given in Table \[tb:effspec\] for the $r$ and $i$ filters, and $\magpeak$ is the SN magnitude at $t_0$. The coefficient $\epsilon_0 = 0.4$ for type Ia and 0.3 for ; this difference in the $\epsilon_0$ values was due to an error in the simulation (§\[subsec:sim\_bugs\]). The efficiency function is nearly flat for bright SNe and then decreases rapidly to zero at the limiting magnitude. A simulated SN is identified if $21.5 < \magpeak^i< \ispeclimit$ and a randomly generated number (0–1) is less than $\effspec^i$, or if the analogous criterion is satisfied for the $r$ band. Since the $\effspec$ parametrization is an educated guess, future simulations should use a more refined parametrization based on the range of epochs in which are expected to be obtained.
[l | lll]{} Filter & $\ell$ & $\magmin$ & $\maglim$\
$r$ & 5 & 16.0 & $\rspeclimit$\
$i$ & 6 & 21.5 & $\ispeclimit$\
\[tb:effspec\]
Bugs {#subsec:sim_bugs}
----
Taking the SN Classifier Challenge {#sec:take_challenge}
==================================
As described in §\[sec:sim\], two independent challenges were generated: one with a host-galaxy for each SN and another without any redshift information. In addition to these challenges based on the entire light curve, there was also an early-epoch challenge motivated by the need to prioritize SNe for follow-up observations; this challenge was based on the first six photometric observations (in any filter) with ${\rm S/N} > 4$. Participants attempted the full light-curve challenges with and without redshift information, but none of the participants attempted the early-epoch challenge, due to time limitations and the increased interest on the full light curve challenge that will eventually impact the cosmology analyses.
The simulated light curves are available at the Web site.[^7] Details on how to analyze the simulated sample are given in [§3]{} of the release note. To fully optimize classification algorithms during the challenge, several participants wanted to know the exact value of the false-tag weight (§\[sec:eval\]) used to determine the figure of merit. On 2010 April 27 we therefore publicly announced that $\wIafalse = \wIafalseVALUE$; while this information clearly helped some participants optimize results for the confirmed subset, it is not clear if the information improved results for the unconfirmed sample.
A total of groups (or individuals) sent submissions to be evaluated. Among the submissions, are based on the , while the remaining are based on the . Photo-$z$ estimates were given by four participants in the and by three participants in the .
Table \[tb:participants\] shows the list of groups and participants, indicates which challenge(s) were taken, and indicates if SN estimates were given. The average processing time is also given for each method, and these times vary from 1 s to $>200$ s per SN using similar processors. A brief description for each method is given in Appendix \[app:methods\].
Among the participants, four general strategies were used to classify SNe. The first and simplest strategy was to fit each light curve to an SN Ia model and use the “duck test” philosophy: if it looks like a duck (i.e., an SN Ia) and quacks like a duck, then it is a duck. Selection cuts, mainly on the minimum $\chi^2$, were used to determine which SNe are type Ia, and there was no attempt to classify a subtype for . This strategy was used by Gonzalez, Portsmouth-$\chi^2$ and SNANA cuts.
The second strategy compares each light curve against both SN Ia and templates, and uses the Bayesian probabilities to determine the most likely SN type. Poz2007 used the simplest Bayesian implementation, with a single Ia and template. Belov & Glazov and Sako used SN Ia templates that depend on stretch and extinction, and they also used several templates. Sako included 8 templates from the , although there was no coordination between his template development for classification and the development of templates for the . Rodney used a variant of this technique by accounting for the fact that templates from observed SNe do not form a complete set. MGU+DU used another variation by using slopes (mag/day) at four different epochs and comparing with slopes expected for type Ia and SNe.
The third strategy used confirmed SNe Ia to parametrize a Hubble diagram, and then identified SN Ia as those SNe that lie near the expected Hubble diagram. Portsmouth-Hub used a high-order polynomial to define the Hubble diagram while JEDI-Hub used the kernel density estimation technique.
In the last strategy (InCA and JEDI-KDE) each light curve was fit with a parametric function such as a spline, and the fitted parameters were used for statistical inferences. Light-curve fitting parameters such as stretch and color were not explicitly used.
Evaluating the {#sec:eval}
===============
Ideally we would like to assign a single number, or figure of merit (FoM), for each submission. We begin the discussion by considering a measurement of the SN Ia rate based on . After selection requirements have been applied, let $\NIatrue$ be the number of correctly typed SNe Ia, and $\NIafalse$ be the number of non-Ia that are incorrectly typed as an SN Ia. A simple classification FoM is the square of the S/N divided by the total number of SNe Ia ($\NIaTOT$) before selection cuts, $$\begin{aligned}
\FoMIa & \equiv &
\frac{1}{\NIaTOT} \times
\frac{({\NIatrue})^2}{\NIatrue + \wIafalse\NIafalse}
\nonumber \\
& & \nonumber \\
& = & \frac{\NIatrue}{\NIaTOT} \times
\frac{\NIatrue}{\NIatrue + \wIafalse\NIafalse} ~,
\nonumber \\
& = & \effIa \times \PurityIa~,
\label{eq:FoM}\end{aligned}$$ where $\effIa = \NIatrue/\NIaTOT$ is the SN Ia efficiency that includes both selection and classification requirements, $\PurityIa$ is the pseudopurity, and $\wIafalse$ is the false-tag weight (penalty factor). Since $\NIaTOT$ is a constant that is independent of the analysis, we have divided out this term so that $0 \le \FoMIa \le 1$, with $\FoMIa = 1$ corresponding to the theoretically optimal analysis.
When $\wIafalse=1$, the denominator in $\PurityIa$ comes from the Poisson noise term in the S/N, and $\PurityIa$ can be interpreted as the traditional purity factor defined as the fraction of classified Ia that really are SNe Ia. In the ideal case where the mean of $\NIafalse$ is perfectly determined,[^8] the naive Poisson is the only contribution to the noise term and therefore $\wIafalse = 1$. In practice, however, in determining the false-tag rate lead to $\wIafalse > 1$. For example, suppose that the estimate of $\NIafalse$ is scaled from a confirmed subset containing a fraction ($\effspec$) of the total number of SNe; in this case, the Poisson noise term is defined by setting $\wIafalse = 1 + \effspec^{-1}$, and $\wIafalse \gg 1$ if the subset is small.
When using SN Ia for cosmological applications, it may be possible to reduce $\wIafalse$ using other methods to determine $\NIafalse$, such as fitting the tails in the distance-modulus residuals. A proper determination of $\wIafalse$ is beyond the scope of this classification challenge, and we have therefore arbitrarily set $\wIafalse = \wIafalseVALUE$. While this value is well below $1/\effspec \sim 15$ based on using the confirmed subset, $\wIafalse$ is notably larger than unity and therefore penalizes incorrect classifications more than rejected SNe.
Results {#sec:results}
=======
Here, we give a relatively brief overview of the main results and comparisons. Ideally, we would fully understand the strengths and weaknesses for each entry, but this level of detail is deferred to future analyses from individual participants. Also, since the results presented here are simply a starting point for these studies, a detailed postchallenge analysis could soon become obsolete as the algorithms are improved. Finally, the most important goal here is not to identify the best method now, but to motivate improvements and then identify the best method appropriate to each SN survey.
We begin by showing the SNe that were misidentified as SNe Ia. For each challenge entry we have computed the fraction of false SN Ia tags corresponding to each SED template: the sum of these fractions equals one for each entry. Fig. \[fig:non1afrac\] shows the false-tag fractions averaged over all entries, and they are sorted from largest to smallest. For both challenges (with and without host-galaxy ), the most frequently misidentified is based on SN 2006ep ( index $=8$; see Table \[tb:nonIa\_subtypes\]), a confirmed SN Ic with a rest-frame $g$-band peak magnitude of $-19.1$ mag. While the generated fraction for each SED template is 1.7% of the total, simulated SNe based on 2006ep account for $\sim 20$% of all misidentified SN Ia. The second most frequently misidentified template, accounting for 8% of all falsely tagged SN Ia, is based on SN 2006ns ( index $=27$), a confirmed type II-P SN with a $g$-band peak magnitude of $-18.3$ mag.
The results from the SN Ia evaluations (§\[sec:eval\]) are shown in Figures \[fig:hostz\_resultsIa\]-\[fig:nohostz\_resultsIa\], corresponding to the challenges with and without host-galaxy information. As a function of the true (generated) redshift, we have plotted the figure-of-merit quantity $\FoMIa$ (Eq. \[eq:FoM\]), efficiency ($\effIa$), pseudopurity ($\PurityIa$), and true purity. For each variable, the redshift dependence is shown separately for the confirmed subset (solid) and the unconfirmed SNe (dashed). The label on each panel indicates the name of the participant or group. The first panel labeled “All Ia tag” is an arbitrary reference in which every SN has been tagged as an SN Ia, thereby ensuring 100% efficiency. The corresponding results for type II classifications are shown in Fig. \[fig:resultsII\].
For the SN Ia classifications, the most notable trend in all of the entries is that the figure of merit ($\FoMIa$) is significantly worse for the unconfirmed sample than for the confirmed subset. Depending on the redshift, the confirmed-unconfirmed differences vary by tens of percent to nearly an order of magnitude. Several methods show improving $\FoMIa$ with redshift. We see this trend for the confirmed “All Ia” entry because at high redshift anything bright enough to obtain a spectrum is likely to be an SN Ia.
For the unconfirmed SN subset, the largest $\FoMIa$ value in any redshift bin is about 0.6, but these entries show at least a factor-of-2 variation in $\FoMIa$ as a function of redshift. The most stable figure of merit versus redshift (for unconfirmed SNe) has $\FoMIa = 0.3$ – 0.45 at all redshifts. The largest variation is $0.1 < \FoMIa < 0.6$.
In spite of the caveats about trying to determine the best method in this first , here we carefully examine the $\FoMIa$ for the unconfirmed sample in the (Fig. \[fig:hostz\_resultsIa\]). The entry with the highest average figure of merit (Sako) has an average SN Ia efficiency of and an average SN Ia purity (i.e., $\wIafalse=1$) of . However, comparing the best figure of merit (vs. redshift) for each strategy shows that three strategies yield similar results: selection cuts, Bayesian probabilities and statistical inference. The remaining Hubble-diagram strategy is somewhat worse at low and high redshifts. Among the entries for a given strategy there is a large variation in the figure of merit, suggesting that the optimum has not been achieved. For participants who applied the same method to both the and the , the average $\FoMIa$ was smaller for the by as little as 6% (Sako and JEDI-KDE) and by as much as a factor of 2.
The residuals are shown in Fig. \[fig:photoz\_resid\] for those entries that include estimates. Here we show residuals only for true SNe Ia that have been correctly typed as an SN Ia. When the host-galaxy is available, the supernova light curve improves the precision for redshifts up to about 0.4. For the , the bias and scatter of the residuals is significantly larger than for the .
After evaluating the classification results and algorithms, two notable problems were identified in the implementations. First, the confirmed subset was generally treated as a random subset, which it clearly is not (§\[subsec:sim\_spec\_subset\]). The magnitude-limited selection of targets resulted in the selection of brighter objects in the training subset. In principle, the brighter objects in the training subset should be re-simulated at higher redshifts so that classification algorithms can be trained on more distant (dimmer) objects for which spectra cannot be obtained.
The second general problem is that several entries did not use all available information from the light curves (most notably, ignoring colors), or effectively added noise to the information. The latter was mainly an artifact from a very poor determination of the epoch of maximum brightness. Specific details of these problems are given in Appendix \[app:methods\].
Updated Simulations {#sec:sim_update}
===================
While we have no plans for another competition-style challenge, we have released updated simulated samples as a public resource for the development of photometric SN classification and estimators.[^9] For these updated samples we have fixed the known bugs (§\[subsec:sim\_bugs\]), made some improvements, provided additional samples corresponding to the LSST [@LSSTSB09] and surveys, and included the answer keys giving the generated type and other parameters for each SN. The answer keys will allow developers to study different confirmed training subsets, and to evaluate their own analysis.
The updated simulations have two main improvements related to the generation of SNe Ia. The first improvement is a more realistic modeling of color variations based on recent results from @Guy2010. The newly measured variation is about 0.05 mag (Gaussian sigma) in the ultraviolet wavelength region and $\sim 0.02$ mag in the other wavelength regions. These variations are significantly smaller than what was used in the , where an independent variation per passband was drawn randomly from a Gaussian distribution with $\sigma_m =0.09$ mag. To obtain a reasonable Hubble scatter in the updated simulations, a 0.12 mag random Gaussian smearing is added coherently to all epochs and passbands. The second improvement is to use more realistic distributions of color and stretch ($x_1$ parameter) for the SNe Ia generated with the model. These distributions include more realistic tails corresponding to dimmer SNe, resulting in fewer -generated SNe Ia satisfying the loose selection criteria. The sample sizes generated from the and models are thus very similar, in contrast to the larger sample in the (§\[subsec:sim\_bugs\]).
Conclusion {#sec:end}
==========
We have presented results from the SN classification challenge that finished 2010 June 1. Among the four basic strategies that were used in the (§\[sec:take\_challenge\]), three strategies show comparable results for the entries with the highest figure of merit. Therefore no particular strategy was notably superior. For all of the entries, the classification performance was significantly better for the training subset than for the unconfirmed sample. The degraded performance on the unconfirmed sample was in part due to participants not accounting for the bias in the training sample.
There is a large variation in the figure of merit and therefore we urge caution in using these evaluations to determine the best method. The quality of each implementation varies significantly between participants (Appendix \[app:methods\]) and therefore some improvements are needed before drawing more clear conclusions. While this article signifies the end of the , we consider this effort to be the start of a new era for developing classification methods with significantly improved simulation tools. The results from this may serve as a reference to assess future progress from using improved algorithms and improved simulations. As described in §\[sec:sim\_update\], these updated simulations, along with the answer keys giving the true type for each SN, are publicly available.
While the optimal classification algorithm can in principle be optimized after a survey has completed, it is advantageous to define the necessary training sample before a survey has started. In particular, is a magnitude-limited training sample adequate (i.e., as used in this ), or is a less biased training sample needed? The latter sample is clearly more desirable for training classification algorithms, but this strategy results in fewer confirmed SNe Ia. As described in §\[sec:sim\_update\], this issue can be investigated more thoroughly by defining arbitrary training subsets for the publicly available simulated samples.
To optimize the use of a magnitude-limited sample, we suggest another strategy that was not tried by any of the participants. In principle the confirmed sample can be used to simulate SNe at higher redshifts to obtain an extended training sample for the classification algorithms. In contrast to an ideal unbiased sample however, this simulation strategy does not account for changes in the relative rates with redshift.
The figure of merit used in this challenge (§\[sec:eval\]) allows for a quantitative comparison between methods, but does not quantify the impact of photometric classification on the inference of cosmological parameters. Therefore, an important next step in using these simulations is to carry out a full analysis that includes the determination of cosmological parameters from a Hubble diagram.
Classification Methods from Participants {#app:methods}
========================================
[^1]:
[^2]:
[^3]:
[^4]: The CTIO history of observing conditions is available in the public package (previous footnote).
[^5]: Although two of us (RK & SK) are members of the DES, we did not include other DES colleagues in any discussions about preparing the challenge, and we made our best efforts to prevent our DES collaborators from obtaining additional information beyond that contained in the release note.
[^6]: The default rest-frame wavelength ranges for 2k2 and are 3200-9500 [Å]{} and 2900-7000 [Å]{}, respectively.
[^7]:
[^8]: The mean $\NIafalse$ value is the average over many independent measurements.
[^9]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In order to develop certain fractional probabilistic analogues of Taylor’s theorem and mean value theorem, we introduce the $n$th-order fractional equilibrium distribution in terms of the Weyl fractional integral and investigate its main properties. Specifically, we show a characterization result by which the $n$th-order fractional equilibrium distribution is identical to the starting distribution if and only if it is exponential. The $n$th-order fractional equilibrium density is then used to prove a fractional probabilistic Taylor’s theorem based on derivatives of Riemann-Liouville type. A fractional analogue of the probabilistic mean value theorem is thus developed for pairs of nonnegative random variables ordered according to the survival bounded stochastic order. We also provide some related results, both involving the normalized moments and a fractional extension of the variance, and a formula of interest to actuarial science. In conclusion we discuss the probabilistic Taylor’s theorem based on fractional Caputo derivatives.
[*MSC 2010*]{}: Primary 60E99; Secondary 26A33, 26A24.
[*Key Words and Phrases*]{}: characterization of exponential distribution, fractional calculus, fractional equilibrium distribution, generalized Taylor’s formula, mean value theorem, survival bounded order.
author:
- '[Antonio Di Crescenzo]{}[^1]'
- '[Alessandra Meoli]{}[^2]'
date: |
**First published in [*Fractional Calculus and Applied Analysis*]{},\
Vol. 19, n. 4, p. 921–939 © 2016 by De Gruyter**
title: |
**ON THE FRACTIONAL PROBABILISTIC\
TAYLOR’S AND MEAN VALUE THEOREMS**
---
Introduction and background {#sec:1}
===========================
Taylor’s theorem is the most important result in differential calculus. In fact, given the derivatives of a function at a single point, it provides insight into the behavior of the function at nearby points. Motivated by the large numbers of its applications, researchers have shown a heightened interest in the extensions of this theorem. For instance, Massey and Whitt [@MasseyWhitt1993] derived probabilistic generalizations of the fundamental theorem of calculus and Taylor’s theorem by making the argument interval random and expressing the remainder terms by means of iterates of the equilibrium residual-lifetime distribution from the theory of stochastic point processes. Lin [@Lin1994] modified Massey and Whitt’s probabilistic generalization of Taylor’s theorem and gave a natural proof by using an explicit form for the density function of high-order equilibrium distribution. In a similar spirit to these probabilistic extensions of Taylor’s theorem, Di Crescenzo [@DiCrescenzo1999] gave a probabilistic analogue of the mean value theorem. The previous results have direct applications to queueing and reliability theory. However, probabilistic generalizations are not the only ones. Indeed, fractional Taylor series have been introduced with the idea of approximating non-integer power law functions. Here we recall the most interesting ones. Trujillo et al. [@Trujillo1999] established a Riemann-Liouville generalized Taylor’s formula, in which the coefficients are expressed in terms of the Riemann-Liouville fractional derivative. On the other hand, Odibat et al. [@Odibat2007] expressed the coefficients of a generalized Taylor’s formula in terms of the Caputo fractional derivatives. In the aforementioned papers, an application of the generalized Taylor’s formula to the resolution of fractional differential equations is also shown. Great emphasis has been placed on fractional Lagrange and Cauchy type mean value theorems too (cf. [@Guo2012], [@Odibat2007], [@Pecaric2005], [@Trujillo1999], for example). Inspired by such improvements, in our present investigation we propose to unify these two approaches by presenting a fractional probabilistic Taylor’s theorem and a fractional probabilistic mean value theorem.
We start by briefly recalling some basic definitions and properties of fractional integrals and derivatives of Riemann-Liouville and Weyl type as well as some notions on a generalized Taylor’s formula that are pertinent to the developments in this paper. For more details on fractional calculus we refer the reader to [@Gorenflo2014].
Background on fractional integrals
----------------------------------
Let $ \Omega = \left [ a,b \right ]$ $\left ( -\infty<a<b\leq+\infty \right ) $ be an interval on the real axis $ \mathbb{R} $. The *progressive* or *right-handed fractional integral* $ I_{a^{+}}^{\alpha }f $ of order $ \alpha $ is defined by $$\label{RL:a+}
I_{a+}^{\alpha }f \left ( x \right ):=\frac{1}{\Gamma \left ( \alpha \right )}\int_{a}^{x}\left ( x-t \right )^{\alpha -1}f\left ( t \right )dt,\qquad a<x<b,\quad \alpha>0.$$ Here $ \Gamma \left ( \alpha \right ) $ is the Gamma function and the function $ f(x) $ is assumed to be well-behaved, in order to ensure the finiteness of the values $ I_{a+}^{\alpha }f \left ( x \right ) $ for $ a<x<b $. A major property of the fractional integral (\[RL:a+\]) is the additive index law *(semi-group property)*, according to which $$I_{a+}^{\alpha }I_{a+}^{\beta }= I_{a+}^{\alpha+\beta },\quad \alpha ,\beta \geq 0,$$ where, for complementation, $ I_{a+}^{0}
:=\mathds{I} $ (identity operator).
The *progressive fractional derivative of order* $ \alpha $ is defined by $$\label{RLderivative}
D_{a+}^{\alpha }f(x):=D^{m}I_{a+}^{m-\alpha }f(x),\qquad \text{$ a<x<b,\; m-1<\alpha \leq m $},$$ where $ m $ is a positive integer, and $ D_{a+}^{0}
:=\mathds{I} $ (identity operator). Furthermore, the *sequential fractional derivative* is denoted by $$D_{a+}^{n\alpha }=\underbrace{D_{a+}^{\alpha }\dots D_{a+}^{\alpha }}_{n\mathrm{\,times}},$$ where $ n\in \mathbb{N}\equiv \{0,1,\dots\} $.
A fractional integral over an unbounded interval can also be defined. Specifically, if the function $ f(x) $ is locally integrable in $ -\infty\leq a<x<+\infty $, and behaves well enough for $x\rightarrow +\infty $, the *Weyl fractional integral of order $ \alpha $* is defined as $$\label{Weyl-}
I_{-}^{\alpha }f(x):=\frac{1}{\Gamma \left ( \alpha \right )}\int_{x}^{+\infty}\left ( t-x \right )^{\alpha -1}f\left ( t \right )dt,\qquad a<x<+\infty,\quad \alpha>0.$$ Also for the Weyl fractional integral the corresponding *semigroup property* holds: $$\label{SemiGroup}
I_{-}^{\alpha }I_{-}^{\beta }= I_{-}^{\alpha +\beta },\quad \alpha ,\beta \geq 0,$$ where, again for complementation $
I_{-}^{0}:=\mathds{I} $.
Background on a generalized Taylor’s formula
--------------------------------------------
Let $ \Omega $ be a real interval and $ \alpha \in [0,1) $. Let $ F\left( \Omega\right) $ denote the space of Lebesgue measurable functions with domain in $ \Omega $ and suppose that $ x_{0}\in \Omega $. Then a function $ f $ is called $ \alpha$-*continuous in* $ x_{0} $ if there exists $ \lambda\in [0,1-\alpha) $ for which the function $ h $ given by $$h\left(x\right)=\left | x-x_{0} \right |^{\lambda }f\left ( x \right )$$ is continuous in $ x_{0} $. Moreover, $ f $ is called $ 1$-*continuous in* $ x_{0} $ if it is continuous in $ x_{0} $, and $ \alpha$-*continuous on* $ \Omega $ if it is $ \alpha$-continuous in $ x $ for every $ x\in\Omega $. We denote, for convenience, the class of $ \alpha$-continuous functions on $ \Omega $ by $ C_{\alpha}\left(\Omega\right) $, so that $ C_{1}\left(\Omega\right)=C\left(\Omega\right) $.
For $ a\in\Omega $, a function $ f $ is called $ a$-*singular of order $ \alpha $* if $$\lim _{x\rightarrow a}\frac{f(x)}{\left | x-a \right |^{\alpha -1}}= k<\infty\quad\mathrm{and}\quad k\neq 0.$$ Let $ \alpha\in\mathbb{R^{+}},\,a\in\Omega$ and let $ E $ be an interval, $ E \subset\Omega $, such that $ a\leq x $ for every $ x\in E$. Then we write $$_{a}\mathbf{I}_{\alpha }\left ( E \right )= \left \{ f\in F \left ( \Omega \right ):I_{a+}^{\alpha }f\left ( x \right )\;\textrm{exists and it is finite}\;\forall x\in E\right \}.$$ Recently, Trujillo et al. (cf. Theorem 4.1 of [@Trujillo1999]) proved the following result.
\[ThTrujillo\] Set $ \alpha \in \left [ 0,1 \right ] $ and $ n\in \mathbb{N} $. Let $ g $ be a continuous function in $ \left ( a,b \right ] $ satisfying the following conditions:
1. $ \forall\, j= 1,\dots,n, D_{a}^{j\alpha }g\in C\left ( \left ( a,b \right ] \right ) $ and $ D_{a}^{j\alpha }g\in \,_{a}\mathbf{I}_{\alpha }\left ( \left [ a,b \right ] \right ) $;
2. $ D_{a}^{(n+1)\alpha }g $ is continuous on $ \left [ a,b \right ] $;
3. If $ \alpha < 1/2 $ then, for each $ j\in \mathbb{N}, 1\leq j \leq n $, such that $ (j+1)\alpha<1 $, $ D_{a}^{\left ( j+1 \right )\alpha }g\left ( x \right ) $ is $ \gamma$-continuous in $ x=a $ for some $ \gamma $, $ 1-\left ( j+1 \right )\alpha\leq \gamma \leq 1$, or $ a$-singular of order $ \alpha $.
Then, $ \forall x\,\in(a,b] $, $$g(x)=\sum_{j=0}^{n}\frac{c_{j}(x-a)^{\left ( j+1 \right )\alpha -1}}{\Gamma (\left ( j+1 \right )\alpha ) }+R_{n}(x,a),$$ with $$R_{n}(x,a)=\frac{D_{a}^{\left ( n+1 \right )\alpha }g\left ( \xi \right )}{\Gamma \left ( \left ( n+1 \right )\alpha +1 \right )}\left( x-a\right)^{\left ( n+1 \right )\alpha },\qquad a\leq \xi\leq x,$$ and $$c_{j}= \Gamma \left ( \alpha \right )\left [ \left( x-a\right) ^{1-\alpha } D_{a}^{j\alpha }g(x)\right ]\left ( a^{+} \right )=I_{a}^{1-\alpha }D_{a}^{j\alpha }g\left ( a^{+} \right )$$ for each $ j\in\mathbb{N},\,0\leq j\leq n $.
Plan of the paper
-----------------
The paper is organized as follows. In Section \[sec:2\], after recalling the notion of equilibrium distribution, we define a fractional extension of the high-order equilibrium distribution. Then, we give an equivalent version by exploiting the semigroup property of the Weyl fractional integral and derive the explicit expression of the related density function. Moreover, by means of the Mellin transform we underline the role played by the fractional equilibrium density in characterizing the exponential distribution. In Section \[sec:3\] we prove a fractional probabilistic Taylor’s theorem by using the expression of the $ n $th-order fractional equilibrium density. Section \[sec:4\] is devoted to the analysis of the fractional analogue of the probabilistic mean value theorem. We first consider pairs of non-negative random variables ordered in a suitable way so as to construct a new random variable, say $ Z_{\alpha} $, which extends the fractional equilibrium operator. The fractional probabilistic mean value theorem indeed is given in terms of $ Z_{\alpha} $. We also discuss some related results, including a formula of interest to actuarial science. We stress that all those results are involving the derivatives of Riemann-Liouville type. However, in some instances they can be restated also under different setting. Indeed, in Section \[sec:5\] we conclude the paper by exploiting a fractional probabilistic Taylor’s theorem in the Caputo sense.
Fractional equilibrium distribution {#sec:2}
===================================
Let $ X $ be a nonnegative random variable with distribution $ F\left( x\right) = \mathbb{P}\left( X\leq x\right) $ for $ x\geq 0 $ and with mean $ \mathbb{E}\left[ X\right]\in (0,+\infty) $. Let $ X_{e} $ be a nonnegative random variable with distribution $$F_{1}\left( x\right) =\mathbb{P}\left ( X_{e}\leq x\right )= \frac{\int_{0}^{x}\overline{F}\left ( y \right )dy}{\mathbb{E}\left[ X\right] },\qquad x\geq 0,$$ or, equivalently, with complementary distribution function $$\overline{F}_{1}\left( x\right)=\mathbb{P}\left ( X_{e}> x\right )= \frac{\int_{x}^{+\infty}\overline{F}\left ( y \right )dy}{\mathbb{E}\left[ X\right] },\qquad x\geq 0,$$ where $ \overline{F}=1-F $. The distribution $ F_{1} $ is called *equilibrium distribution with respect to F*. Further, suppose $ \mathbb{E}\left[ X^{2}\right]<+\infty $. Then the equilibrium distribution with respect to $ F_{1} $ is well-defined and it reads $$F_{2}\left ( x \right )= \mathbb{P}\left ( X_{e}^{\left ( 2 \right )} \leq x\right )=\frac{\int_{0}^{x}\overline{F}_{1}\left ( y \right )dy}{\mathbb{E}\left [ X_{e} \right ]},\qquad x\geq 0.$$ $ F_{2} $ is known as the second order equilibrium distribution with respect to $ F $. Continuing $ n-2 $ more iterates of this transformation, it is possible to obtain the $ n $th-order equilibrium distribution with respect to $ F $, denoted by $ F_{n} $, provided the required moments of $ X $ are finite.
Hereafter we introduce a fractional version of the $ n $th-order equilibrium distribution. Let $ \alpha\in\mathbb{R}^{+} $ and let $ X $ be a nonnegative random variable with distribution $ F\left( t\right) = \mathbb{P}\left( X\leq t\right) $ for $ t\geq 0 $ and with moment $ \mathbb{E}\left[ X^{\alpha}\right]\in (0,+\infty) $. Then we define a random variable $ X_{\alpha}^{(1)} $ whose complementary distribution function is $$\label{FractionalEquilibrium}
\overline{F}_{\,1}^{\,\alpha}\left ( t \right )= \mathbb{P}\left ( X_{\alpha }^{\left ( 1 \right )}>t \right )= \frac{\Gamma \left ( \alpha +1 \right )}{\mathbb{E}\left [ X^{\alpha } \right ]}I_{-}^{\alpha}\overline{F}\left ( t \right ),\qquad t\geq 0,$$ where $ I_{-}^{\alpha} $ is the Weyl fractional integral (\[Weyl-\]) and $ \overline{F}=1-F $. We call the distribution of $ X_{\alpha}^{(1)} $ *fractional equilibrium distribution with respect to $ F $*.
Recalling (\[Weyl-\]), from (\[FractionalEquilibrium\]) we obtain the following suitable probabilistic interpretation of the distribution function of $ X_{\alpha }^{\left ( 1 \right )} $ in terms of $X$. In fact, $$\mathbb{P}\left ( X_{\alpha }^{\left ( 1 \right )}\leq t \right )
=\frac{\alpha}{\mathbb{E}\left[ X^{\alpha}\right]}\int_{0}^{+\infty}y^{\alpha -1}\mathbb{P}\left ( y<X\leq y+t \right )dy.$$
Further, suppose $ \mathbb{E}\left[ X^{2\alpha}\right]<+\infty $. Then the *second-order fractional equilibrium distribution with respect to $ F $* is well defined and its complementary distribution function reads $$\overline{F}_{\,2}^{\,\alpha}\left ( t \right )= \mathbb{P}\left ( X_{\alpha }^{\left ( 2 \right )}>t \right )=\frac{\Gamma \left ( 2\alpha +1 \right )}{\Gamma \left ( \alpha +1 \right )}\frac{\mathbb{E}\left [ X^{\alpha } \right ]}{\mathbb{E}\left [ X^{2\alpha } \right ]}I_{-}^{\alpha }\overline{F}_{\,1}^{\,\alpha}\left ( t \right ),\qquad t\geq 0.$$ Generally, we can recursively define the *$ n $th-order fractional complementary equilibrium distribution with respect to $ F $* by $$\overline{F}_{\,n}^{\,\alpha }\left ( t \right )=\mathbb{P}\left ( X_{\, \alpha }^{\, \left ( n \right )}>t \right ) = \frac{\Gamma \left ( n\alpha +1 \right )}{\Gamma \left (\left ( n-1 \right )\alpha +1\right )}\frac{\mathbb{E}\left [ X^{\left ( n-1 \right )\alpha } \right ]}{\mathbb{E}\left [X^{n\alpha } \right ]}\, I_{-}^{\alpha }\overline{F}_{\,n-1}^{\,\alpha }\left( t\right),\qquad t\geq 0,$$ provided that all the moments $ \mathbb{E}[X^{n\alpha}]$, for $ n\in\mathbb{N}$, $n \geq 1$, are finite.
Interestingly enough, $ \overline{F}_{\,n}^{\,\alpha } $ can be alternatively expressed in terms of $ \overline{F} $. Indeed, the following proposition holds.
Let $ \alpha\in\mathbb{R}^{+} $ and let $ X $ be a non-negative random variable with distribution $ F\left( t\right) $ for $ t\geq 0 $. Moreover, suppose that $ \mathbb{E}\left[ X^{n\alpha}\right]\in (0,+\infty)$, with $ n\in\mathbb{N} $, $n \geq 1$. Then the *$ n $th-order fractional complementary equilibrium distribution with respect to $ F $* reads $$\label{AlternativeEquilibrium}
\overline{F}_{\,n}^{\,\alpha }\left ( t \right )= \frac{\Gamma \left ( n\alpha +1 \right )}{\mathbb{E}\left [ X^{n\alpha } \right ]}I_{-}^{n\alpha }\overline{F}\left ( t \right ),\qquad t\geq 0.$$
The proof is by induction on $ n $. In fact, when $ n=1 $ formula (\[AlternativeEquilibrium\]) is true due to definition (\[FractionalEquilibrium\]). Now let us assume that Eq. (\[AlternativeEquilibrium\]) holds for some $ n $; then, for $ t\geq 0 $, $$\begin{aligned}
\overline{F}_{\,n+1}^{\,\alpha }\left ( t \right )&= \frac{\Gamma \left ( (n+1)\alpha +1 \right )}{\Gamma \left (n\alpha +1\right )}\frac{\mathbb{E}\left [ X^{n\alpha } \right ]}{\mathbb{E}\left [X^{(n+1)\alpha } \right ]}\, I_{-}^{\alpha }\overline{F}_{\,n}^{\,\alpha }\left( t\right)\nonumber \\
&=\frac{\Gamma \left ( (n+1)\alpha +1 \right )}{\Gamma \left (n\alpha +1\right )}\frac{\mathbb{E}\left [ X^{n\alpha } \right ]}{\mathbb{E}\left [X^{(n+1)\alpha } \right ]}I_{-}^{\alpha}\frac{\Gamma \left ( n\alpha +1 \right )}{\mathbb{E}\left [ X^{n\alpha } \right ]}I_{-}^{n\alpha }\overline{F}\left ( t \right )\nonumber \\
&=\frac{\Gamma \left( \left ( n+1 \right )\alpha +1 \right )}{\mathbb{E}\left [X^{(n+1)\alpha } \right ]}I_{-}^{\left ( n+1 \right )\alpha }\overline{F}\left ( t \right ).\end{aligned}$$ The last equality is valid due to the linearity of the integral and to semigroup property (\[SemiGroup\]). So the validity of Eq. (\[AlternativeEquilibrium\]) for $ n $ implies its validity for $ n+1 $. Therefore it is true for all $ n\in\mathbb{N} $, $n\geq 1$. [$\Box$\
]{}
In order to obtain the explicit expression of the density function of the $ n $th-order fractional equilibrium distribution, we premise the following lemma, which is a generalization of Proposition 4 of [@Cheng1999].
\[GeneralizedChengPai\] Let $ X $ be a nonnegative random variable whose moment of order $ n\alpha $ is finite for $ \alpha\in\mathbb{R}^{+} $ and $ n\in\mathbb{N}$, $n\geq 1$. Then $$\lim_{x\rightarrow +\infty}\left ( x-t \right )^{n\alpha}\overline{F}\left ( x \right )=0,\quad \forall t\geq 0.$$
Because the moment of order $ n\alpha $ of $ X $ is finite, we have $$\label{FinitenessOfMoment}
\lim_{x\rightarrow +\infty}\int_{x}^{+\infty}y^{n\alpha}dF\left ( y \right )= 0.$$ Hence, due to a generalized Markov’s inequality, $$\lim_{x\rightarrow +\infty}\left ( x-t \right )^{n\alpha}\overline{F}\left ( x \right )\leq \lim_{x\rightarrow +\infty}x^{n\alpha}\overline{F}\left ( x \right )\leq \lim_{x\rightarrow+\infty}\int_{x}^{+\infty}y^{n\alpha}dF\left ( y \right )= 0,$$ this completing the proof. [$\Box$\
]{}
Here and throughout the paper, we denote, for convenience, $ \left ( x \right )^{\alpha-1}_{+}= (x)^{\alpha-1}\mathds{1}_{\{x>0\}} $. The following result concerns the probability density function associated with $ F_{\,n}^{\,\alpha}\left ( t \right ) $.
\[PropositionFractionalDensity\] Let $ X $ be a non-negative random variable with distribution function $ F $ and let $ \mathbb{E}[X^{n\alpha}]<+\infty $ for some integer $ n\geq 1 $ and $ \alpha\in\mathbb{R}^{+} $. Then the density function of $ X_{\, \alpha }^{\, \left ( n \right )} $ is $$\label{FractionalDensity}
f_{n}^{\alpha }\left ( t \right )= \frac{n\alpha \mathbb{E}\left [\left( X-t\right) _{+}^{ n\alpha -1}\right ]}{\mathbb{E}\left [ X^{n\alpha } \right ]},\qquad t\geq 0.$$
By virtue of (\[AlternativeEquilibrium\]) and (\[Weyl-\]) we have: $$\begin{aligned}
\label{Stop-loss}
\overline{F}_{\,n}^{\,\alpha }\left ( t \right )&= \frac{\Gamma \left ( n\alpha +1 \right )}{\mathbb{E}\left [ X^{n\alpha } \right ]}I_{-}^{n\alpha }\overline{F}\left ( t \right )\nonumber\\
&=\frac{n\alpha \: \Gamma \left ( n\alpha \right )}{\mathbb{E}\left [ X^{n\alpha } \right ]}\frac{1}{\Gamma \left ( n\alpha \right )}\int_{t}^{+\infty}\left ( x-t \right )^{n\alpha -1}\overline{F}\left ( x \right )dx.\nonumber
$$ Due to integration by parts and making use of Lemma \[GeneralizedChengPai\], we have: $$\begin{aligned}
\overline{F}_{\,n}^{\,\alpha }\left ( t \right )&=-\frac{1}{\mathbb{E}\left [ X^{n\alpha } \right ]}\int_{t}^{+\infty}\left ( x-t \right )^{n\alpha }d\overline{F}\left ( x \right )\nonumber\\
&=\frac{1}{\mathbb{E}\left [ X^{n\alpha } \right ]}\int_{t}^{+\infty}\left ( x-t \right )^{n\alpha }dF\left ( x \right )\nonumber\\
&=\Bigg(\frac{\mathbb{E}\left [ \left ( X-t \right )_{+}^{n\alpha }\right ]}{\mathbb{E}\left [ X^{n\alpha } \right ]}\Bigg)\\
&= \frac{n\alpha }{\mathbb{E}\left [ X^{n\alpha } \right ]}\int_{t}^{+\infty}dF\left ( x \right )\int_{t}^{x}\left ( x-y \right )^{n\alpha -1}dy\nonumber\\
&=\frac{n\alpha }{\mathbb{E}\left [ X^{n\alpha } \right ]}\int_{t}^{+\infty}dy\int_{y}^{+\infty}\left ( x-y \right )^{n\alpha -1}dF\left ( x \right )\nonumber\\
&=\frac{n\alpha }{\mathbb{E}\left [ X^{n\alpha } \right ]}\int_{t}^{+\infty}\mathbb{E}\left [ \left ( X-y \right )_{+}^{n\alpha -1} \right ]dy,\nonumber\end{aligned}$$ this giving the density function (\[FractionalDensity\]). [$\Box$\
]{}
We observe that in Proposition \[PropositionFractionalDensity\] the random variable $ X $ is not necessarily absolutely continuous, unlike $ X_{\, \alpha }^{\, \left ( n \right )} $.
Formula (\[Stop-loss\]) is useful in showing that $ \overline{F}_{\,n}^{\,\alpha }\left ( t \right ) $ is a proper complementary distribution function. Indeed,
- $ \overline{F}_{\,n}^{\,\alpha }\left ( 0 \right )=\frac{\mathbb{E}\left [ \left ( X-t \right )_{+}^{n\alpha }\right ]}{\mathbb{E}\left [ X^{n\alpha } \right ]}\biggl|_{t=0}=1; $
- $ \overline{F}_{\,n}^{\,\alpha }\left ( t \right ) $ is decreasing and continuous in $ t\geq 0 $;
- $ \overline{F}_{\,n}^{\,\alpha }\left ( t \right )\rightarrow 0, $ when $ t\rightarrow +\infty $. In fact, due to (\[FinitenessOfMoment\]), we have $$\lim_{t\rightarrow +\infty}\int_{t}^{+\infty}\left ( x-t \right )^{n\alpha }dF\left ( x \right )\leq \lim_{t\rightarrow +\infty}\int_{t}^{+\infty}x^{n\alpha}dF\left ( x \right )= 0.$$
We now prove a characterization result concerning the fractional equilibrium density (\[FractionalDensity\]). In fact, if $ X $ is a non-negative random variable with probability density function $ f $, the $ n $th-order fractional equilibrium density associated with $ f $ coincides with $ f $ if and only if $ X $ is exponentially distributed. This extends the well-known result concerning case $ \alpha=1 $.
Let $ X $ be a non-negative random variable with probability density function $ f $. Then, for every $ n\in\mathbb{N} $ and $ \alpha\in\mathbb{R}^{+} $ $$f_{n}^{\alpha }\left ( t \right )=f(t),\;t\geq 0\qquad\Leftrightarrow\qquad X\sim \mathcal{E}(\lambda),$$ where $ f_{n}^{\alpha }\left ( t \right ) $ is the $ n $th-order fractional equilibrium density (\[FractionalDensity\]) and $ \mathcal{E}(\lambda) $ is the exponential distribution with parameter $ \lambda\in \mathbb{R}^{+}$.
First, let us assume that $ X $ is exponentially distributed with parameter $ \lambda $. Since $$\mathbb{E}[X^{n\alpha}]=\frac{\Gamma(n\alpha+1)}{\lambda^{n\alpha}}\quad\mathrm{and}\quad \mathbb{E}\left [ \left( X-t \right )_{+}^{ n\alpha -1}\right]=\lambda^{\-n\alpha}e^{-t\lambda}\Gamma(n\alpha),\quad t\geq 0,$$ by virtue of (\[FractionalDensity\]) the assertion “if” is trivially proved. Conversely, suppose that the density and the $ n $th-order fractional equilibrium density of $ X $ coincide for every $ n\in\mathbb{N} $ and $ \alpha\in\mathbb{R}^{+} $, that is $$f_{n}^{\alpha }\left ( t \right )=f(t),\qquad t\geq 0.$$ Due to (\[FractionalDensity\]), the last equality can be rewritten as $$n\alpha \int_{t}^{+\infty}\left ( x-t \right )^{n\alpha -1}f\left ( x \right )dx=f\left ( t \right )\int_{0}^{+\infty}x^{n\alpha }f\left ( x \right )dx,$$ and, on account of (\[Weyl-\]), as $$\Gamma(n\alpha+1)I_{-}^{n\alpha }f(t)=f\left ( t \right )\int_{0}^{+\infty}x^{n\alpha }f\left ( x \right )dx.$$ Taking the Mellin transform of both sides of this equation yields the functional equation $$\label{TransformedEquation}
\frac{f^{*}\left ( s+n\alpha \right )}{\Gamma \left ( s+n\alpha \right )}= \frac{f^{*}\left ( n\alpha +1 \right )}{\Gamma \left ( n\alpha +1 \right )}\frac{f^{*}\left ( s \right )}{\Gamma \left ( s \right )},\qquad \Re(s)>0,$$ where $$f^{*}(s)=\int_{0}^{\infty}x^{s-1}f(x)dx,$$ is the Mellin transform of a function $ f(x) $ (cf. (C.3.21) and (C.3.22) of [@Gorenflo2014]). By reducing Eq. (\[TransformedEquation\]) to a well-known Cauchy equation, we observe that its nontrivial measurable solution (cf. [@Hewitt1965] for instance) is $$f^{*}(s)=a^{c(s-1)}\Gamma(s)\qquad a>0,\,c\in\mathbb{R}.$$ By performing the Mellin inversion, we have $$f(t)=a^{-c}e^{-a^{-c}t},\qquad t\geq 0.$$ As a consequence, $ X\sim \mathcal{E}(\lambda) $, having set $ \lambda=a^{-c} $, and then the “only if” part of the theorem is proved. [$\Box$\
]{}
In the next proposition, we give the expression of the moments of a random variable following the $ n $th-order fractional equilibrium distribution (\[FractionalDensity\]). We omit the proof since it is straightforwardly derived after some cumbersome computations.
For $ \alpha\in\mathbb{R}^{+} $ and $ n\in\mathbb{N} $, if $ \mathbb{E}[X^{n\alpha}]<\infty $, then $$\label{moments}
\mathbb{E}\left [\left( X_{\, \alpha }^{\, \left ( n \right )}\right)^{r} \right ]= \frac{n\alpha\,B(n\alpha,r+1) }{\mathbb{E}[X^{n\alpha}]}\, \mathbb{E}[X^{n\alpha+r}],\qquad r\in\mathbb{R}^{+},$$ where $ B(x,y) $ is the Beta function.
Clearly, when $ \alpha=1 $ the moments (\[moments\]) identify with the expression for the iterated stationary-excess variables given in the Lemma of Massey and Whitt [@MasseyWhitt1993] and in Theorem 2.3 of Harkness and Shantaram [@Harkness1969].
Fractional probabilistic Taylor’s theorem {#sec:3}
=========================================
We now derive a probabilistic generalization of the Riemann-Liouville generalized Taylor’s formula shown in Theorem \[ThTrujillo\]. For convenience, let us denote $$I_{F}=\bigcup_{n=2}^{\infty}\left [ 0,F^{-1}\left ( 1-\frac{1}{n} \right ) \right ],
\label{eq:IF}$$ the smallest interval containing both $ {0} $ and the support of the distribution $ F $. Additionally, without loss of generality, we consider the expansion of a function $ g $ about $ t=0 $.
\[ThTaylor\] Let $ 0<\alpha\leq 1 $ and let $ X $ be a nonnegative random variable with distribution $ F $, with moment $ \mathbb{E}[X^{(n+1)\alpha}]<+\infty $ for some integer $ n\geq 0 $ and moments $ \mathbb{E}\left [ X^{\left ( j+1 \right )\alpha -1} \right ]<+\infty $ for all $ j\in\mathbb{N},0\leq j\leq n $. Suppose that $ g $ is a function defined on $I_{F}$ and satisfying the hypoteses (i),(ii) and (iii) of Theorem \[ThTrujillo\] in $I_{F}$. Assume further $ \mathbb{E}\left [ \left | D_{0}^{\left ( n+1 \right )\alpha }g\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right ) \right |\right ]<+\infty $. Then $ \mathbb{E}\left [ g\left ( X \right ) \right ]<+\infty $ and $$\begin{gathered}
\label{Thesis}
\mathbb{E}\left [ g\left ( X \right ) \right ]= \sum_{j=0}^{n}\frac{c_{j}}{\Gamma \left ( \left ( j+1 \right ) \alpha \right )}\mathbb{E}\left [ X^{\left ( j+1 \right )\alpha -1} \right ] \\
+\frac{\mathbb{E}[X^{(n+1)\alpha}]}{\Gamma \left ( \left ( n+1 \right )\alpha +1 \right )}\,\mathbb{E}\left [ D_{0}^{\left ( n+1 \right )\alpha }g\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right ) \right ],\end{gathered}$$ with $ c_{j}=\Gamma(\alpha)[x^{1-\alpha}D_{0}^{j\alpha}g(x)](0^{+}) $ for each $ j\in\mathbb{N}, 0\leq j\leq n $, where $ X_{\,\alpha }^{\left ( n+1 \right )} $ has density (\[FractionalDensity\]).
To begin with, we recall a Riemann-Liouville generalized Taylor’s formula with integral remainder term (cf. formula (4.1) of [@Trujillo1999]), that is, for $ x\in I_{F} $, $$\label{Taylor}
g(x)=\sum_{j=0}^{n}\frac{c_{j}x^{\left ( j+1 \right )\alpha -1}}{\Gamma (\left ( j+1 \right )\alpha ) }+ R_{n}(x),$$ where $$\begin{aligned}
\label{resto}
R_{n}(x)&=I_{0}^{\left ( n+1 \right )\alpha }D_{0}^{\left ( n+1 \right )\alpha }g\left ( x \right ) \nonumber\\
&=\frac{1}{\Gamma \left ( \left ( n+1 \right )\alpha \right )}\int_{0}^{x}\left ( x-t \right )^{\left ( n+1 \right )\alpha -1}D_{0}^{\left ( n+1 \right )\alpha }g\left ( t \right )dt.\end{aligned}$$ Since $ R_{n}(x) $ is continuous on $ I_{F} $, $ R_{n}(X) $ is a true random variable. Therefore, from (\[Taylor\]) we have $$\label{stella}
\mathbb{E}[g\left ( X \right )]= \sum_{j=0}^{n}\frac{c_{j}}{\Gamma \left ( \left ( j+1 \right ) \alpha \right )}\mathbb{E}[X^{\left ( j+1 \right )\alpha -1}]+\mathbb{E}[R_{n}\left ( X \right )],$$ where, from (\[resto\]), $$\mathbb{E}[R_{n}\left ( X \right )]=\frac{1}{\Gamma((n+1)\alpha)}\int_{0}^{+\infty}\int_{0}^{x}D_{0}^{\left ( n+1 \right )\alpha } g\left ( t \right )\left ( x-t \right )^{\left ( n+1 \right )\alpha -1}dt\, dF\left ( x \right ).$$ By making use of Fubini’s theorem, the equality above becomes $$\mathbb{E}[R_{n}\left ( X \right )]=\frac{1}{\Gamma((n+1)\alpha)}\int_{I_{F}}D_{0}^{\left ( n+1 \right )\alpha }g\left ( t \right )\mathbb{E}\left [ X-t \right ]_{+}^{\left ( n+1 \right )\alpha -1}dt,$$ and in turn, owing to (\[FractionalDensity\]), $$\begin{aligned}
\label{doppiastella}
\mathbb{E}[R_{n}\left ( X \right )]&=\frac{\mathbb{E}[X^{(n+1)\alpha}]}{\left ( n+1 \right )\alpha\, \Gamma \left ( \left ( n+1 \right )\alpha \right )}\int_{I_{F}}D_{0}^{\left ( n+1 \right )\alpha }g\left ( t \right )f_{n+1}^{\alpha }\left ( t \right )dt\nonumber\\
&=\frac{\mathbb{E}[X^{(n+1)\alpha}]}{ \Gamma \left ( \left ( n+1 \right )\alpha +1 \right )}\mathbb{E}\left [ D_{0}^{\left ( n+1 \right )\alpha }g\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right ) \right ].\end{aligned}$$ Finally, observing that the condition $ \mathbb{E}\left [ \left | D_{0}^{\left ( n+1 \right )\alpha }g\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right ) \right |\right ]<+\infty $ is equivalent to $ \int_{I_{F}} |D_{0}^{\left ( n+1 \right )\alpha }g\left ( t \right ) |\mathbb{E}\left [ X-t \right ]_{+}^{\left ( n+1 \right )\alpha -1}dt<+\infty $, and making use of (\[stella\]) and (\[doppiastella\]), the proof of (\[Thesis\]) is thus completed. [$\Box$\
]{}
Equation (\[Thesis\]) can be seen as a fractional version of the probabilistic generalization of Taylor’s theorem studied in [@Lin1994] and in [@MasseyWhitt1993].
\[Oss\] We observe that for $ n=0 $ formula (\[Thesis\]) becomes $$\label{RemarkTaylor}
\mathbb{E}\left [ g\left ( X \right ) \right ]=\frac{c_{0}}{\Gamma \left ( \alpha \right )}\mathbb{E}\left [ X^{\alpha -1} \right ]+\frac{\mathbb{E}\left [ X^{\alpha} \right ]}{\Gamma \left ( \alpha +1\right )}\mathbb{E}\left [ D_{0}^{\alpha }g\left ( X_{\alpha }^{\left ( 1 \right )} \right ) \right ],$$ with $ c_{0}=\Gamma \left ( \alpha \right )\left [ x^{1-\alpha} g(x)\right ](0^{+}) $, this being useful to prove Theorem \[Lagrange\] below.
In the recent years much attention has been paid to the study of the fractional moments of distributions. See, for instance, [@Mikoschetal2013] and references therein. Motivated by this, in the next corollary we consider the case $ g(x)=x^{\beta},\,\beta\in\mathbb{R} $. From Theorem \[ThTaylor\] we have the following result.
Let $ 0<\alpha\leq 1 $, $ \beta\geq\alpha $ and $ n\leq \frac{\beta-\alpha}{\alpha},\, n\in \mathbb{N} $. Moreover, let $ X $ be a nonnegative random variable with distribution $ F $, with moment $ \mathbb{E}[X^{(n+1)\alpha}]<+\infty $ and moments $ \mathbb{E}[X^{(j+1)\alpha-1}]<+\infty $ for all $ j\in\mathbb{N}, 0\leq j\leq n $. Assume further $ \mathbb{E}\left [ \left | D_{0}^{\left ( n+1 \right )\alpha }\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right )^{\beta} \right |\right ]<+\infty $. Then $ \mathbb{E}[X^{\beta}]<+\infty $ and $$\mathbb{E}[X^{\beta}]=\frac{\mathbb{E}[X^{(n+1)\alpha}]}{\Gamma \left ( \left ( n+1 \right )\alpha +1 \right )}\,\frac{\Gamma(1+\beta)}{\Gamma(1-(n+1)\alpha + \beta)}\mathbb{E}\left [ (X_{\,\alpha }^{\left ( n+1 \right )})^{\beta-(n+1)\alpha}\right],$$ where $ X_{\,\alpha }^{\left ( n+1 \right )} $ has density (\[FractionalDensity\]).
Since, in general, for $ k\in\mathbb{N} $ $$D_{0}^{ k \alpha }x^{\beta }=\frac{\Gamma(1+\beta)}{\Gamma(1-k\alpha+\beta)}x^{\beta-k\alpha},$$ we have $$c_{j}=\Gamma(\alpha)\left[\frac{\Gamma(1+\beta)}{\Gamma(1-j\alpha+\beta)}x^{1-(j+1)\alpha+\beta}\right](0^{+})=0, \qquad 0\leq j\leq n.$$ Furthermore, assumption $ \mathbb{E}\left [ \left | D_{0}^{\left ( n+1 \right )\alpha }\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right )^{\beta} \right |\right ]<+\infty $ ensures the finiteness of $ \mathbb{E}\left [ (X_{\,\alpha }^{\left ( n+1 \right )})^{\beta-(n+1)\alpha}\right] $. Therefore, formula (\[Thesis\]) reduces to the sole remainder term, and hence the thesis. [$\Box$\
]{}
Fractional probabilistic mean value theorem {#sec:4}
===========================================
In this section we develop the probabilistic analogue of a fractional mean value theorem. To this purpose we first recall some stochastic orders and introduce a relevant random variable, $ Z_{\alpha} $.
Let $ X $ be a random variable with distribution function $ F_{X} $ and let $ a=\inf \left \{ x|F_{X}\left ( x \right )>0 \right \} $ and $ b=\sup \left \{ x|F_{X}\left ( x \right )<1 \right \} $. We set for every real $ \alpha>0 $ $$F_{X}^{\left (\alpha \right ) }\left ( t \right )=
\left\{
\begin{array}{ll}
\dfrac{\mathbb{E}\left [ \left(t-X \right)_{+}^{\alpha -1}\right ]}{\Gamma \left ( \alpha \right )} &\mbox{if } t>a\\
0& \text{if }t\leq a,
\end{array}
\right.$$ and $$\overline{F}_{X}^{\left (\alpha \right ) }\left ( t \right )=
\left\{
\begin{array}{ll}
\dfrac{\mathbb{E}\left [\left( X-t\right)_{+}^{\alpha -1} \right ]}{\Gamma \left ( \alpha \right )}&\mbox{if } t<b\\
0& \text{if }t\geq b.
\end{array}
\right.$$
Ortobelli et al. [@Ortobelli2008] define a stochastic order as follows:
Let $ X $ and $ Y $ be two random variables. For every $ \alpha>0 $, $ X $ dominates $ Y $ with respect to the $ \alpha $-*bounded stochastic dominance order* $ (X\overset{b}{\underset{\alpha}{\geq}}Y) $ if and only if $ F_{X}^{\left (\alpha \right ) }\left ( t \right )\leq F_{Y}^{\left (\alpha \right ) }\left ( t \right ) $ for every $ t $ belonging to $ \mathit{supp}\{X,Y\}\equiv \left [ a,b \right ] $, where $ a,b\in\mathbb{\overline{R}} $ and $ a=\inf\left \{ x|F_{X} (x)+F_{Y}(x)>0\right \} $, $ b=\sup\left \{ x|F_{X} (x)+F_{Y}(x)<2\right \} $.
Similarly, Ortobelli et al. [@Ortobelli2008] define a *survival bounded order* as follows:
\[SurvivalBoundedOrder\] For every $ \alpha>0 $, we write $ X \overset{a}{\underset{\textit{sur $ \alpha $}}{\geq}} Y$ if and only if $\overline{F}_{X}^{\left (\alpha \right ) }\left ( t \right )\leq\overline{F}_{Y}^{\left (\alpha \right ) }\left ( t \right ) $ for every $ t $ belonging to $ \mathit{supp}\{X,Y\}$.
We remark that certain random variables cannot be compared with respect to these orders. For example, Ortobelli et al. [@Ortobelli2008] proved that for any pair of bounded (from above or/and from below) random variables $ X $ and $ Y $ that are continuous on the extremes of their support, there is no $ \alpha\in (0,1) $ such that $ F_{X}^{\left (\alpha \right ) }\left ( t \right )\leq F_{Y}^{\left (\alpha \right ) }\left ( t \right ) $ for all $ t\in\mathit{supp}\{X,Y\} $. However, although $ \alpha $-bounded orders with $ \alpha\in (0,1) $ are not applicable in many cases, they could be useful to rank truncated variables and financial losses, thus resulting of interest from a financial point of view.
We outline that for $ \alpha>1 $ the survival bounded order given in Definition \[SurvivalBoundedOrder\] is equivalent to the extension to all real $ \alpha>0 $ of the order $ \leq_{c}^{\alpha} $ defined in 1.7.1 of [@Stoyan1983]. Moreover, when $ \alpha=2 $, it is equivalent to the increasing convex order $ \leq_{icx} $ (cf. Section 4.A of Shaked and Shantikumar [@Shaked2007]).
The following result comes straightforwardly.
\[ZetaAlfa\] Let $ X $ and $ Y $ be non-negative random variables such that $ \mathbb{E}\left[ X^{\alpha}\right]<\mathbb{E}\left[ Y^{\alpha}\right]<+\infty $ for some $ \alpha>0 $. Then $$\label{Transformation}
f_{Z_{\alpha }}\left ( t \right )=\alpha \frac{\mathbb{E}\left [ \left ( Y-t \right )_{+}^{\alpha -1} \right ]-\mathbb{E}\left [ \left ( X-t \right )_{+}^{\alpha -1} \right ]}{\mathbb{E}\left [ Y^{\alpha } \right ]-\mathbb{E}\left [ X^{\alpha } \right ]},\qquad t\geq 0,$$ is the probability density of an absolutely continuous non-negative random variable, say $ Z_{\alpha} $, if and only if $ X \overset{0}{\underset{\textit{sur $ \alpha $}}{\geq}} Y$.
We remark that condition $ X \overset{0}{\underset{\textit{sur $ \alpha $}}{\geq}} Y $ ensures that $ \mathbb{E}[X^{\alpha}]\leq \mathbb{E}[Y^{\alpha}] $ for $ \alpha>0 $. Moreover, it is interesting to note that $ Z_{\alpha} $ is necessarily absolutely continuous, in contrast with $ X $ and $ Y $.
Let $ X $ and $ Y $ be exponential random variables having means $ \mu _{X} $ and $ \mu_{Y} $, $ \mu_{Y}>\mu _{X} >0$, and let $ \alpha\geq 1 $. From (\[Transformation\]) we obtain the following expression for the density of $ Z_{\alpha } $: $$f_{Z_{\alpha }}\left ( t \right )= \frac{\mu _{Y}^{\alpha -1}e^{-\frac{t}{\mu _{Y}}}-\mu _{X}^{\alpha -1}e^{-\frac{t}{\mu _{X}}}}{\mu _{Y}^{\alpha }-\mu _{X}^{\alpha }},\qquad t\geq 0.$$
Let $ Y $ be a random variable taking values in $ [0,b) $, with $ b\in(0,+\infty] $, and let $ \mathbb{E}[Y^{\alpha}] $ and $ \mathbb{E}[(Y-t)_{+}^{\alpha-1}] $ be finite, $ 0\leq t\leq b $. Furthermore, we define a random variable $ X $ with cumulative distribution function $$F_{X}(x):=\left\{
\begin{array}{ll}
0, & x<0, \\
p+(1-p)F_{Y}(x), & 0\leq x\leq b, \\
1, & x\geq b,
\end{array}
\right.$$ where $ F_{Y}(x) $ is the cumulative distribution function of $ Y $ and $ 0<p\leq 1 $. We remark that $ X $ can be viewed as a $ 0 $-inflated version of $ Y $, i.e. $ X=I\cdot Y $, where $ I $ is a Bernoulli r.v. independent of Y. It is easily ascertained that $$f_{Z_{\alpha }}\left ( t \right )=\frac{\alpha\mathbb{E}\left [ \left ( Y-t \right )_{+}^{\alpha -1} \right ] }{\mathbb{E}\left [ Y^{\alpha } \right ]}\equiv f_{Y_{\alpha }^{\left ( 1 \right )}}(t)\equiv f_{X_{\alpha }^{\left ( 1 \right )}}(t),\quad t\geq 0.$$
We note that the density of $ Z_{\alpha } $ given in (\[Transformation\]) is related to the densities of the fractional equilibrium variables $ X_{\alpha}^{(1)} $ and $ Y_{\alpha}^{(1)} $ which, by virtue of (\[FractionalDensity\]), are respectively given by $$f_{X_{\alpha }^{\left ( 1 \right )}}(t)= \frac{\alpha\mathbb{E}\left [ \left ( X-t \right )_{+}^{\alpha -1} \right ] }{\mathbb{E}\left [ X^{\alpha } \right ]}\quad\mathrm{and}\quad f_{Y_{\alpha }^{\left ( 1 \right )}}(t)= \frac{\alpha\mathbb{E}\left [ \left ( Y-t \right )_{+}^{\alpha -1} \right ] }{\mathbb{E}\left [ Y^{\alpha } \right ]},\quad t\geq 0.$$
Indeed, from (\[Transformation\]) the following generalized mixture holds: $$\label{Non-convexCombination}
f_{Z_{\alpha }}\left ( t \right )= cf_{Y_{\alpha }^{\left ( 1 \right )}}\left ( t \right )+\left ( 1-c \right )f_{X_{\alpha }^{\left ( 1 \right )}}\left ( t \right ),$$ where $$\label{Coefficient}
c=\frac{\mathbb{E}\left [ Y^{\alpha } \right ]}{\mathbb{E}\left [ Y^{\alpha } \right ]-\mathbb{E}\left [ X^{\alpha } \right ]}\geq 1.$$ Such representation is useful to find an expression for the moments of $ Z_{\alpha} $. In fact, from (\[moments\]), (\[Non-convexCombination\]) and (\[Coefficient\]), Proposition \[MomentsZetaAlpha\] follows immediately.
\[MomentsZetaAlpha\] Let $ \alpha \in \mathbb{R}^{+}$ and suppose that X and Y are two non-negative random variables such that $ \mathbb{E}\left[ X^{\alpha}\right]<\mathbb{E}\left[ Y^{\alpha}\right]<+\infty $, and $ X \overset{0}{\underset{\textit{sur $ \alpha $}}{\geq}} Y $. Then $$\label{momentsZalpha}
\mathbb{E}\left [ Z_{\alpha } ^{r}\right ]= \frac{\alpha B\left ( \alpha ,r+1 \right )}{\mathbb{E}\left [ Y^{\alpha } \right ]-\mathbb{E}[X^{\alpha }]}\left \{ \mathbb{E}\left [ Y^{\alpha+r } \right ]-\mathbb{E}[X^{\alpha+r }] \right \},\qquad r\in\mathbb{R}^{+}.$$
With the notation of 1.C(3) of [@Marshall2007], let $ \lambda_{\alpha}(X) $ denote the normalized moment of a random variable $ X $, i.e. $$\label{NormalizedMoments}
\lambda_{\alpha} (X)=\frac{\mathbb{E}[X^{\alpha}]}{\Gamma(\alpha+1)},\qquad \alpha>0.$$
We are now ready to prove the main result of this section.
\[Lagrange\] Let $ 0<\alpha\leq 1 $. Suppose that X and Y are two non-negative random variables such that $ \mathbb{E}\left[ X^{\alpha}\right]<\mathbb{E}\left[ Y^{\alpha}\right]<+\infty $, and $ X \overset{0}{\underset{\textit{sur $ \alpha $}}{\geq}} Y $. Moreover, let Theorem \[ThTaylor\] hold for some function $ g $. Then $$\begin{gathered}
\label{FracLagrange}
\mathbb{E}\left [ g\left ( Y \right ) \right ]-\mathbb{E}\left [ g\left ( X \right ) \right ]= \frac{c_{0}}{\Gamma \left ( \alpha \right )}\left \{ \mathbb{E}\left [ Y^{\alpha -1} \right ] -\mathbb{E}\left [ X^{\alpha -1} \right ] \right \}\\
+\{\lambda_{\alpha}(Y)-\lambda_{\alpha}(X)\}\mathbb{E}\left [ D_{0}^{\alpha }g\left ( Z_{\alpha } \right ) \right ],\end{gathered}$$ where $ Z_{\alpha } $ is a random variable whose density is defined in (\[Transformation\]), and $ c_{0}=\Gamma \left ( \alpha \right )\left [ x^{1-\alpha} g(x)\right ](0^{+}) $.
By applying Theorem \[ThTaylor\] for $ n=0 $, cf. formula (\[RemarkTaylor\]), we have $$\begin{gathered}
\mathbb{E}\left [ g\left ( Y \right ) \right ]-\mathbb{E}\left [ g\left ( X \right ) \right ]= \frac{c_{0}}{\Gamma \left ( \alpha \right )}\left \{ \mathbb{E}\left [ Y^{\alpha -1} \right ] -\mathbb{E}\left [ X^{\alpha -1} \right ] \right \}\\
+\frac{1}{\Gamma \left ( \alpha +1 \right )}\left \{ \mathbb{E}\left [ Y^{\alpha } \right ]\mathbb{E}\left [ D_{0}^{\alpha } g\left ( Y_{\alpha }^{\left ( 1 \right )} \right )\right ]-\mathbb{E}\left [ X^{\alpha } \right ]\mathbb{E}\left [ D_{0}^{\alpha } g\left ( X_{\alpha }^{\left ( 1 \right )} \right )\right ] \right \}.\nonumber\end{gathered}$$ From (\[Non-convexCombination\]), (\[Coefficient\]) and (\[NormalizedMoments\]) the theorem is straightforwardly proved. [$\Box$\
]{}
Under certain hypotheses, the Lagrange’s Theorem guarantees the existence of a mean value belonging to the interval of interest. With regard to Theorem \[Lagrange\], one might therefore expect that a probabilistic analogue of this relation holds too. However, the relation $ X^{\alpha}\leq_{st}Z_{\alpha}\leq_{st} Y^{\alpha} $ does not hold in general. It can be satisfied only when $ \mathbb{E}[X^{\alpha}]\leq\mathbb{E}[Z_{\alpha}]\leq\mathbb{E}[Y^{\alpha}] $, which is case $ (ii) $ of the next Proposition. For simplicity’s sake, if $ X $ is a random variable with $ \mathbb{E}[X^{\alpha+1}]<+\infty $, we set $$\label{FractionalVariance}
V_{\alpha }\left ( X \right ):= \mathbb{E}\left [ X^{\alpha +1} \right ]-\alpha \left ( \mathbb{E} \left [ X^{\alpha} \right ]\right )^{2}, \qquad \alpha\in\mathbb{R}^{+},$$ which turns out to be a fractional extension of the variance of $ X $.
Let $ 0<\alpha\leq 1 $ and let $ X $ and $ Y $ satisfy the assumptions of Theorem \[Lagrange\], with $ \mathbb{E}[X^{\alpha+1}] $ and $ \mathbb{E}[Y^{\alpha+1}] $ finite. Then,
1. $ \begin{multlined}[t]
\mathbb{E}[Z_{\alpha}]\leq\mathbb{E}[X^{\alpha}]\\ \Leftrightarrow\;V_{\alpha }\left ( Y \right )-V_{\alpha }\left ( X \right )\leq -\left \{\mathbb{E}[Y^{\alpha }] -\mathbb{E}[X^{\alpha}]\right \}\left \{ \alpha \mathbb{E}[Y^{\alpha }] -\mathbb{E}[X^{\alpha}] \right \};
\end{multlined}$
2. $ \begin{multlined}[t]
\mathbb{E}[X^{\alpha}]\leq\mathbb{E}[Z_{\alpha}]\leq\mathbb{E}[Y^{\alpha}]\\ \Leftrightarrow\;\left\{\begin{array}{lr}
V_{\alpha }\left ( Y \right )-V_{\alpha }\left ( X \right )\geq -\left \{\mathbb{E}[Y^{\alpha }] -\mathbb{E}[X^{\alpha}]\right \}\left \{ \alpha \mathbb{E}[Y^{\alpha }] -\mathbb{E}[X^{\alpha}] \right \} \\
V_{\alpha }\left ( Y \right )-V_{\alpha }\left ( X \right )\leq \left \{\mathbb{E}[Y^{\alpha }] -\mathbb{E}[X^{\alpha}]\right \}\left \{\mathbb{E}[Y^{\alpha }] -\alpha \mathbb{E}[X^{\alpha}]\right \};
\end{array}\right.
\end{multlined}$
3. $\begin{multlined}[t]
\mathbb{E}[Y^{\alpha}]\leq \mathbb{E}[Z_{\alpha}]\\ \Leftrightarrow\;V_{\alpha }\left ( Y \right )-V_{\alpha }\left ( X \right )\geq \left \{\mathbb{E}[Y^{\alpha }] -\mathbb{E}[X^{\alpha}]\right \}\left \{\mathbb{E}[Y^{\alpha }] -\alpha \mathbb{E}[X^{\alpha}]\right \};
\end{multlined}$
4. $\begin{multlined}[t] \mathbb{E}[Z_{\alpha }]= \dfrac{2\alpha }{\alpha +1}\, \dfrac{\mathbb{E}[X^{\alpha }]+\mathbb{E}[Y^{\alpha }]}{2}\;\Leftrightarrow \; V_{\alpha }(Y)=V_{\alpha }(X)\end{multlined} $,
where $ V_{\alpha} $ has been defined in (\[FractionalVariance\]).
It easily follows from the identity $$\frac{\mathbb{E}\left [ Z_{\alpha } \right ]-\mathbb{E}\left [ X^{\alpha } \right ]}{\mathbb{E}[Y^{\alpha }]-\mathbb{E}\left [ X^{\alpha } \right ]}= \frac{1}{\alpha +1}\left \{ \frac{\alpha \mathbb{E}\left [ Y^{\alpha } \right ]-\mathbb{E}\left [ X^{\alpha } \right ]}{\mathbb{E}\left [ Y^{\alpha } \right ]-\mathbb{E}\left [ X^{\alpha } \right ]} +\frac{V_{\alpha }(Y)-V_{\alpha }(X)}{\left ( \mathbb{E}\left [ Y^{\alpha } \right ]-\mathbb{E}\left [ X^{\alpha } \right ] \right )^{2}}\right \},$$ which is a consequence of (\[momentsZalpha\]) written for $ r=1 $. [$\Box$\
]{}
\[PreApplication\] Let $ 0<\alpha\leq 1 $. Suppose that X and Y are two non-negative random variables such that $ \mathbb{E}\left[ X^{\alpha}\right]<\mathbb{E}\left[ Y^{\alpha}\right]<+\infty $, and $ X \overset{0}{\underset{\textit{sur $ \alpha $}}{\geq}} Y $. Moreover, if $ \beta>\alpha-1 $, let Theorem \[ThTaylor\] hold for some function $ g(x)\sim x^{\beta}$, $ x\rightarrow 0^{+}$. Then $$\mathbb{E}\left [ g\left ( Y \right ) \right ]-\mathbb{E}\left [ g\left ( X \right ) \right ]=\left \{ \lambda_{\alpha}(Y)-\lambda_{\alpha}(X) \right \}\mathbb{E}\left [ D_{0}^{\alpha }g\left ( Z_{\alpha } \right ) \right ],$$ where $ Z_{\alpha } $ is a random variable whose density is defined in (\[Transformation\]).
We observe that the first term in formula (\[FracLagrange\]) vanishes, since $ c_{0}\sim x^{1-\alpha+\beta}\big |_{x=0^{+}}=0$, and hence the thesis. [$\Box$\
]{}
As application, we now show a result of interest to actuarial science. A *deductible* is a treshold amount, denoted $ d $, which must be exceeded by a loss in order for a claim to be paid. If $ X $ is the severity random variable representing a single loss event, $ X>0 $, and if the deductible is exceeded (that is, if $ X>d $), then the amount paid is $ X-d $. Therefore, for $ d>0 $, the claim amount random variable $ X_{d} $ is defined to be $$\label{PaymentPerLoss}
X_{d}:=(X-d)_{+}=\left\{
\begin{array}{ll}
0,&\text{for $ X\leq d $,}\\
X-d,&\text{for $ X>d $}.
\end{array}
\right.$$ It is clear from equation (\[PaymentPerLoss\]) that $ X_{d} $ has a mixed distribution. In particular, such random variable has an atom at zero representing the absence of payment because the loss did not exceed $ d $. The interested reader is referred to [@Kellison2011] and [@Klugman2012] for further information. Let $ b=\sup \left \{ x|F_{X}\left ( x \right )<1 \right \} $. Bearing in mind Definition \[ZetaAlfa\], the next Proposition immediately follows from Corollary \[PreApplication\].
\[Appication\] Let $ 0<\alpha\leq 1 $ and $ 0<r<s<b $. With reference to (\[PaymentPerLoss\]), suppose that $ X_{s} \overset{0}{\underset{\textit{sur $ \alpha $}}{\geq}} X_{r} $ and $ \lambda_{\alpha} (X_{s})<\lambda_{\alpha} (X_{r})<+\infty$. Moreover, let $ g $ satisfy the assumptions of Corollary \[PreApplication\]. Then $$\mathbb{E}\left [ g\left ( X_{r} \right ) \right ]-\mathbb{E}\left [ g\left ( X_{s} \right ) \right ]=[\lambda_{\alpha}(X_{r})-\lambda_{\alpha}(X_{s})]\mathbb{E}\left [ D_{0}^{\alpha }g\left ( Z_{\alpha } \right ) \right ],$$ where $ Z_{\alpha } $ is a random variable with density $$f_{Z_{\alpha}}(z)=\alpha \frac{\mathbb{E}\left [ \left ( X_{r}-z \right )_{+}^{\alpha -1} \right ]-\mathbb{E}\left [ \left ( X_{s}-z \right )_{+}^{\alpha -1} \right ]}{\mathbb{E}\left [ X_{r}^{\alpha } \right ]-\mathbb{E}\left [ X_{s}^{\alpha } \right ]},\qquad z\geq 0.$$
We conclude this section with the following example.
Let $ 0<\alpha\leq 1 $ and $ 0<r<s $.
1. Let $ X $ be an exponential random variable with parameter $ \lambda $. Owing to (\[NormalizedMoments\]) and Proposition \[Appication\], we have $$\mathbb{E}\left [ g\left ( X_{r} \right ) \right ]-\mathbb{E}\left [ g\left ( X_{s} \right ) \right ]=\lambda^{-\alpha}\left(e^{-\lambda r}-e^{-\lambda s}\right)\mathbb{E}\left [ D_{0}^{\alpha }g\left ( Z\right ) \right ],$$ where $ Z $ turns out to be exponentially distributed with parameter $ \lambda $ as well. In this case, it is interesting to note that for $ 0<r<s $ and $ 0<u<v $ it results: $$\frac{\mathbb{E}\left [ g\left ( X_{r} \right ) \right ]-\mathbb{E}\left [ g\left ( X_{s} \right ) \right ]}{\mathbb{E}\left [ g\left ( X_{u} \right ) \right ]-\mathbb{E}\left [ g\left ( X_{v} \right ) \right ]}=\frac{e^{-\lambda r}-e^{-\lambda s}}{e^{-\lambda u}-e^{-\lambda v}},$$ which is independent of $ g $.
2. Now let $ X $ be a 2-phase hyperexponential random variable with phase probabilities $ p $ and $ 1-p $, $ 0<p<1 $, and rates $ \lambda_{1} $ and $ \lambda_{2} $. Similarly, it holds $$\begin{gathered}
\mathbb{E}\left [ g\left ( X_{r} \right ) \right ]-\mathbb{E}\left [ g\left ( X_{s} \right ) \right ]=\left \{ p\lambda_{1}^{-\alpha}\left(e^{-\lambda_{1} r}-e^{-\lambda_{1} s}\right)\right.\\
\left. +\left(1-p\right)\lambda_{2}^{-\alpha}\left(e^{-\lambda_{2} r}-e^{-\lambda_{2} s}\right)\right \} \mathbb{E}\left [ D_{0}^{\alpha }g\left ( Z_{\alpha } \right ) \right ],\end{gathered}$$ where, the density of $ Z_{\alpha } $ is, for $ z\geq 0 $,
$$f_{Z_{\alpha }}\left ( z \right )=\frac{p\lambda_{1}^{1-\alpha}e^{-\lambda_{1}z}\left(e^{-\lambda_{1} r}-e^{-\lambda_{1} s}\right)+\left(1-p\right)\lambda_{2}^{1-\alpha}e^{-\lambda_{2}z}\left(e^{-\lambda_{2} r}-e^{-\lambda_{2} s}\right)}{p\lambda_{1}^{-\alpha}\left(e^{-\lambda_{1} r}-e^{-\lambda_{1} s}\right)+\left(1-p\right)\lambda_{2}^{-\alpha}\left(e^{-\lambda_{2} r}-e^{-\lambda_{2} s}\right)}.$$
Concluding remarks {#sec:5}
==================
The overall aim of this research is to present a novel Taylor’s theorem from a probabilistic and a fractional perspective at the same time and to discuss other related findings. It is meaningful to note that, while the coefficients of our formula (\[Thesis\]) are expressed in terms of the Riemann-Liouville fractional derivative, it is possible to establish a fractional probabilistic Taylor’s theorem in the Caputo sense too. The Caputo derivative, denoted by $_{*}D_{a+}^{\alpha } $, is defined by exchanging the operators $ I_{a+}^{m-\alpha } $ and $ D^{m} $ in the classical definition (\[RLderivative\]). Taking the paper of Odibat et al. [@Odibat2007] as a starting point, the following theorem, which is in some sense the equivalent of Theorem \[ThTaylor\], can be effortlessly proved.
Let $ \alpha\in \left(0,1\right] $ and let $ X $ be a nonnegative random variable with distribution $ F $ and moment $ \mathbb{E}\left[X^{(n+1)\alpha}\right]<+\infty $ for some integer $ n\geq 0 $. Assume that $g$ is a function defined on $I_{F}$, with $I_{F}$ defined in (\[eq:IF\]), and suppose that $_{*}D_{0+}^{\alpha }g(x)\in C \left(I_{F}\right) $ for $ k=0,1,\dots, n+1 $ and $ \mathbb{E}\left [ \left | _{*}D_{0}^{\left ( n+1 \right )\alpha }g\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right ) \right |\right ]<+\infty $. Then $ \mathbb{E}\left [ g\left ( X \right ) \right ]<+\infty $ and $$\begin{gathered}
\mathbb{E}\left [ g(X) \right ]= \sum_{i=0}^{n}\frac{\left(_{*}D_{0+}^{i\alpha }f\right)(0)}{\Gamma \left ( i\alpha +1 \right )}\mathbb{E}\left [ X^{i\alpha } \right ]\\
+\frac{\mathbb{E}[X^{(n+1)\alpha}]}{\Gamma \left ( \left ( n+1 \right )\alpha +1 \right )}\,\mathbb{E}\left [ _{*}D_{0}^{\left ( n+1 \right )\alpha }g\left ( X_{\,\alpha }^{\left ( n+1 \right )} \right ) \right ],\end{gathered}$$ where $ _{*}D_{0+}^{n\alpha }\,=\,_{*}D_{0+}^{\alpha }\cdot\,_{*}D_{0+}^{\alpha }\,\cdots\,_{*}D_{0+}^{\alpha } $ ($ n $ times) and $ X_{\,\alpha }^{\left ( n+1 \right )} $ has density (\[FractionalDensity\]).
Acknowledgements {#acknowledgements .unnumbered}
================
The research that led to the present paper was partially supported by a grant of the group GNCS of INdAM.
[99]{}
Y. Cheng, J.S. Pai, The maintenance properties of nth stop-loss order. In: *Proceedings of the 30th International ASTIN Colloquium/9th International AFIR Colloquium*, (1999), 95–118. A. Di Crescenzo, A probabilistic analogue of the mean value theorem and its applications to reliability theory. *J. Appl. Probab.* **36**, No 3 (1999), 706–719. doi:10.1239/jap/1032374628 R. Gorenflo, A.A. Kilbas, F. Mainardi, S.V. Rogosin, *Mittag-Leffler functions, related topics and applications*. Springer, Heidelberg (2014). doi:10.1007/978-3-662-43930-2 W.L. Harkness, R. Shantaram, Convergence of a sequence of transformations of distribution functions. *Pacific J. Math.* **31**, No 2 (1969), 403–415. E. Hewitt, K. Stromberg, *Real and abstract analysis. A modern treatment of the theory of functions of a real variable*. Third printing. Graduate Texts in Mathematics, No 25. Springer-Verlag, New York-Heidelberg (1975). P. Guo, C. Li, G. Chen, On the fractional mean-value theorem. *Int. J. Bifurcation Chaos* **22**, 1250104 (2012). doi: 10.1142/S0218127412501040 S.G. Kellison, R.L. London, *Risk Models and Their Estimation*, ACTEX Publications, Winsted, CT (2011). S.A. Klugman, H.H. Panjer, G.E. Willmot, *Loss Models: From Data to Decisions*. Fourth edition, Wiley, Hoboken, NJ (2012). G.D. Lin, On a probabilistic generalization of Taylor’s theorem. *Statist. Probab. Lett.* **19**, No 3 (1994), 239–243. doi:10.1016/0167-7152(94)90110-4 A.W. Marshall, I. Olkin, *Life Distributions: Structure of Nonparametric, Semiparametric, and Parametric Families*. Springer-Verlag, New York (2007). doi:10.1007/978-0-387-68477-2 W.A. Massey, W. Whitt, A probabilistic generalization of Taylor’s theorem. *Statist. Probab. Lett.* **16**, No 1 (1993), 51–54. doi:10.1016/0167-7152(93)90122-Y T. Mikosch, G. Samorodnitsky, L. Tafakori, Fractional moments of solutions to stochastic recurrence equations. *J. Appl. Probab.* **50**, No 4 (2013), 969–982. doi:10.1239/jap/1389370094 Z.M. Odibat, N.T. Shawagfeh, Generalized Taylor’s formula. *Appl. Math. Comput.* **186**, No 1 (2007), 286–293. doi:10.1016/j.amc.2006.07.102 S. Ortobelli, S. Rachev, H. Shalit, F.J. Fabozzi, Orderings and Risk Probability Functionals in Portfolio Theory. *Probab. Math. Statist.* **28**.2 (2008), 203–234. J.E. Pe$\check{\rm c}$arić, I. Perić, H.M. Srivastava, A family of the Cauchy type mean-value theorems. *J. Math. Anal. Appl.* **306**, No 2 (2005), 730–739. doi:10.1016/j.jmaa.2004.10.018 M. Shaked, J.G. Shanthikumar, *Stochastic orders*. Springer, New York (2007). doi:10.1007/978-0-387-34675-5 D. Stoyan, *Comparison methods for queues and other stochastic models*. Wiley, Chichester (1983). J.J. Trujillo, M. Rivero, B. Bonilla, On a Riemann–Liouville generalized Taylor’s formula. *J. Math. Anal. Appl.* **231**, No 1 (1999), 255–265. doi:10.1006/jmaa.1998.6224
[^1]: Dipartimento di Matematica, Università degli Studi di Salerno, Via Giovanni Paolo II, 132, 84084 Fisciano (SA), Italy, email: [email protected]
[^2]: Dipartimento di Matematica, Università degli Studi di Salerno, Via Giovanni Paolo II, 132, 84084 Fisciano (SA), Italy, email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a pseudoconvex hypersurface in ${\mathbb{C}}^n$ and an arbitrary weight, we show the existence of local coordinates in which the polynomial model contains a particularly simple sum of squares of monomials. Our second main result provides a normalization of a part of any Catlin boundary system. We illustrate by an example that this normalization cannot be extended to the rest of the boundary system due to the existence of what we refer to as torsion.'
address:
- 'Department of Mathematics, Statistics and Computer Science, University of Wisconsin-Stout, Menomonie, WI 54751'
- 'School of Mathematics, Trinity College Dublin, Dublin 2, Ireland'
author:
- Alexander Basyrov
- 'Andreea C. Nicoara'
- Dmitri Zaitsev
title: ' Sums of squares in pseudoconvex hypersurfaces and torsion phenomena for Catlin’s boundary systems '
---
Introduction {#intro}
============
The goal of this paper is to gain new geometric insight into the tools developed for establishing global regularity and subelliptic estimates for the $\bar\d$-Neumann problem. We are focusing on the approach by Catlin in [@C84; @C87]. However, we expect our methods to also shed light on Kohn’s multiplier ideal technique initiated in [@kohnacta] and continued more recently in [@fribourgcatda; @siunote; @ra; @kimzaitsev; @siunew] as well as on other research related to the $\bar\d$-equation (see [@dimatensors 1.3]) and potentially more general PDE’s, as evidenced by the program pioneered by Siu in [@siunew].
Recall that Catlin established global regularity and subelliptic estimates for the $\bar\d$-Neumann problem as consequences of his Property (P) type conditions. The only known proofs of Property (P) type conditions for general smooth pseudoconvex finite type domains in the sense of D’Angelo [@opendangelo] rely on the techniques of multitype, polynomial models, and boundary systems introduced in [@catlinbdry].
One of the motivations in this paper is reducing the complexity in Catlin’s techniques by a more explicit use of pseudoconvexity and a precise normalization of the geometry. Our first result is showing the existence of so-called positive balanced terms in polynomial models of pseudoconvex hypersurfaces. We call a monomial $z_1^{\a_1} \z_1^{\b_1} \ldots z_n^{\a_n} \z_n^{\b_n} $ [*balanced*]{} if $\a_j=\b_j$ for all $j$. We shall use the (standard) [*lexicographic order*]{} for multiweights and the [*reverse lexicographic order*]{} for multidegrees. We prove the following:
\[main\] Let $M\subset {\mathbb{C}}^n$ be a pseudoconvex real smooth hypersurface with $0\in M$. Let $z=(z_1, \ldots, z_n)$ be local holomorphic coordinates in a neighborhood of $0$ and $$\mu= (\mu_1, \mu_2, \mu_3, \ldots, \mu_n),
\quad
1=\mu_1> \mu_2\ge \mu_3 \ge \ldots \ge \mu_n > 0,$$ be a (multi-)weight such that $M$ is given by $$r(z,\bar z) =0,
\quad
d r \ne 0,
\quad
r = O_\mu(1),$$ i.e. the expansion of $r$ contains only terms of weight greater or equal to $1$.
Then after a weighted homogenous polynomial change of coordinates, the defining function $r$ of $M$ admits a decomposition $$\label{main-eq}
r(z,\bar z) =-2\operatorname{Re}z_1
+ p (z_2, \ldots, z_n, \bar z_2, \ldots, \bar z_n)
+ o_\mu(1),
$$ where $p$ is a weighted homogeneous polynomial of weight $1$ containing, as part of its expansion, the sum of squares \[bal\] A\_2 |z\_2|\^[2k\_[22]{}]{} + A\_3 |z\_2|\^[2k\_[32]{}]{} |z\_3|\^[2k\_[33]{}]{} + … + A\_n |z\_2|\^[2k\_[n2]{}]{} |z\_n|\^[2k\_[nn]{}]{} with $A_j \ge 0$, $k_{jj} >0$ for all $j=2,\ldots, n,$ such that the (total) degree of $p$ in each $(z_j, \bar z_j)$ is not greater than $2k_{jj}$, and $o_\mu(1)$ stands for a smooth function in $(\operatorname{Im}z_1, z_2, \ldots, z_n, \bar z_2, \ldots, \bar z_n)$ whose formal Taylor series expansion contains only terms of weight greater than $1$. In addition, for each $j$, the multidegree of each term of in $(z_2,\z_2), \ldots, (z_j,\z_j)$ is maximal among all balanced monomials in $p$ in the reverse lexicographic order.
Furthermore, either all $A_j$ can be chosen positive, or the weight $\mu$ can be lowered lexicographically.
The following example illustrates that pseudoconvexity is an essential assumption:
Let $M$ be the non-pseudoconvex domain given by $r<0$ with $$r= -2\operatorname{Re}z_1 + 2\operatorname{Re}(z_2^2\z_3^3).$$ Then no weighted polynomial change of variables transforms $M$ into a form satisfying the conclusion of Theorem \[main\]. In fact, no terms as in can be obtained (even with $A_j$ negative).
On the other hand, when $M$ is pseudoconvex, a natural question arises whether the statement could be improved by reducing the polynomial $p$ to just a sum of the terms in . It is not possible in general, however, as the following simple example illustrates:
Let $M\subset{\mathbb{C}}^2$ be the tube given by $r=0$ with $$r=-2\operatorname{Re}z_1 + (\operatorname{Re}z_2)^2.$$ Then the polynomial $(\operatorname{Re}z_2)^2$ is invariant under weighted homogeneous polynomial coordinate changes, and hence cannot be reduced to contain the terms only.
Furthermore, it is generally not possible to reduce $r$ to the sum only, even when $r$ can be written as a sum of squares of holomorphic functions:
Let $p,q \in {\mathbb{N}},$ $p,q \ge 2.$ Consider $$r_0 =
-2\operatorname{Re}z_1 + |z_2|^{2p}+|z_3|^{2q} +
2\epsilon \operatorname{Re}z_2^p \bar z_3^q$$ with $|\epsilon|<1$. Then $r_0$ determines a real-algebraic pseudoconvex hypersurface that can be written as a sum of squares of holomorphic functions: \[sq\] r\_0 = -2z\_1 +|z\_2\^p +z\_3\^q|\^2 + (1-\^2) | z\_3|\^[2q]{}. By a direct computation (or using e.g. [@martinIMRN Theorem 4.1]), it can be seen, however, that no biholomorphic change of variables can transform $r_0$ into a sum of squares of the form $$|a_2z_2|^{2k_{22}} + A_3 |z_2|^{2k_{32}} |z_3|^{2k_{33}}.$$
The following example illustrates that the terms cannot be in general reduced to sums of single powers $|z_j|^{2k_{jj}}$. Let $M\subset{\mathbb{C}}^3$ be given by $r=0$ with \[eqq\] r=-2z\_1 + |z\_2|\^8 + |z\_2|\^4|z\_3|\^6, a weighed homogeneous polynomial in $(z_1, z_2, z_3)$ and conjugates with corresponding weights $(1,8,12)$. Then the only weighted homogeneous changes of coordinates are the linear dilations $(z_1,z_2,z_3)\mapsto (a_1z_1,a_2z_2,a_3z_3)$ that clearly preserve up to a change of coefficients. In particular, it is not possible to obtain a conclusion similar to Theorem \[main\] with consisting of sums of single powers $|z_j|^{2k_{jj}}$ only.
Applying Theorem \[main\] to the Catlin multitype yields the following:
\[catlincase\] Let $M$ be a pseudoconvex smooth real hypersurface in ${\mathbb{C}}^n$ with $0 \in M$ and of the Catlin multitype $$\Lambda= (1, \lambda_2, \lambda_3, \ldots, \lambda_n),
\quad
\lambda_n < + \infty$$ at $0$.
Then there exists a holomorphic change of coordinates at $0$ preserving the multitype so that the defining function for $M$ in the new coordinates is given by $$\label{m-eq}
r = -2\operatorname{Re}z_1 + p(z_2, \ldots, z_n, \bar z_2, \ldots, \bar z_n)
+ o_{\L^{-1}}(1),$$ where $p$ is a weighted homogeneous polynomial of weight $1$ that contains the sum of squares $$\label{a-terms}
|z_2|^{2k_{22}} + |z_2|^{2k_{32}} |z_3|^{2k_{33}} + \ldots + |z_2|^{2k_{n2}} \cdots |z_n|^{2k_{nn}},
$$ where $k_{jj} >0$ for all $j$ and the (total) degree of $p$ in each $(z_j, \bar z_j)$ is not greater than $k_{jj}.$
It follows from the weighted homogeneity of that \[lj\] \_[l=2]{}\^j =1, j=1,…, n. In particular, since $k_{jj}\ne 0$, the weights $\l_l$ are uniquely determined from .
A better understanding of the Catlin multitype has been obtained when $M$ is a boundary of a [*convex*]{} domain, see [@bs; @mcneal; @jiyeyu]. However, even in this case, Thereom \[main\] seems to be new.
For a pseudoconvex hypersurface, the multitype and the commutator multitype coincide as Catlin established in [@catlinbdry]. Without pseudoconvexity, however, it might not be possible to obtain terms of the type $$|z_2|^{2k_{j2}} \cdots |z_n|^{2k_{jj}}$$ via a change of variables, and the two multitypes may differ, as the following example due to Bloom shows:
Consider $$r_0 =
\operatorname{Re}z_1 + (\operatorname{Re}z_2 +|z_3|^2)^2.$$ The multitype at $0$ is given by ${\ensuremath{\mathfrak M}}=(1,2,4),$ whereas the commutator multitype at $0$ is lexicographically strictly larger: ${\ensuremath{\mathfrak C}}=(1,2, + \infty).$ Catlin proved that ${\ensuremath{\mathfrak M}}\leq {\ensuremath{\mathfrak C}}$ for any domain, and this example shows nothing more can be expected to hold in the absence of pseudoconvexity.
Our second main result is a normalization of any boundary system for a pseudoconvex domain via a change of variables:
\[flatnessprop\] Let $M$ be a pseudoconvex smooth real hypersurface in ${\mathbb{C}}^n$ with $0 \in M,$ Levi rank $s_0$ at $0,$ and of the Catlin multitype $$\Lambda= (1, \underbrace{2, \dots, 2}_{s_0}, \lambda_{s_0+2}, \ldots, \lambda_n)
$$ at $0$, where $$2<\lambda_{s_0+2}=\dots = \lambda_{s_0+s_1+1}< \lambda_{s_0+s_1+2} < +\infty.$$ Then for any boundary system at $0$, $${\ensuremath{\mathfrak B}}_n (0) = \{r_1, r_{s_0+2},\dots, r_n; L_2, \dots, L_n\},$$ there exists a holomorphic change of coordinates at $0$ preserving the multitype and transforming ${\ensuremath{\mathfrak B}}_n(0)$ into $$\tilde {\ensuremath{\mathfrak B}}_n (0) = \{\tilde r_1,\tilde r_{s_0+2}, \dots, \tilde r_n; \tilde L_2, \dots, \tilde L_n\},$$ satisfying the normalization $$\label{flatness-eq}
\begin{aligned}
\tilde r_j &= \operatorname{Re}z_j
+o(
\lambda_j^{-1}
),
\quad s_0+2 \leq j \leq s_0+s_1+1,\\
\tilde L_k&=\d_{z_k}
+o(
\lambda_k^{-1}
),
\quad 2\le k\le s_0+s_1+2, \\
\end{aligned}$$ where the partial derivatives $\d_{\z_j}$ are counted with weight $-\l_j^{-1}$.
Therefore, the first function in the boundary system beside the defining function can always be brought to the simplest possible form. In case several entries of the Catlin multitype are equal to the first $\lambda_j >2,$ all of their corresponding functions and vector fields in the boundary system can be normalized. This normalization process cannot be carried out on the subsequent functions, however. In fact, we provide a [*counterexample*]{} in the final section of the paper. This non-existence of a complete normalization provides a very important insight into the behavior of Catlin’s boundary systems that we call a [*torsion phenomenon*]{}.
In particular, for $M$ of Levi rank $0$ in ${\mathbb{C}}^3,$ this flattening result yields a simplified geometric picture, which we state as a corollary:
\[C3case\] Let $M$ be a pseudoconvex smooth real hypersurface in ${\mathbb{C}}^3$ with $0 \in M,$ Levi rank $0,$ and of the Catlin multitype $$\Lambda= (1, \lambda_2, \lambda_3),
\quad
\lambda_3 < + \infty$$ at $0$. After a holomorphic change of variables, any boundary system at $0$ $${\ensuremath{\mathfrak B}}_3 (0) = \{r_1, r_2, r_3; L_2, L_3\}$$ becomes $$\tilde {\ensuremath{\mathfrak B}}_3 (0) = \{\tilde r_1, \tilde r_2, \tilde r_3; \tilde L_2, \tilde L_3\}$$ with $$\label{flatness-eqc3}
\begin{aligned}
\tilde r_{2} &= \operatorname{Re}z_2 +o (\lambda_2^{-1}),\\
\tilde L_j&=\partial_{z_j}+
o( \lambda_j^{-1} ),
\quad j =2,3,\\
\end{aligned}$$
The paper is organized as follows: Section \[examples\] provides additional examples illustrating various phenomena. Section \[notation\] gives the relevant definitions and notation. Section \[bdrysystems\] defines Catlin’s multitype and boundary systems. Section \[estimates\] presents some elementary auxiliary results such as one-dimensional estimates for non-negative homogeneous polynomials and a several variables version proven via scaling using the Newton polygon. Section \[normalforms\] contains the proofs of Theorem \[main\] and Corollary \[catlincase\] carried out in a sequence of lemmas. In the same section, it is shown that the first function in the boundary system can always be normalized, thus establishing Theorem \[flatnessprop\] and Corollary \[C3case\]. Section \[r3counterexample\] demonstrates by example that the same type of normalization cannot be carried out on subsequent functions in the boundary system.
Further motivation and examples {#examples}
===============================
The grandfather of all examples is a strongly pseudoconvex hypersurface given by $r=0$ with $$r=-2\operatorname{Re}z_1 + |z_2|^2 +\ldots + |z_n|^2 + o_{\L^{-1}}(1),
\quad
\L=(1, 2, \ldots, 2),$$ where the whole leading polynomial (the quadric), is already of the form , once diagonalized. Of course, the diagonalization here is a special property of quadrics that does not extend to higher degree polynomials.
More generally, it was shown by the third author in [@dimatensors] that any pseudoconvex hypersurface admits the form $r=0$ with \[p4\] r = -2z\_1 + |z\_2|\^2 + …+ |z\_[q-1]{}|\^2 + p\_4(z\_[\[q+2,n\]]{}, \_[\[q+2,n\]]{} ) + o\_[Ł\^[-1]{}]{}(1), with the inverse weights $
\L= (1, 2, \ldots, 2, 4, \ldots, 4),
$ where the number of $2$’s equals the Levi form rank $q$, and $p_4$ represents the CR invariant quartic tensor defined in [@dimatensors]. In view of this general fact, it suffices to only study terms arising from $p_4$ and $o_{\L^{-1}}(1)$.
Any real hypersurface $M\subset{\mathbb{C}}^2$ of finite type $m$ admits the form $r=0$ with $$r = -2\operatorname{Re}z_1 + p_m(z_2,\z_2) + o_{\L^{-1}}(1),
\quad
p_m = \sum_{j+k=m} a_{jk} z_2^j \z_2^{k},$$ where $\L=(1,m)$ and $p_m$ is not harmonic. In this (well-known) case, if $M$ is pseudoconvex, Theorem \[main\] implies that $m=2l$ is even and $p_m$ contains a nontrivial term $a_{ll} |z_2|^{2l}$, $a_l\ne 0$. Since $p_m$ is a (tensor) invariant, it is clear that other terms $a_{jk} z_2^j \z_2^{k}$ with $j\ne k$ cannot be eliminated.
For a general pseudoconvex hypersurface $M\subset{\mathbb{C}}^3$ of the form in its multitype coordinates, Corollary \[catlincase\] implies that after a linear change of coordinates, $p$ contains nontrivial terms \[p22\] p = |z\_2|\^[2k\_[22]{}]{} + |z\_2|\^[2k\_[32]{}]{} |z\_3|\^[2k\_[33]{}]{} + p(z\_2,z\_3, \_2,\_3), k\_[22]{}, k\_[33]{} 1. where $\2p$ consists of all remaining terms. In particular, the multitype $\L=(1, m_2, m_3)$ is uniquely determined from the identities $$m_2= 2k_{22},
\quad
\frac{2k_{32}}{m_2} + \frac{2k_{33}}{m_3} = 1$$ expressing the property that the first two terms in are of weight $1$.
Furthermore, the additional degree property in Theorem \[main\] asserts that the degree of $p$ in $(z_3,\z_3)$ equals $2k_{33}$. This property puts additional restrictions on the terms in $p$ and makes the choice of the first two terms in canonical. For instance, if $$p= |z_2|^4 + |z_2z_3|^2 + |z_3|^4,$$ the degree condition forces us to choose the terms $|z_2|^4 + |z_3|^4$ rather than $|z_2|^4 + |z_2z_3|^2$, because the latter choice would violate the property that the degree of $p$ in $(z_3,\z_3)$ is $2$.
Let $M\subset{\mathbb{C}}^4$ be given by $r=0$ with $$r = - 2\operatorname{Re}z_1 + |z_2|^4 + |z_2|^2 |z_3|^2 + ( |z_2|^2 + |z_3|^2 ) |z_4|^2.$$ Then the degree property in Theorem \[main\] implies that the sum of squares in must be \[sum0\] |z\_2|\^4 + |z\_2|\^2 |z\_3|\^2 + |z\_3|\^2 |z\_4|\^2, since the remaining square term $|z_2|^2 |z_4|^2$ has the same degree in $(z_4,\z_4)$ as $|z_3|^2 |z_4|^2$ but lower degree in $(z_3,\z_3)$. That last restriction again determines uniquely the terms in .
Corollary \[catlincase\] allows us to estimate how many inverse weights can arise as multitypes ${\ensuremath{\mathfrak M}}=(m_1, \dots,m_n)$ at $0$ of a pseudoconvex hypersurface $M\subset{\mathbb{C}}^n$ of finite $1$-type $m.$ We first recall that by part (4) of Catlin’s Main Theorem in [@catlinbdry], $m_n \leq m.$ It is always the case that $m_1=1,$ $m_2$ is an integer, and $2 \leq m_2 \leq \cdots \leq m_n.$
Since $m$ is rational in general, consider its floor (or integral part) $\lfloor m \rfloor$. By equation , $m_2 = 2k_{22} \leq m,$ so there are at most $\lfloor\frac{m}{2}\rfloor$ values for $$m_2 = 2, 4, \ldots, \left\lfloor\frac{m}{2}\right\rfloor.$$ Once again from equation we obtain \[m23\] + = 1 with $k_{33}\neq 0$, implying $$0 \leq k_{32}<\frac{m}{2}.$$ If $\frac{m}{2} \not\in {\mathbb{Z}},$ then there are $\lfloor\frac{m}{2}\rfloor+1$ choices for $k_{32},$ namely the integers from $0$ to $\lfloor\frac{m}{2}\rfloor.$ If $\frac{m}{2}\in {\mathbb{Z}},$ then $0 \leq k_{32} \leq \frac{m}{2}-1,$ so there are $\frac{m}{2}$ choices for $k_{32}.$ In both cases, we get at most $\lfloor\frac{m}{2}\rfloor+1$ choices for $k_{32}.$ As for $k_{33},$ implies $$0<2k_{33} \leq m_3 \le m,$$ so there are at most $\lfloor\frac{m}{2}\rfloor$ choices for $k_{33}.$ Since the choice of $k_{32}$ and $k_{33}$ determines $m_3,$ we have $\left(\lfloor\frac{m}{2}\rfloor+1\right)\lfloor\frac{m}{2}\rfloor$ choices for $m_3$, without accounting for different equations yielding the same solution. To determine the number of choice for $m_4,$ we use the equation $$\frac{2k_{42}}{m_2} + \frac{2k_{43}}{m_3}+ \frac{2k_{44}}{m_4} = 1.$$ The same analysis gives us $\lfloor\frac{m}{2}\rfloor+1$ choices for each of $k_{42}$ and $k_{43}$ and $\lfloor\frac{m}{2}\rfloor$ choices for $k_{44}$ due to the condition $k_{44}\neq 0.$ We thus have at most $\left(\lfloor\frac{m}{2}\rfloor+1\right)^2\lfloor\frac{m}{2}\rfloor$ choices for $m_4.$ In general, there are $\left(\lfloor\frac{m}{2}\rfloor+1\right)^{j-2}\lfloor\frac{m}{2}\rfloor$ choices for $m_j,$ where $2 \leq j \leq n.$ Altogether, we have obtained $$\left(\left\lfloor\frac{m}{2}\right\rfloor+1\right)^{0+1+ \cdots +(n-2)}\left(\left\lfloor\frac{m}{2}\right\rfloor\right)^{n-1}=\left(\left\lfloor\frac{m}{2}\right\rfloor+1\right)^{\frac{(n-2)(n-1)}{2}}\left(\left\lfloor\frac{m}{2}\right\rfloor\right)^{n-1}$$ possible multitypes at $0$ of a pseudoconvex hypersurface $M\subset{\mathbb{C}}^n$ of finite $1$-type $m.$ This estimate significantly improves the one in [@ra].
Notation
========
For an $n$-tuple $z = (z_1, \ldots, z_n)$, we shall use the short-hand notation $$z_{[k,m]} := (z_k, z_{k+1}, \ldots, z_m),
\quad
1\le k \le m\le n.$$ We use the extended sets of nonnegative rationals and reals $$\1{\mathbb{Q}}_{\ge0} := {\mathbb{Q}}_{\ge0}\cup \{+\infty\},
\quad
\1{\mathbb{R}}_{\ge0} := {\mathbb{R}}_{\ge0}\cup \{+\infty\},$$ and consider real nonnegative $n$-tuples of weights, or simply weights \[reg-wt\] = (1, \_2, …, \_n), 1=\_1> \_2…\_n0, \_i \_[0]{}. Following Catlin’s notation [@catlinbdry], we also consider [*inverse weights*]{} $$\Lambda= (1, \lambda_2, \lambda_3, \ldots, \lambda_n),
\quad
\lambda_i \in \1{\mathbb{Q}}_{\ge0}.$$ For every weight $\mu= (1, \mu_2, \mu_3, \ldots, \mu_n),$ we have its associated inverse weight given by reciprocals $$\Lambda= (1, \lambda_2, \ldots, \lambda_n)
= (1, \mu_2^{-1}, \ldots, \mu_n^{-1})$$ with the convention that $0^{-1} = +\infty$, $(+\infty)^{-1}=0$. Let $$(\alpha | \mu) := \alpha_1 \mu_1 + \alpha_2 \mu_2 + \cdots + \alpha_n \mu_n$$ for $\alpha = (\alpha_1, \alpha_2, \dots, \alpha_n)$ a multi-index. Given a smooth real function $r(z,\bar z)$ defined in a neighborhood of $0$ in ${\mathbb{C}}^n$, a weight $\mu$ as in and nonnegative constant $C\ge0$ we write $$r = O_\mu(C),
\quad \text{ resp. }
r = o_\mu(C),$$ whenever $(\a+\b|\mu)\ge C$, resp.$(\a+\b|\mu)> C$ holds for any nonzero monomial $r_{\a\b}z^\a\bar z^\b$ in the Taylor expansion of $r$ at $0$.
Let $M\subset{\mathbb{C}}^n$ be an oriented smooth real hypersurface defined in a neighborhood of a point $p=0$ by $r=0$ with $dr\ne 0$, such that $r<0$ is the negative side with respect to the orientation. Recall that $M$ is pseudoconvex if and only if the restriction of the complex Hessian of $r$ to the complex tangent space of $M$ is positive semidefinite.
We have the following elementary properties, provided with short proofs for the reader’s convenience.
Let $$ r = -2 \operatorname{Re}z_1 + f(z_{[2,n]}, \z_{[2,n]}),$$ where $f$ is any smooth function. Then the domain given by $\{r<0\}$ is pseudoconvex if and only if \[f\] \_[j,k=2]{}\^n f\_[z\_j\_k]{} a\_j |a\_k 0 for all $(a_2, \ldots, a_n)\in {\mathbb{C}}^{n-1}$.
The Levi form of the boundary $M := \{r=0\}$ is given by the restriction of the complex Hessian of $r$, given by the left-hand side of , to the complex tangent subbundle of $M$. Since the latter projects surjectively to $\{0\}\times {\mathbb{C}}^{n-1}$, the Levi form is positive semidefinite if and only if holds, proving the statement.
Given a weight $\mu$ as in , let $$ r = -2 \operatorname{Re}z_1 + p(z_{[2,n]}, \z_{[2,n]}) + o_\mu(1),$$ where $p$ is a weighted homogeneous polynomial of weight $1$. Assume that the domain given by $\{r<0\}$ is pseudoconvex. Then the model domain given by $\{r_0<0\}$, where $$r_0 := -2 \operatorname{Re}z_1 + p(z_{[2,n]}, \z_{[2,n]}),$$ is also pseudoconvex. \[pscmodel\]
The proof is obtained by a simple weighted scaling argument. Consider the weighted dilation $$T_t(z_1, \ldots, z_n):= (z_1, t^{\mu_2} z_2, \ldots, t^{\mu_n} z_n).$$ Then $r_0$ is invariant under composition with $T_t$, whereas for any function $f(z,\z)=o_\mu(1)$, the rescaled function $f(T_t(z), \1{T_t(z)})$ converges to $0$ uniformly on compacta as $t\to 0$. The statement follows from the continuity of the complex tangent bundles and the Levi form under the limit $t\to 0$.
Finally, we shall write \[sim\] A\~B whenever there is a nonzero constant $c$ with $A=cB$.
Catlin Multitype and Boundary Systems {#bdrysystems}
=====================================
We devote this section to defining the multitype notion as Catlin introduced in [@catlinbdry] in order to characterize the vanishing order of the defining function in different directions.
In the previous section, we defined completely general weights and inverse weights, but Catlin restricts the inverse weights he considers to only those that could represent the vanishing of the defining function. We restrict the set of weights via two natural definitions:
An inverse weight $\Lambda= (1, \lambda_2, \lambda_3, \ldots, \lambda_n)$ is called [*admissible*]{} if for every $i,$ $1 \leq i \leq n,$ either $\lambda_i = + \infty$ or there exists a set of non-negative integers $a_1, \dots , a_i$ with $a_i >0$ such that $\displaystyle\sum_{j=1}^i {a_j}\lambda_j^{-1}=1.$ Let $\Gamma_n$ be the set of all admissible inverse weights ordered lexicographically.
Consider a smooth domain $\Omega \subset {\mathbb{C}}^n$ with defining function $r.$ An admissible inverse weight $\Lambda=(\lambda_1, \dots, \lambda_n) \in \Gamma_n$ is called [*distinguished*]{} at $z_0\in \d\Omega$ if there exist holomorphic coordinates $(z_1, \dots, z_n)$ about $z_0$ with $z_0$ mapped to the origin such that if $\sum_{i=1}^n \: \frac{\alpha_i+ \bar \beta_i}{\lambda_i}<1,$ then $D^\alpha \bar D^{\bar \beta} r(0)=0,$ where $D^\alpha= \frac{\partial^{|\alpha|}}{\partial z^{\alpha_1}_1 \cdots \partial z^{\alpha_n}_n}$ and $\bar D^{\bar\beta}= \frac{\partial^{|\bar\beta|}}{\partial \bar z^{\bar\beta_1}_1 \cdots \partial \bar z^{\bar\beta_n}_n}.$ Let $\tilde \Gamma_n (z_0)$ be the set of distinguished weights at $z_0.$ \[distinguishedweightdef\]
\[multitypedef\] The multitype ${\ensuremath{\mathfrak M}}(z_0)$ is defined to be lexicographically the smallest admissible weight ${\ensuremath{\mathfrak M}}(x_0) = (m_1, \dots, m_n)$ such that ${\ensuremath{\mathfrak M}}(z_0) \geq \Lambda$ for every admissible distinguished weight $\Lambda \in \tilde \Gamma_n (z_0).$
Definitions \[distinguishedweightdef\] and \[multitypedef\] together prompt the following natural question:
[**Question:**]{} Let the multitype ${\ensuremath{\mathfrak M}}(z_0) = (m_1, \dots, m_n)$ be such that $m_n < + \infty.$ What are the multi-indices $\alpha$ and $\bar\beta$ satisfying $\sum_{i=1}^n \: \frac{\alpha_i+ \bar \beta_i}{\lambda_i}=1$ such that $D^\alpha \bar D^{\bar \beta} r(0)\neq0$ after a holomorphic change of variables mapping $z_0$ to the origin?
Corollary \[catlincase\], which we shall prove, not only gives an answer to this question but also identifies balanced terms in the defining function responsible for the condition $D^\alpha \bar D^{\bar \beta} r(0)\neq0.$
A priori, Definition \[multitypedef\] gives no indication how to compute the multitype ${\ensuremath{\mathfrak M}}(z_0).$ To achieve that, Catlin introduced in [@catlinbdry] the commutator multitype ${\ensuremath{\mathfrak C}}(z_0)$ computed by differentiating the Levi form along certain lists of vector fields arising from a geometric object called a [*boundary system*]{}. He was then able to prove that this commutator multitype ${\ensuremath{\mathfrak C}}(z_0)$ equals the multitype for a pseudoconvex domain.
Recall from [@catlinbdry] that a boundary system is a collection of vector fields and real-valued functions $${\ensuremath{\mathfrak B}}_\nu (z_0)= \{r_1, r_{p+2}, \dots, r_\nu; L_2, \dots, L_\nu\}$$ for some $\nu \leq n$. The first function in the boundary system is $r_1=r,$ the defining function. Let $p$ be the rank of the Levi form of $b \Omega$ at $z_0.$ Since $r=0$ defines a manifold, we can choose the vector field $L_1$ such that $L_1(r)=1.$ Recall from the beginning of this section that the multitype seeks to capture the vanishing order of the defining function in different directions. From the information we have so far, if ${\ensuremath{\mathfrak C}}(z_0)= (1, c_2, c_3, \ldots, c_n),$ then the first entry comes from the condition $L_1(r)=1,$ and subsequently, $c_2 = \dots = c_{p+1}=2,$ an entry of $2$ for every non-zero eigenvalue of the Levi form. We choose vector fields of type $(1,0)$ $L_2, \dots, L_{p+1}$ such that $L_i(r)=\partial r (L_i)\equiv 0$ and the $p \times p$ Hermitian matrix $\partial \bar \partial r (L_i, L_j)(x_0)$ is nonsingular for $2 \leq i,j \leq p+1.$ We have kept Catlin’s notation of round parentheses for the evaluation of forms on vector fields. If $p+1=n,$ our construction is finished; otherwise, we need to make sense of vanishing orders higher than two. Let us denote by $T^{\, (1,0)}_{p+2}$ the bundle composed of $(1,0)$ vector fields $L$ such that $\partial r (L)=0$ and $\partial \bar \partial r (L, \bar L_j)=0$ for $j=2, \dots, p+1.$ For $l \geq 3$ we denote by ${\ensuremath{\mathcal L}}$ a list of vector fields ${\ensuremath{\mathcal L}}= \{L^1, \dots, L^l\}$ and by ${\ensuremath{\mathcal L}}\partial r$ the function $${\ensuremath{\mathcal L}}\partial r (z) = L^1 \cdots L^{l-2} \, \partial r \, ([L^{l-1},L^l]) (z)$$ for $z \in b \Omega.$ We are interested in lists ${\ensuremath{\mathcal L}}$ such that ${\ensuremath{\mathcal L}}\partial r(z_0) \neq 0$ that are chosen in the most natural way possible. For every $j$ such that $p+2 \leq j \leq n,$ we will pick a $(1,0)$ vector field $L_j \in T^{\, (1,0)}_{p+2}$ and a corresponding real-valued function $r_j$ such that $L_j r_j \neq 0$ but $L_j r_i = 0$ whenever $i<j.$ The process is inductive. When $j = p+2,$ the simplest possible list ${\ensuremath{\mathcal L}}$ of $(1,0)$ vectors in $T^{\, (1,0)}_{p+2}$ that can yield ${\ensuremath{\mathcal L}}\partial r(z_0) \neq 0$ consists of a smooth $(1,0)$ vector field $L_{p+2}$ and its conjugate $\bar L_{p+2}.$ If no such list exists, we set $c_{p+2}= \dots =c_n =+\infty,$ and we have finished our construction of the commutator multitype. If such a list exists, however, we choose a list ${\ensuremath{\mathcal L}}_{p+2} = \{ L^1, \dots, L^l \}$ of minimal length $l$ and set $c_{p+2}=l.$ Let ${\ensuremath{\mathcal L}}'_{p+2} = \{ L^2, \dots, L^l \}.$ Set $r_{p+2}= \operatorname{Re}({\ensuremath{\mathcal L}}'_{p+2} \partial r)$ or $r_{p+2}= \operatorname{Im}({\ensuremath{\mathcal L}}'_{p+2} \partial r)$ so that the condition $L_{p+2} r_{p+2} \neq 0$ holds. Define $S_{p+2}=\{L_{p+2}, \bar L_{p+2}\}.$
Now assume that for some integer $j-1$ with $p+2 \leq j-1<n$ we have already constructed finite positive numbers $c_1, \dots, c_{j-1}$ as well as real-valued functions $r_1,$ $r_{p+2}, \dots, r_{j-1};$ linearly independent smooth $(1,0)$ vector fields $L_2, \dots, L_{j-1};$ and lists ${\ensuremath{\mathcal L}}_{p+2}, \dots, {\ensuremath{\mathcal L}}_{j-1}$ such that the following properties hold:
1. ${\ensuremath{\mathcal L}}_i \partial r(z_0) \neq 0$ for every $i,$ $p+2 \leq i \leq j-1;$
2. If ${\ensuremath{\mathcal L}}_i = \{ L^1, \dots, L^l \},$ then ${\ensuremath{\mathcal L}}'_i = \{ L^2, \dots, L^l \}$ and $r_i= \operatorname{Re}({\ensuremath{\mathcal L}}'_i \partial r)$ or $r_i= \operatorname{Im}({\ensuremath{\mathcal L}}'_i \partial r)$ in order that the condition $L_i r_i \neq 0$ holds;
3. $L_i r_k =0$ for $p+2 \leq k < i \leq j-1;$
4. Each of the lists ${\ensuremath{\mathcal L}}_i = \{ L^1, \dots, L^l \}$ is
(a) i-[*admissible*]{} (in Catlin’s terminology) meaning $L^1 \in S_i =\{L_i, \bar L_i\}$ and if $l_k$ is the number of times a vector from $S_k$ occurs in the list, then $$\displaystyle \sum_{k=p+2}^{i-1} \frac{l_k}{c_k} <1$$
and
(b) [*ordered*]{} meaning $L^k \in S_{\alpha_k}$ for every $1 \leq k \leq l$ and $\alpha_1 \geq \alpha_2 \geq \dots \geq \alpha_l;$
5. If $l^i_k$ equals the number of times $L_k$ and $\bar L_k$ occur in ${\ensuremath{\mathcal L}}_i,$ then $l^i_k=0$ whenever $k>i$ and $\displaystyle \sum_{k=p+2}^{j-1} \frac{l^i_k}{c_k} = 1;$
6. All lists ${\ensuremath{\mathcal L}}_i$ are of minimal length, namely if ${\ensuremath{\mathcal L}}= \{ L^1, \dots, L^l \}$ is any ordered list, $l_k$ equals the number of times $L_k$ and $\bar L_k$ occur in ${\ensuremath{\mathcal L}},$ and $\displaystyle \sum_{k=p+2}^{j-1} \frac{l_k}{c_k} <1,$ then ${\ensuremath{\mathcal L}}\partial r(z_0) = 0.$
We will show that we can chose a positive rational number $c_j$ such that properties (1)-(6) are fulfilled with $j$ replacing $j-1.$ Let $T^{\, (1,0)}_j$ denote the set of $(1,0)$ smooth vector fields $L$ such that $\partial \bar \partial r (L, \bar L_i)=0$ for $i=2, \dots, p+1$ and $L(r_k)=0$ for $k=1,p+2,\dots,j-1.$ For any smooth vector field $L_j \in T^{\, (1,0)}_j,$ we consider all ordered j-admissible lists ${\ensuremath{\mathcal L}}.$ If for all such lists, ${\ensuremath{\mathcal L}}\partial r(z_0) = 0,$ then we set $c_j = \dots = c_n = + \infty;$ otherwise, there exists at least one such list ${\ensuremath{\mathcal L}}$ for which ${\ensuremath{\mathcal L}}\partial r(z_0) \neq 0.$ We choose one of minimal length and denote it ${\ensuremath{\mathcal L}}_j.$ The vector field $L_j$ used in its construction gets added to the collection $L_2, \dots, L_{j-1}.$ For $p+2\leq k \leq j$ let $l_k$ be the number of times a vector from $S_k=\{L_k, \bar L_k\}$ occurs in the list ${\ensuremath{\mathcal L}}_j,$ and let $c({\ensuremath{\mathcal L}})$ denote the solution to the equation $$\sum_{k=p+2}^{j-1} \frac{l_k}{c_k}+\frac{l_j}{c({\ensuremath{\mathcal L}})}=1.$$ Since ${\ensuremath{\mathcal L}}_j$ is j-admissible, $c({\ensuremath{\mathcal L}})\in {\mathbb{Q}}^+.$ Let $c_j = c({\ensuremath{\mathcal L}})$ and let $r_j= Re ({\ensuremath{\mathcal L}}'_j \partial r)$ or $r_j= Im ({\ensuremath{\mathcal L}}'_j \partial r)$ so that the condition $L_j r_j \neq 0$ holds. All properties (1)-(6) are thus fulfilled. We continue this process until we have generated ${\ensuremath{\mathfrak C}}(z_0)= (1, c_2, c_3, \ldots, c_n),$ the [*commutator multitype*]{} at $z_0.$ Let $\nu \leq n$ be the highest index for which the entry $c_\nu$ is finite. The collection $${\ensuremath{\mathfrak B}}_\nu (x_0) = \{r_1, r_{p+2}, \dots, r_\nu; L_2, \dots, L_\nu\}$$ of functions and vector fields that we have generated in the process of computing ${\ensuremath{\mathfrak C}}(z_0)$ is called a [*boundary system*]{} of rank $p$ and codimension $n-\nu.$
Estimates for Non-negative Homogeneous Polynomials {#estimates}
==================================================
One Variable Case
-----------------
Let $P(z,\bar z)\ge0$ be a real homogeneous polynomial of degree $2m$ in ${\mathbb{C}}$, $$P(z, \bar z) =
\sum_{k=-m}^m C_k z^{m+k} \bar z^{m-k}.$$ Then $C_0 \ge 0$ and $ |C_{k}| \le C_0$ for all $k$. Furthermore, if $P(z,\bar z) \not\equiv 0$, then $C_0 > 0.$ \[estdegree1\]
Considering the values of $z$ on $|z| = 1$, we observe that $\bar z = 1/z$, and $$P = \sum_{k=-m}^m C_k z^{2k} \ge 0, \quad |z| = 1.$$ We parametrize $|z|=1$ as $\gamma (\theta) = e^{i \theta}$ on $[-\pi, \pi]$ and observe that $$0 \le \frac{1}{2\pi } \int_{-\pi}^\pi P(\theta) d \theta = C_0,$$ whence $C_0 \ge 0$. Note that if $P > 0$ for some $z$ in a neighbourhood of $0,$ then $\displaystyle \int_{-\pi}^\pi P(\theta) d \theta >0$ so $C_0 > 0.$
Since for any $k \neq 0$ and $z$ with $|z|=1$, $$0 \le (2\operatorname{Re}z^k)^2 = (z^k + \bar z^k)^2 = (z^k + 1/z^k)^2 = z^{2k} + z^{-2k} + 2$$ and $$0 \ge (2i\operatorname{Im}z^k)^2 = (z^k - \bar z^k)^2 = (z^k - 1/z^k)^2 = z^{2k} + z^{-2k} - 2,$$ we have $$0 \le \frac{1}{2\pi } \int_{-\pi}^\pi (e^{i\theta k} + e^{-i\theta k})^2 P(\theta) d \theta = C_{-k} + C_{k} + 2C_{0}$$ and $$0 \ge \frac{1}{2\pi }\int_{-\pi}^\pi (e^{i\theta k} - e^{-i\theta k})^2 P(\theta) d \theta = C_{-k} + C_{k} - 2C_{0},$$ which immediately gives $ |C_{-k} + C_{k}| \le 2C_{0} $. Since $C_{k} = \bar C_{-k}$, we have $$\label{eq:real_part_is_smaller}
|\operatorname{Re}C_{-k} | = |\operatorname{Re}C_{k} | \le C_{0}.$$
Furthermore, by rotating $z$ we may assume $C_k\in{\mathbb{R}}$, and hence the above inequality yields $|C_k|\le C_0$ as desired.
Inspired by these estimates, we seek to divide the terms of a non-negative homogenous polynomial $P(z, \bar z)$ in ${\mathbb{C}}^n$ into the terms that control others and those that are controlled.
Let a universal constant $M>0$ be given, and let $P(z,\bar z)$ be a non-negative homogeneous polynomial of degree $2m$ in ${\mathbb{C}}^n.$ If $$P(z, \bar z) = \sum_{|\alpha|+|\bar\beta|=2m} C_{\alpha \bar\beta} \, z^\alpha {\bar z}^{\bar \beta},$$ then a coefficient $C_{\alpha \bar\beta}$ is called $M$-dominant if $|C_{\alpha' \bar\beta'}| \le M\, |C_{\alpha \bar\beta}|$ for all $\alpha', \bar\beta'$ such that $|\alpha'|+|\bar\beta'|=2m.$
Newton Polygon Lemma
--------------------
\[newton\] Let $P(x,y)$ be a non-negative homogeneous polynomial of degree $2m$ in ${\mathbb{R}}^p \times {\mathbb{R}}^q$ for $x \in {\mathbb{R}}^p$ and $y \in {\mathbb{R}}^q.$ If $$P(x, y) = \sum_{p+q=2m} P_{pq},$$ where $P_{pq}(\lambda x, \mu y)= \lambda^p \mu^q P(x,y),$ then $P_{p_0 q_0}\geq 0$ if either $p_0 = \max p$ or $q_0 = \max q.$ \[newtonpolygon\]
[**Proof:**]{} We use a scaling argument. $P(x,y) \geq 0$ for all $(x,y) \in {\mathbb{R}}^p \times {\mathbb{R}}^q$ implies for any $t \in {\mathbb{R}},$ $t \neq 0,$ we have that $P(t x, t^{-1} y) \geq 0.$ Letting $t \to \infty$ shows $P_{p_0 q_0}\geq 0$ when $p_0 = \max p.$ Letting $t \to 0$ shows $P_{p_0 q_0}\geq 0$ when $q_0 = \max q.$
The statement of Lemma \[newtonpolygon\] also holds for weighted homogenous polynomials.
Normal forms in the pseudoconvex case {#normalforms}
=====================================
Setup
-----
Consider a smooth real hypersurface $M$ passing through $0$, defined near $0$ by $r=0$, where $$r = -2\operatorname{Re}z_1 + p(z_{[2,n]},\bar z_{[2,n]}) +
o_\mu(1),$$ $p$ is a weighted homogeneous polynomial of weight $1$, and where the components $z_1,\ldots, z_n$ and their conjugates are assigned the corresponding weights $1, \mu_2, \mu_3, \ldots, \mu_n$. In other words, we can write $$\displaystyle p(z_{[2,n]},\bar z_{[2,n]}) =
\sum_{(\alpha + \beta | \mu) =1}
C_{\alpha \beta} z^\alpha \bar z^{ \beta}.$$ Together with $M,$ we consider its model hypersurface $M_0$ defined by the weighted homogeneous part $$r_0=0,
\quad r_0: = -2\operatorname{Re}z_1 + p(z_{[2,n]}, \bar z_{[2,n]}).$$ Note that by elementary scaling argument, if $M$ is pseudoconvex, then so is its model $ M_0$ as seen in Lemma \[pscmodel\].
As customary, we shall assume that $p(z_{[2,n]}, \bar z_{[2,n]})$ does not contain any pure (harmonic) monomials of the form $z^\alpha$ or $\bar z^\beta$. Otherwise, they can be always eliminated by a holomorphic transformation $$(z_1, z_{[2,n]}) \mapsto (z_1 + h(z_{[2,n]}), z_{[2,n]}).$$
In what follows, homogeneity will be gauged with respect to a regular weight $\mu$
First Step
----------
\[5.1\] Let $M_0\subset {\mathbb{C}}^n$, $n\ge 2$, be a pseudoconvex model hypersurface defined by $$r_0 =0, \quad r_0= -2\operatorname{Re}z_1 + p(z_{[2,n]}, \bar z_{[2,n]}),$$ where $p$ is a weighted homogeneous polynomial of weight $1$ with respect to some weights $$\mu= (\mu_1, \mu_2, \mu_3, \ldots, \mu_n),
\quad
1=\mu_1>\mu_2\ge \mu_3 \ge \ldots \ge \mu_n \ge0.$$ Let $s$ be such that $$\mu_2 = \ldots = \mu_s > \mu_{s+1},$$ and assume that \[p2r\] p\_[\[2,s\]]{}(z\_[\[2,s\]]{}, |z\_[\[2,s\]]{}) := p(z\_[\[2,s\]]{},0, |z\_[\[2,s\]]{}, 0) 0.
Then after a unitary change of the variables $(z_2, \dots, z_s)$, we have \[p2-def\] p\_2(z\_2, |z\_2) := p(z\_2,0, |z\_2, 0) 0, which is a plurisubharmonic homogeneous polynomial of even degree $2 k_{22}$, where $k_{22} := \frac{1}{2 \mu_2}$. Furthermore, $p_2$ has the form \[p2\] p\_2 = \_[j=-k\_[22]{}+1]{}\^[k\_[22]{}-1]{} C\_[2,j]{} z\_2\^[k\_[22]{}+j]{} |z\_2\^[k\_[22]{}-j]{}, where $C_{2,0}>0
$ is $\frac{1}{2 \mu_2}$-dominant among all the coefficients of $p_2$. \[firststeprotation\]
By our assumptions, $p_{[2,s]}$ is a nonzero weighted homogenous polynomial in $(z_{[2,s]}, \bar z_{[2,s]})$. Then clearly after a unitary transformation of $z_{[2,s]}$, we may assume that holds with $p_2$ of the form . Using pseudoconvexity of $M_0$ we obtain $$\partial_{z_2 \bar z_2} p_2 \ge 0,$$ where $p_2$ is defined by . Since $p_2$ does not contain any harmonic terms by our assumption, each monomial in $(z_2, \z_2)$ will contribute to $\partial_{z_2 \bar z_2} p_2$. We compute $$\partial_{z_2 \bar z_2} p_2 = \sum_{j=-k_{22}+1}^{k_{22}-1} C_{2,j}(k_{22}+j)(k_{22}-j) \, z_2^{k_{22}+j-1} \bar z_2^{k_{22}-j-1} .$$ By Lemma \[estdegree1\] applied to $\partial_{z_2 \bar z_2} p_2,$ we conclude $C_{2,0} > 0,$ and
$$|C_{2,j}| \leq
\frac{k_{22}^2}
{k_{22}^2-j^2} C_{2,0}
\le
\frac{k_{22}^2}
{2k_{22}-1} C_{2,0}
<
k_{22} C_{2,0}
,
\quad
j \ne 0.$$
Note that $C_{2,0}$ is the coefficient of the balanced term $C_{2,0} \, z_2^{k_{22}} \bar z_2^{k_{22}}$ in $p_2.$ Since $r_0$ has weight $1$ with respect to the weight $\mu,$ we have $
k_{22} =\frac{1}{ 2\mu_2}
$ and $2 k_{22}$ is the degree of $p_2$.
We shall now apply Lemma \[5.1\] with the weights given by the Catlin multitype. Using the previous lemma, we can normalize the first non-trivial function in the Catlin boundary system:
Let $M$ be a pseudoconvex hypersurface defined by $r=0$, where $$r =
-2\operatorname{Re}z_1 + p(z_{[2,n]}, \bar z_{[2,n]})
+ o_\mu(1),$$ such that $p$ is a weighted homogeneous polynomial of weight $1$ with respect to the Catlin multitype at $0$, $$\Lambda= (1, \lambda_2, \lambda_3, \ldots, \lambda_n).$$ Assume the Levi rank at $0$ is $s_0$ and $2<\lambda_{s_0+2} = \dots = \lambda_{s_0+s_1+1}< \lambda_{s_0+s_1+2}< +\infty.$ Let $${\ensuremath{\mathfrak B}}_n (0) = \{r_1, r_{s_0+2},\dots, r_n; L_2, \dots, L_n\}$$ be any boundary system at $0.$ Then there exists a holomorphic change of coordinates at $0$ preserving the multitype so that the boundary system in the new coordinates $$\tilde {\ensuremath{\mathfrak B}}_n (0) = \{\tilde r_1,\tilde r_{s_0+2}, \dots, \tilde r_n; \tilde L_2, \dots, \tilde L_n\},$$ has the simplest possible first functions $r_{s_0+2}, \dots, r_{s_0+s_1+1}$ given by r\_j = z\_j+o() \[firststepfixr2\] for $s_0+2 \leq j \leq s_0+s_1+1.$
By a Chern-Moser type argument [@CM], we may assume that at $0$ the Levi rank $s_0=0.$ Furthermore, it is easy to see this lemma reduces to proving that at the level of the model hypersurface $r_0=0,$ where $$r_0 =
-2\operatorname{Re}z_1 + p(z_{[2,n]}, \bar z_{[2,n]}),$$ for $s_0+2 \leq j \leq s_0+s_1+1$ we can bring $r_j$ to the form $r_j=\operatorname{Re}z_j$ via a holomorphic polynomial change of the variables. We start by noting that the assumption must hold for the Catlin’s multitype of the model hypersurface, which is the same as that of the original hypersurface. Indeed, otherwise we could increase the inverse weights $$\lambda_2 =\ldots = \lambda_{s_1+1}$$ and decrease the remaining bigger inverse weights, still keeping the total weighted degrees of the terms in $p$ greater or equal $1$, which would contradict that assumption that $(1, \lambda_2, \lambda_3, \ldots, \lambda_n)$ is the Catlin multitype at $0$. For the moment, assume $s_1=1.$
We can apply Lemma \[firststeprotation\] to obtain a decomposition $$r_0 = -2\operatorname{Re}z_1 + p_2(z_2, \bar z_2) + q_2(z_{[2,n]}, \bar z_{[2,n]}),$$ where $p_2$ satisfies and $$q_2(z_2,0,\bar z_2,0) \equiv 0.$$ We would like to make another change of variables that would ensure the function $r_2$ in the boundary system has the simplest possible expression, namely $r_2 = \operatorname{Re}z_2.$ Note that the boundary system contains a function $r_2$ due to the assumption $2<\lambda_2 < +\infty.$ To determine what change of variables needs to be made, we notice that regardless of what form $r_2$ initially assumes, after the change of variables mandated by the application of Lemma \[firststeprotation\], $$\label{rp+2prep}
\frac{1}{(k'-1)! k''!}
\partial^{k' - 1}_{z_2} \partial^{k''}_{\bar z_2} r_0
= C_{2,0} \, z_2 + C_{2,-1} \, \bar z_2
+ T(z_{[3,n]},\bar z_{[3,n]} ),$$ where $k'+k''=2 k_{22}=\lambda_{2}$ and $ T(z_{[3,n]},\bar z_{[3,n]} )$ is the sum of terms coming from differentiation of $q_2$ in the expression for $r_0.$ Note that $T$ cannot depend on $z_2$ due to the weight restriction.
It is beneficial to split $q_2$ as $$q_2 =
\sum_{k+l = 2k_{22}-1} z_2^k \bar z_2^l T_{kl}(z_{[3,n]}, \bar z_{[3,n]})\\
+
S_2(z_{[2,n]}, \bar z_{[2,n]}),
$$ where all monomials in $S_2$ have their total degree in $(z_2, \bar z_2)$ less than $2k_{22}-1$.
By the pseudoconvexity of $M_0$, we see that for $j \ge 3$, $$\partial_{z_j \bar z_j} r_0 =
\sum_{k+l = 2k_{22}-1} z_2^k \bar z_2^l
\partial_{z_j \bar z_j}T_{kl}(z_{[3,n]}, \bar z_{[3,n]})
+ \partial_{z_j \bar z_j} S_2 \ge 0.$$ Choosing $|z_2| >> |z_j|$ for all $j \ge 3$, we conclude $$\sum_{k+l = 2k_{22}-1} z_2^k \bar z_2^l
\partial_{z_j \bar z_j}T_{kl}(z_{[3,n]}, \bar z_{[3,n]})
\ge 0,
\quad
j \ge 3.$$ Since the left-hand side changes the sign when $z_2$ does, we obtain $$\partial_{z_j \bar z_j} T_{kl} = 0, \ j \ge 3.$$ Since the $z_j$-direction can be rotated arbitrarily by a change of coordinates, by a similar argument, the whole Levi form of each $T_{kl}$ must vanish identically, which means that each $T_{kl}$ is harmonic, i.e. a sum of holomorphic and antiholomorphic functions in $z_{[3,n]}$.
In particular, after the change of variables mandated by the application of Lemma \[firststeprotation\], in the boundary system $$L_2:= \d_{z_2} + p_{z_2} \d_{z_1},
\quad
{\ensuremath{\mathcal L}}= \{\bar L_2, L_2, \bar L_2, \ldots, L_2, \bar L_2\},$$ where $L_2$ and $\bar L_2$ appear $k'-1$ and $k''$ times respectively. Note that if $\lambda_2 = \lambda_3 = \dots = \lambda_{s_1+1},$ another linear change of variables to gather all terms coming from $z_3, \dots, z_{s_1+1}$ into $z_2$ may be required in order to achieve the expression for $L_2$ claimed above.
Since $p$ does not depend on $z_1$, we can ignore differentiations in that direction in the expressions of ${\ensuremath{\mathcal L}}\d r$. Hence we can compute $r_2$ up to a constant as $$r_2 = {\ensuremath{\mathcal L}}\d r \sim \operatorname{Re}\partial^{k' - 1}_{z_2} \partial^{k''}_{\bar z_2} r_0
\sim \operatorname{Re}( C_{2,0} \, z_2 + C_{2,-1} \, \bar z_2 + T),$$ where $$T = \phi_2(z_{[3,n]}) + \overline{\psi_2(z_{[3,n]})}$$ with $\phi_2$ and $\psi_2$ holomorphic. By a rotation in $z_2$, we can assume $C_{2,-1}$ real and nonnegative. Since $C_{2,0}>0$, we can consider the change of variables, $$\label{eq:change_of_vars_z2}
z_2 = \frac{1}{(C_{2,0} + \bar C_{2,-1})} z'_2 - (\phi_2 + \psi_2)$$ leading to $r_2 \sim \operatorname{Re}z_2$ and finally by scaling to $$r_2 = \operatorname{Re}z_2.$$ Now let $s_1>1.$ For every $j$ such that $3 \leq j \leq s_1+1,$ we carry out the same procedure noting that due to weight considerations the change of variables required to transform $r_j$ into $r_j = \operatorname{Re}z_j$ will not affect variables $z_1, \dots, z_{j-1}.$
A similar argument cannot be carried out to normalize the next boundary system function $r_{s_0+s_1+2}$ if $\lambda_{s_0+s_1+1} < \lambda_{s_0+s_1+2}.$ A counterexample is given at the end of the paper in Section \[r3counterexample\].
[**Proof of Theorem \[flatnessprop\]:**]{} At the level of the model hypersurface, $r_0=0,$ with $$r_0 =
-2\operatorname{Re}z_1 + p(z_{[2,n]}, \bar z_{[2,n]}),$$ Lemma \[firststepfixr2\] shows that after a change of variables $\tilde r_j = \operatorname{Re}z_j$ for $s_0+2 \leq j \leq s_0+ s_1+1.$ Therefore, $\tilde L_j=\frac{\partial}{\partial z_j}$ and $
\tilde L_{s_0+ s_1+2} =\frac{\partial}{\partial z_{s_0+ s_1+1}}+o\left(-\frac{1}{\lambda_{s_0+ s_1+2}}\right)$ for the model hypersurface as $\tilde L_{s_0+ s_1+2}$ is chosen so that $\tilde L_{s_0+ s_1+2} \tilde r_{s_0+ s_1+1} = 0$, and by Catlin own normalization result in [@catlinbdry], Proposition 5.3, $$r_0=2\operatorname{Re}z_1+\sum_{j=2}^{s_0+1} |z_j|^2+f_1(z_{s_0+2}, \dots, z_n)+ 2 \operatorname{Re}\left(\sum_{j=2}^{s_0+1} z_j f_j(z_{s_0+2}, \dots, z_n)\right).$$ The proposition follows.
[**Proof of Corollary \[C3case\]:**]{} We apply Proposition \[flatnessprop\] with $s_0=0$ and $n=3.$
Second Step
-----------
Let $M_0$ be a pseudoconvex hypersurface with polynomial defining function $$r_0 = -2\operatorname{Re}z_1 + p(z_{[2,n]}, \bar z_{[2,n]}).$$ such that $p$ is a weighted homogeneous polynomial in $z_2, \dots, z_n$ of total weight $1$ with respect to the weights $$\mu= (1, \mu_2, \mu_3, \ldots, \mu_n).$$ Let $s$ be such that $$\mu_3 = \ldots = \mu_s > \mu_{s+1},$$ and assume that $$\label{eq:nondegeneracy_z3}
p(z_{[2,s]},0, \bar z_{[2,s]},0) \not\equiv p_2(z_2, \bar z_2),$$ where $p_2$ is given by .
Then after a unitary change of variables $z_3, \dots, z_s$, the polynomial $p$ admits a decomposition $$\label{eq:decomposition_z3}
p=p_2(z_2, \bar z_2) + p_3(z_{[2,3]}, \bar z_{[2,3]}) + q_3(z_{[2,n]}, \bar z_{[2,n]}),$$ where $p_3$ is of degrees $2k_{23}$ and $2k_{33}$ in $(z_2,\bar z_2)$ and $(z_3, \bar z_3)$ respectively, $q_3$ has only terms of degree less than $2k_{33}$ in $(z_3, \bar z_3)$, and $p_3$ contains a non-zero term $$C |z_2|^{2k_{23}} |z_3|^{2k_{33}} \ge 0,
\quad k_{33}>0.$$
\[secondsteprotation\]
[**Proof:**]{} Just like at the beginning of the proof of Lemma \[firststepfixr2\], we write $r_0$ as $$r_0 = -2\operatorname{Re}z_1 + p_2(z_2, \bar z_2) + q_2(z_{[2,n]}, \bar z_{[2,n]} ).$$ After a possible unitary change of variables in $z_{[3,s]}$, we can assume that $$q_2(z_{[2,3]},0, \bar z_{[2,3]},0)\not\equiv 0.$$ Using Lemma \[newtonpolygon\] for guidance, we next identify the non-zero terms in $q_2$ of the highest (total) degree $d_3$ in $(z_3, \bar z_3)$ and denote their sum by $p_3(z_{[2,3]}, \bar z_{[2,3]})$. Then $p$ can be decomposed as $$p = p_2(z_2, \bar z_2)
+
p_3(z_{[2,3]}, \bar z_{[2,3]})
+ q_3(z_{[2,n]}, \bar z_{[2,n]} ),$$ where all monomials in $q_3(z_{[2,3]}, 0, \bar z_{[2,3]}, 0)$ have degree less than $d_3$ in $(z_3, \bar z_3)$.
We shall consider the inequality \[diag\] (\_[z\_2]{} + t\_[z\_3]{}) (\_[|z\_2]{} + t\_[|z\_3]{}) (p\_2(z\_2, |z\_2) + p\_3(z\_[\[2,3\]]{}, |z\_[\[2,3\]]{}) + q\_3(z\_[\[2,3\]]{},0, |z\_[\[2,3\]]{},0)) 0, that follows from the pseudoconvexity of $M_0$, where $t$ is an arbitrary parameter. Identifying terms of the highest degree in $(z_3, \bar z_3)$ we obtain $$\d_{z_2\bar z_2} p_3 \ge 0.$$
We first assume that \[z22\] \_[z\_2|z\_2]{} p\_3 0, i.e. all terms in $p_3$ are harmonic in $z_2$. Then again identifying terms of the highest degree in $(z_3, \bar z_3)$ under this assumption, we obtain $$2t \operatorname{Re}\d_{z_2\bar z_3} p_3 + \d_{z_2\bar z_2} \tilde q_3 \ge 0,$$ where $\tilde q_3$ is the sum of certain terms from $q_3$. Since $t$ is arbitrary, we must have $$\d_{z_2\bar z_3} p_3 \equiv 0.$$ Since $p_3$ has no harmonic terms and all terms are harmonic in $z_2$, the only possibility remaining is that $p_3$ is independent of $z_2$. But then, since $p_3$ is nonzero and has no harmonic terms, we must have \[d33\] \_[z\_3|z\_3]{} p\_3 0.
On the other hand, if does not hold, we obtain the polynomial $\d_{z_2\bar z_2} p_3\ge 0$, which for any generic fixed $z_2$, is non-constant and homogeneous in $(z_3,\bar z_3)$, which again implies .
Thus in all cases, we must have . Applying Lemma \[estdegree1\] to $\d_{z_3\bar z_3} p_3$ for fixed $z_2$ (when it does not identically vanish), we conclude that the degree $d_3$ in $(z_3, \bar z_3)$ is even, $d_3=2k_{33}$, and $\d_{z_3\bar z_3} p_3$ contains nonzero terms of the form $$\tilde p(z_2,\bar z_2)z_3^{k_{33}}\bar z_3^{k_{33}} \ge 0.$$ Applying again Lemma \[estdegree1\], this time to $\tilde p$, we conclude that $p_3$ contains a nonzero term $$C z_2^{k_{23}} z_3^{k_{33}} \bar z_2^{k_{23}} \bar z_3^{k_{33}} \ge 0,$$ as desired.
Inductive Step
--------------
The general inductive step will be obtained from the following result.
Let $M_0$ be a pseudoconvex (model) hypersurface through $0$ with polynomial defining function $$r = -2\operatorname{Re}z_1 + p(z_{[2,n]}, \bar z_{[2,n]}).$$ such that $p$ is a weighted homogeneous polynomial in $z_2, \dots, z_n$ of weight $1$ with respect to the weights $$\mu= (1, \mu_2, \ldots, \mu_n),
\quad
1>\mu_2\ge\ldots\ge \mu_n.$$ Assume that we have already shown that $$\begin{aligned}
p(z_{[2,n]}, \bar z_{[2,n]}) &=
p_2(z_2, \bar z_2)
+
p_3(z_{[2,3]}, \bar z_{[2,3]})
+
\cdots
\\
&
+
p_{m-1}(z_{[2,m-1]},\bar z_{[2,m-1]})
+ q_{m-1}(z_{[2,n]}, \bar z_{[2,n]}).
\end{aligned}$$ where $$q_{m-1}(z_{[2,m-1]},0,\bar z_{[2,m-1]},0) \equiv 0.
$$ Let $s$ be such that $$\mu_m = \ldots = \mu_s > \mu_{s+1},$$ and assume that $$\label{eq:nondegeneracy_z_n}
q_{m-1}(0,z_{[m,s]},0, 0,\bar z_{[m,s]},0) \not\equiv 0.
$$
Then after a unitary change of the variables $(z_m, \dots, z_s)$, $q_{m-1}$ admits the decomposition $$\label{eq:decomposition_z_n}
q_{m-1} = p_m(z_{[2,m]}, \bar z_{[2,m]}) + q_m(z_{[2,n]}, \bar z_{[2,n]}),$$ where $p_m$ is a homogeneous polynomial of degree $2k_{mm}>0$ in $(z_m, \bar z_m)$ whose expansion contains a term $$C |z_2|^{2 k_{2m}} |z_3|^{2 k_{3m}} \cdots |z_m|^{2 k_{mm}},
\quad C > 0,
$$ and $q_m$ has only terms of degree less than $2k_{mm}$ in $(z_m, \bar z_m)$. \[inductivesteprotation\]
[**Proof:**]{} After a possible unitary change of variables in $z_{[m,s]}$, we can assume that $$q_{m-1}(0,z_m,0, 0,\bar z_m,0)\not\equiv 0.$$ We next identify the non-zero terms in $q_{m-1}$ of the highest degree $d_m>0$ in $(z_m, \bar z_m)$ and denote their sum by $p_m(z_{[2,m]}, \bar z_{[2,m]})$. We thus write $$q_{m-1} = p_m(z_{[2,m]}, \bar z_{[2,m]}) + q_m(z_{[2,n]}, \bar z_{[2,n]}),
$$ where all monomials in $q_{m}(z_{[2,m]},0, \bar z_{[2,m]},0)$ have degree less than $d_m$ in $(z_m, \bar z_m)$ by construction.
We shall now use the pseudoconvexity assumption on $M_0.$ For any $j,$ $2\leq j<m$, and any arbitrary real parameter $t,$ consider \[diagm\]
(\_[z\_j]{} + t\_[z\_m]{}) (\_[|z\_j]{} + t\_[|z\_m]{}) & ( p\_2(z\_2, |z\_2) + p\_3(z\_[\[2,3\]]{}, |z\_[\[2,3\]]{}) +\
&+ p\_[m-1]{}(z\_[\[2,m-1\]]{},|z\_[\[2,m-1\]]{}) + p\_[m]{}(z\_[\[2,m\]]{},|z\_[\[2,m\]]{})\
& + q\_[m]{}(z\_[\[2,n\]]{}, |z\_[\[2,n\]]{}) ) 0.
Identifying the highest degree terms in $(z_m, \bar z_m),$ we obtain that \[pos\] \_[z\_j|z\_j]{} p\_m 0, j<m.
We first assume $\partial_{z_j\bar z_j} p_m \equiv 0$ for all $j<m$, i.e. all terms of $p_m$ are harmonic in $z_j$. Then looking at the highest degree terms in $(z_m, \bar z_m)$ in yields $$2t \operatorname{Re}\d_{z_j\bar z_m} p_m + \d_{z_j\bar z_j} \tilde q_m \ge 0,$$ where $\tilde q_m$ consists of the sum of the terms of degree $d_m-1$ in $(z_m, \bar z_m)$ from $p_2+\cdots+ p_m + q_m$. Given that $t$ is arbitrary, we conclude $$\d_{z_j\bar z_m} p_m \equiv 0.$$ Note that $p_m$ contains no harmonic terms and by our assumption, all terms of $p_m$ are harmonic in $z_j$ for all $j<m$. Hence any nonzero term of $p_m$ must have both $z_m$ and $\bar z_m$, that is we must have that \[mmDpm\] \_[z\_m|z\_m]{} p\_m 0.
Now, on the contrary, assume that $\partial_{z_j\bar z_j} p_m \not\equiv 0$ for some $j<m$. For any generic fixed $z_2, \dots, z_{m-1}$, the polynomial $\partial_{z_j\bar z_j} p_m$ is non-constant, non-negative by , and homogeneous in $(z_m, \bar z_m)$ of degree $d_m>0$. Clearly, must hold in this case as well. Therefore, regardless of the case, holds.
We claim that by Lemma \[estdegree1\] inductively applied to $\partial_{z_m\bar z_m} p_m,$ the expansion of the polynomial $\partial_{z_m\bar z_m} p_m$ contains a term \[c-term\] C |z\_2|\^[2 \_2]{}|z\_3|\^[2 \_3]{} |z\_m|\^[2 \_m]{}, \_m>0, C> 0. Indeed, first keep $z_2, \dots, z_{m-1}$ fixed and apply Lemma \[estdegree1\] to $\partial_{z_m\bar z_m} p_m,$ a non-constant, non-negative, and homogeneous polynomial in $(z_m, \bar z_m)$ of degree $d_m-2$. We conclude that the sum of the terms in $p_m$ having equal degrees in $z_m$ and $\bar z_m$ is of the form $$P^{m-1}(z_2, \dots, z_{m-1}, \bar z_2, \dots, \bar z_{m-1}) |z_m|^{2 \alpha_m} \geq 0,
\quad
\alpha_m = d_m/2 > 0,$$ where $P^{m-1}$ is a nonzero weighted homogeneous polynomial.
Let $l$ be the highest index among $2, \dots, m-1$ for which $P^{m-1}$ has degree $d_l >0$ in $(z_l, \bar z_l).$ If no such $l$ exists, then $P^{m-1}$ is constant and positive, and we are done; otherwise we extract the sum $\tilde P^{m-1}$ of terms of the top degree $d_l$ in $(z_l, \bar z_l)$, which is not identically zero, and nonnegative in view of Lemma \[newton\], keep $z_2, \dots, z_{l-1}$ fixed and apply Lemma \[estdegree1\] to $\tilde P^{m-1}$ viewed as a homogeneous polynomial in $(z_l, \bar z_l).$ Proceeding inductively, we see that $
p_m$ contains a non-zero term maximizing the multidegree in $(z_2,\z_2), \ldots, (z_m, \z_m)$ in the reversed lexicographic order as claimed.
[**Proof of Theorem \[main\]:**]{} Our Main Theorem is a consequence of Lemmas \[firststeprotation\], \[secondsteprotation\], and \[inductivesteprotation\]. In fact, at each step, either the nonvanishing assumption in the Lemmas holds, and hence we obtain a positive $A_j$ or we can lower the weight $\mu$ lexicographically starting with $\mu_j$. In the latter case, either we regain the nonvanishing assumption for a lower $\mu_j$ and keep applying the Lemmas with lower weights, or no further term is left and the proof is complete.
[**Proof of Corollary \[catlincase\]:**]{} This result follows from Theorem \[main\].
Counterexample to a boundary system normalization {#r3counterexample}
=================================================
We would like to show via an example that the kind of normalization of function $r_2$ in Catlin’s boundary system that we carried out in Theorem \[flatnessprop\] fails for $r_3.$ Let the defining function of the domain be given by $$r_0 = -2\operatorname{Re}z_1 + p(z_{[2,4]}, \bar z_{[2,4]}),$$ where $p$ is a weighted homogeneous polynomial chosen so that the weight of $z_2^2$ equals the weight of $z_3^3.$ For example, the polynomial $$\begin{split}
p(z_{[2,4]}, \bar z_{[2,4]}) &= |z_2|^6
+ |z_2|^2 |z_3|^6
+ |z_2|^4 |z_3|^2 |z_4|^2
+|z_2|^2|z_3|^4 |z_4|^4\\&
+ 2\epsilon \operatorname{Re}( |z_2|^2 z_3^2 \bar z_3^3 |z_4|^2)
+ |z_3|^8|z_4|^2
\end{split}$$
is homogeneous of weight $1$ with respect to
$$\Lambda = \left(1, \frac16, \frac19, \frac1{18}\right).$$ Note that $$f:=\d_{z_2}\d_{\bar z_2} \d_{z_3}^2 \d_{\bar z_3}^3 p
= c_1 z_3 + c_2 |z_4|^2,$$ and hence $r_3=\operatorname{Re}f$ cannot be tranformed into $c\operatorname{Re}z_3$ by any holomorphic coordinate change.
We would like to show that $p$ is plurisubharmonic when $\epsilon$ is small, where $0<\epsilon<1.$ First, we observe that if $z_2=0,$ $z_3=0,$ or $z_4=0,$ the term $2\epsilon \operatorname{Re}( |z_2|^2 z_3^2 \bar z_3^3 |z_4|^2)$ vanishes, and thus $ p$ is a sum of squares, making it automatically plurisubharmonic. Therefore, without the loss of generality, we can assume simultaneously that $z_2\neq0,$ $z_3\neq0,$ and $z_4\neq0.$ As a result, we can compute the Levi form in terms of vectors fields $$X=\sum_{j=2}^4 a_j \, z_j \, \frac{\partial}{\partial z_j}$$ and $\bar X,$ which keep the weight of each term of $p$ unchanged.
Rather than writing out the Levi form in full in terms of $X,\bar X$ as a quadratic form in $a_j,$ we observe that by Cauchy-Schwarz, $$\label{cs1}
\left|2\operatorname{Re}( |z_2|^2 z_3^2 \bar z_3^3 |z_4|^2)\right| \le ( |z_2|^2 |z_3|^6 + |z_2|^2|z_3|^4 |z_4|^4)$$ and $$\label{cs2}
\left|2\operatorname{Re}( |z_2|^2 z_3^2 \bar z_3^3 |z_4|^2)\right| \le ( |z_2|^4 |z_3|^2 |z_4|^2 + |z_3|^8|z_4|^2).$$
For the right-hand side expression in , its kernel is given by the simultaneous vanishing of $\{a_2+3a_3 =0\}$ and $\{a_2+2a_3+2a_4 =0\}.$ For the right-hand side expression in , its kernel is given by the simultaneous vanishing of $\{2a_2+a_3+a_4 =0\}$ and $\{4a_3+a_4 =0\}.$ The intersection of these kernels is just the origin, so for small $\epsilon,$ $p$ is indeed plurisubharmonic. Therefore, $r_0$ defines a pseudoconvex domain.
[KoMZ14]{}
T. Bloom and I. Graham. On “type” conditions for generic real submanifolds of [${\bf
C}^{n}$]{}. , 40(3):217–243, 1977.
H. P. Boas and E. J. Straube. On equality of line type and variety type of real hypersurfaces in ${\mathbb{C}}^n$. [*J. Geom. Anal.*]{} [**2**]{} (1992), no. 2, 95–98.
D. W. Catlin. Global regularity of the $\bar\d$-Neumann problem. Complex analysis of several variables (Madison, Wis., 1982), 39–49, Proc. Sympos. Pure Math., 41, Amer. Math. Soc., Providence, RI, 1984.
D. W. Catlin. Boundary invariants of pseudoconvex domains. , 120(3):529–586, 1984.
D. W. Catlin. Subelliptic estimates for the $\bar\d$-Neumann problem on pseudoconvex domains. [*Ann. of Math. (2)*]{}, 126 (1): 131–191, (1987).
D. W. Catlin and J. P. D’Angelo. Subelliptic estimates. In [*Complex analysis*]{}, Trends Math., pages 75–94. Birkhäuser/Springer Basel AG, Basel, 2010.
S. S. Chern and J. K. Moser. Real hypersurfaces in complex manifolds. , 133:219–271, 1974.
J. P. D’Angelo. Real hypersurfaces, orders of contact, and applications. , 115(3):615–637, 1982.
M. Kolar. The Catlin multitype and biholomorphic equivalence of models. [*Int. Math. Res. Not. IMRN*]{} 2010, no. 18, 3530–3548.
M. Kolar and F. Meylan. Chern-[M]{}oser operators and weighted jet determination problems. In [*Geometric analysis of several complex variables and related topics*]{}, volume 550 of [*Contemp. Math.*]{}, pages 75–88. Amer. Math. Soc., Providence, RI, 2011.
M. Kolar, F. Meylan, and D. Zaitsev. Chern-[M]{}oser operators and polynomial models in [CR]{} geometry. , 263:321–356, 2014.
J. J. Kohn. Subellipticity of the [$\bar \partial $]{}-[N]{}eumann problem on pseudo-convex domains: sufficient conditions. , 142(1-2):79–122, 1979.
S.Y. Kim and D. Zaitsev. Jet vanishing orders and effectivity of [K]{}ohn’s algorithm in dimension $3$. Preprint 2017. https://arxiv.org/abs/1702.06908.
J. D. McNeal. Convex domains of finite type. [*J. Funct. Anal.*]{} 108 (1992), no. 2, 361–373.
A. C. Nicoara. irect [P]{}roof of [T]{}ermination of the [K]{}ohn [A]{}lgorithm in the [R]{}eal-[A]{}nalytic [C]{}ase. https://arxiv.org/abs/1409.0963v1.
Y.-T. Siu. Effective termination of [K]{}ohn’s algorithm for subelliptic multipliers. , 6(4, Special Issue: In honor of Joseph J. Kohn. Part 2):1169–1241, 2010.
Y.-T. Siu. New procedure to generate multipliers in complex [N]{}eumann problem and effective [K]{}ohn algorithm. , 60(6):1101–1128, 2017.
J. Yu. Multitypes of convex domains. , 41(3):837–849, 1992.
D. Zaitsev. A geometric approach to [C]{}atlin’s boundary systems. Preprint 2017. https://arxiv.org/abs/1704.01808.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the $\mathcal{PT} $-symmetry breaking for the scattering problem in a one-dimensional (1D) non-Hermitian tight-binding lattice model with balanced gain and loss distributed on two adjacent sites. In the scattering process the system undergoes a transition from the exact $\mathcal{PT} $-symmetry phase to the phase with spontaneously breaking $\mathcal{PT} $-symmetry as the amplitude of complex potentials increases. Using the S-matrix method, we derive an exact discriminant, which can be used to distinguish different symmetry phases, and analytically determine the exceptional point for the symmetry breaking. In the $\mathcal{PT} $-symmetry breaking region, we also confirm the appearance of the unique feature, i.e., the coherent perfect absorption Laser, in this simple non-Hermitian lattice model. The study of the scattering problem of such a simple model provides an additional way to unveil the physical effect of non-Hermitian $\mathcal{PT} $-symmetric potentials.'
author:
- Baogang Zhu
- Rong Lü
- Shu Chen
date: today
title: '$\mathcal{PT} $ symmetry breaking for the scattering problem in a one-dimensional non-Hermitian lattice model'
---
Introduction
============
It is known that one of the fundamental principles of quantum mechanics is that physical observables must be presented by Hermitian operators in the Hilbert space to guarantee real observables and probability conservation. However, in 1988 Bender and Boettcher found that a large class of non-Hermitian Hamiltonians can exhibit all real eigenvalues if these systems have parity-time ($\mathcal{PT} $) symmetry [@Bender]. The notable feature of these systems is that they undergo a transition from an exact $\mathcal{PT} $-symmetric (unbroken) phase to a spontaneously $\mathcal{PT} $-symmetry breaking phase, which is distinguished by the eigenvalues changing from all real to complex correspondingly. This leads to interesting phenomena in many fields, such as in quantum field theories and mathematical physics [@Bender04], open quantum systems [@Rotter09], Anderson models for disorder systems [@Goldsheid98; @Heinrichs01; @Molinari09]. The non-Hermitian systems have been experimentally realized in optical materials [@Regensburger13], waveguide arrays [@Guo09; @Ruter10], microresonators [@WenJianming; @LanYang14], acoustic sensor [@Romain15; @Zhangxiang2014] and LRC circuits [@Schindler11].
Although whether the $\mathcal{PT} $ symmetric non-Hermitian Hamiltonian can define real quantum systems is not clear, efforts have been made to detect the effects of these non-Hermitian parts both in theoretical and experimental ways. Up to now, most of these studies rely on the basis that the scalar paraxial approximation of Maxwell’s equation shares the same form with Schrödinger equation, in which the axial wave vector plays the role of energy [@ZinLin]. In order to find the original physical meaning of these non-Hermitian systems, a metric-operator theory is developed to map the non-Hermitian Hamiltonian to an equivalent Hermitian one [@Mosta04]. Based on these fundamental issues, the effects of the non-Hermitian terms have been investigated under different boundary conditions. Under the open boundary condition (OBC), the Bethe ansatz solution of a 1D tight-binding chain with two conjugated imaginary potentials at two end sites confirms the non-Hermitian $\mathcal{PT} $-symmetric quantum theory for the discrete systems [@Song09; @Yuce15; @Yogesh10]. Under the periodic boundary condition (PBC), the topological invariance is derived to find the topological properties in the quantum evolution of non-Hermitian models [@LiangShidong13]. The scattering propagation has also been investigated in various models, in which the scattering coefficients calculated by the transfer matrix method show the interesting unidirectional invisibility and self-emission effects [@Mostafa09; @LiGe12; @YDChong11; @YDChong10; @Mostafa13; @Rotter13; @Zhangxiang14; @Longhi2010; @Longhi10; @Ahmed14].
In this work we study the scattering propagation in the 1D tight-binding non-Hermitian lattice model whose balanced gain and loss are distributed on two adjacent sites to ensure the $\mathcal{PT} $-symmetry of the model. While the $\mathcal{PT} $-symmetry breaking in the eigenvalue problem of non-Hermitian lattice model has been extensively studied [@Song09; @Yuce15; @Yogesh10], the $\mathcal{PT} $-symmetry breaking in the lattice scattering problem is not well understood yet. The study of the simple lattice model can illustrate the most fundamental principle in the non-Hermitian scattering process and helps us to understand the more complicated phenomenon in other models. With the help of the transfer matrix $S $ we can analyze the stationary problem in the scattering propagation and define the form of the stationary wavefunction, thus getting the reflection and transmission coefficients. The $S $-matrix theory shows that the eigenvalues of the $S $ matrix remain unimodular in the exact $\mathcal{PT} $-symmetric phase whereas they have reciprocal moduli in the broken $\mathcal{PT} $-symmetric phase [@YDChong11]. The critical point is called the exceptional point (EP), which signifies the breaking of the symmetry. In this paper we give the clear analytical results of eigenvalues of the $S $-matrix and the exceptional point for our studied model. We also derive a discriminant to find the EPs and observe the relationship between the real and complex potentials. Moreover, the significant property - coherent perfect absorption (CPA) Laser - can also be found in this simple model and the overall output coefficient confirms the results to be the typical CPA indeed. This model can be easily implemented on many experiments, such as optical material, waveguide arrays, and acoustic sensor. The paper is organized as follows. In Sec. , the model is presented. In Sec. , we investigate the symmetry-breaking transition of the scattering process with the help of the $S $-matrix theory. In Sec. , we focus on the CPA Laser and the overall output coefficient. A summary is given in Sec. .
Physical model
==============
We consider the simplest 1D $\mathcal{PT} $-symmetric model which consists of two lattice sites embedded with $\mathcal{PT} $-symmetric complex on-site chemical potentials in the infinite tight-binding linear main chain. As schematically shown in Fig. 1, the model can be described by the Hamiltonian: $$\begin{aligned}
H &=&\sum J\hat{\phi} _{n+1}^{\dagger } \hat{\phi}_{n}+h.c. \nonumber\\
&&+(U+i\gamma )\hat{\phi} _{0}^{\dagger } \hat{\phi}_{0}+(U-i\gamma )\hat{\phi} _{1}^{\dagger } \hat{\phi}_{1}\label{Hamiltonian},\end{aligned}$$ where $\hat{\phi}_{n}$ ($\hat{\phi}_{n}^{\dagger}$) denotes the annihilation (creation) operator for annihilating (creating) a mode state $|\phi_{n} \rangle $ at $n$th site. The main chain is an infinite isotropic tight-binding chain with $J $ being the nearest-neighbor hopping strength. The hopping terms provide the fluent channel of the continuum spectrum for the propagation of plane waves with dispersion $\omega = 2J\cos k $. The two defects are added with complex on-site chemical potentials, leading to the resonance at some income states during the propagation (as shown below). The real part of the potential $U $ is set to be $\left\vert U\right\vert < 2J $ to fulfill the resonance condition. For convenience, the hopping amplitude $J$ is set to be the unit of the energy ($J=1$).
{width="8.5cm"}
In general, $\mathcal{P}$ and $\mathcal{T}$ are defined as the space-reflection (parity) operator and the time-reversal operator, whose effects are $\mathcal{P}$: $p\rightarrow -p$, $x\rightarrow -x$ and $\mathcal{T}$: $p\rightarrow -p$, $x\rightarrow x$, $i\rightarrow -i$. A Hamiltonian is said to be $\mathcal{PT} $ symmetric if it obeys the commutation relation $[\mathcal{PT},H]=0 $. In this discrete model the effect of $\mathcal{P} $ operator is $\mathcal{P} \hat{\phi}_{n}\mathcal{P}= \hat{\phi}_{-n+1}$ with the vertical line between site 0 and 1 as the mirror axis, and the effect of $\mathcal{T} $ operator is $\mathcal{T}i \mathcal{T}= -i $. So it’s easy to prove that the Hamiltonian (\[Hamiltonian\]) is invariant under the combined operation $\mathcal{PT} $.
The scattering problem and $S $-matrix properties
=================================================
Using the $S $-matrix method, we now study the transport properties of the non-Hermitian $\mathcal{PT}$-symmetric system and analyze the $\mathcal{PT}$ symmetry breaking during the scattering process. If we expand the wave function of the system as $|\psi \rangle = \sum_{n} \phi_{n}(\tau) |\phi_n \rangle $ with $|\phi_n \rangle=\hat{\phi}_n^{\dagger} |0\rangle$, from the equation $i\partial_{\tau} |\psi \rangle = H |\psi \rangle$, we can derive the coupled-mode equation for expansion coefficients $\phi_{n}(\tau)$: $$\begin{aligned}
i\dot{\phi}_{n} &=&\phi _{n-1}+\phi _{n+1}+(U+i\gamma )\phi_{0}\delta_{n0}+(U-i\gamma )\phi_{1}\delta_{n1},\end{aligned}$$ where the overdot stands for the derivative of time $\tau$. The stationary solution can be expressed as the following form $$\phi _{n}(\tau)=A_{n}e^{-i\omega \tau },$$ and then we obtain the algebraic relationship of the amplitudes on each site: $$\begin{aligned}
\omega A_{n}&=& A_{n-1}+A_{n+1}\nonumber\\
&+&(U+i\gamma )A_{0}\delta_{n0}+(U-i\gamma )A_{1}\delta_{n1}.\end{aligned}$$ The above effective defect equation reveals that the wave propagates along the 1D chain with scattering center at sites 0 and 1 with the localized complex potentials. For the scattering problem, the wave function can be expressed in the form $$A_{n}=\left\{
\begin{array}{c}
F_{f}^{-}e^{ikn}+F_{b}^{-}e^{-ikn},n<0, \\
F_{f}^{+}e^{ikn}+F_{b}^{+}e^{-ikn},n>1,
\end{array}
\right.$$ where $F_{f(b)}^{+(-)}$ stands for the coefficient of the two compositions in the whole wave function, the superscript $- $ stands for the left side of the scattering center whereas $+ $ the right side of the scattering center, and $f $ stands for the forward-going composition whereas $b $ the backward-going composition. $F_{f}^{-} $ is the so-called left-injection coefficient and $F_{b}^{+}$ is the right-injection coefficient. The momentum $k $ is assumed to be positive, and it is related to the incoming wave frequency $\omega $ by the dispersion equation $\omega=2 \cos k$. By substituting the expression $A_n$ into the effective defect equation (2) and applying the $S $-matrix method, we obtain the expression of all the coefficients in the wave function as $$\left(
\begin{array}{c}
F_{b}^{-} \\
F_{f}^{+}%
\end{array}%
\right) =S\left(
\begin{array}{c}
F_{b}^{+} \\
F_{f}^{-}%
\end{array}%
\right) ,$$ with the elements of $S$-matrix as $$\begin{aligned}
S_{11} &=&-2i\sin ke^{-ik}/\Gamma , \notag \\
S_{12} &=&[-U^{2}-\gamma ^{2}+2\gamma \sin k+2U\cos k]/\Gamma , \notag \\
S_{21} &=&[-U^{2}-\gamma ^{2}-2\gamma \sin k+2U\cos k]e^{-2ik}/\Gamma ,
\notag \\
S_{22} &=&-2i\sin ke^{-ik}/\Gamma , \label{Smatrix}\end{aligned}$$and$$\Gamma =(U-e^{-ik})^{2}+\gamma ^{2}-1.$$
From the elements of $S $ matrix one can easily obtain the transport properties of the system. In the case of only injection from the left side, the elements $S_{12}$ and $S_{22}$ are the left reflection and transmission amplitudes defined as $r_{L}=\frac{F_{b}^{-}}{F_{f}^{-}}$, and $t_{L}=\frac{F_{f}^{+}}{F_{f}^{-}}$, respectively. While in the case of only injection from the right side, the matrix elements $S_{11}$ and $S_{21}$ are the right transmission and reflection amplitudes defined as $t_{R}=\frac{F_{b}^{-}}{F_{b}^{+}}$, and $r_{R}=\frac{F_{f}^{+}}{F_{b}^{+}}$, respectively. So the $S $ matrix can be expressed as $$S=\left(
\begin{array}{cc}
t_{R} & r_{L} \\
r_{R} & t_{L}
\end{array}
\right) . \label{SS}$$ With the help of Eq. (\[Smatrix\]), one can show that $t_{L}=t_{R}=t $ for a linear non-magnetic system [@ZinLin; @Mostafa09]. For the $\mathcal{PT} $-symmetric system, one can prove the relation between transmission and reflection amplitudes as $r_{L}r_{R}^{\ast }= 1-\left\vert t\right\vert ^{2} $. Furthermore, defining the reflection and transmission coefficient as $R_{L,R}=\left\vert r_{L,R} \right\vert ^{2}$ and $T=\left\vert t\right\vert ^{2}$, we find that the generalized conservation relation $\sqrt{R_{L}R_{R}}=\left\vert T-1\right\vert $ holds in this $\mathcal{PT} $-symmetric non-Hermitian system, instead of the traditional relation $R+T=1$ for the Hermitian system [@LiGe12; @YDChong11]. The transmission coefficient of this $\mathcal{PT} $-symmetric non-Hermitian model is found to be $$\begin{aligned}
&&T \\
&=&\frac{4-\omega ^{2}}{[U^{2}-U\omega +1]^{2}+(\gamma
^{2}-1)[2U^{2}-2U\omega +\omega ^{2}+\gamma ^{2}-3]}.\end{aligned}$$From the above equation, one can show that in the Hermitian case ($\gamma=0$), the resonant transport occurs when the income wave frequency equals to the potential strength. In this case, the central defects provide a path to guarantee a perfect transmission, i.e., $T=1 $ when $\omega=U $. This resonant state plays a significant role in the $\mathcal{PT} $ symmetry breaking.
According to the $S $-matrix theory, the breaking of $\mathcal{PT} $ symmetry may occur when we tune the strength of the non-Hermitian term in the scattering process. The system undergos a transition from the exact $\mathcal{PT} $ symmetric phase to the spontaneous $\mathcal{PT} $ symmetry breaking phase, which is distinguished by the eigenvalues of the $S $-matrix. In the exact $\mathcal{PT} $ symmetric phase, the eigenvectors of the $S $-matrix is also $\mathcal{PT} $ symmetric, and thus the eigenvalues are unimodular. While in the $\mathcal{PT} $ symmetry breaking phase, the eigenvectors are not $\mathcal{PT} $ symmetric, and they transform into each other under the $\mathcal{PT} $ symmetry operator. Consequently, the two eigenvalues have reciprocal moduli, instead of being unimodular [@YDChong11].
For the model (1), we can derive the analytical form of the eigenvalues of $S $-matrix. From Eq. (\[SS\]), we have $$\begin{aligned}
s_{1,2} &=&t\pm \sqrt{r_{L}r_{R}} . \nonumber\end{aligned}$$ By using the property of the $S$ matrix, i.e., $S^*=S^{-1}$, we have $r_L=- \frac{t}{t^*}r^*_L$ and thus get $$\begin{aligned}
s_{1,2}&=&t(1\pm \sqrt{\frac{T-1}{T}}) \nonumber \\
&=&t(1\pm \sqrt{g(\gamma ,U,\omega )\Delta })\label{s12},\end{aligned}$$ where $g(\gamma ,U,\omega )$ and $\Delta $ are given by $$g(\gamma ,U,\omega )=-[\gamma ^{2}+U^{2}]/[4-\omega ^{2}]$$ and $$\Delta =(\omega -U)^{2}+\frac{\gamma ^{2}(\gamma ^{2}+U^{2}-4)}{\gamma
^{2}+U^{2}}, \label{delta}$$ respectively. While $g(\gamma ,U,\omega )$ is always negative for the permitted frequency $|\omega| \leq 2$, $\Delta $ serves as a discriminant to determine the moduli of the eigenvalues. Eq. (\[s12\]) indicates that the eigenvalues are unimodular for $T<1 $, i.e., $$\left\vert s_{1,2}\right\vert =1 ,$$ whereas for $T>1 $, there exists $$\left\vert s_{1}\right\vert =1/ \left\vert s_{2}\right\vert.$$ So $T=1 $ can be taken as a criterion to signify the phase transition from $\mathcal{PT} $ symmetric phase to $\mathcal{PT} $ symmetry breaking phase, which is called the exceptional point (EP). The phase transition and EP can be also determined by the discriminant $\Delta $. One can easily show that the $\mathcal{PT} $ symmetric phase exists when $\Delta >0 $, while $\mathcal{PT} $ symmetry breaking phase exists when $\Delta <0 $. The criterion $T=1 $ is equivalent to the discriminant $\Delta=0 $.
From Eq. (\[delta\]), we find that under the condition $$\begin{aligned}
\gamma ^{2}+U^{2}-4<0, \label{sss}\end{aligned}$$ there exist two EPs at $$\omega _{\pm }=U\pm \sqrt{-\gamma ^{2}(\gamma ^{2}+U^{2}-4)/(\gamma ^{2}+U^{2}) } ,$$ which is determined by $\Delta=0 $. For $\omega_{-}<\omega<\omega_{+} $, the system is in the $\mathcal{PT} $ symmetry breaking phase, whereas it is in the $\mathcal{PT} $ symmetry phase for $\omega<\omega_{-}$ or $\omega>\omega_{+} $. On the other hand, the system undergos no phase transition when $\gamma ^{2}+U^{2}-4 \geq 0 $ due to $\Delta >0 $, and thus the system is always in the $\mathcal{PT} $ symmetric phase. The positions of EPs $\omega_{\pm} $ also show that the symmetry breaking starts from the resonance state $\omega=U $ and spreads out to both sides when increasing the strength of non-Hermitian potential $\gamma $. This behavior indicates that the symmetry breaking is quite relevant to the resonant transport during the scattering propagation. Actually it is such the case because the $\mathcal{PT} $ symmetry breaking can also be signified by the transmission coefficient $T=1 $, and the system is in the perfect transmission in the resonant condition.
{width="8.5cm"}
To give concrete examples, in Fig. 2 we show the numerical results of eigenvalues of the $S $ matrix as a function of $\omega $ for a certain $U=0.5 $. In Fig. 2(a) the strength of non-Hermitian potential is taken as $\gamma=0.5 $ which fulfils the condition given by Eq. (\[sss\]). The results show that the system undergos two phase transitions at two EPs $\omega_{\pm}= 0.5 \pm \sqrt{1.75} $ as the incoming wave frequency $\omega $ changes. The system is in the $\mathcal{PT} $ symmetry breaking phase when the frequency $\omega_{-}<\omega<\omega_{+} $, which is signified by the reciprocal but non-unit moduli relation of $s_{1}$ and $s_{2}$. As shown in Fig. 2(a), blue dash-dotted and red dashed lines represent $\log _{10}\left\vert s_{1}\right\vert ^{2} $ and $\log _{10}\left\vert s_{2}\right\vert ^{2} $, respectively, and they are definitely zero when $\omega<\omega_{-}$ and $ \omega > \omega_{+} $, indicating that the system is in the $\mathcal{PT} $ symmetric phase. We also show the transmission coefficient $T$ denoted by the green solid line as a function of $\omega$ in the same figure. It is shown that $T<1$ in the $\mathcal{PT} $ symmetric phase, whereas $T>1$ in the broken $\mathcal{PT} $ symmetric phase. $T=1 $ defines the EPs just as $s_{1,2}$ defines them by the bifurcation of their modular. The dotted black lines mark the positions of two EPs in Fig. 2(a). In Fig. 2(b), we show the eigenvalues of the $S $ matrix as a function of $\omega $ for a certain $\gamma=1.95 $. In this case, the system is in the $\mathcal{PT} $ symmetric phase in the whole $\omega $ region and no phase transition occurs, as $\gamma=1.95 $ obeys the condition $\gamma ^{2}+U^{2}-4 >0 $. Correspondingly, $T$ is always below unit in the whole $\omega $ region.
CPA-Laser
=========
In the breaking $\mathcal{PT} $ symmetric phase of the non-Hermitian system, there is a unique absorption feature caused by the $\mathcal{PT} $ symmetry. As the simultaneous gain and loss are balanced embedded, the growth and decay modes are tightly linked each other in a unique manner. When the system permits self lasing at some frequency, it permits coherent perfect absorption (CPA) for a certain amplitude plane wave at the same frequency [@Longhi10]. The unique feature is known as CPA-Laser.
In this discrete model, we rewrite the scattering equation by the transfer matrix $M $, $$\left(
\begin{array}{c}
F_{f}^{+} \\
F_{b}^{+}
\end{array}
\right) = M\left(
\begin{array}{c}
F_{f}^{-} \\
F_{b}^{-}
\end{array}
\right) \label{MM} .\\$$ Supposing that the CPA-Laser occurs at frequency $\omega_{0} =2 \cos k_{0}$, the $\mathcal{PT} $ symmetry requires that the elements of the transfer matrix $M $ obey the relation $M_{11}(k_{0})=M_{22}^{\ast }(k_{0})=0 $ at this frequency $\omega_{0}$. In this discrete model, the elements of $M $ matrix are found to be $$\begin{aligned}
M_{11} &=&[1-(U-e^{ik})^{2}-\gamma ^{2}]e^{-ik}/(-2i\sin k) \\
M_{12} &=&[2U\cos k-U^{2}-\gamma ^{2}-2\gamma \sin k]e^{-ik}/(-2i\sin k) \\
M_{21} &=&[U^{2}+\gamma ^{2}-2U\cos k-2\gamma \sin k]e^{ik}/(-2i\sin k) \\
M_{22} &=&[(U-e^{ik})^{2}+\gamma ^{2}-1]e^{ik}/(-2i\sin k) .\end{aligned}$$ The frequency for CPA-Laser in this model can be calculated following the relation discussed above, which leads to the restriction of the strength of non-Hermitian potential, $$\begin{aligned}
\omega _{0} &=&2\cos k_{0}=2U, \\
\gamma &=&\sqrt{2-U^{2}}. \label{CPA}\end{aligned}$$ Despite that the $\mathcal{PT} $-symmetry breaking phase may exist when $\left\vert U\right\vert <2$, the CPA-Laser exists only under the condition $\left\vert U\right\vert < 1$ according to the above restriction, while the CPA-Laser does not exist when $1 \leq \left\vert U\right\vert <2$.
The CPA-Laser can be carved by an overall output coefficient $\Theta (k,\sigma )$ in $\mathcal{PT} $ symmetric system, defined as [@Longhi10] $$\begin{aligned}
\Theta (k,\sigma)&=&\frac{\left\vert F_{b}^{-}\right\vert
^{2}+\left\vert F_{f}^{+}\right\vert ^{2}}{\left\vert F_{f}^{-}\right\vert
^{2}+\left\vert F_{b}^{+}\right\vert ^{2}} \nonumber \\
&=&\frac{\left\vert 1+\sigma M_{12}(k)\right\vert ^{2}+\left\vert \sigma -M_{21}(k)\right\vert ^{2}}
{(1+\left\vert \sigma \right\vert ^{2})\left\vert M_{22}(k)\right\vert ^{2}},\end{aligned}$$ in which $\sigma =F_{b}^{+}/F_{f}^{-} $ represents the ratio of the injected signals from two sides, and in the second row we have used the property of the $M$-matrix, i.e., $M^*=M^{-1}$. It is noted that the output coefficient $\Theta (k,\sigma )$ is a useful sign to detect the simultaneous CPA-Laser, that is, the vanishing of coefficient $\Theta $ is the signature of coherent perfect absorption and $\Theta $ vanishes only for particular values of $\sigma=M_{21}(k_{0}) $, i.e., $\Theta (k_{0},\sigma=M_{21}(k_{0}))=0 $. For simplicity, this particular phenomenon can be understood from Eq. (\[MM\]). Since the CPA requires that at the momentum $k_{0}$ there is $F_{f}^{-}, F_{b}^{+} \neq 0 $ and $F_{b}^{-}, F_{f}^{+} =0 $, so it’s clear from the relation $F_{b}^{+}=M_{21}(k_{0})F_{f}^{-}+M_{22}(k_{0})F_{b}^{-} $ that $\sigma =F_{b}^{+}/F_{f}^{-} =M_{21}(k_{0}) $ is the equal criterion for CPA and Laser.
{width="4cm"}\
{width="8.5cm"}
Fig. 3(a) shows the numerical results of $\Theta (k,\sigma )$ for $\sigma=M_{21}(k) $ as a function of frequency $\omega=2\cos k $ under the parameter $U=0.5 $. Two kinds of amplitude $\gamma$ are calculated. The blue solid line indicates that at the coherent $\gamma = \sqrt{2-0.5^2} $ required by Eq. (\[CPA\]), the CPA-Laser occurs at frequency $\omega _{0} = 2\cos k_{0}=1$, which is signed by the vanish of $\Theta $. While for the other $\gamma $ marked by the red dashed line, there is no $k_{0}$ for CPA-Laser in the whole region as $\gamma $ can not follow the Eq. (\[CPA\]). There is another method to characterize the CPA-Laser by the eigenvalues of $S $ matrix. Fig. 3(b) shows that there is a sharp peak of the curve of $\log_{10}|s_{1,2}|^2 $ which indicates the appearance of a lasing mode at a certain $\gamma $ and $\omega $. Meanwhile there is also a deep peak indicating that it’s an absorption mode at the same time. Fig. 3(c) gives an aerial view of Fig. 3(b) from the top and the results show that peaks appears at the frequency $\omega _{0} = 1, \gamma= \sqrt{2-0.5^2} $, which coincides with the results given by the output coefficient $\Theta $.
Summary
=======
In summary, we have studied the symmetry-breaking transition of the scattering process in a simple 1D $\mathcal{PT} $-symmetric non-Hermitian lattice model. With the help of the $S $-matrix method, we have analytically determined the exceptional points and given a discriminant to distinguish the $\mathcal{PT} $-symmetric phase and the breaking phase. In addition, we have shown that the unique feature of CPA-Laser can also be achieved in this simple model and obtained the condition to decide the appearance of CPA-Laser, which can be verified by the overall output coefficient $\Theta $ and the appearance of sharp peaks in the eigenvalues of $S $ matrix individually. This simple model can be easily implemented on experiments, such as the optical materials, the waveguide arrays, and the acoustic sensor. Thus our work provides an additional way to explore the physical effect of the $\mathcal{PT}$-symmetric lattice system both in analytical and experimental fields and help us to understand the physical meaning of $\mathcal{PT}$ symmetry breaking in the lattice scattering problem.
This work is supported by NSF of China under Grants No. 11425419, No. 11374354 and No. 11421092. R. L. is supported by the NSFC under Grant No. 11274195 and the National Basic Research Program of China (973 Program) Grant No. 2011CB606405 and No. 2013CB922000.
[99]{}
C. M. Bender and S. Boettcher, Phys. Rev. Lett. **80**, 5243 (1998). C. M. Bender, D. C. Brody, and H. F. Jones, Phys. Rev. D **70**, 025001 (2004); H. F. Jones, J. Phys. A **39**, 10123 (2006). I. Rotter, J. Phys. A **42**, 153001 (2009). I. Y. Goldsheid and B. A. Khoruzhenko, Phys. Rev. Lett. **80**, 2897 (1998). J. Heinrichs, Phys. Rev. B **63**, 165108 (2001). L. G. Molinari, J. Phys. A **42**, 265204 (2009). A. Regensburger, M.-Ali Miri, C. Bersch, J. Näger, G. Onishchukov, D. N. Christodoulides, and U. Peschel, Phys. Rev. Lett. **110**, 223902 (2013). A. Guo, G. J. Salamo, D. Duchesne, R. Morandotti, M. Volatier-Ravat, V. Aimez, G. A. Siviloglou, and D. N. Christodoulides, Phys. Rev. Lett. **103**, 093902 (2009). C. E. Rüter, K. G. Makris, R. El-Ganainy, D. N. Christodoulides, M. Segev, and D. Kip, Nat. Phys. **6**, 192 (2010). L. Chang, X. S. Jiang, S. Y. Huai, C. Yang, J. M. Wen, L. Jiang, G. Y. Li, G. Z. Wang, and M. Xiao, Nature Photonics **8**, 524¨C529(2014). B. Peng, S. K. Özdemir, F. Lei, F. Monifi, M. Gianfreda, G. L. Long, S. Fan, F. Nori, C. M. Bender and L. Yang, Nature Physics **10**, 394 (2014). R. Fleury, D. Sounals, and A. Alu, Nat. Commun. **6**, 5905 (2015). X. Zhu, H. Ramezani, C. Shi, J. Zhu, and X. Zhang, Phys. Rev. X **4**, 031042 (2014). J. Schindler, A. Li, M. C. Zheng, F. M. Ellis, and T. Kottos, Phys. Rev. A **84**, 040101(R) (2011). Z. Lin, H. Ramezani, T. Eichelkaut, T. Kottos, H. Cao and D. N. Christodoulides, Phys. Rev. Lett. **106**,213901 (2011). A. Mostafazadeh and A. Batal, J. Math. A. **37**, 11645 (2004); L. Jin and Z. Song, Phys. Rev. A **80**, 052107 (2009). C. Yuce, Phys. Lett. A **379**, 1213 (2015). Y. N. Joglekar, D. Scott, M. Babbey and A. Saxena, Phys. Rev. A **82**, 030103(R) (2010). S. D. Liang and G. Y. Huang, Phys. Rev. A **87**, 012118 (2013). A. Mostafazadeh, Phys. Rev. Lett. **102**, 220402 (2009). L. Ge, Y. D. Chong and A. D. Stone, Phys. Rev. A **85**, 023802 (2012). Y. D. Chong, L. Ge and A. D. Stone, Phys. Rev. Lett. **106**, 093902 (2011). Y. D. Chong, L. Ge, H. Cao and A. D. Stone, Phys. Rev. Lett. **105**, 053901(2010). A. Mostafazadeh, Phys. Rev. A **87**,012103 (2013). P. Ambichl, K. G. Makris, L. Ge, Y. D. Chong, A. D. Stone, and S. Rotter, Phys. Rev. X **3**, 041030 (2013). H. Ramezani, H. K. Li, Y. Wang, and X. Zhang, Phys. Rev. Lett. **113**, 263905 (2014). S. Longhi, Phys. Rev. Lett. **105**, 013903 (2010). S. Longhi, Phys. Rev. A **82**, 031801(R) (2010). Z. Ahmed, J. Phys. A: Math. Theor. **47** 385303 (2014).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Junjie Mao, Jelle Kaastra'
- 'N. R. Badnell'
date: 'Received date / Accepted date'
title: 'The electron energy-loss rate due to radiative recombination'
---
[For photoionized plasmas, electron energy-loss rates due to radiative recombination (RR) are required for thermal equilibrium calculations, which assume a local balance between the energy gain and loss. While many calculations of total and/or partial RR rates are available from literature, specific calculations of associated RR electron energy-loss rates are lacking. ]{} [Here we focus on electron energy-loss rates due to radiative recombination of H-like to Ne-like ions for all the elements up to and including zinc ($Z=30$), over a wide temperature range.]{} [We use the AUTOSTRUCTURE code to calculate the level-resolved photoionization cross section and modify the ADASRR code so that we can simultaneously obtain level-resolved RR rate coefficients and associated RR electron energy-loss rate coefficients. The total RR rates and electron energy-loss rates of and are compared with those found in literature. Furthermore, we utilize and parameterize the weighted electron energy-loss factors (dimensionless) to characterize total electron energy-loss rates due to RR.]{} [The RR electron energy-loss data are archived according to the Atomic Data and Analysis Structure (ADAS) data class [*adf48*]{}. The RR electron energy-loss data are also incorporated into the SPEX code for detailed modelling of photoionized plamsas. ]{}
Introduction {#sct:intro}
============
Astrophysical plasmas observed in the X-ray band can roughly be divided into two subclasses: collisional ionized plasmas and photoionized plasmas. Typical collisional ionized plasmas include stellar coronae (in coronal/collisional ionization equilibrium), supernova remnants (SNRs, in non-equilibrium ionization) and the intracluster medium (ICM). In a low-density, high-temperature collisional ionized plasma, e.g. ICM, collisional processes play an important role [for a review, see e.g. @kaa08]. In contrast, in a photoionized plasma, photoionization, recombination and fluorescence processes are also important in addition to collisional processes. Both the equations for the ionization balance (also required for a collisional ionized plasma) and the equations of the thermal equilibrium are used to determine the temperature of the photoionized plasma. Typical photoionized plasmas in the X-ray band can be found in X-ray binaries (XRBs) and active galactic nuclei (AGN).
For collisional ionized plasmas, various calculations of total radiative cooling rates are available in the literature, such as @cox71, @ray76, @sut93, @sch09, @fos12 and @lyk13. These calculations take advantage of full plasma codes like SPEX [@kaa96] and APEC [@smi01], and do not treat individual energy-loss (cooling) processes separately. Total radiative cooling rates include the energy-loss of both the line emission and the continuum emission. The latter includes the energy-loss due to radiative recombination (RR). Even more specifically, the energy-loss due to RR can be separated into the electron energy-loss and the ion energy-loss.
On the other hand, for photoionized plasmas, the electron energy-loss rate due to RR is one of the fundamental parameters for thermal equilibrium calculations, which assume a local balance between the energy gain and loss. Energy can be gained via photoionization, Auger effect, Compton scattering, collisional ionization, collisional de-excitation and so forth. Energy-loss can be due to radiative recombination, dielectronic recombination, three body recombination, inverse Compton scattering, collisional excitation, bremsstrahlung, etc., as well as the line/continuum emission following these atomic processes. In fact, the energy-loss/gain of all these individual processes need to be known. The calculations of electron energy-loss rates due to RR in the Cloudy code [@fer98; @fer13] are based on hydrogenic results [@fer92; @lam01]. In this manuscript, we focus on improved calculations of the electron energy-loss due to radiative recombination, especially providing results for He-like to Ne-like isoelectronic sequences.
While several calculations of RR rates, including the total rates and/or detailed rate coefficients, for different isoelectronic sequences are available, e.g. @gu03 and @bad06, specific calculations of the associated electron energy-loss rate due to RR are limited. The pioneering work was done by @sea59 for hydrogenic ions using the asymptotic expansion of the Gaunt factor for photoionization cross sections (PICSs).
By using a modified semi-classical Kramers formula for radiative recombination cross sections (RRCSs), @kim83 calculated the total RR electron energy-loss rate for a few ions in a relatively narrow temperature range.
@fer92 used the $nl$-resolved hydrogenic PICSs provided by @sto91 to calculate both $n$-resolved RR rates ($\alpha_{\rm i}^{\rm RR}$) and electron energy-loss rates ($L_{\rm i}^{\rm RR}$). Contributions up to and include $n=1000$ are taken into account.
Using the same $nl$-resolved hydrogenic PICSs provided by @sto91, @hum94 calculated the RR electron energy-loss rates for hydrogenic ions in a wide temperature range. In addition, @hum98 calculated PICSs of (photoionizing ion) for $n \leq 25$ with their close-coupling ***R***-matrix calculations. Together with hydrogenic [@sto91] PICSs for $n > 25$ (up to $n=800$ for low temperatures), the RR electronic energy-loss rate coefficient of (recombined ion) was obtained.
Later, @lam01 used the exact PICSs from the Opacity Project [@sea92] for $n < 30$ and PICSs of @ver96 for $n \ge 30$ to obtain $n$-resolved RR electron energy-loss rates for hydrogenic ions in a wide temperature range. The authors introduced the ratio of $\beta / \alpha$ (dimensionless), with $\beta = L / kT$ and $L$ the RR electron energy-loss rate. The authors also pointed out that $\beta / \alpha$ changes merely by 1 dex in a wide temperature range meanwhile $\alpha$ and $\beta$ change more than 12 dex.
In the past two decades, more detailed and accurate calculations of PICSs of many isoelectronic sequences have been carried out [e.g. @bad06], which can be used to calculate specifically the electron energy-loss rates due to RR.
Currently, in the SPEX code [@kaa96], the assumption that the mean kinetic energy of a recombining electron is $3kT/4$ [@kal82] is applied for calculating the electron energy-loss rate due to RR. Based on the level-resolved PICSs provided by the AUTOSTRUCTURE[^1] code [v24.24.3, @bad86], the electron energy-loss rates due to RR are calculated for the H-like to Ne-like isoelectronic sequences for elements up to and including Zn ($Z=30$) in a wide temperature range. Subsequently, the electron energy-loss rate coefficients ($\beta = L / kT$) are weighted with respect to the total RR rates ($\alpha_{\rm t}$), yielding the weighted electron energy-loss factors ($f= \beta / \alpha_{\rm t}$, dimensionless). The weighted electron energy-loss factors can be used, together with the total RR rates, to update the description of the electron energy-loss due to RR in the SPEX code or other codes.
In Sect. \[sct:mo\], we describe the details of the numerical calculation from PICSs to the electron energy-loss rate due to RR. Typical results are shown graphically in Sect. \[sct:res\]. Parameterization of the weighted electron energy-loss factors is also illustrated in Sect. \[sct:par\]. The detailed RR electron energy-loss data are archived according to the Atomic Data and Analysis Structure (ADAS) data class [*adf48*]{}. Full tabulated (unparameterized and parameterized) weighted electron energy-loss factors are available in CDS. Comparison of the results for and can be found in Section \[sct:cf\_0101\_0202\]. The scaling of the weighted electron energy-loss factors with respect to the square of the ionic charge of the recombined ion can be found in Section \[sct:sca\_z2\]. We also discuss the electron and ion energy-loss due to RR (Section \[sct:rrc\]) and the total RR rates (Section \[sct:rr\_tot\]).
Throughout this paper, we refer to the recombined ion when we speak of the radiative recombination of a certain ion, since the line emission following the radiative recombination comes from the recombined ion. Furthermore, only RR from the ground level of the recombining ion is discussed here.
Methods {#sct:mo}
=======
Cross sections {#sct:cs}
--------------
The AUTOSTRUCTURE code is used for calculating level-resolved non-resonant PICSs under the intermediate coupling (“IC") scheme [@bad03]. The atomic and numerical details can be found in @bad06, we briefly state the main points here. We use the Slater-Type-Orbital model potential to determine the radial functions. PICSs are calculated first at zero kinetic energy of the escaping electron, and subsequently on a $z$-scaled logarithmic energy grid with three points per decade, ranging from $\sim z^2 10^{-6}$ to $z^2 10^2$ ryd, where $z$ is the ionic charge of the photoionizing ion/atom. PICSs at even higher energies are at least several orders of magnitude smaller compared to PICSs at zero kinetic energy of the escaping electron. Nonetheless, it still can be important, especially for the $s$- and $p$-orbit, to derive the RR data at the high temperature end. We take advantage of the analytical hydrogenic PICSs [calculated via the dipole radial integral, @bur65] and scale them to the PICS with the highest energy calculated by AUTOSTRUCTURE to obtain PICSs at very high energies. Note that fully $nLSJ$-resolved PICSs for those levels with $n\le15$ and $l\le3$ are calculated specifically. For the rest of the levels, we use the fast, accurate and recurrence hydrogenic approximation [@bur65]. Meanwhile, bundled-$n$ PICSs for $n=$ 16, 20, 25, 35, 45, 55, 70, 100, 140, 200, 300, 450, 700, 999 are also calculated specifically in order to derive the total RR and electron energy-loss rates (interpolation and quadrature required as well).
The inverse process of dielectronic and radiative recombination is resonant and non-resonant photoionization, respectively. Therefore, radiative recombination cross sections (RRCSs) are obtained through the Milne relation under the principle of detailed balance (or microscopic reversibility) from non-resonant PICSs.
Rate coefficients {#sct:rc}
-----------------
The RR rate coefficient is obtained by $$\small
\alpha_{i}(T) = \int_0^{\infty} v~\sigma_{i}(v)~f(v,~T)~dv~,
\label{eq:rr_def}$$ where $v$ is the velocity of the recombining electron, $\sigma_{i}$ is the individual detailed (level/term/shell-resolved) RRCS, $f(v,~T)$ is the probability density distribution of the velocity of the recombining electrons for the electron temperature $T$. The Maxwell-Boltzmann distribution for the free electrons is adopted throughout the calculation, with the same quadrature approach as described in @bad06. Accordingly, the total RR rate per ion/atom is $$\small
\alpha_{\rm t}(T) = \sum_{i} \alpha_{i}(T)~.
\label{eq:rr_ioa}$$ Total RR rates for all the isoelectronic sequences, taking contributions up to $n=10^3$ into account (see its necessity in Section \[sct:res\]).
The RR electron energy-loss rate coefficient is defined as [e.g. @ost89] $$\small
\beta_{i}(T) = \frac{1}{kT}~\int_0^{\infty} \frac{1}{2}~m~v^3~\sigma_{i}(v)~f(v,~T)~dv~,
\label{eq:el_def}$$ The total electron energy-loss rate due to RR is obtained simply by adding all the contributions from individual captures, $$\small
L_{\rm t}(T) = \sum_{i}~L_i = kT~\sum_{i}~\beta_i~,$$ which can be identically derived via $$\small
L_{\rm t}(T) = kT~\alpha_{\rm t}(T)~f_{\rm t}(T)~,
\label{eq:el_ioa}$$ where $$\small
f_{\rm t}(T) = \frac{\sum_{i}~\beta_{i}(T)}{\alpha_{\rm t}(T)}~,
\label{eq:beta_ioa}$$ is defined as the weighted electron energy-loss factor (dimensionless) hereafter.
The above calculation of the electron energy-loss rates is realized by adding Equation (\[eq:el\_def\]) into the archival post-processor FORTRAN code ADASRR[^2] (v1.11). Both the level-resolved and bundled-$n/nl$ RR data and the RR electron energy-loss data are obtained. The output files have the same format of [*adf48*]{} with RR rates and electron energy-loss rates in the units of ${\rm cm^3~s^{-1}}$ and ${\rm ryd~cm^3~s^{-1}}$, respectively. Note that ionization potentials of the ground level of the recombined ions from NIST[^3] (v5.3) are adopted to correct the conversion from PICSs to RRCSs at low kinetic energy for low-charge ions. We should point out that although the level-resolved and bundled-$nl$/$n$ RR data are, in fact, available on OPEN ADAS[^4], given the fact that we use the latest version of the AUTOSTRUCTURE code and a modified version of the ADASRR code, here we re-calculate the RR data, which are used together with the RR electron energy-loss data to derive the weighted electron energy-loss factor $f_{\rm t}$ for consistency. In general, our re-calculate RR data are almost identical to those on OPEN ADAS, except for a few many-electron ions at the the high temperature end, where our re-calculated data differ by a few percent. Whereas, both RR data and electron energy-loss data are a few orders of magnitude smaller compared to those at the lower temperature end, thus, the above mentioned difference has negligible impact on the accuracy of the weighted electron energy-loss factor (see also in Section \[sct:rr\_tot\]).
For all the isoelectronic sequences discussed here, the conventional ADAS 19-point temperature grid $z^2(10-10^7)$ K is used.
Results {#sct:res}
=======
For each individual capture due to radiative recombination, when $kT \ll I$, where $I$ is the ionization potential, the RR electron energy-loss rate $L_i$ is nearly identical to $kT~\alpha_i$, since the Maxwellian distribution drops exponentially for $E_{\rm k} \gtrsim kT$, where $E_{\rm k}$ is the kinetic energy of the free electron before recombination. On the other hand, when $kT \gg I$, the RR electron energy-loss rate is negligible compared with $kT~\alpha_i$. As in an electron-ion collision, when the total energy in the incident channel nearly equals that of a closed-channel discrete state, the channel interaction may cause the incident electron to be captured in this state [@fan68]. That is to say, those electrons with $E_{\rm k} \simeq I$ are preferred to be captured, thus, $L_i \sim I~\alpha_i$. Figure \[fig:rr\_welf\_0212\] shows the ratio of $\beta_i / \alpha_i = L_i / (kT \alpha_i )$ for representative $nLSJ$-resolved levels (with $n \le 8$) of He-like .
![For He-like , the ratio between level-resolved electron energy-loss rates $L_i$ and the corresponding radiative recombination rates times the temperature of the plasma, i.e. $\beta_i / \alpha_i$ (not be confused with $\beta_i / \alpha_{\rm t}$), where $i$ refers to the $nLSJ$-resolved levels with $n \le 8$ (shown selectively in the plot). []{data-label="fig:rr_welf_0212"}](welf_as_0212_lev_plt.eps){width="\hsize"}
In terms of capturing free electrons into individual shells (bundled-$n$), due to the rapid decline of the ionization potentials for those very high-$n$ shells, the ionization potentials can be comparable to $kT$, if not significantly less than $kT$, at the low temperature end. Therefore we see the significant difference between the top panel (low-$n$ shells) and middle panel (high-$n$ shells) of Figure \[fig:rr\_welf\_0426\]. In order to achieve adequate accuracy, contributions from high-$n$ shells (up to $n \le 10^3$) ought to be included. The middle panel of Figure \[fig:rr\_welf\_0426\] shows clearly that even for $n=999$ (the line at the bottom), at the low temperature end, the ratio between $\beta_{n=999}$ and $\alpha_{n=999}$ does not drop to zero. Nevertheless, the bottom panel of Figure \[fig:rr\_welf\_0426\] illustrates the advantage of weighting the electron energy-loss rate coefficients with respect to the total RR rates, i.e. $\beta_{i} / \alpha_{\rm t}$, which approaches zero more quickly. At least, for the next few hundreds shells following $n=999$, their weighted electron energy-loss factors should be no more than $10^{-5}$, thus, their contribution to the total electron energy-loss rate should be less than 1%.
![Ratios of $\beta_i / \alpha_i$ for Be-like (upper and middle panel) and ratios of $\beta_i / \alpha_{\rm t}$ (bottom panel) where $i$ refers to the shell number. Low- and high-$n$ shell results are shown selectively in the plot. The upper panel shows all the shells with $n \le 8$. The middle panel shows shells with $n=$100, 140, 200, 300, 450, 700 and 999. In the lower panel the shells are $n=2,~8,~16,~49,~100,~300,~999$. []{data-label="fig:rr_welf_0426"}](welf_as_0426_shn_plt.eps){width="\hsize"}
The bottom panels of Figure \[fig:rr\_welf\_02ies\] and \[fig:rr\_welf\_ins26\] illustrate the weighted electron energy-loss factors for He-like isoelectronic sequences (He, Si and Fe) and Fe isonuclear sequence (H-, He-, Be- and N-like), respectively. The deviation from (slightly below) unity at the lower temperature end is simply due to the fact that the weighted electron energy-loss factors of the very high-$n$ shells are no longer close to unity (Figure \[fig:rr\_welf\_0426\], middle panel). The deviation from (slightly above) zero at the high temperature end is because the ionization potentials of the first few low-$n$ shells can still be comparable to $kT$, while sum of these $n$-resolved RR rates are more or less a few tens of percent of the total RR rates.
Due to the non-hydrogenic screening of the wave function for low-$nl$ states in low-charge many-electron ions, the characteristic high-temperature bump is present in not only the RR rates [see Figure 4 in @bad06 for an example] but also in the electron energy-loss rates. The feature is even enhanced in the weighted electron energy-loss factor.
![The total RR rates $\alpha_{\rm t}$ (top), electron energy-loss rates $L_{\rm t}$ (middle) and weighted electron energy-loss factors $f_{\rm t}$ (bottom) of He-like isoelectronic sequences for ions including (black), (red) and (orange). The temperature is down-scaled by $z^2$, where $z$ is the ionic charge of the recombined ion, to highlight the discrepancy between hydrogenic and non-hydrogenic. The captures to form the shows non-hydrogenic feature in the bottom panel. []{data-label="fig:rr_welf_02ies"}](welf_as_02ies_ioa_plt.eps){width="\hsize"}
![Top panel is total RR rates $\alpha_{\rm t}$ of the Fe isonuclear sequence, including H- (black), He- (red), Be- (orange) and N-like (blue); Middle panel is the RR electron energy-loss rates $L_{\rm t}$; And bottom panel is the weighted electron energy-loss factors $f_{\rm t}$. The temperature of the plasma is down-scaled by $z^2$, as in Figure \[fig:rr\_welf\_02ies\]. []{data-label="fig:rr_welf_ins26"}](welf_as_ins26_ioa_plt.eps){width="\hsize"}
Parameterization {#sct:par}
----------------
We parameterize the ion/atom-resolved radiative recombination electron energy-loss factors using the same fitting strategy described in @mao16, with the model function of $$\small
f_{\rm t}(T) = a_{0}~T^{-b_{0} - c_{0}\log{T}}~\left(\frac{1 + a_2 T^{-b2}}{1 + a_1 T^{-b1}}\right)~,
\label{eq:fit_m16}$$ where the electron temperature $T$ is in units of eV, $a_{0}~{\rm and}~b_{0}$ are primary fitting parameters, $c_0,~a_{1,~2}$ and $b_{1,~2}$ are additional fitting parameters. The additional parameters are frozen to zero if they are not used. Furthermore, we constrain $b_{0-2}$ to be within -10.0 to 10.0 and $c_0$ between 0.0 and 1.0. The initial values of the two primary fitting parameters $a_{0}~{\rm and}~b_{0}$ are set to unity together with the four additional fitting parameters $a_{1,~2}~{\rm and}~b_{1,~2}$ if they are thawn. Conversely, the initial value of $c_0$, if it is thawn, is set to either side of its boundary, i.e. $c_0 = 0.0$ or $c_0 = 1.0$ (both fits are performed).
In order to estimate the goodness of fit, the fits are performed with a set of artificial relative errors $(r)$. We started with $r=0.625\%$, following with increasing the artificial relative error by a factor of two, up to and including $2.5\%$. The chi-squared statistics adopted here are $$\small
\chi^2 = \sum_{i = 1}^{N} \left(\frac{n_i - m_i}{r~n_i}\right)^2~,
\label{eq:chi2}$$ where $n_i$ is the $i$th numerical calculation result and $m_i$ is the $i$th model prediction (Equation \[eq:fit\_m16\]).
For the model selection, we first fit the data with the simplest model (i.e. all the five additional parameters are frozen to zero), following with fits with free additional parameters step by step. Thawing one additional parameter decreases the degrees of freedom by one, thus, only if the obtained statistics ($\chi^2$) of the more complicated model improves by at least 2.71, 4.61, 6.26, 7.79 and 9.24 for one to five additional free parameter(s), respectively, the more complicated model is favored (at a 90% nominal confidence level).
Parameterizations of the ion/atom-resolved RR weighted electron energy-loss factors for individual ions/atoms in H-like to Ne-like isoelectronic sequences were performed. A typical fit for non-hydrogenic systems is shown in Figure \[fig:rr\_welf\_0726\] for N-like iron (). The fitting parameters can be found in Table \[tbl:fit\_case\]. Again, the weighted energy-loss factor per ion/atom is close to unity at low temperature end and drops towards zero rapidly at the high temperature end.
![The RR weighted electron energy-loss factor for N-like iron (). The black dots in both panels (associated with artificial error bars of 2.5% in the upper one) are the calculated weighted electron energy-loss factor. The red solid line is the best-fit. The lower panel shows the deviation (in percent) between the best-fit and the original calculation. []{data-label="fig:rr_welf_0726"}](welf_as_0726_ioa_fit.eps){width="\hsize"}
In Figure \[fig:welf\_ioa\_stat\] we show the histogram of maximum deviation $\delta_{\rm max}$ (in percent) between the fitted model and the original calculation for all the ions considered here. In short, our fitting accuracy is within 4%, and even accurate ($\lesssim 2.5\%$) for the more important H-like, He-like and Ne-like isoelectronic sequences.
![The histogram of maximum deviation in percent ($\delta_{\rm max}$) for all the ions considered here, which reflects the overall goodness of our parameterization. The dashed-histogram is the statistics of the more important H-like, He-like and Ne-like isoelectronic sequences, while the solid one is the statistics of all the isoelectronic sequences. []{data-label="fig:welf_ioa_stat"}](welf_as_ioa_stat.eps){width="\hsize"}
In addition, we also fit specifically the Case A ($f_{\rm A} = \beta_{\rm t}/\alpha_{\rm t}$) and Case B [@bak38 $f_{\rm B} = \beta_{n\ge2}/\alpha_{n\ge2}$] RR weighted electron energy-loss factors of (Figure \[fig:rr\_welf\_0101case\]) and (Figure \[fig:rr\_welf\_0202case\]). Typical unparameterized factors ($f_{\rm A}$ and $f_{\rm B}$) and fitting parameters can be found in Table \[tbl:unfit\_case\] and \[tbl:fit\_case\], respectively.
![The Case A (solid line, filled circles) and Case B (dashed line, empty diamonds) RR weighted electron energy-loss factor ($f_{\rm A/B}$) for . The black dots in both panels (associated with artificial error bars in the upper one) are the calculated weighted electron energy-loss factor. The red solid line is the best-fit. The lower panel shows the deviation (in percent) between the best-fit and the original calculation. []{data-label="fig:rr_welf_0101case"}](welf_as_0101case_ioa_fit.eps){width="\hsize"}
![Similar to Figure \[fig:rr\_welf\_0101case\] but for . []{data-label="fig:rr_welf_0202case"}](welf_as_0202case_ioa_fit.eps){width="\hsize"}
--------- -------- -------- -------- -------- -------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --
$T/z^2$
K Case A Case B Case A Case B Case A
$10^1$ 0.911 0.895 0.899 0.882 0.869
$10^2$ 0.879 0.851 0.871 0.844 0.845
$10^3$ 0.841 0.786 0.847 0.797 0.797
$10^4$ 0.780 0.668 0.813 0.701 0.678
$10^5$ 0.642 0.470 0.816 0.578 0.460
$10^6$ 0.392 0.268 0.637 0.486 0.246
$10^7$ 0.172 0.123 0.303 0.265 0.113
--------- -------- -------- -------- -------- -------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --
: Unparameterize of RR weighted electron energy-loss factors for , and . For the former two, both Case A and Case B results are treated seperately.[]{data-label="tbl:unfit_case"}
$s$ $Z$ Case $a_0$ $b_0$ $c_0$ $a_1$ $b_1$ $a_2$ $b_2$ $\delta_{\rm max}$
----- ----- ------ ----------- ----------- ----------- ----------- ----------- ----------- ----------- -------------------- -- -- -- -- -- -- -- -- -- -- --
1 1 A 8.655E+00 5.432E-01 0.000E+00 1.018E+01 5.342E-01 0.000E+00 0.000E+00 1.2%
1 1 B 2.560E+00 4.230E-01 0.000E+00 2.914E+00 4.191E-01 0.000E+00 0.000E+00 2.1%
2 2 A 2.354E+00 3.367E-01 0.000E+00 6.280E+01 8.875E-01 2.133E+01 5.675E-01 1.5%
2 2 B 1.011E+04 1.348E+00 4.330E-03 1.462E+04 1.285E+00 0.000E+00 0.000E+00 3.5%
7 26 A 2.466E+01 4.135E-01 0.000E+00 2.788E+01 4.286E-01 0.000E+00 0.000E+00 2.1%
Discussions {#sct:dis}
===========
Comparison with previous results for and {#sct:cf_0101_0202}
-----------------------------------------
![The comparison of the RR data for among results from this work (black), @sea59 [blue], @fer92 [orange] and @hum94 [red]. Both results of case A (solid lines) and case B (dashed lines) are shown. The total RR rates ($\alpha_{\rm A/B}^{\rm RR}$) and electron energy-loss rates ($L_{\rm A/B}^{\rm RR}$) are shown in the top two panels. The RR weighted electron energy-loss factors ($f_{\rm A/B}$) are shown in the middle panel. The ratios of $f_{\rm A/B}$ from this work and previous works with respect to the fitting results (Equation \[eq:fit\_m16\] and Table \[tbl:fit\_case\]) of this work, i.e. ${f_{\rm A/B}^{\rm other}}/{f_{\rm A/B}^{\rm present}}$, are shown in the bottom two panels.[]{data-label="fig:0101_cf_ref"}](welf_as_0101_ioa_cf_ref.eps){width="\hsize"}
Figure \[fig:0101\_cf\_ref\] shows a comparison of RR rates ($\alpha_{t}^{\rm RR}$), electron energy-loss rates ($L_{t}^{\rm RR}$), weighted electron energy-loss factors ($f_{t}^{\rm RR}$) from this work, @sea59 [blue], @fer92 [orange] and @hum94 [red]. Since both @fer92 and @hum94 use the same PICSs [@sto91], the two results are highly consistent as expected. The Case A and Case B results of this work are also consistent within 1% at the low temperature end, and increase to $\sim$5% (underestimation). For the high temperature end ($T\gtrsim 0.1$ keV), since the ion fraction of is rather low (almost completely ionized), the present calculation is still acceptable. A similar issue towards to the high temperature end is also found in the Case A results of @sea59, with a relatively significant overestimation ($\gtrsim5\%$) from the other three calculations.
![Similar to Figure \[fig:0202\_cf\_ref\] but for between this work (black) and @hum98. The latter one only provides data with $T \le 10^{4.4}~{\rm K}$. []{data-label="fig:0202_cf_ref"}](welf_as_0202_ioa_cf_ref.eps){width="\hsize"}
Likewise, the comparison for between this work and @hum98 is presented in Figure \[fig:0202\_cf\_ref\]. The Case A and Case B results from both calculations agree well (within $2\%$), at the low temperature end ($T\lesssim 2.0$ eV). At higher temperatures with $T\gtrsim 2$ eV, the RR rate and electron energy-loss rate for are not available in @hum98.
Scaling with $z^2$ {#sct:sca_z2}
------------------
Previous studies of hydrogenic systems, @sea59 [@fer92; @hum98], all use $z^2$ scaling for $\alpha_{\rm t}^{\rm RR}$. That is to say, $\alpha_{\rm t}^{X}= z^2~\alpha_{\rm t}^{\rm H}$, where $z$ is the ionic charge of the recombined ion $X$. The same $z^2$ scaling also applies for $\beta_{\rm t}^{\rm RR}$ (or $L_{\rm t}^{\rm RR}$). @lam01 also pointed out that the shell-resolved ratio of $f_{n}^{RR}$ (=$\beta_{n}^{\rm RR}/\alpha_{n}^{\rm RR}$) can also be scaled with $z^2 / n^2$, i.e. $f_{n}^{X}= \frac{z^2}{n^2}~f_{n}^{\rm H}$ with $n$ refers to the principle quantum number.
In the following, we merely focus on the scaling for the ion/atom-resolved data set. We show in the top panel of Figure \[fig:welf\_01ies\_sca\] the ratios of $f_t/z^2$ for H-like ions. Apparently, from the bottom panel of Figure \[fig:welf\_01ies\_sca\], the $z^2$ scaling for the H-like isoelectronic sequence is accurate within 2%. For the rest of the isoelectronic sequences, for instance, the He-like isoelectronic sequence shown in Figure \[fig:welf\_02ies\_sca\], the $z^2$ scaling applies at the low temperature end, whereas, the accuracies are poorer toward the high temperature end. We also show the $z^2$ scaling for the Fe isonuclear sequence in Figure \[fig:welf\_ins26\_sca\].
![The $z^2$ scaling for the H-like isoelectronic sequence (Case A), including (black), (red), (orange) and (green). The top panel shows the ratios of $f_t/z^2$ as a function of electron temperature ($T$). The bottom panel is the ratio of $(f_{\rm t}/z^2)^{X}$ for ion $X$ with respect to the ratio of $(f_{\rm t}/z^2)^{\rm H}$ for H. []{data-label="fig:welf_01ies_sca"}](welf_as_01ies_ioa_sca.eps){width="\hsize"}
![Similar to Figure \[fig:welf\_01ies\_sca\] but for the $z^2$ scaling for the He-like isoelectronic sequences. []{data-label="fig:welf_02ies_sca"}](welf_as_02ies_ioa_sca.eps){width="\hsize"}
![The $z^2$ scaling for the Fe isonuclear sequence. The top panel shows the ratios of $f_{\rm t}/z^2$ as a function of electron temperature ($T$). The bottom panel is the ratio of $(f_{\rm t}/z^2)^{X-{\rm like}}$ for $X$-like Fe with respect to the ratio of $(f_{\rm t}/z^2)^{\rm H-like}$ for H-like .[]{data-label="fig:welf_ins26_sca"}](welf_as_ins26_ioa_sca.eps){width="\hsize"}
Radiative recombination continua {#sct:rrc}
--------------------------------
We restrict the discussion above for the RR energy-loss of the electrons in the plasma only. The ion energy-loss of the ions due to RR can be estimated as $P^{\rm RR} \sim I_i~\alpha_i$, where $I_i$ is the ionization potential of the level/term the free electron is captured into, and $\alpha_i$ is the corresponding RR rate coefficient. Whether to include the ionization potential energies as part of the total internal energy of the plasma is not critical, as long as the entire computation of the net energy gain/loss is self-consistent [see a discussion in @gna12]. On the other hand, when interpreting the emergent spectrum due to RR, such as the radiative recombination continua (RRC) for a low-density plasma, the ion energy-loss of the ion is essentially required. The RRC emissivity [@tuc66] can be obtained via $$\begin{aligned}
\small
\frac{dE^{\rm RRC}}{dt~dV} &=& \int_{0}^{\infty} n_{\rm e}~n_{\rm i}~\left(I + \frac{1}{2} m v^2\right)~v~\sigma(v)~f(v,~kT) dv \nonumber \\
&=& n_{\rm e}~n_{\rm i}I~\left(1 + f_{\rm t}~kT/I\right)~\alpha_{\rm t}~,
\label{eq:rrc}\end{aligned}$$ where $n_{\rm e}$ and $n_{\rm i}$ are the electron and (recombining) ion number density, respectively. Generally speaking, the ion energy-loss of the ion dominates the electron energy-loss of the electrons, since $f_{\rm t}$ is of the order of unity while $kT \lesssim I$ holds for those X-ray photoionizing plasmas in XRBs [@lie96], AGN [@kin02] and recombining plasmas in SNRs [@oza09]. Figure \[fig:rr\_welf\_rrc\] shows the threshold temperature above which the electron energy-loss via RR cannot be neglected compared to the ion energy-loss. For hot plasmas with $kT \gtrsim 2$ keV, the electron energy-loss is comparable to the ion energy-loss for $Z>5$. It is necessary to emphasize that we refer to the electron temperature $T$ of the plasma here, which is not necessarily identical to the ion temperature of the plasma, in particular, in the nonequilibrium ionization scenario.
![The threshold temperature above which the electron energy-loss via RR cannot be neglected, compared to the ion energy-loss, for H-like (solid lines) and He-like ions (dashed lines). []{data-label="fig:rr_welf_rrc"}](welf_as_rrc_ioa_plt.eps){width="\hsize"}
Total radiative recombination rate {#sct:rr_tot}
----------------------------------
Various calculations of (total or shell/term/level-resolved) RR data are available from the literature. Historically, different approaches have been used for calculating the total RR rates, including the Dirac-Hartree-Slater method [@ver93] and the distorted-wave approximation [@gu03; @bad06]. Additionally, Nahar and coworkers [e.g. @nah99] obtained the total (unified DR + RR) recombination rate for various ions with their ***R***-matrix calculations. Different approaches can lead to different total RR rates [see a discussion in @bad06], as well as the Individual term/level-resolved RR rate coefficients, even among the most advanced ***R***-matrix calculations. Nevertheless, the bulk of the total RR rates for various ions agrees well among each other. As for the detailed RR rate coefficients, consequently, the detailed RR electron energy-loss rate, as long as the difference among different methods are within a few percent and given the fact that each individual RR is $\lesssim 10\%$ of the total RR rate for a certain ion/atom, the final difference in the total weighted electron energy-loss factors $f_{\rm t}$ are still within 1%. In other words, although we used the re-calculated total RR rate (Section \[sct:rc\]) to derive the weighted electron energy-loss factors, we assume these factors can still be applied to other total RR rates.
J.M. acknowledges discussions and support from M. Mehdipour, A. Raassen, L. Gu and M. O’Mullane. We thank the referee, G. Ferland, for the valuable comments on the manuscript SRON is supported financially by NWO, the Netherlands Organization for Scientific Research.
Badnell, N. R. 1986, Journal of Physics B Atomic Molecular Physics, 19, 3827
Badnell, N. R., & Seaton, M. J. 2003, Journal of Physics B Atomic Molecular Physics, 36, 4367
Badnell, N. R. 2006, , 167, 334
Baker, J. G., & Menzel, D. H. 1938, , 88, 52
Burgess, A. 1965, , 69, 1
Cox, D. P., & Daltabuit, E. 1971, , 167, 113
Hummer, D. G. 1994, , 268, 109
Hummer, D. G., & Storey, P. J. 1998, , 297, 1073
Fano, U., & Cooper, J. W. 1968, Reviews of Modern Physics, 40, 441
Ferland, G. J., Peterson, B. M., Horne, K., Welsh, W. F., & Nahar, S. N. 1992, , 387, 95
Ferland, G. J., Korista, K. T., Verner, D. A., et al. 1998, , 110, 761
Ferland, G. J., Porter, R. L., van Hoof, P. A. M., et al. 2013, , 49, 137
Foster, A. R., Ji, L., Smith, R. K., & Brickhouse, N. S. 2012, , 756, 128
Gnat, O., & Ferland, G. J. 2012, , 199, 20
Gu, M. F. 2003, , 589, 1085
Kaastra, J. S., Mewe, R., & Nieuwenhuijzen, H. 1996, UV and X-ray Spectroscopy of Astrophysical and Laboratory Plasmas, 411
Kaastra, J. S., Bykov, A. M., Schindler, S., et al. 2008, , 134, 1
Kallman, T. R., & McCray, R. 1982, , 50, 263
Kim, Y. S., & Pratt, R. H. 1983, , 27, 2913
Kinkhabwala, A., Sako, M., Behar, E., et al. 2002, , 575, 732
LaMothe, J., & Ferland, G. J. 2001, , 113, 165
Liedahl, D. A., & Paerels, F. 1996, , 468, L33
Lykins, M. L., Ferland, G. J., Porter, R. L., et al. 2013, , 429, 3133
Mao, J., & Kaastra, J. 2016, , 587, A84
Nahar, S. N. 1999, , 120, 131
Osterbrock, D. E. 1989, Research supported by the University of California, John Simon Guggenheim Memorial Foundation, University of Minnesota, et al. Mill Valley, CA, University Science Books, 1989, 422 p.,
Ozawa, M., Koyama, K., Yamaguchi, H., Masai, K., & Tamagawa, T. 2009, , 706, L71
Raymond, J. C., Cox, D. P., & Smith, B. W. 1976, , 204, 290
Schure, K. M., Kosenko, D., Kaastra, J. S., Keppens, R., & Vink, J. 2009, , 508, 751
Seaton, M. J. 1959, , 119, 81
Seaton, M. J., Zeippen, C. J., Tully, J. A., et al. 1992, , 23, 19
Smith, R. K., Brickhouse, N. S., Liedahl, D. A., & Raymond, J. C. 2001, , 556, L91
Storey, P. J., & Hummer, D. G. 1991, Computer Physics Communications, 66, 129
Sutherland, R. S., & Dopita, M. A. 1993, , 88, 253
Tucker, W. H., & Gould, R. J. 1966, , 144, 244
Verner, D. A., Yakovlev, D. G., Band, I. M., & Trzhaskovskaya, M. B. 1993, Atomic Data and Nuclear Data Tables, 55, 233
Verner, D. A., & Ferland, G. J. 1996, , 103, 467
[^1]: http://amdpp.phys.strath.ac.uk/autos/
[^2]: http://amdpp.phys.strath.ac.uk/autos/ver/misc/adasrr.f
[^3]: http://physics.nist.gov/PhysRefData/ASD/ionEnergy.html
[^4]: http://open.adas.ac.uk/adf48
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents new efficient solutions to the rolling shutter camera absolute pose problem. Unlike the state-of-the-art polynomial solvers, we approach the problem using simple and fast linear solvers in an iterative scheme. We present several solutions based on fixing different sets of variables and investigate the performance of them thoroughly. We design a new alternation strategy that estimates all parameters in each iteration linearly by fixing just the non-linear terms. Our best 6-point solver, based on the new alternation technique, shows an identical or even better performance than the state-of-the-art R6P solver and is two orders of magnitude faster. In addition, a linear non-iterative solver is presented that requires a non-minimal number of 9 correspondences but provides even better results than the state-of-the-art R6P. Moreover, all proposed linear solvers provide a single solution while the state-of-the-art R6P provides up to 20 solutions which have to be pruned by expensive verification.'
author:
- Zuzana Kukelova^1^Cenek Albl^2^Akihiro Sugimoto^3^Tomas Pajdla^2^
title: 'Linear solution to the minimal absolute pose rolling shutter problem[^1]'
---
Introduction
============
Rolling shutter (RS) cameras are omnipresent. They can be found in smartphones, consumer, professional, and action cameras and even in self-driving cars. RS cameras are cheaper, and easier to produce, than global shutter cameras. They also posses other advantages over the global shutter cameras, such as higher achievable frame-rate or longer exposure times.
There is, however, a significant drawback when using them for computer vision applications. When the scene or the camera is moving during image capture, images produced by RS cameras will become distorted. The amount and type of distortion depends on the type and speed of camera motion and on the depth of the scene. It has been shown that RS image distortions can cause problems for standard computer vision methods such as Structure from Motion [@Hedborg2012], visual SLAM [@Klein2009] or multi-view dense stereo [@Saurer2013]. Therefore, having a special camera model for rolling shutter cameras is desirable.
The camera absolute pose computation is a fundamental problem in many computer vision tasks such as Structure from Motion, augmented reality, visual SLAM, and visual localization. The problem is to compute the camera pose from 3D points in the world and their 2D projections into an image. The minimal number of correspondences necessary to solve the absolute pose problem for a perspective calibrated camera is three. The first solution to this problem was introduced by Grunert [@Grunert-1841] and since then it was many times revisited [@Haralick1991; @Ameller02camerapose; @Fischler-Bolles-ACM-1981]. Other work has focused on computing the absolute pose from a larger than the minimal number of correspondences [@Lepetit-IJCV-2009; @Quan1999; @Triggs1999; @Wu2006; @Zhi2002]. All of the previous work consider a perspective camera model, which is not suitable for dynamic RS cameras.
Recently, as RS cameras became more and more common, the focus turned to computing camera absolute pose from images containing RS effects. First, several RS camera motion models were introduced in [@Meingast2005]. A solution to RS absolute pose using non-minimal (eight and half) number of points was presented in [@Aitaider2006]. It relied on a non-linear optimization and required a planar scene.
In [@Hedborg2011], video sequences were exploited and the absolute camera pose was computed sequentially using a non-linear optimization starting from the previous camera pose. Another approach using video sequences was used for visual SLAM in [@Klein2009] where the camera motion estimated from previous frames was used to compensate the RS distortion in the next frame prior to the optimization.
A polynomial solution that is globally optimal was presented in [@Magerand2012]. It uses Gloptipoly [@Henrion2009] solver to find a solution from 7 or more points. Authors show that the method provides better results than [@Aitaider2006], but the runtime is in the order of seconds, making it impractical for typical applications such as RANSAC.
The first minimal solution to the rolling shutter camera absolute pose problem was presented in [@Albl-CVPR-2015]. It uses the minimal number of six 2D to 3D point correspondences and the method to generate an efficient solver. The proposed R6P is based on the constant linear and angular velocity model as in [@Aitaider2006; @Magerand2012; @Hedborg2012] but it uses the first order approximation to both the camera orientation and angular velocity, and, therefore, it requires an initialization of the camera orientation, e.g., from P3P [@Fischler-Bolles-ACM-1981]. Paper [@Albl-CVPR-2015] has shown that R6P solver significantly outperforms perspective P3P solver in terms of camera pose precision and the number of inliers captured in the RANSAC loop.
Motivation
----------
It has been demonstrated in the literature that RS camera absolute pose is beneficial and often necessary when dealing with RS images from moving camera or dynamic scene. Still, until now, all the presented solutions have significant drawbacks that make them impractical for general use.
The state-of-the-art solutions require a non-minimal or a larger number of points [@Aitaider2006; @Magerand2012], planar scene [@Aitaider2006], video sequences [@Hedborg2011; @Hedborg2012; @Klein2009], are very slow [@Magerand2012] and provide too many solutions [@Albl-CVPR-2015]. If one requires a practical algorithm similar to P3P, but working on RS images, the closest method available is R6P [@Albl-CVPR-2015]. However, R6P still needs around $1.7ms$ to compute the camera pose, compared to around $3\mu s$ for P3P. Therefore, in typical applications where P3P is used, one would suffer a several orders of magnitude slowdown compared to P3P. This makes it hard to use for real-time applications such as augmented reality. In addition, R6P provides up to 20 real solutions, which need to be verified. This makes tasks like RANSAC, which uses hundreds or thousands of iterations and verifies all solutions, extremely slow compared to P3P. This motivates us to create a solution to RS absolute pose problem with similar performance to R6P [@Albl-CVPR-2015] and runtime comparable to P3P.
Contribution
------------
In this work we present solutions that remove previously mentioned drawbacks of the state-of-the-art methods and provide practical and fast rolling shutter camera absolute pose solvers. We take a different approach to formulating the problem and propose linear solutions to rolling shutter camera absolute pose. Specifically, we present the following RS absolute camera pose solvers:
- a 6-point linear iterative solver, which provides identical or even better solutions than R6P in $10\mu s$, which is up to $170\times$ faster than R6P. This solver is based on a new alternating method;
- two 6-point linear iterative solvers that outperform R6P for purely translational motion;
- a 9-point linear non-iterative solver that provides more accurate camera pose estimates than R6P in $20\mu s$;
All solvers are easy to implement and they return a single solution. We formulate the problem of RS camera absolute pose in Section \[sec:formulation\]. Derivations of all new solvers are in Section \[sec:lin\_solvers\]. Section \[sec:experiments\] contains experiments verifying the feasibility of the proposed solvers and it compares them against P3P and R6P [@Albl-CVPR-2015].
Problem formulation {#sec:formulation}
===================
For calibrated perspective cameras, the projection equation can be written as $$\lambda_i {{\mathtt{x}}}_i = {\mathtt{R}}{{\mathtt{X}}}_i+{{\mathtt{C}}},
\label{eq:persp_proj}$$ where ${\mathtt{R}}$ and ${{\mathtt{C}}}$ are the rotation and translation bringing a 3D point ${{\mathtt{X}}}_i$ from a world coordinate system to the camera coordinate system with ${{\mathtt{x}}}_i = \left[r_i,c_i,1\right]^{\top}$, and scalar $\lambda_i \in \mathbb{R}$. For RS cameras, every image row is captured at different time and hence at a different position when the camera is moving during the image capture. Camera rotation ${\mathtt{R}}$ and translation ${{\mathtt{C}}}$ are therefore functions of the image row $r_i$ being captured $$\lambda_i {{\mathtt{x}}}_i= \lambda_i {\left[\!\!{\begin{array}{l} r_i\\c_i\\1\end{array}}\!\!\right]} = {\mathtt{R}}(r_i){{\mathtt{X}}}_i+{{\mathtt{C}}}(r_i).
\label{eq:proj_rs}$$
In recent work [@Albl-CVPR-2015; @Saurer2013; @Magerand2012; @Meingast2005; @Hedborg2012; @Aitaider2006], it was shown that for the short time-span of a frame capture, the camera translation ${{\mathtt{C}}}(r_i)$ can be approximated with a simple constant velocity model as $${{\mathtt{C}}}(r_i) = {{\mathtt{C}}} + (r_i-r_0){{\mathtt{t}}},
\label{eq:C_ri}$$ where ${{\mathtt{C}}}$ is the camera center corresponding to the perspective case, i.e. when $r_i=r_0$, and ${{\mathtt{t}}}$ is the translational velocity. The camera rotation ${\mathtt{R}}(r_i)$ can be decomposed into two rotations to represent the camera initial orientation by ${\mathtt{R}}_{{{\mathtt{v}}}}$ and the change of orientation during frame capture by ${\mathtt{R}}_{{{\mathtt{w}}}}(r_i-r_0)$.
In [@Magerand2012; @Albl-CVPR-2015], it was observed that it is usually sufficient to linearize ${\mathtt{R}}_{{{\mathtt{w}}}}(r_i-r_0)$ around the initial rotation ${\mathtt{R}}_{{\mathtt{v}}}$ using the first order Taylor expansion such that $$\lambda_i {\left[\!\!{\begin{array}{l} r_i\\c_i\\1\end{array}}\!\!\right]} = \left({\mathtt{I}}+(r_i-r_0)[{{\mathtt{w}}}]_\times\right){\mathtt{R}}_{{\mathtt{v}}}{{\mathtt{X}}}_i +{{\mathtt{C}}}+(r_i-r_0){{\mathtt{t}}},
\label{eq:model_lin}$$ where $[{{\mathtt{w}}}]_\times$ is a skew-symmetric matrix of vector ${{\mathtt{w}}}$. The model (\[eq:model\_lin\]), with linearized rolling shutter rotation, will deviate from the reality with increasing rolling shutter effect. Still, it is usually sufficient for most of the rolling shutter effects present in real situations.
In [@Albl-CVPR-2015], a linear approximation to the camera orientation ${\mathtt{R}}_{{\mathtt{v}}}$ was used to solve the rolling shutter absolute pose problem from a minimal number of six 2D-3D point correspondences. This model has the form $$\lambda_i {\left[\!\!{\begin{array}{l} r_i\\c_i\\1\end{array}}\!\!\right]} = \left({\mathtt{I}}+(r_i-r_0)[{{\mathtt{w}}}]_\times\right)\left({\mathtt{I}}+[{{\mathtt{v}}}]_\times\right){{\mathtt{X}}}_i +{{\mathtt{C}}}+(r_i-r_0){{\mathtt{t}}}.
\label{eq:model_double_lin}$$ The drawback of the model (\[eq:model\_double\_lin\]) is that ${\mathtt{R}}_{{{\mathtt{v}}}}$ is often not small and thus cannot be linearized. Therefore, the accuracy of the model is dependent on the initial orientation of the camera in the world frame. In [@Albl-CVPR-2015], it was shown that the standard P3P algorithm [@Fischler-Bolles-ACM-1981] is able to estimate camera orientation with sufficient precision even for high camera rotation velocity and therefore P3P can be used to bring the camera rotation matrix ${\mathtt{R}}_{{\mathtt{v}}}$ close to the identity, where (\[eq:model\_double\_lin\]) works reasonably.
The model (\[eq:model\_double\_lin\]) leads to a system of six quadratic equations in six unknowns. This system has 20 solutions and it was solved in [@Albl-CVPR-2015] using the method [@Cox-Little-etal-05; @Kukelova-ECCV-2008]. The solver [@Albl-CVPR-2015] for the R6P rolling shutter problem requires the G-J elimination of a $ 196 \times 216$ matrix and computing the eigenvectors of a $20 \times 20$ matrix. The R6P solver runs for about 1.7ms and thus is too slow in many practical situations.
We will next show how to simplify this model by linearizing equation (\[eq:model\_double\_lin\]) and yet still obtaining a similar performance as the R6P absolute pose solver [@Albl-CVPR-2015] for the original model (\[eq:model\_double\_lin\]).
Linear rolling shutter solvers {#sec:lin_solvers}
==============================
We present here several linear iterative solvers to the minimal absolute pose rolling shutter problem. All these solvers start with the model (\[eq:model\_double\_lin\]) and they use six 2D-3D image point correspondences to estimate 12 unknowns ${{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{w}}}$, and ${{\mathtt{t}}}$. The proposed solvers differ in the way how the system (\[eq:model\_double\_lin\]) is linearized. Additionally we propose a linear non-iterative 9 point absolute pose rolling shutter solver.
$\textbf{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}}}^{{{\mathtt{w}}},{{\mathtt{t}}}}$ solver {#sec:solver1}
-------------------------------------------------------------------------------------
The solver is based on the idea of alternating between two linear solvers. The first $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}}}$ solver fixes the rolling shutter parameters ${{\mathtt{w}}}$ and ${{\mathtt{t}}}$ in (\[eq:model\_double\_lin\]) and estimates only the camera parameters ${{\mathtt{v}}}$ and ${{\mathtt{C}}}$. The second $\text{R6P}_{{{\mathtt{w}}},{{\mathtt{t}}}}$ solver fixes the camera parameters ${{\mathtt{v}}}$ and ${{\mathtt{C}}}$ and estimates only the rolling shutter parameters ${{\mathtt{w}}}$ and ${{\mathtt{t}}}$. Both these partial solvers results in 12 linear equations in 6 unknowns that can be solved in the least square sense. The motivation for this solver comes from the fact that even for larger rolling shutter speed, the camera parameters ${{\mathtt{v}}}$ and ${{\mathtt{C}}}$ can be estimated quite accurately.
The solver starts with ${{\mathtt{w}}}_0 = {{\mathtt{0}}}$ and ${{\mathtt{t}}}_0 = {{\mathtt{0}}}$ and, in the first iteration, uses linear $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}}}$ solver to estimate ${{\mathtt{v}}}_1$ and ${{\mathtt{C}}}_1$. Using the estimated ${{\mathtt{v}}}_1$ and ${{\mathtt{C}}}_1$, the linear solver $\text{R6P}_{{{\mathtt{w}}},{{\mathtt{t}}}}$ estimates ${{\mathtt{w}}}_1$ and ${{\mathtt{t}}}_1$. This process is repeated until the desired precision is obtained or a maximum number of iterations is reached.
The solver does not perform very well in our experiments, which we account to the fact that it never estimates the pose parameters ${{\mathtt{v}}}$,${{\mathtt{C}}}$ and the motion parameters ${{\mathtt{w}}}$,${{\mathtt{t}}}$ together in one step. Nevertheless, we present this solver as a logical first step when considering the iterative approach to RS absolute pose problem.
$\textbf{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{t}}}}^{{{\mathtt{w}}}}$ and $\textbf{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{t}}}}^{{{\mathtt{w}}},{{\mathtt{t}}}}$ solver {#sec:solver2}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
To avoid problems of the solver, we introduce the solver. The solver alternates between two solvers, i.e. the linear $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{t}}}}$ solver, which fixes only the rolling shutter rotation ${{\mathtt{w}}}$ and estimates ${{\mathtt{v}}},{{\mathtt{C}}}$ and ${{\mathtt{t}}}$, and the $\text{R6P}_{{{\mathtt{w}}}}$ solver that estimates only the rolling shutter rotation ${{\mathtt{w}}}$ using the fixed ${{\mathtt{v}}},{{\mathtt{C}}}$ and ${{\mathtt{t}}}$. The $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{t}}}}$ solver solves 12 linear equations in 9 unknowns and the $\text{R6P}_{{{\mathtt{w}}}}$ solver solves 12 linear equations in 3 unknowns in the least square sense. Since the first $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{t}}}}$ solver assumes unknown rolling shutter translation, the camera parameters are estimated with better precision than in the case of the $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}}}$ solver. Moreover, in many applications, e.g. cameras on a car, cameras often undergo only a translation motion, and therefore ${{\mathtt{w}}}$ is negligible. In such situations, the first iteration of the $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{t}}}}$ solver already provides very precise estimates of the camera parameters.
Another approach is to use only the ${{\mathtt{v}}}$ and ${{\mathtt{C}}}$ estimated by $\text{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{t}}}}$ solver and in the second step re-estimate the rolling shutter translation ${{\mathtt{t}}}$ together with the rolling shutter rotation ${{\mathtt{w}}}$ using the linear $\text{R6P}_{{{\mathtt{w}}},{{\mathtt{t}}}}$ solver. The solver based on this strategy will be referred to as .
The resulting solvers and , again, alternate between the two linear solvers until the desired precision is obtained or a maximum number of iterations is reached. We show in the experiments that those solvers outperform R6P in the case of pure translational motion.
$\textbf{R6P}_{{{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{w}}},{{\mathtt{t}}}}^{[{{\mathtt{v}}}]_\times}$ solver
-------------------------------------------------------------------------------------------------------------
The solver estimates all unknown parameters ${{\mathtt{v}}},{{\mathtt{C}}},{{\mathtt{w}}}$ and ${{\mathtt{t}}}$ together in one step. To avoid non-linearity in (\[eq:model\_double\_lin\]), the solver fixes \[${{\mathtt{v}}}]_\times$ that appears in the nonlinear term $[{{\mathtt{w}}}]_\times [{{\mathtt{v}}}]_\times$ in (\[eq:model\_double\_lin\]). Thus the solver solves equations $$\lambda_i {\left[\!\!{\begin{array}{l} r_i\\c_i\\1\end{array}}\!\!\right]} = \left({\mathtt{I}}+(r_i-r_0)[{{\mathtt{w}}}]_\times\right){{\mathtt{X}}}_i + [{{\mathtt{v}}}]_\times{{\mathtt{X}}}_i + (r_i-r_0)[{{\mathtt{w}}}]_\times[{{\mathtt{\hat{v}}}}]_\times{{\mathtt{X}}}_i +{{\mathtt{C}}}+(r_i-r_0){{\mathtt{t}}},
\label{eq:model_double_lin2}$$ where ${{\mathtt{\hat{v}}}}$ is a fixed vector.
In the first iteration ${{\mathtt{\hat{v}}}}$, is set to the zero vector and the term $(r_i-r_0)[{{\mathtt{w}}}]_\times[{{\mathtt{\hat{v}}}}]_\times{{\mathtt{X}}}_i$ in (\[eq:model\_double\_lin2\]) disappears. This is usually a sufficient approximation. The explanation for this is as follows. After the initialization with P3P the camera rotation is already close to the identity and in real applications the rolling shutter rotation ${{\mathtt{w}}}$ during the capture is usually small. Therefore, the nonlinear term $[{{\mathtt{w}}}]_\times[{{\mathtt{{v}}}}]_\times$ is small, sometimes even negligible, and thus it can be considered to be zero in the first iteration.
In the remaining iterations we fix ${{\mathtt{\hat{v}}}}$ in the $(r_i-r_0)[{{\mathtt{w}}}]_\times[{{\mathtt{\hat{v}}}}]_\times{{\mathtt{X}}}_i$ term to be equal to the ${{\mathtt{v}}}_i$ estimated in the previous iteration of the solver. Note that we fix only ${{\mathtt{v}}}$ that appears in the nonlinear term $[{{\mathtt{w}}}]_\times[{{\mathtt{{v}}}}]_\times$ and there is still another term with ${{\mathtt{v}}}$ in (\[eq:model\_double\_lin2\]) from which a new ${{\mathtt{v}}}$ can be estimated. Therefore, all parameters are estimated at each step which is a novel alternating strategy. To our knowledge, all existing algorithms that are based on the alternating optimization approach completely fix a subset of the variables, meaning that they cannot estimate all the variables in one step.
The in each iteration solves only one system of 12 linear equations in 12 unknowns and is therefore very efficient. In experiments we will show that the provides very precise estimates already after 1 iteration and after 5 iterations it has virtually the same performance as the state-of-the-art R6P solver [@Albl-CVPR-2015].
R9P
---
Our final solver is a non-iterative solver that uses a non-minimal number of nine 2D-3D point correspondences. We note that the projection equation (\[eq:model\_double\_lin2\]) can be rewritten as $$\lambda_i {\left[\!\!{\begin{array}{l} r_i\\c_i\\1\end{array}}\!\!\right]} = \left({\mathtt{I}}+[{{\mathtt{v}}}]_\times\right){{\mathtt{X}}}_i +{{\mathtt{C}}} + (r_i-r_0)([{{\mathtt{w}}}]_\times({\mathtt{I}}+[{{\mathtt{{v}}}}]_\times){{\mathtt{X}}}_i + {{\mathtt{t}}}).
\label{eq:model_double_lin3}$$ We can substitute the term $[{{\mathtt{w}}}]_\times({\mathtt{I}}+[{{\mathtt{{v}}}}]_\times)$ in (\[eq:model\_double\_lin3\]) with a $3 \times 3$ unknown matrix ${\mathtt{R}}_{\text{RS}}$. After eliminating the scalar values $\lambda_i$ by multiplying equation (\[eq:model\_double\_lin3\]) from the left by the skew symmetric matrix for vector ${\left[\!\!{\begin{array}{ccc} r_i & c_i & 1\end{array}}\!\!\right]}^\top$ and without considering the internal structure of the matrix ${\mathtt{R}}_{\text{RS}}$, we obtain three linear equations in 18 unknowns, i.e. ${{\mathtt{v}}}, {{\mathtt{C}}}, {{\mathtt{t}}}$, and 9 unknowns in ${\mathtt{R}}_{\text{RS}}$. Since only two from these tree equations are linearly independent we need nine 2D-3D point correspondences to solve this problem.
Note that the original formulation (\[eq:model\_double\_lin\]) was an approximation to the real rolling shutter camera model and therefore the formulation with a general $3 \times 3$ matrix ${\mathtt{R}}_{\text{RS}}$ is yet a different approximation to this model.
Experiments {#sec:experiments}
===========
We tested the proposed solvers on a variety of synthetic and real datasets and compared the results with the original R6P solver [@Albl-CVPR-2015] as well as P3P. We followed the general pattern of experiments used in [@Albl-CVPR-2015] in order to provide consistent comparison on the additional factor of experiments that are specific to our iterative solvers such as their convergence.
To analyze the accuracy of the estimated camera poses and velocities, we used synthetic data in the following setup. A random set of 3D points was generated in a cubic region with $x,y,z\in[-1;1]$ and a camera with a distance $d\in[2;3]$ from the origin and pointing towards the 3D points. The camera was set to be calibrated, i.e. ${\mathtt{K}}={\mathtt{I}}$ and the field of view was set to 45 degrees. Rolling shutter projections were created using a constant linear velocity and a constant angular velocity with various magnitudes.
Using the constant angular velocity model for generating the data ensures that our data is not generated with the same model as the one that is estimated by the solvers (linear approximation to a rotation). Although the used model is just an approximation of the real rolling shutter model and we could have chosen another one, e.g. constant angular acceleration, we consider the constant angular velocity model as a reasonable description of the camera motion during the short time period of frame capture.
We used 6 points for the original R6P and all proposed R6P iterative solvers. In order to provide P3P with the same data, we used all possible triplets from the 6 points used by R6P and then chose the best result. For R9P we used 9 points. Unless stated otherwise, all iterative solvers were run for maximum 5 iterations in the experiments.
Synthetic data {#sec:exp_synth}
--------------
In the first experiment, we gradually increased the camera velocities during capture. The maximum translational velocity was 0.3 per frame and the maximum angular velocity was 30 degrees per frame. Figure \[fig:synth\_inc\_w\_t\_pose\] shows the results, from which we can see how the increasing RS deformation affects the estimated camera pose and also estimated camera velocities in those solvers.
![Experiment on synthetic data focusing on the precision of estimated camera poses and velocities. Notice that the performance of is identical to R6P. In terms of camera pose these two solvers are slightly outperformed by R9P. Other linear solvers perform very poorly in all respects.[]{data-label="fig:synth_inc_w_t_pose"}](figs/c_err_RI_w_t.pdf "fig:"){width="0.45\columnwidth"} ![Experiment on synthetic data focusing on the precision of estimated camera poses and velocities. Notice that the performance of is identical to R6P. In terms of camera pose these two solvers are slightly outperformed by R9P. Other linear solvers perform very poorly in all respects.[]{data-label="fig:synth_inc_w_t_pose"}](figs/v_err_RI_w_t.pdf "fig:"){width="0.45\columnwidth"}\
#### [**Rotational and translational motion:**]{} In agreement with [@Albl-CVPR-2015], R6P provides much better results than P3P thanks to the RS camera model. The newly proposed solver provides almost identical results to R6P at much lower computation cost (cf. Table \[tab:timings\]). The best estimates of the camera pose are provided by R9P at the cost of using more than minimal number of points. The other 6-point iterative solutions are performing really bad, often providing worse results than P3P. In the next experiment we tested the sensitivity of the proposed solvers to increasing levels of image noise. Figure \[fig:exp\_tr\_only\] right shows that the new solvers have approximately the same noise sensitivity as R6P [@Albl-CVPR-2015].
#### [**Translational motion only:**]{} The advantage of solvers and is when the motion of the camera is purely translational, or close to it, which is a common scenario in, e.g., a moving car or a moving train. In such cases, both original R6P and provide significantly worse estimates of the camera pose. We explain this by the fact that and are constrained to estimate only camera translation in the initial step, whereas R6P and try to explain the image noise by the camera rotation. See Figure \[fig:exp\_tr\_only\] left. This fact can be used to create a ”joined solver” that runs both and and gives better performance than R6P [@Albl-CVPR-2015] while still being significantly faster.
![(Left) Purely translational camera motion, increasing on the x axis. Image noise with $\sigma$ 1pix. Notice that and now outperform all the others. (Right) Performance on general camera motion with increasing image noise.[]{data-label="fig:exp_tr_only"}](figs/v_err_inc_t_w0_nonplanar_noise1pix_RI.pdf "fig:"){width="0.40\columnwidth"} ![(Left) Purely translational camera motion, increasing on the x axis. Image noise with $\sigma$ 1pix. Notice that and now outperform all the others. (Right) Performance on general camera motion with increasing image noise.[]{data-label="fig:exp_tr_only"}](figs/v_err_inc_t_inc_noise_1pix_p3p_init.pdf "fig:"){width="0.45\columnwidth"}
#### [**Convergence:**]{} For , , and , the maximum 5 iterations might not be enough to converge to a good solution, whereas seems to perform at its best. We thus increased the maximum number of iterations. Figure \[fig:synth\_convergence\] (left) shows that the performance of , , and is improved by increasing the maximum number of iterations to 50. However, it is still far below the performance of R6P, , and R9P. performs as well as the R6P even with a single iteration, making it two orders of magnitude faster alternative. The algebraic error, evaluated on the equations (\[eq:model\_double\_lin\]), of the three viable solvers converges within 8 stpdf on average, see Figure \[fig:synth\_convergence\] (right).
![Testing the convergence of the iterative solvers. All iterative solvers have been run with 1, 5 and 50 iterations on data with ${\mathtt{R}}={\mathtt{I}}$ and increasing RS effect (left). Convergence of the algebraic error using the three viable iterative solvers (right).[]{data-label="fig:synth_convergence"}](figs/v_err_over_itreations.pdf "fig:"){width="0.4\columnwidth"} ![Testing the convergence of the iterative solvers. All iterative solvers have been run with 1, 5 and 50 iterations on data with ${\mathtt{R}}={\mathtt{I}}$ and increasing RS effect (left). Convergence of the algebraic error using the three viable iterative solvers (right).[]{data-label="fig:synth_convergence"}](figs/convergence.pdf "fig:"){width="0.54\columnwidth"}
#### [**The effect of linearized camera rotation model:**]{} Since all the proposed solvers have a linearized form of the camera orientation, in the same way as R6P [@Albl-CVPR-2015], we tested how being further from the linearization point affects the performance (Fig. \[fig:synth\_inc\_R\]). The camera orientation was set to be at a certain angle from ${\mathtt{R}}={\mathtt{I}}$. The camera velocities were set to 0.15 per frame for the translation and 15 degrees per frame for the rotation. In [@Albl-CVPR-2015] the authors show that R6P outperforms P3P in terms of camera center estimation up to 6 degrees away from the initial ${\mathtt{R}}$ estimate and up to 15 degrees away from ${\mathtt{R}}$ for the camera orientation estimate. Our results in Figure \[fig:synth\_inc\_R\] show similar behavior and identical results of R6P and . R9P performs comparable to both, even slightly outperforming them in terms of camera orientation estimation.
![Experiment showing the effect of the linearized camera pose which is present in all models. The further the camera orientation is from the linearization point, the worse are the results. matches the results of R6P and so does R9P.[]{data-label="fig:synth_inc_R"}](figs/c_err_Rinc_w_t.pdf "fig:"){width="0.45\columnwidth"} ![Experiment showing the effect of the linearized camera pose which is present in all models. The further the camera orientation is from the linearization point, the worse are the results. matches the results of R6P and so does R9P.[]{data-label="fig:synth_inc_R"}](figs/v_err_Rinc_w_t.pdf "fig:"){width="0.45\columnwidth"}
![Increasing the camera motion and estimating camera pose with all solvers being initialized with P3P. and R9P now provide consistently excellent results, comparable or outperforming those of R6P at a fraction of the computation cost. , and with 50 iterations now perform better than P3P, but still not as good as the other RS solvers. []{data-label="fig:synth_inc_R_p3p_init"}](figs/c_err_w_t_p3p_init_50.pdf){width="0.45\columnwidth"}
[**Using P3P as initial estimate:**]{} Last synthetic experiment shows the performance of the solvers when using the initial estimate of ${\mathtt{R}}$ from the result of P3P. The camera orientation was randomly generated and the camera motion was increased as in the first experiment. P3P was computed first and the 3D scene was pre-rotated using ${\mathtt{R}}$ from P3P. This shows probably the most practical usage among all R6P solvers. To make the figure more informative, we chose the number of iterations for , , and to be 50 as the 5 iterations already proved to be insufficient, see Figure \[fig:synth\_inc\_w\_t\_pose\]. We also set the maximum number of iterations for to 1, to demonstrate the potential of this solver.
As seen in Figure \[fig:synth\_inc\_R\_p3p\_init\], provides at least as good, or even better, results than R6P after only a single iteration. This is a significant achievement since the computational cost of is two orders of magnitude less than of R6P. With 50 iterations the other iterative solvers perform better than P3P, but considering the computational cost of 50 iterations, which is even higher than that of a R6P, we cannot recommend using them in such a scenario.
solver P3P R6P R9P
----------------------- ---------- -------------- ------------ ------------ ------------ ------------ ------------
time per iteration $3\mu s$ $1700 \mu s$ $10 \mu s$ $24 \mu s$ $30 \mu s$ $27 \mu s$ $20 \mu s$
max $\#$ of solutions 4 20 1 1 1 1 1
: Average timings on 2.5GHz i7 CPU per iteration for all used solvers.[]{data-label="tab:timings"}
#### [**Computation time:**]{} The computation times for all the tested solvers are shown in Table \[tab:timings\]. One iteration of is two orders of magnitude faster than R6P. According to the experiments, even one iteration of provides very good results, comparable with R6P and 5 iterations always match the results of R6P or even outperform them at 34$\times$ the speed. Note that R9P can be even faster than because it is non-iterative and runs only once and is therefore as fast as 2 iterations of . One iteration of , and is around three times slower than but still almost two orders of magnitude faster than R6P.
Real data {#sec:exp_real}
---------
We used the publicly available datasets from [@Hedborg2012] and we show the results of the same frames shown in [@Albl-CVPR-2015] (seq1, seq8, seq20 and seq22) in order to make a relevant comparison. We also added one more real dataset (House), containing high RS effects from a fast moving drone carrying a GoPro camera. The 3D-2D correspondences were obtained in the same way as in [@Albl-CVPR-2015] by reconstructing the scene using global shutter images and then matching the 2D features from the RS images to the reconstructed 3D points.
We performed RANSAC with 1000 iterations for each solver to estimate the camera pose and calculated the number of inliers. The inlier threshold was set to 2 pixels in the case of the data from [@Hedborg2012] which was captured by handheld iPhone at 720p and to 8 pixels for the GoPro footage which was recorder in 1080p. The higher threshold in the second case allowed to capture a reasonable number of inliers even for such fast camera motions.
\
\
\
\
\
The results in Figure \[fig:real\_inliers\] show the number of inliers captures over the sequences of images. We see that the performance of with 5 iterations is virtually identical to R6P. The results of and are also very similar and often outperform R6P and , except for the most challenging images in the House dataset.
The performance of is unstable, sometimes performing comparable to or below P3P. In seq20 in particular, there is almost exclusively a fast translational camera motion. The drop in performance can therefore be explained by being the only solver that does not estimate the translational velocity ${{\mathtt{t}}}$ in the first step. R9P performs solidly across all the experiments and on the most challenging House dataset it even provides significantly better results.
![Histogram of reprojection errors on the Aruco markers in the augmented reality experiment. The rolling shutter absolute pose solvers (R6P in magenta, in green, in cyan) keep the cube in place during camera motion whereas P3P (red) reprojects the cube all over the place.[]{data-label="fig:real_aruco_reproj"}](figs/cube_1.jpg "fig:"){height="0.295\columnwidth"} ![Histogram of reprojection errors on the Aruco markers in the augmented reality experiment. The rolling shutter absolute pose solvers (R6P in magenta, in green, in cyan) keep the cube in place during camera motion whereas P3P (red) reprojects the cube all over the place.[]{data-label="fig:real_aruco_reproj"}](figs/cube_3.jpg "fig:"){height="0.295\columnwidth"}
To test another useful case of camera absolute pose, which is augmented reality, we created an environment filled with Aruco [@aruco] markers in known positions. We set up the markers in such a way that they covered three perpendicular walls. The scene was recorded with a camera performing translational and rotational motion, similar to what a human does when looking around or shaking the head.
All solvers were used in RANSAC with 100 iterations to allow some robustness to outliers and noise. Note that 100 iterations of RANSAC would take at least 200ms for R6P excluding the inlier verification. That makes R6P not valuable for real time purposes (in practice only less than 10 iterations of R6P would give realtime performance). On the other hand, 100 runs of with 5 iterations take around 5ms (200fps) and takes around 12.5ms (80fps). We did not test solvers , and R9P in this experiment. This is because the performance of is unstable, the performance of is almost identical to and with R9P we do not have a way to extract the camera motion parameters and the reprojection without these parameters does not provide fair comparison.
We evaluated the reprojection error in each frame on all the detected markers. The results are shown in Figure \[fig:real\_aruco\_reproj\]. All the rolling shutter solvers outperform P3P in terms of precision of the reprojections. again provides identical performance to R6P. has a slight edge over the others, which is interesting, considering its poor performance on the synthetic data.
Figure \[fig:real\_aruco\_reproj\] gives a visualization of the estimated camera pose by reprojecting a cube in front of the camera. There is a significant misalignment between the cube and the scene during camera motion when using P3P pose estimate. In comparison, all the rolling shutter solvers keep the cube much more consistent with respect to the scene.
Conclusions
===========
We revisited the problem of rolling shutter camera absolute pose and proposed several new practical solutions. The solutions are based on iterative linear solvers that improve the current state-of-the-art methods in terms of speed while providing the same precision or better. The practical benefit of our solvers is also the fact that they provide only a single solution, compared to up to 20 solutions of R6P [@Albl-CVPR-2015].
The overall best performing solver needs only a single iteration to provide similar performance to R6P while being approximately 170x faster. At 5 iterations the performance of R6P is matched while the new solver is still approximately 34x faster than R6P. This allows for much broader applicability, especially in the area of augmented reality, visual SLAM and other real-time applications.
We also proposed 3 other iterative linear solvers (, , ) that alternate between estimating different camera pose and velocity parameters. These three solvers are slower than but still almost two orders of magnitude faster than R6P. While not as precise as R6P or in the synthetic experiments, they proved usefulness on the real data, providing more inliers and better reprojections than P3P and even R6P. We presented these three solvers mainly because they follow the concept of making the rolling shutter absolute pose equations linear by alternatively fixing some variables and then others. Although and do not offer the fastest and most precise results, they performed best in some of the experiments, especially for purely translational motion, and we think they are worth mentioning.
Last but not least we presented a non-iterative linear solver that uses 9 correspondences. This solver is as fast as 2 iterations of and proved to be the most precise in terms of estimated camera pose in the synthetic experiments and provided solid performance on the real data.
Altogether, this paper presents a big step forward in practical computation of rolling shutter camera absolute pose, making it more available in real world applications.
[^1]: This work was supported by the European Regional Development Fund under the project IMPACT (reg. no. CZ.02.1.01/0.0/0.0/15\_003/0000468), EC H2020-ICT-731970 LADIO project, ESI Fund, OP RDE programme under the project International Mobility of Researchers MSCA-IF at CTU No. CZ.02.2.69/0.0/0.0/17\_050/0008025, and Grant-in-Aid for Scientific Research (Grant No. 16H02851) of the Ministry of Education, Culture, Sports, Science and Technology of Japan. A part of this work was done when Zuzana Kukelova was visiting the National Institute of Informatics (NII), Japan, funded in part by the NII MOU grant.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop the theory of variable exponent Hardy spaces $H^{{p(\cdot)}}$. Analogous to the classical theory, we give equivalent definitions in terms of maximal operators. We also show that $H^{{p(\cdot)}}$ functions have an atomic decomposition including a “finite” decomposition; this decomposition is more like the decomposition for weighted Hardy spaces due to Strömberg and Torchinsky [@MR1011673] than the classical atomic decomposition. As an application of the atomic decomposition we show that singular integral operators are bounded on $H^{{p(\cdot)}}$ with minimal regularity assumptions on the exponent ${{p(\cdot)}}$.'
address:
- 'Department of Mathematics, Trinity College'
- 'Department of Mathematics, Trinity College'
author:
- 'David Cruz-Uribe, SFO'
- 'Li-An Daniel Wang'
bibliography:
- 'hardyvar.bib'
date: 'November 15, 2012'
title: Variable Hardy Spaces
---
[^1]
Introduction
============
Variable Lebesgue spaces are a generalization of the classical $L^p$ spaces, replacing the constant exponent $p$ with an exponent function ${{p(\cdot)}}$: intuitively, they consist of all functions $f$ such that $$\int_{{{\mathbb R}^n}}|f(x)|^{p(x)}\,dx < \infty.$$ These spaces were introduced by Orlicz [@0003.25203] in 1931, but they have been the subject of more intensive study since the early 1990s, because of their intrinsic interest, for their use in the study of PDEs and variational integrals with nonstandard growth conditions, and for their applications to the study of non-Newtonian fluids and to image restoration. (See [@cruz-fiorenza-book; @diening-harjulehto-hasto-ruzicka2010] and the references they contain.)
In this paper we extend the theory of variable Lebesgue spaces by studying the variable exponent Hardy spaces, or more simply, the variable Hardy spaces $H^{{p(\cdot)}}$. The classical theory of $H^p$ spaces, $0<p\leq 1$, is well-known (see [@MR0447953; @garcia-cuerva-rubiodefrancia85; @grafakos08b; @MR1342077; @stein93]) and our goal is to replicate that theory as much as possible in this more general setting. This has been done in the context of analytic functions on the unit disk by Kokilashvili and Paatashvili [@MR2294576; @MR2464049]. We are interested in the theory of real Hardy spaces in all dimensions. Here we give a broad overview of our techniques and results; we will defer the precise statement of definitions and theorems until the body of the paper.
Given an exponent function ${{p(\cdot)}}: {\mathbb R}^n \rightarrow (0,\infty)$, we define the space ${L^{p(\cdot)}}$; this is a quasi-Banach space. In the study of variable Lebesgue spaces it is common to assume that the exponent ${{p(\cdot)}}$ satisfies log-Hölder continuity conditions locally and at infinity. While these conditions will be sufficient for us, we prefer to work with a much weaker hypothesis: that there exists $p_0>0$ such that the Hardy-Littlewood maximal operator is bounded on $L^{{{p(\cdot)}}/p_0}$. This approach was first introduced in [@MR2210118] and developed systematically in [@cruz-fiorenza-book]. While in certain cases weaker hypotheses are possible, this appears to be the “right” universal condition for doing harmonic analysis in the variable exponent setting.
The variable Hardy space $H^{{p(\cdot)}}$ consists of all tempered distributions $f$ such that ${\mathcal{M}_N}f\in {L^{p(\cdot)}}$, where ${\mathcal{M}_N}$ is the grand maximal operator of Fefferman and Stein. We show that an equivalent definition is gotten by replacing the grand maximal operator with a maximal operator defined in terms of convolution with a single Schwartz function or with the non-tangential maximal operator defined using the Poisson kernel. This proof follows the broad outline of the argument in the classical case, but differs in many technical details. Here we make repeated use of the fact that the maximal operator is bounded on $L^{{{p(\cdot)}}/p_0}$.
We next prove an atomic decomposition for distributions in $H^{{p(\cdot)}}$. Given ${{p(\cdot)}}$ and $q$, $1<q\leq \infty$, we say that a function $a(\cdot)$ is a $({{p(\cdot)}},q)$ atom if there is a ball $B$ such that $\operatorname{supp}(a)\subset B$, $$\|a\|_q \leq |B|^{1/q}\|\chi_B\|^{-1}_{{p(\cdot)}}, \qquad
\text{and }
\qquad \int a(x)x^\alpha\,dx =0$$ for all multi-indices $\alpha$ such that $|\alpha|$ is not too large. We then show that $f\in H^{{p(\cdot)}}$ if and only if for $q$ sufficiently large there exist $({{p(\cdot)}},q)$ atoms $a_j$ such that $$\label{eqn:intro1}
f = \sum_j \lambda_j a_j,$$ and $$\label{eqn:intro2}
\|f\|_{H^{{p(\cdot)}}} \approx \inf\left\{
\bigg\| \sum_j \lambda_j \frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}
\bigg\|_{{p(\cdot)}}: f = \sum_j \lambda_j a_j \right\},$$ where the infimum is taken over all possible atomic decompositions of $f$. This is very different from the classical atomic decomposition; it is based on the atomic decomposition developed for weighted Hardy spaces by Strömberg and Torchinsky [@MR1011673]. A comparable decomposition in the classical case is due to Uchiyama: see Janson and Jones [@MR671315]. Moreover, we are able to prove that for $q<\infty$, if the summation in is finite, the infimum in can be taken over finite decompositions. This “finite” atomic decomposition is a generalization of the result of Meda, [*et al.*]{} [@MR2399059] in the classical case. As part of our work we also prove a finite atomic decomposition theorem for weighted Hardy spaces, extending the results in [@MR1011673].
To construct our atomic decomposition we first adapt the Calderón-Zygmund decomposition of classical Hardy spaces to give a $({{p(\cdot)}},\infty)$ atomic decomposition. Here a key tool is a vector-valued inequality for the maximal operator, which in turn depends on the boundedness of maximal operator. For the case $q<\infty$ we also rely on the theory of weighted Hardy spaces and on the Rubio de Francia extrapolation theory for variable Lebesgue spaces developed in [@MR2210118] (see also [@cruz-fiorenza-book; @cruz-martell-perezBook]). Neither approach was sufficient in itself in this case. We were not able to extend the classical approach to prove half the equivalence in . On the other hand, while such an equivalence exists in the weighted case, extrapolation requires careful density arguments and we could not, [*a priori*]{}, find the requisite dense subsets needed to prove both inequalities in . Again, in applying extrapolation the key hypothesis is the boundedness assumption on the maximal operator.
Finally, we prove that convolution type Calderón-Zygmund singular integral operators with sufficiently smooth kernels are bounded on $H^{{p(\cdot)}}$. In our proof we make extensive use of the finite atomic decomposition in weighted Hardy spaces; this allows us to avoid the more delicate convergence arguments that are often necessary when using the “infinite” atomic decomposition (e.g., see [@garcia-cuerva-rubiodefrancia85]).
The remainder of this paper is organized as follows. In Section \[section:prelim\] we give precise definitions of variable Lebesgue spaces and state a number of results we will need in the subsequent sections. In Section \[section:maximal\] we characterize $H^{{p(\cdot)}}$ in terms of maximal operators. In Sections \[section:density\] and \[section:CZ\] we prove two technical results: that $L^1_{loc}$ is dense in $H^{{p(\cdot)}}$ and the Calderón-Zygmund decomposition for distributions in $H^{{p(\cdot)}}$. In Section \[section:p-infty-atoms\] we construct the $({{p(\cdot)}},\infty)$ atomic decomposition, and in Section \[section:pq-atoms\] we construct the atomic decomposition for $q<\infty$ and prove the finite atomic decompositions for both the variable and weighed Hardy spaces. This second decomposition is used in Section \[section:sio\], where we prove that singular integrals are bounded on $H^{{p(\cdot)}}$.
As we were completing this project we learned that the variable Hardy spaces had been developed independently by Nakai and Sawano [@Nakai20123665]. They prove the equivalent definitions in terms of maximal operators using another approach. They also define an atomic decomposition but one which is weaker than ours. They show that $\|f\|_{H^{{p(\cdot)}}}$ is equivalent to the infimum of $$\bigg\| \bigg(\sum_j \bigg(\lambda_j^{p_*}
\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}^{p_*}}\bigg)\bigg)^{1/p_*}
\bigg\|_{{p(\cdot)}},$$ where $p_*=\min(1, \operatorname*{ess\,inf}p(x) )$. In particular, if ${{p(\cdot)}}$ takes on values less than 1, this quantity is larger than that in . They prove that this is equivalent to only when $q=\infty$ and with the further assumption that ${{p(\cdot)}}$ is log-Hölder continuous. Using their atomic decomposition they prove that singular integrals are bounded, but again they must assume that ${{p(\cdot)}}$ is log-Hölder continuous.
Preliminaries {#section:prelim}
=============
In this section we give without proof some basic results about the variable Lebesgue spaces. Unless otherwise specified, we refer the reader to [@cruz-fiorenza-book; @dcu-af-crm; @diening-harjulehto-hasto-ruzicka2010; @MR1866056; @MR1134951] for proofs and further information. Let ${\mathcal P}={\mathcal P}({\mathbb R}^n)$ denote the collection of all measurable functions ${{p(\cdot)}}:
{\mathbb R}^n \rightarrow [1,\infty]$. Given a measurable set $E$, let $$p_-(E) = \operatorname*{ess\,inf}_{x\in E} p(x), \qquad p_+(E) = \operatorname*{ess\,sup}_{x\in E}
p(x).$$ For brevity we will write $p_-=p_-({\mathbb R}^n)$ and $p_+=p_+({\mathbb R}^n)$. Define the set $\Omega_\infty = \{ x\in {\mathbb R}^n : p(x)=\infty \}$. Then for ${{p(\cdot)}}\in {\mathcal P}$, the space ${L^{p(\cdot)}}={L^{p(\cdot)}}({\mathbb R}^n)$ is the collection of all measurable functions $f$ such that for some $\lambda>0$, $$\rho(f/\lambda) = \int_{{\mathbb R}^n \setminus \Omega_\infty}
\left(\frac{|f(x)|}{\lambda}\right)^{p(x)}\,dx + \lambda^{-1}\|f\|_{L^\infty(\Omega_\infty)}< \infty.$$ This becomes a Banach function space when equipped with the Luxemburg norm $$\|f\|_{{p(\cdot)}}= \inf\left\{ \lambda > 0 : \rho(f/\lambda)\leq 1
\right\}.$$ Given ${{p(\cdot)}}\in {\mathcal P}$, define the conjugate exponent ${{p'(\cdot)}}$ by the equation $$\frac{1}{p(x)}+\frac{1}{p'(x)} = 1,$$ with the convention that $1/\infty = 0$.
\[lemma:holder\] Given ${{p(\cdot)}}\in {\mathcal P}$, if $f\in {L^{p(\cdot)}}$ and $g\in L^{{p'(\cdot)}}$, $$\int_{{{\mathbb R}^n}}|f(x)g(x)|\,dx \leq C({{p(\cdot)}}) \|f\|_{{p(\cdot)}}\|g\|_{{p'(\cdot)}}.$$ Conversely for all $f\in {L^{p(\cdot)}}$, $$\|f\|_{{p(\cdot)}}\leq C({{p(\cdot)}})\sup \int_{{{\mathbb R}^n}}f(x)g(x)\,dx,$$ where the supremum is taken over all $g\in L^{{p'(\cdot)}}$ such that $\|g\|_{{p'(\cdot)}}\leq 1$.
\[lemma:imbed\] Let $E\subset {\mathbb R}^n$ be such that $|E|<\infty$. If ${{p(\cdot)}},\,{{q(\cdot)}}\in{\mathcal P}$ satisfy $p(x)\leq q(x)$ a.e., then $$\|f\chi_E\|_{{p(\cdot)}}\leq (1+|E|)\|f\chi_E\|_{{q(\cdot)}}.$$
To define the variable Hardy spaces we need to extend the collection of allowable exponents. For simplicity we restrict ourselves to spaces where ${{p(\cdot)}}$ is bounded. Let ${\mathcal P}_0$ denote the collection of all measurable functions ${{p(\cdot)}}: {\mathbb R}^n \rightarrow (0,\infty)$ such that $p_+<\infty$. With the same definition of the modular $\rho$ as above, we again define ${L^{p(\cdot)}}$ as the collection of measurable functions $f$ such that for some $\lambda>0$, $\rho(f/\lambda)<\infty$. We define $\|\cdot\|_{{p(\cdot)}}$ as before; if $p_-<1$ (the case we are primarily interested in) this is not a norm: it is a quasi-norm and ${L^{p(\cdot)}}$ becomes a quasi-Banach space. We will abuse terminology and refer to it as a norm.
The next four lemmas are proved exactly as in the case when $p_-\geq 1$.
\[lemma:homog-exp\] Given ${{p(\cdot)}}\in {\mathcal P}_0$, $p_+<\infty$, then for all $s>0$, $$\||f|^s\|_{{p(\cdot)}}= \|f\|_{s{{p(\cdot)}}}^s.$$
\[lemma:norm-mod\] Suppose ${{p(\cdot)}}\in {\mathcal P}_0$. Given a sequence $\{f_k\}\subset {L^{p(\cdot)}}$, $$\int_{{{\mathbb R}^n}}|f_k(x)|^{p(x)}\,dx \rightarrow 0$$ as $k\rightarrow \infty$ if and only if $\|f_k\|_{{p(\cdot)}}\rightarrow 0$.
\[lemma:monotone\] Suppose ${{p(\cdot)}}\in {\mathcal P}_0$. Given a sequence $\{f_k\}$ of ${L^{p(\cdot)}}$ functions that increase pointwise almost everywhere to a function $f$, $$\lim_{k\rightarrow \infty} \|f_k\|_{{p(\cdot)}}= \|f\|_{{p(\cdot)}}.$$
\[lemma:norm-mod2\] Suppose ${{p(\cdot)}}\in {\mathcal P}_0$. Given $f\in {L^{p(\cdot)}}$, if $\|f\|_{{p(\cdot)}}\leq 1$, $$\rho(f)^{1/p_-} \leq \|f\|_{{p(\cdot)}}\leq \rho(f)^{1/p_+};$$ if $\|f\|_{{p(\cdot)}}\geq 1$, $$\rho(f)^{1/p_+} \leq \|f\|_{{p(\cdot)}}\leq \rho(f)^{1/p_-}.$$
\[lemma:minkowski-low\] Given ${{p(\cdot)}}\in {\mathcal P}_0$, $p_-\leq 1$, then for all $f,\,g\in {L^{p(\cdot)}}$, $$\|f+g\|_{{p(\cdot)}}^{p_-} \leq \|f\|_{{p(\cdot)}}^{p_-}+\|g\|_{{p(\cdot)}}^{p_-}$$
Since ${{p(\cdot)}}/p_-\in {\mathcal P}$, by Lemma \[lemma:homog-exp\], convexity and Minkowski’s inequality for the variable Lebesgue spaces, $$\begin{gathered}
\|f+g\|_{{p(\cdot)}}^{p_-} = \||f+g|^{p_-}\|_{{{p(\cdot)}}/p_-} \leq
\||f|^{p_-}+|g|^{p_-}\|_{{{p(\cdot)}}/p_-} \\ \leq
\||f|^{p_-}\|_{{{p(\cdot)}}/p_-}+\||g|^{p_-}\|_{{{p(\cdot)}}/p_-} = \|f\|_{{p(\cdot)}}^{p_-}
+\|g\|_{{p(\cdot)}}^{p_-}.\end{gathered}$$
This lemma is false if $p_->1$, but in this case $\|\cdot\|_{{p(\cdot)}}$ is a norm and so Minkowski’s inequality holds. This will cause minor technical problems in the proofs below; we will generally consider the case $p_-\leq 1$ in detail and sketch the changes required for the other case.
The Hardy-Littlewood maximal operator
-------------------------------------
Given a function $f\in L^1_{loc}$ we define the maximal function of $f$ by $$Mf(x) =\sup_{Q\ni x} \avgint_Q |f(y)|\,dy,$$ where $\avgint_Q g\, dy = |Q|^{-1}\int_Q g\, dy$, and the supremum is taken over all cubes whose sides are parallel to the coordinate axes. Throughout, we will make use of the following class of exponents.
\[defn:MP0\] Given ${{p(\cdot)}}\in \mathcal{P}_0$, we say ${{p(\cdot)}}\in {M\mathcal{P}_0}$ if $p_- > 0$ and there exists $p_0$, $0 < p_0 <
p_-$, such that $\|Mf\|_{{{p(\cdot)}}/p_0} \leq C(n,{{p(\cdot)}},p_0)\|f\|_{{{p(\cdot)}}/p_0}$.
A useful sufficient condition for the boundedness of the maximal operator is log-Hölder continuity: for a proof, see [@cruz-fiorenza-book; @diening-harjulehto-hasto-ruzicka2010].
Given ${{p(\cdot)}}\in {\mathcal P}$, such that $1<p_-\leq p_+<\infty$, suppose that ${{p(\cdot)}}$ satisfies the log-Hölder continuity condition locally, $$\label{eqn:log-Holder}
|p(x) - p(y)| \leq \frac{C_0}{-\log(|x-y|)}, \qquad |x-y| < 1/2,$$ and at infinity: there exists $p_\infty$ such that $$\label{eqn:decay}
|p(x) - p_\infty| \leq \frac{C_\infty}{\log(e+|x|)}.$$ Then $ \|Mf\|_{{p(\cdot)}}\leq C(n,{{p(\cdot)}})\|f\|_{{p(\cdot)}}$.
We want to stress that while in practice it is common to assume that the exponent ${{p(\cdot)}}$ satisfies the log-Hölder continuity conditions, we will not assume this in our main results. For a further discussion of sufficient conditions for the maximal operator to be bounded, see [@cruz-fiorenza-book; @diening-harjulehto-hasto-ruzicka2010] and the references they contain.
\[lemma:max-up\] Given ${{p(\cdot)}}\in {\mathcal P}$, if the maximal operator is bounded on ${L^{p(\cdot)}}$, then for every $s>1$, it is bounded on $L^{s{{p(\cdot)}}}$.
This follows at once from Hölder’s inequality and Lemma \[lemma:homog-exp\]: $$\|Mf\|_{s{{p(\cdot)}}} = \|(Mf)^s\|_{{p(\cdot)}}^{1/s} \leq \|M(|f|^s)\|_{{p(\cdot)}}^{1/s}
\leq
C^{1/s} \||f|^s\|_{{p(\cdot)}}^{1/s}=C^{1/s}\|f\|_{s{{p(\cdot)}}}.$$
The following necessary condition is due to Kopaliani [@MR2341730]. It should be compared to the Muckenhoupt $A_p$ condition from the study of weighted norm inequalities. (See [@duoandikoetxea01; @garcia-cuerva-rubiodefrancia85].)
\[lemma:kopaliani\] Given ${{p(\cdot)}}\in {\mathcal P}$, if the maximal operator is bounded on ${L^{p(\cdot)}}$, then for every ball $B\subset {\mathbb R}^n$, $$\|\chi_B\|_{{p(\cdot)}}\|\chi_B\|_{{p'(\cdot)}}\leq C|B|.$$
The maximal operator also satisfies a vector-valued inequality. This result was proved using extrapolation in [@MR2210118]. (See also [@cruz-fiorenza-book; @cruz-martell-perezBook].)
\[lemma:max-vector\] Given ${{p(\cdot)}}\in {\mathcal P}$ such that $p_+<\infty$, if the maximal operator is bounded on ${L^{p(\cdot)}}$, then for any $r$, $1<r<\infty$, $$\bigg\| \bigg(\sum_k (Mf_k)^r\bigg)^{1/r}\bigg\|_{{p(\cdot)}}\leq C(n,{{p(\cdot)}},r) \bigg\| \bigg(\sum_k |f_k|^r\bigg)^{1/r}\bigg\|_{{p(\cdot)}}.$$
Our final lemma is a deep result due to Diening [@MR2166733; @diening-harjulehto-hasto-ruzicka2010].
\[lemma:diening\] Given ${{p(\cdot)}}\in {\mathcal P}$ such that $1< p_- \leq p_+<\infty$, the maximal operator is bounded on $L^{{p(\cdot)}}$ if and only if it is bounded on $L^{{p'(\cdot)}}$.
The maximal characterization {#section:maximal}
============================
In this section we define the variable Hardy spaces and give equivalent characterizations in terms of maximal operators. To state our results, we need a few definitions. Let ${\mathcal{S}}$ be the space of Schwartz functions and let ${\mathcal{S}}'$ denote the space of tempered distributions. We will say that a tempered distribution $f$ is bounded if $f*\Phi\in L^\infty$ for every $\Phi\in {\mathcal{S}}$. For complete information on distributions, see [@MR1681462; @MR0304972]. Define the family of semi-norms on $\|\cdot\|_{\alpha,\beta}$, $\alpha$ and $\beta$ multi-indices, on ${\mathcal{S}}$ by $$\|f\|_{a,b} = \sup_{x\in {\mathbb R}^n} |x^\alpha D^\beta f(x)|,$$ and for each integer $N>0$ let $${\mathcal{S}}_N = \big\{ f \in {\mathcal{S}}: \|f\|_{\alpha,\beta} \leq 1,
|\alpha|,|\beta| \leq N \big\}.$$ Given $\Phi$ and $t>0$, let $\Phi_t(x)=t^{-n}\Phi(x/t)$. We define three maximal operators: given $\Phi\in {\mathcal{S}}$ and $f\in {\mathcal{S}}'$, define the radial maximal operator $${M_{\Phi, 0}}f = \sup_{t > 0} |f \ast \Phi_t (x)|,$$ and for each $N>0$ the grand maximal operator, $${\mathcal{M}_N}f (x) = \sup_{\Phi \in {\mathcal{S}}_N} {M_{\Phi, 0}}f(x).$$ Finally, define the non-tangential maximal operator $${\mathcal{N}}f(x) = \sup_{|x-y|<t} |P_t*f(y)|,$$ where $P$ is the Poisson kernel $$P(x) =
\frac{\Gamma\left(\frac{n+1}{2}\right)}{\pi^{\frac{n+1}{2}}}\frac{1}{(1+|x|^2)^{\frac{n+1}{2}}}.$$
Our main result in this section is the following.
\[Thm-Max\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, for every $f\in {\mathcal{S}}'$ the following are equivalent:
1. there exists $\Phi\in {\mathcal{S}}$, $\int \Phi(x) dx \neq 0$, such that ${M_{\Phi, 0}}f \in {L^{p(\cdot)}}$;
2. for all $N>n/p_0 +n+1$, ${\mathcal{M}_N}f\in {L^{p(\cdot)}}$;
3. $f$ is a bounded distribution and ${\mathcal{N}}f \in {L^{p(\cdot)}}$.
Moreover, the quantities $\|{M_{\Phi, 0}}f\|_{{p(\cdot)}}$, $\|{\mathcal{M}_N}f\|_{{p(\cdot)}}$ and $\|{\mathcal{N}}f\|_{{p(\cdot)}}$ are comparable with constants that depend only on ${{p(\cdot)}}$ and $n$ and not on $f$.
If we choose $N$ sufficiently large, then by Theorem \[Thm-Max\] we can use any of these three maximal operators to given an equivalent definition of the variable Hardy spaces. To be definite we will use the grand maximal operator, but in the rest of the paper we will move between these three norms without comment.
\[Hp-defn\] Let ${{p(\cdot)}}\in {M\mathcal{P}_0}$. For $N>n/p_0 +n+1$, define the space $H^{{p(\cdot)}}$ to be the collection of $f\in {\mathcal{S}}'$ such that $\|f\|_{H^{{p(\cdot)}}}=\|{\mathcal{M}_N}f\|_{{p(\cdot)}}<\infty$.
The proof is similar to that of the corresponding result for real Hardy spaces: cf. [@MR1342077; @stein93]. The most difficult step is the implication $(1)\Rightarrow (2)$ which we will prove in Sections \[subsec:12a\] and \[sec:apriori\]. We will then prove $(2)\Rightarrow (1)$ in Section \[subsec:21\] and $(2)
\Rightarrow (3) \Rightarrow (1)$ in Section \[subsec:2-3\].
The implication $(1)\Rightarrow (2)$ {#subsec:12a}
------------------------------------
The proof requires two supplemental operators: the non-tangential maximal operator with aperture $1$, $${M_{\Phi, 1}}f(x) = \sup_{\substack{|x - y| < t\\ t > 0}} |f \ast \Phi_t (y)|,$$ and the tangential maximal operator, $${M_{\Phi,T}}f(x) = \sup_{\substack{y \in {\mathbb R}^n\\ t > 0}} |\Phi_t \ast f(x - y)|
\left( 1 + \frac{|y|}{t} \right)^{-T}.$$ Note that $T$ is a parameter in the definition of ${M_{\Phi,T}}$ and not just notation indicating the that this is a “tangential” operator.
We will prove this implication by proving three norm inequalities. First, if $N\geq T+n+1$, we will show that $$\label{eqn:Step-a}
\|{\mathcal{M}_N}f\|_{{p(\cdot)}}\leq C(n,\Phi) \|{M_{\Phi,T}}f\|_{{p(\cdot)}}.$$ Second, if $T>n/p_0$, we will show that $$\label{eqn:Step-cd}
\|{M_{\Phi,T}}f \|_{{p(\cdot)}}\leq C(n,T,{{p(\cdot)}},p_0)\|{M_{\Phi, 1}}f \|_{{p(\cdot)}}.$$ Finally, we will show that $$\label{eqn:Step-e}
\|{M_{\Phi, 1}}f\|_{{p(\cdot)}}\leq C({{p(\cdot)}},T)\|{M_{\Phi, 0}}f\|_{{p(\cdot)}}.$$ To prove this we will first make the [*a priori*]{} assumption that ${M_{\Phi, 1}}f\in {L^{p(\cdot)}}$; we will then show that this is always the case by showing that if ${M_{\Phi, 0}}f\in
{L^{p(\cdot)}}$, then holds with a constant that depends on $f$. This proof parallels the proof we just sketched; to emphasize this we will defer it to Section \[sec:apriori\] and organize it similarly.
### Proof of inequality {#proof-of-inequality .unnumbered}
The proof requires a lemma from [@MR1342077 Lemma 2.1].
\[lemma:lu\] Let $\Phi \in {\mathcal{S}}$, $\int \Phi(x)\,dx \neq 0$. Then for any $\Psi \in
{\mathcal{S}}$ and $T>0$, there exist functions $\Theta^s \in {\mathcal{S}}$, $0<s<1$, such that $$\Psi(x) = \int_0^1 \Phi_t * \Theta^s(x)\,dx$$ and for all $m\geq T+1$, $$\int_{{{\mathbb R}^n}}(1+|x|)^T |\Theta^s(x)|\,dx \leq C(\Phi,n) s^T
\|\Psi\|_{m+n,m}.$$
Fix $N\geq T+n+1$ and fix $\Psi \in S_N$. Then by the definition of the tangential maximal operator, by making the change of variables $w=z/t$, and by Lemma \[lemma:lu\], we get $$\begin{aligned}
|f*\Psi_t(x)|
& \leq \int_0^1 \int_{{{\mathbb R}^n}}|f*\Phi_{st}(x-z)||\Theta^s(z/t)|t^{-n}\,dz\,ds \\
& = \int_0^1 \int_{{{\mathbb R}^n}}|f*\Phi_{st}(x-z)|\left(1+\frac{|z|}{st}\right)^{-T} \\
& \qquad \qquad \times \left(1+\frac{|z|}{st}\right)^{T}|\Theta^s(z/t)|t^{-n}\,dz\,ds \\
& \leq {M_{\Phi,T}}f(x) \int_0^1 \int_{{{\mathbb R}^n}}\left(\frac{1}{s}+\frac{|z|}{st}\right)^{T}|\Theta^s(z/t)|t^{-n}\,dz\,ds \\
& \leq {M_{\Phi,T}}f(x) \int_0^1 s^{-T} \int_{{{\mathbb R}^n}}(1+|w|)^T|\Theta^s(w)|\,dw\,ds \\
& \leq C(\Phi,n) {M_{\Phi,T}}f(x)\|\Psi\|_{T+n+1,T+1} \\
& \leq C(\Phi,n) {M_{\Phi,T}}f(x).\end{aligned}$$ Given this pointwise inequality, we immediately get inequality .
### Proof of inequality {#proof-of-inequality-1 .unnumbered}
Our proof is adapted from [@MR1342077 Lemma 3.1]. Fix $x,\,y\in {\mathbb R}^n$ and $t>0$. Then for all $z\in B(x-y,t)$, $$|f*\Phi_t(x-y)| \leq {M_{\Phi, 1}}f(z).$$ Let $q=n/T>0$. Since $B(x-y,t) \subset B(x,|y|+t)$, we have that $$\begin{gathered}
|f*\Phi_t(x-y)|^q
\leq \avgint _{B(x-y,t)} {M_{\Phi, 1}}f(z)^q\,dz \\
\leq \frac{B(x,|y|+t)}{B(x-y,t)|} \avgint _{B(x,|y|+t)} {M_{\Phi, 1}}f(z)^q\,dz
\leq \left(1+\frac{|y|}{t}\right)^n M({M_{\Phi, 1}}(f)^q)(x).\end{gathered}$$ If we rearrange terms, then by our choice of $q$ we have that $$\left|f*\Phi_t(x-y)\left(1+\frac{|y|}{t}\right)^{-T}\right|^q \leq
M({M_{\Phi, 1}}(f)^q)(x).$$ Taking the supremum over all $y$ and $t$ we get that $${M_{\Phi,T}}f(x)^q \leq M({M_{\Phi, 1}}(f)^q)(x).$$ Therefore, by Lemmas \[lemma:homog-exp\] and \[lemma:max-up\], since ${{p(\cdot)}}\in {M\mathcal{P}_0}$ and $q=n/T<p_0$, $$\begin{gathered}
\|{M_{\Phi,T}}f\|_{{p(\cdot)}}= \|({M_{\Phi,T}}f)^q\|_{{{p(\cdot)}}/q}^{1/q} \leq
\|M({M_{\Phi, 1}}(f)^q)\|_{{{p(\cdot)}}/q}^{1/q} \\
\leq C({{p(\cdot)}},p_0,n,q) \|{M_{\Phi, 1}}(f)^q\|_{{{p(\cdot)}}/q}^{1/q} = C({{p(\cdot)}},p_0,n,q)\|{M_{\Phi, 1}}f\|_{{p(\cdot)}}.\end{gathered}$$ Since $q$ depends on $n$ and $T$, this gives us inequality .
### Proof of inequality {#proof-of-inequality-2 .unnumbered}
As we remarked above, we first assume that ${M_{\Phi, 1}}f \in {L^{p(\cdot)}}$. Our argument is very similar to that in Stein [@stein93 pp. 95–98].
Let $\lambda > 0$ be some large number; the precise value will be fixed below. Define $F = F_{\lambda} = \{ x : {\mathcal{M}_N}f (x) \leq \lambda
{M_{\Phi, 1}}f (x) \}$. Then by inequalities and , $$\| {M_{\Phi, 1}}(f) \cdot \chi_{F^c} \|_{{{p(\cdot)}}}
\leq \frac{1}{\lambda} \| {\mathcal{M}_N}(f) \chi_{F^c} \|_{{{p(\cdot)}}}
\leq \frac{1}{\lambda} \| {\mathcal{M}_N}(f) \|_{{{p(\cdot)}}}
\leq \frac{C_0}{\lambda} \| {M_{\Phi, 1}}(f) \|_{{{p(\cdot)}}},$$ where $C_0=C_0(n,\Phi,T,{{p(\cdot)}},p_0)$. Therefore, by Lemma \[lemma:minkowski-low\] (if $p_-<1$; the other case is treated similarly), $$\begin{gathered}
\| {M_{\Phi, 1}}f \|_{{{p(\cdot)}}}^{p_-}
\leq \| {M_{\Phi, 1}}(f) \cdot \chi_{F} \|_{{{p(\cdot)}}}^{p_-} + \| {M_{\Phi, 1}}(f) \cdot \chi_{F^c} \|_{{{p(\cdot)}}}^{p_-} \\
\leq \| {M_{\Phi, 1}}(f) \cdot \chi_F \|_{{{p(\cdot)}}}^{p_-} + \left( \frac{C_0}{\lambda} \right)^{p_-} \| {M_{\Phi, 1}}f \|_{{{p(\cdot)}}}^{p_-}.\end{gathered}$$ Fix $\lambda = 2^{1/p_-}C_0$; since we assumed that ${M_{\Phi, 1}}f \in
L^{{{p(\cdot)}}}$, we can rearrange terms to get $$\| {M_{\Phi, 1}}f \|_{{{p(\cdot)}}} \leq 2 \| {M_{\Phi, 1}}(f) \cdot \chi_F
\|_{{{p(\cdot)}}}.$$
To estimate the right-hand side, we will use the fact that there exists $c = c(p_0,\Phi,n,N,\lambda)$ such that for all $x\in F$, $$\label{Mone-MHL}
{M_{\Phi, 1}}f (x) \leq c M (({M_{\Phi, 0}}f)^{p_0}) (x) ^{1/p_0}.$$ (See [@stein93 p. 96].) Then again by Lemma \[lemma:homog-exp\] and since ${{p(\cdot)}}\in {M\mathcal{P}_0}$, $$\begin{gathered}
\| {M_{\Phi, 1}}f \cdot \chi_F \|_{{{p(\cdot)}}}
\leq c \| (M(({M_{\Phi, 0}}f)^{p_0}) )^{1/p_0} \|_{{{p(\cdot)}}} \\
\leq \|M(({M_{\Phi, 0}}f)^{p_0}) \|_{{{p(\cdot)}}/p_0}^{1/p_0}
\leq C \| ({M_{\Phi, 0}}f)^{p_0} \|_{{{p(\cdot)}}/p_0}^{1/p_0} = C\| {M_{\Phi, 0}}f \|_{{{p(\cdot)}}},\end{gathered}$$ where $C=C(p_0,\Phi,n,N,\lambda,{{p(\cdot)}},p_0)$. This completes the proof of inequality with the [*a priori*]{} assumption that ${M_{\Phi, 1}}f \in {L^{p(\cdot)}}$.
Proof that ${M_{\Phi, 1}}f\in {L^{p(\cdot)}}$ {#sec:apriori}
---------------------------------------------
To show that this [*a priori*]{} assumption holds, we adapt the argument briefly sketched in [@stein93 p. 97]. Fix a tempered distribution $f$ such that ${M_{\Phi, 0}}f\in {L^{p(\cdot)}}$. We define truncated versions of the operators used in the previous section. Hereafter, $\epsilon>0$ will be a positive parameter that will tend to $0$, and $L>0$ will be a constant that will depend on $f$ but will be independent of $\epsilon$. Define $$\begin{aligned}
{M_{\Phi, 0}^{\epsilon, L}}& = \sup_{t\geq 0} |f\ast \Phi_t(x)| \frac{t^L}{(t + \epsilon + \epsilon|x|)^L}, \\
{\mathcal{M}_N^{\epsilon, L}}f(x)
&= \sup_{\Phi \in S_N} M_{\Phi, 0}^{\epsilon, L} f(x), \\
{M_{\Phi, 1}^{\epsilon, L}}f(x)
&= \sup_{\substack{0 < t < 1/\epsilon\\ |x - y| < t}} |f \ast \Phi_t (y)| \frac{t^L}{(t + \epsilon + \epsilon|y|)^L} \\
{M_{\Phi, T}^{\epsilon, L}}f(x)
&= \sup_{\substack{y \in {\mathbb R}^n\\ 0 < t < 1/\epsilon}}
|\Phi_t \ast f(x - y)| \left( 1 + \frac{|y|}{t} \right)^{-T} \frac{t^L}{(t + \epsilon + \epsilon|x - y|)^L}.\end{aligned}$$ For every $t$ and $L$, $ \frac{t^L}{(t + \epsilon + \epsilon |x|)^L}$ increases to $1$ as $\epsilon\rightarrow 0$; in particular ${M_{\Phi, 1}^{\epsilon, L}}f$ increases pointwise to ${M_{\Phi, 1}}f$.
Our proof proceeds as follows. We start by showing that there exists $L=L(f,n,p_0)>0$ such that for every $\epsilon$, $0<\epsilon<1/2$, ${M_{\Phi, 1}^{\epsilon, L}}f \in {L^{p(\cdot)}}$. Next we will prove three inequalities. First, we will show that there exists $N= T+L+n+1$ such that for all $\epsilon>0$, $$\label{eqn:Step-a1}
\| {\mathcal{M}_N^{\epsilon, L}}f\|_{{p(\cdot)}}\leq C(\phi,n)\|{M_{\Phi, T}^{\epsilon, L}}f\|_{{p(\cdot)}}.$$ We will then show that $$\label{eqn:Step-cd1}
\|{M_{\Phi, T}^{\epsilon, L}}f\|_{{p(\cdot)}}\leq C(n,T,{{p(\cdot)}},p_0)\|{M_{\Phi, 1}^{\epsilon, L}}f \|_{{p(\cdot)}}.$$ Finally, we will show that if $$x \in F = F_\lambda^{\epsilon,L} = \{ x : {\mathcal{M}_N^{\epsilon, L}}f(x) < \lambda
{M_{\Phi, 1}^{\epsilon, L}}f(x) \},$$ then $$\label{eqn:Step-e1}
{M_{\Phi, 1}^{\epsilon, L}}f(x) \leq C(p_0,\Phi,n,N,L,\lambda)M({M_{\Phi, 0}}(f)^{p_0})(x)^{1/p_0}.$$ We can then repeat the argument used to prove inequality . First, using , and the fact that ${M_{\Phi, 1}^{\epsilon, L}}\in {L^{p(\cdot)}}$, we show that there exists $\lambda=\lambda(\Phi,n,T,{{p(\cdot)}},p_0)$ such that $$\|{M_{\Phi, 1}^{\epsilon, L}}f\|_{{p(\cdot)}}\leq 2\|{M_{\Phi, 1}}(f) \chi_F\|_{{p(\cdot)}}.$$ Then we can use to show that $$\|{M_{\Phi, 1}^{\epsilon, L}}f\|_{{p(\cdot)}}\leq C\|{M_{\Phi, 0}}f\|_{{p(\cdot)}}< \infty.$$ The constant $C$ is independent of $\epsilon$, and so by Fatou’s lemma, we get that $$\|{M_{\Phi, 1}}f\|_{{p(\cdot)}}\leq C(f)\|{M_{\Phi, 0}}f\|_{{p(\cdot)}}< \infty.$$ This completes the proof.
### Construction of the constant $L=L(f)$ {#construction-of-the-constant-llf .unnumbered}
Since $f\in {\mathcal{S}}'$, it is a continuous linear functional on ${\mathcal{S}}$. In particular, arguing as in Folland [@MR1681462 Proposition 9.10], we have that there exists $m>0$ (depending only on $f$) such that $$|f*\Phi_t(y)| \leq C(\Phi,f) \left(1+\frac{|y|}{t}\right)^m.$$ Assume $L>2m$. If $x$, $y$ and $t$ are such that $|x-y|<t<1/\epsilon$, then we have that $$\left(1+\frac{|y|}{t}\right)^m\frac{t^L}{(t+\epsilon+\epsilon|y|)^L}
\leq \epsilon^{-L}\left(\frac{1}{\epsilon}+\frac{|y|}{t}\right)^{m-L} \leq
\epsilon^{-L}\left(\frac{1}{\epsilon}+\frac{|y|}{t}\right)^{-L/2}.$$ But by the triangle inequality, $$1+\epsilon|x| < 1+\frac{|x|}{t} < 2+\frac{|y|}{t} <
\frac{1}{\epsilon}+\frac{|y|}{t}.$$ Combining these inequalities we get that $$|f*\Phi_t(y)| \frac{t^L}{(t+\epsilon+\epsilon|y|)^L} \leq
\epsilon^{-L} C(\Phi,f) (1+\epsilon|x|)^{-L/2}.$$ Fix $x$ and take the supremum over all such $y$ and $t$; this shows that $${M_{\Phi, 1}^{\epsilon, L}}f(x) \leq \epsilon^{-L} C(\Phi,f) (1+\epsilon|x|)^{-L/2}.$$ Finally, recall that $$(1+\epsilon|x|)^{-n} \leq \epsilon^{-n}(1+|x|)^{-n}
\leq \epsilon^{-n}C(n) M(\chi_{B(0,1)})(x);$$ hence, $${M_{\Phi, 1}^{\epsilon, L}}f(x) \leq \epsilon^{-3L/2} C(\Phi,f,n)
M(\chi_{B(0,1)})(x)^{\frac{L}{2n}}.$$ Fix $L$ so that $\frac{L}{2n}>\frac{1}{p_0}$. Since ${{p(\cdot)}}\in {M\mathcal{P}_0}$, by Lemmas \[lemma:homog-exp\] and \[lemma:max-up\], $$\|{M_{\Phi, 1}^{\epsilon, L}}f\|_{{p(\cdot)}}\leq
C\|M(\chi_{B(0,1)}\|_{\frac{L}{2n}{{p(\cdot)}}}^{\frac{L}{2n}} \leq
C\|\chi_B(0,1)\|_{\frac{L}{2n}{{p(\cdot)}}}^{\frac{L}{2n}}<\infty,$$ where $C=C(\Phi,f,n,\epsilon,L,{{p(\cdot)}},p_0)$.Even though this constant depends on $\epsilon$ it does not effect the above argument, which only used the qualitative fact that $\|{M_{\Phi, 1}^{\epsilon, L}}f\|_{{p(\cdot)}}<\infty$.
### Proof of inequality {#proof-of-inequality-3 .unnumbered}
We begin with an auxiliary estimate. Fix $s$, $0<s<1$ and as we did above, assume that $\epsilon<1/2$. Then we have that $$\begin{gathered}
\left( \frac{(s t + \epsilon + \epsilon |x - z|)^L}{(s t)^L} \right) \cdot
\frac{t^L}{(t + \epsilon + \epsilon |x|)^L}
= \left( \frac{ s t + \epsilon + \epsilon |x - z|}{s}
\cdot \frac{1}{t + \epsilon + \epsilon |x|} \right)^L \\
\leq s^{-L} \left( \frac{t + \epsilon + \epsilon |x|}{t + \epsilon +
\epsilon |x|}
+ \frac{\epsilon|z|}{t + \epsilon + \epsilon |x|} \right)^L
\leq\left( \frac{1}{s} + \frac{|z|}{st} \right)^L.\end{gathered}$$
Given this estimate we can argue exactly as in proof of inequality : if $N\geq T+L+n+1$ and $\Psi\in {\mathcal{S}}_N$, then we get that $$\begin{aligned}
& |f*\Psi_t(x)|\frac{t^L}{(t+\epsilon+\epsilon|x|)^L} \\
& \qquad \leq \int_0^1 \int_{{{\mathbb R}^n}}|f*\Psi_{st}(x-z)|\frac{t^L}{(t+\epsilon+\epsilon|x|)^L}
\frac{(st)^L}{(st +\epsilon+\epsilon|x-z|)^L} \\
& \qquad \qquad \times
\left(1+\frac{|z|}{st}\right)^{-T} \frac{(st +\epsilon+\epsilon|x-z|)^L}{(st)^L}
\left(1+\frac{|z|}{st}\right)^{T} \Theta^s(z/t)t^{-n}\,dz\,ds \\
& \qquad \leq {M_{\Phi, T}^{\epsilon, L}}f(x) \int_0^1 \int_{{{\mathbb R}^n}}\left(\frac{1}{s}+\frac{|z|}{st}\right)^{L+T}
\Theta^s(z/t)t^{-n}\,dz\,ds \\
& \qquad \leq C(\Phi,n) {M_{\Phi, T}^{\epsilon, L}}f(x) \|\Psi\|_{T+L+n+1,T+L+1} \\
& \qquad \leq C(\Phi,n) {M_{\Phi, T}^{\epsilon, L}}f(x).\end{aligned}$$ The desired inequality follows immediately.
### Proof of inequality {#proof-of-inequality-4 .unnumbered}
This proof is a straightforward modification of the proof of inequality . As before, for all $x,\,y\in {\mathbb R}^n$, $t>0$ and $\epsilon>0$, we have that $$|\Phi_t*f(x-y)| \frac{t^L}{(t+\epsilon+\epsilon|x-y|)^L} \leq
{M_{\Phi, 1}^{\epsilon, L}}f(z)$$ for all $x\in B(x-y,t)$. The proof now proceeds as before.
### Proof of inequality {#proof-of-inequality-5 .unnumbered}
Fix $ x \in F = \{ x : {\mathcal{M}_N^{\epsilon, L}}f(x) < \lambda
{M_{\Phi, 1}^{\epsilon, L}}f(x) \}$. Then by the definition of the truncated maximal operator, there exists $(y, t)$ with $t <1/\epsilon$ and $|x - y| <
t$, such that $$\label{eqn:saturate}
{M_{\Phi, 1}^{\epsilon, L}}f (x) \leq 2\,|f \ast \Phi_t (y)| \left(
\frac{t^L}{(t + \epsilon + \epsilon |y|)^L} \right).$$ Let $r>0$ be small; its precise value will be fixed below. If $x' \in
B(y, rt)$, then by the Mean Value Theorem, $$|f(x', t) - f(y, t)| \leq rt \sup_{|z - y| < rt} |\nabla_z f(z,
t)|,$$ where for brevity we write $f(y,t)=\Phi_t*f(y)$.
Inequality follows if we can prove that there exists $c = c(N, L,n,\Phi)$ such that $$\begin{aligned}
\label{Grad}
t \sup_{|z - y| < rt} |\nabla_z f(z, t)|
\leq c {\mathcal{M}_N^{\epsilon, L}}f (x) \cdot \frac{(t + \epsilon + \epsilon |y|)^L}{t^L}.\end{aligned}$$ For if holds, then for $x \in F$, and $x' \in B(y,rt)$, $$|f(x', t) - f(y, t)|
\leq c r{\mathcal{M}_N^{\epsilon, L}}f (x) \cdot \frac{(t + \epsilon + \epsilon |y|)^L}{t^L}
\leq c r \lambda {M_{\Phi, 1}^{\epsilon, L}}f (x) \cdot \frac{(t + \epsilon + \epsilon |y|)^L}{t^L}.$$ Now fix $r = r(N,L,n,\Phi, \lambda)$ so small that $cr\lambda \leq 1/4$. Then we have $$\begin{gathered}
|f(x', t)|
\geq \left| \ |f(y, t)| - c r\lambda {M_{\Phi, 1}^{\epsilon, L}}f (x)
\cdot \frac{(t + \epsilon + \epsilon |y|)^L}{t^L} \right| \\
\geq \left( \frac{1}{2} - c r\lambda \right) {M_{\Phi, 1}^{\epsilon, L}}f (x)
\geq \frac{1}{4} {M_{\Phi, 1}^{\epsilon, L}}f(x).
\end{gathered}$$ We can now get by taking the average over all such points $x'$: $$\begin{aligned}
{M_{\Phi, 1}^{\epsilon, L}}f (x)^{p_0}
&\leq 4^{p_0} |f(x', t)|^{p_0}\\
& = 4^{p_0} \frac{1}{|B(y, rt)|} \int_{B(y, rt)} |f(x', t)|^{p_0} dx' \\
&\leq 4^{p_0} \left(\frac{r + 1}{r} \right)^n \frac{1}{|B(x, (1 + r)t)|} \int_{B(x, (1 + r)t)} |f(x', t)|^{p_0} dx' \\
&\leq c(p_0, r, n) \frac{1}{|B(x, (1 + r)t)|} \int_{B(x, (1 + r)t)} {M_{\Phi, 0}}f (x')^{p_0} dx' \\
&\leq c(p_0, r, n) M ({M_{\Phi, 0}}(f)^{p_0}) (x).
\end{aligned}$$
To complete the proof it remains to show . We begin with some notation: if we set $\Phi^{(i)}= \frac{{\partial}\Phi}{{\partial}z_i}$, and $\Phi_t^{(i)} (z) = (\Phi^{(i)})_t (z)$, then $\frac{{\partial}}{{\partial}z_i}
(\Phi_t) (z) = \frac{1}{t} \Phi_t^{(i)} (z)$. Since $f \ast \Phi \in
C^{\infty}$ whenever $f \in \mathcal{S}'$ and $\Phi \in \mathcal{S}$, differentiating the convolution gives $$\frac{{\partial}}{{\partial}z_i} [f(z, t)] = f \ast \frac{{\partial}}{{\partial}z_i}(\Phi_t)
(z) = \frac{1}{t} f \ast \Phi^{(i)}_t (z).$$ Hence, we can rewrite the gradient term as $$\begin{aligned}
|t \nabla_z (f) (z, t)|
&= \left( \sum_{i = 1}^n |f \ast \Phi_t^{(i)} (z)|^2 \right)^{1/2}.
\end{aligned}$$ We multiply and divide the left-hand side by the terms needed to obtain the truncated operator: $$t|\nabla_z f (z, t)|
= \underbrace{t|\nabla_z f (z, t)| \cdot \frac{t^L}{(t +
\epsilon + \epsilon |z|)^L}}_{S(z, t)}
\cdot \underbrace{ \left( \frac{(t + \epsilon + \epsilon |z|)^L}{(t +
\epsilon + \epsilon |y|)^L} \right)}_{R(z, y)^L}
\cdot \left( \frac{(t + \epsilon + \epsilon |y|)^L}{t^L} \right).$$ Recall that we have fixed $x \in {\mathbb R}^n$ and $(y, t)$ so that holds, and fixed $z \in B(y, rt)$. Without loss of generality, we may assume that $r\leq 1$. We first estimate $S(z,t)$: $$\begin{gathered}
S(z, t) = |t \nabla_z f(z, t)| \cdot \frac{t^L}{(t + \epsilon + \epsilon |z|)^L}
= \left( \sum_{i = 1}^n |f \ast \Phi_t^{(i)} (z)|^2 \right)^{1/2}
\cdot \frac{t^L}{(t + \epsilon + \epsilon |z|)^L} \\
= \left( \sum_{i = 1}^n \left| f \ast \Phi_t^{(i)} (z)
\frac{t^L}{(t + \epsilon + \epsilon |z|)^L} \right|^2
\right)^{1/2} \\
\leq C(N,\Phi)\left(\sum_{i = 1}^n |M_{\Phi^{(t)}, 1}^{\epsilon, L} f (x)|^2
\right)^{1/2}
\leq c(N, \Phi,n) {\mathcal{M}_N^{\epsilon, L}}f(x).\end{gathered}$$ To see the first inequality, define the set of functions $\Psi=\Psi^{i,h}$ by $\Psi^{i,h}(x)= \Phi^{(i)}(x+h)$, $1\leq i \leq n$, $|h|\leq 2$. Since $z=x+th$ for some $h$ such that $|h|\leq 1+r \leq 2$, we have that $f*\Phi_t^{i}(z)=f*\Psi^{i,h}_t(z)$. Moreover, since the collection of functions $\Psi^{i,h}$ is sequentially compact in ${\mathcal{S}}$, there exists a constant $c=c(\Phi,N)$ such that $\|\Psi^{i,h}\|_{\alpha,\beta}\leq c$, $|\alpha|,\,|\beta|\leq N$. Hence, $c^{-1}\Psi^{i,h}\in {\mathcal{S}}_N$ and the desired inequality follows.
To estimate $R(z, y)$ we note that if $z \in B(y, rt)$, then $|z| <
|y| + rt$. Then, since we may assume that $\epsilon, r < 1$, $$R(z, y) \leq \left( \frac{t + \epsilon + \epsilon(|y| + rt)}{t +
\epsilon + \epsilon |y|} \right)
= 1 + \frac{\epsilon rt}{t + \epsilon + \epsilon |y|}
\leq 1 + \frac{\epsilon r t}{t} = 1 + \epsilon r \leq 2.$$ Taking the supremum over $z$, we get $$\sup_{|z - y| < rt} t |\nabla_z f(z, t)| \leq C 2^L {\mathcal{M}_N^{\epsilon, L}}f (x) \cdot \frac{(t + \epsilon + \epsilon |y|)^L}{t^L},$$ where $C=C(N,\Phi,n)$. This gives us .
The implication $(2)\Rightarrow(1)$ {#subsec:21}
-----------------------------------
Given any $\Phi \in {\mathcal{S}}$, there exists $c=c(\Phi)$ such that $c\Phi
\in {\mathcal{S}}_N$. Therefore, the radial maximal operator is always dominated pointwise by a constant multiple of the grand maximal operator; hence, $$\label{eqn:2imp1}
\|{M_{\Phi, 0}}f\|_{{p(\cdot)}}\leq C(\Phi)\|{\mathcal{M}_N}f\|_{{p(\cdot)}}.$$
The implication $(2) \Rightarrow (3) \Rightarrow (1)$ {#subsec:2-3}
-----------------------------------------------------
Suppose first that $(2)$ holds. Fix $f\in {H^{p(\cdot)}}$ and let $\Phi \in
{\mathcal{S}}$; then by and , ${M_{\Phi, 1}}f
\in {L^{p(\cdot)}}$. Moreover, for all $x\in {\mathbb R}^n$, $$\begin{gathered}
|f*\Phi(x)|^{p_0} \leq \inf_{|x-y|\leq 1} {M_{\Phi, 1}}f(y)^{p_0}
\leq C(n)\int_{B(x,1)} {M_{\Phi, 1}}f(y)^{p_0}\,dy \\
\leq C(n,{{p(\cdot)}},p_0) \|({M_{\Phi, 1}}f)^{p_0}\|_{{{p(\cdot)}}/p_0}\|\chi_{B(x,1)}\|_{({{p(\cdot)}}/p_0)'}.\end{gathered}$$ By Lemma \[lemma:imbed\], if $q = \operatorname*{ess\,sup}_x (p(x)/p_0)'$, then $$\|\chi_{B(x,1)}\|_{({{p(\cdot)}}/p_0)'} \leq (1+|B(x,1)|)
\|\chi_{B(x,1)}\|_q \leq C(n,{{p(\cdot)}},p_0).$$ Therefore, $f*\Phi \in L^\infty$; since this is the case for every $\Phi\in {\mathcal{S}}$, $f$ is a bounded distribution.
To show that ${\mathcal{N}}f\in {L^{p(\cdot)}}$, we use the fact [@stein93 p. 98] that the Poisson kernel can be written as $$P(x) = \sum_{k=0}^\infty 2^{-k} \Phi_{2^k}^k(x),$$ where $\{\Phi^k\}$ is a family functions in ${\mathcal{S}}$ with uniformly bounded seminorms. Fix $x$ and $y$ such that $|x-y|<t$. Then $$|f*P_t(y)| \leq \sum_{k=0}^\infty 2^{-k} |f*\Phi^k_{2^kt}(y)| \leq
\sum_{k=0}^\infty 2^{-k} M_{\Phi^k,1}f(x).$$ Taking the supremum over all such $y$ and $t$ we get that ${\mathcal{N}}f(x)$ is dominated by the right-hand side. Since the functions $\Phi^k$ are uniformly bounded, by the same argument as we used to prove we have that this inequality holds for $\Phi^k$ with a constant independent of $k$. Therefore, if $p_1\leq
1$, by Lemma \[lemma:minkowski-low\] and , $$\|Nf\|_{{p(\cdot)}}^{p_-} \leq \sum_{k=0}^\infty 2^{kp_-}\|M_{\Phi^k,1}f\|_{{p(\cdot)}}^{p_-}
\leq C(n,{{p(\cdot)}},p_0)\|{\mathcal{M}_N}f\|_{{p(\cdot)}}^{p_-}.$$ If $p_->1$, the same argument holds if we omit $p_-$ and use Minkowski’s inequality.
Now suppose that $(3)$ holds. Then there exists $\Phi\in
{\mathcal{S}}$, $\int \Phi(x)\,dx=1$ such that ${M_{\Phi, 0}}f(x) \leq c{\mathcal{N}}f(x)$ (see [@stein93 p. 99]). Condition $(1)$ follows immediately.
Density of $L^1$ in ${H^{p(\cdot)}}$ {#section:density}
====================================
To prove the atomic decomposition we need two facts about variable Hardy spaces that are of interest in their own right.
\[prop:complete\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, $H^{{{p(\cdot)}}}$ is complete with respect to $\|\cdot\|_{H^{{p(\cdot)}}}$.
First, if a sequence $\{f_k\}$ converges in $H^{{{p(\cdot)}}}$ with respect $\|\cdot\|_{H^{{p(\cdot)}}}$, then it converges in ${\mathcal{S}}'$. To see this, fix $\Phi\in
{\mathcal{S}}$; then $$\begin{gathered}
| \langle f, \Phi \rangle |^{p_0} = |f*\Phi(0)|^{p_0}
\leq \inf_{|y|\leq 1} {M_{\Phi, 1}}f(y) ^{p_0}
\leq C(n)\avgint_{B(0,1)} {M_{\Phi, 1}}f(y) ^{p_0}\,dy \\
\leq C(n,{{p(\cdot)}},p_0) \|{M_{\Phi, 1}}f\|_{{p(\cdot)}}^{p_0}
\|\chi_{B(0,1)}\|_{({{p(\cdot)}}/p_0)'}
\leq C(n,{{p(\cdot)}},p_0) \|{M_{\Phi, 1}}f\|_{{p(\cdot)}}^{p_0} .\end{gathered}$$ Our assertion follows at once.
To show that $H^{{p(\cdot)}}$ is complete we will consider the case $p_-\leq
1$; the case $p_->1$ is proved in essentially the same way. It will suffice to show that $H^{{p(\cdot)}}$ has the Riesz-Fisher property: given any sequence $\{f_k\}$ in $H^{{{p(\cdot)}}}$ such that $$\label{eqn:complete1}
\sum_k \|f_k\|_{H^{{p(\cdot)}}}^{p_-} < \infty,$$ the series $\sum f_k$ converges in $H^{{p(\cdot)}}$. (Cf. Bennett and Sharpley [@bennett-sharpley88]. The argument there is for normed spaces but holds for quasi-norms with the introduction of the exponent $p_-$ in .) Let $$F_j = \sum_{k=1}^j f_k;$$ then by Lemma \[lemma:minkowski-low\] and , the sequence $\{F_j\}$ is Cauchy in $H^{{p(\cdot)}}$ and so in ${\mathcal{S}}'$. Therefore, it converges in ${\mathcal{S}}'$ to a tempered distribution $f$. Moreover, we have that $$\|f\|_{H^{{p(\cdot)}}}^{p_-} = \big\|\sum_k f_k\big\|_{H^{{p(\cdot)}}}^{p_-} \leq
\sum_k \|f_k\|_{H^{{p(\cdot)}}}^{p_-} < \infty,$$ and so $f\in H^{{p(\cdot)}}$. Finally, $$\|f-F_j \|_{H^{{p(\cdot)}}}^{p_-} \leq \sum_{k\geq j+1} \|f_k\|_{H^{{p(\cdot)}}}^{p_-},$$ and since the right-hand side tends to $0$ as $j\rightarrow \infty$, the series converges to $f$ in $H^{{p(\cdot)}}$.
\[Prop-Density\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, $H^{{{p(\cdot)}}} \cap L_{loc}^1$ is dense in $H^{{{p(\cdot)}}}$.
Given $f \in H^{{{p(\cdot)}}}$, by Theorem \[Thm-Max\], $f$ is a bounded distribution. Hence, for any $t > 0$, $f \ast P_t \in
C^{\infty}\subset L_{loc}^1$. Therefore, it will suffice to show $f
\ast P_t \rightarrow f$ in $H^{{{p(\cdot)}}}$. Again by Theorem \[Thm-Max\] it will be enough to show that as $t\rightarrow 0$, $$\|{\mathcal{N}}(f*P_t-f)\|_{{p(\cdot)}}\rightarrow 0.$$ Since $p_+<\infty$, by Lemma \[lemma:norm-mod\] this will follow if $$\label{eqn:density1}
\int_{{{\mathbb R}^n}}{\mathcal{N}}(f*P_t-f)(x)^{p(x)}\,dx \rightarrow 0.$$ Since $P_s \ast P_t = P_{s + t}$, we immediately have $${\mathcal{N}}(f*P_t-f)(x)^{p(x)}= \sup_{s > 0} |P_s \ast (P_t \ast f - f) (x)|^{p(x)} \leq 2^{p_+}{\mathcal{N}}f(x)^{p(x)} \in L^1.$$ Thus, follows from the dominated convergence theorem if for almost every $x$, $$\label{DCT-pw}
\lim_{t \rightarrow 0} \left( \sup_{s > 0} |P_s \ast( P_t \ast f - f)(x)|\right) = 0 .$$ To prove this, let $u(x,s)=P_s*f(x)$. Arguing as above we have that ${\mathcal{N}}u \in L^{{p(\cdot)}}$, and so $u(x,s)$ is non-tangentially bounded almost everywhere. Therefore, for almost every $x$, $$\lim_{s\rightarrow 0} u(x,s)$$ exists. (See [@garcia-cuerva-rubiodefrancia85 Theorem 4.21].) Moreover, by Lemmas \[lemma:kopaliani\] and \[lemma:norm-mod2\] we have that for all $s$ large, $$\begin{aligned}
|u(x,s)|^{p_0}= |P_s * f(x)|^{p_0}
& \leq \avgint_{B(x,s)} {\mathcal{N}}f(y)^{p_0}\,dy \\
& \leq C({{p(\cdot)}}) |B(x,s)|^{-1}\|Nf\|_{{p(\cdot)}}^{p_0} \|\chi_{B(x,s)}\|_{({{p(\cdot)}}/p_0)'} \\
& \leq C({{p(\cdot)}},p_0) \|\chi_{B(x,s)}\|_{{{p(\cdot)}}/p_0}^{-1}\|Nf\|_{{p(\cdot)}}^{p_0} \\
& \leq C({{p(\cdot)}},p_0) |B(x,s)|^{-p_0/p_+}\|Nf\|_{{p(\cdot)}}^{p_0} \\
& \leq C(n,{{p(\cdot)}},p_0) s^{-np_0/p_+}\|Nf\|_{{p(\cdot)}}^{p_0}.\end{aligned}$$ Thus, $$\lim_{s\rightarrow\infty} u(x,s)=0.$$ These two limits, combined with the fact that $u$ is continuous, show that for almost every $x$, $u(x,s)$ is uniformly continuous in $s$. The limit now follows immediately, and this completes our proof.
The Calderón-Zygmund decomposition {#section:CZ}
==================================
\[Thm-CZ\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, fix $f \in
L_{loc}^1 \cap H^{{p(\cdot)}}$. For each $\lambda > 0$ define $\Omega_{\lambda} = \{ x : {\mathcal{M}_N}f (x) > \lambda \}$. Then there exists a set of cubes $\{Q_k^{\ast} \}$ such that $$\begin{gathered}
\Omega_{\lambda} = \bigcup_{k} Q_k^{\ast} \label{eqn:CZ1}\\
\intertext{and}
\sum_k \chi_{Q_k^*}(x) \leq C. \label{eqn:CZ2}\end{gathered}$$ Moreover, we can write $f=g + b$, where $|g(x)| \leq c \lambda$, $b=\sum b_k$, $\operatorname{supp}(b_k)\subset
Q_k^*$, $\int b_k \ dx = 0$, and $$\label{Mr-Mg}
\| {M_{\Phi, 0}}b_k \|_{L^{{{p(\cdot)}}}({\mathbb R}^n)} \leq C\| {\mathcal{M}_N}f \|_{L^{{{p(\cdot)}}}(Q_k^{\ast})}.$$
Our proof is adapted from the proof for constant exponents in Stein [@stein93]. Since $\Omega_{\lambda}$ is open, we can form the Whitney decomposition of $\Omega_\lambda$. This gives us a set of cubes $\{Q_k\}$ with mutually disjoint interiors. Further, if we let $x_k$ and $\ell_k$ be the center and side length of $Q_k$, then there exist constants $1 < \tilde{a} < a^{*}$ such that if $\tilde{Q}_k = \tilde{a}Q_k$ and $Q_k^{*} = a^* Q_k$, then $Q_k \subset \tilde{Q}_k \subset Q_k^*$ and and hold. Let $P_0 = [-1/2, 1/2]^n$ and let $\zeta$ be a smooth function such that $\zeta \big|_{Q_0} = 1$ and $\zeta=0$ outside $\tilde{a} Q_0$. Define $\zeta_k (x) = \zeta(\frac{x -
x_k}{\ell_k})$ and $\eta_k = \zeta_k / (\sum_j \zeta_j)$; then $\{ \eta_k \}$ is a partition of unity for $\Omega_{\lambda}$ subordinate to the cover $\{ \tilde{Q}_k \}$. Lastly, define $\tilde{\eta}_k =
\eta_k / (\int \eta_k dx)$.
Let $d=\lfloor n(1/p_0-1)\rfloor$. We first consider the case $d\leq 0$; then $p_-> p_0 > \frac{n}{n + 1}$, and by Lemma \[lemma:max-up\] the maximal operator is bounded on $L^{{{p(\cdot)}}\frac{n+1}{n}}$. Let $c_k = \langle f, \tilde{\eta}_k
\rangle$ and define $b_k = (f - c_k) \eta_k$, $b = \sum_k b_k$ and $g
= f - b$. Then $\int b_k\,dx =0$. Moreover (see Stein [@stein93 pp. 102-3]), $|g(x)| \leq c \lambda$ and if $x \in Q_k^*$, $$\begin{aligned}
{M_{\Phi, 0}}b_k (x) &\leq C {\mathcal{M}_N}f (x); \label{CZ-bk} \\
\intertext{if $x\in {\mathbb R}^n \setminus Q_k^*$,}
{M_{\Phi, 0}}b_k (x) &\leq C\lambda \frac{\ell_k^{n + 1}}{|x - x_k|^{n + 1}}. \label{CZ-bk2}\end{aligned}$$
It remains to prove . By Lemma \[lemma:minkowski-low\], $$\begin{aligned}
\| {M_{\Phi, 0}}b_k \|_{L^{{{p(\cdot)}}} ({\mathbb R}^n)}^{p_-}
&\leq \| {M_{\Phi, 0}}b_k \chi_{Q_k^*} \|_{{{p(\cdot)}}}^{p_-} +
\| {M_{\Phi, 0}}b_k \chi_{^c Q_k^*} \|_{{{p(\cdot)}}}^{p_-} = I_1 + I_2.\end{aligned}$$ By we immediately get $I_1 \leq C\| {\mathcal{M}_N}f \chi_{Q_k^*}
\|_{{{p(\cdot)}}}^{p_-} = C\|M_N f\|_{{L^{p(\cdot)}}(Q_k^*)}^{p_-}$.
To estimate $I_2$, let $B_0$ be the ball centered at $x$ with radius $c_n |x - x_k|$, with $c_n$ a dimensional constant such that $Q_k^{\ast} \subset B_0$. Then by the definition of $M$, $$\label{eqn:maxest-far}
M (\chi_{Q_k^*}) (x) \geq \frac{1}{|B_0|} \int_{B_0} \chi_{Q_k^*}
dx = \frac{|Q_k^*|}{|B_0|} \geq c(n) \frac{\ell_k^{n }}{|x - x_k|^{n }}.$$ Therefore, by inequality , Lemma \[lemma:homog-exp\], the boundedness of the maximal operator, and the fact that $Q_k^*\subset E_\lambda$, $$\begin{gathered}
I_2
= \| {M_{\Phi, 0}}b_k \chi_{Q_k^*} \|_{{{p(\cdot)}}}^{p_-}
\leq C \lambda \left\| \frac{\ell_k^{n + 1}}{|x - x_k|^{n + 1}}
\chi_{{\mathbb R}^n\setminus Q_k^*} \right\|_{{{p(\cdot)}}}^{p_-}
\leq C \lambda \left\| M (\chi_{Q_k^*}) \chi_{{\mathbb R}^n\setminus Q_k^*}
\right\|_{{{p(\cdot)}}\frac{n + 1}{n}}^{\frac{n + 1}{n}p_-} \\
\leq C \lambda \left\| \chi_{Q_k^*} \right\|_{{{p(\cdot)}}\frac{n + 1}{n}}^{\frac{n + 1}{n}p_-}
= C\lambda \| \chi_{Q_k^*} \|_{{{p(\cdot)}}}^{p_-}
\leq C \lambda \left\| \frac{{\mathcal{M}_N}f}{\lambda} \chi_{Q_k^*}
\right\|_{{{p(\cdot)}}}^{p_-}
= C \left\| {\mathcal{M}_N}f \chi_{Q_k^{\ast}} \right\|_{{{p(\cdot)}}}^{p_-}.\end{gathered}$$
Now suppose $d\geq 1$; We modify the above construction as follows. We have that the maximal operator is bounded on $L^{{{p(\cdot)}}\frac{n+d+1}{n}}$. Let $\mathcal{H}_d$ be the space of polynomials of degree at most $d$, considered as a subspace of the Hilbert space $L^2(Q_K^{\ast},
\tilde{\eta}_k dx)$. Let $c_k$ be the projection of $f$ onto $\mathcal{H}_d$: then for all $q\in \mathcal{H}_d$, $\langle f - c_k, q \eta_k \rangle = 0$. We again define $b_k = (f - c_k) \eta_k$, $b=\sum b_k$, and $g=f-b$. Then we have (see Stein [@stein93 pp. 104-5]) that $|g(x)|\leq
C\lambda$ and $$\begin{aligned}
{M_{\Phi, 0}}b_k (x) &\leq c {\mathcal{M}_N}f (x) \\
\intertext{ if $x \in Q_k^* $, and}
{M_{\Phi, 0}}b_k (x) &\leq c\lambda \frac{\ell_k^{n + d + 1}}{|x - x_k|^{n + d + 1}} \end{aligned}$$ if $x\in {\mathbb R}^n\setminus Q_k^*$. We now repeat the argument above. The estimate for $I_1$ is the same; the estimate for $I_2$ is nearly so: $$\begin{gathered}
I_2 = \| {M_{\Phi, 0}}b_k \cdot \chi_{Q_k^*} \|_{{{p(\cdot)}}}^{p_-}
\leq C \lambda \left\| \frac{\ell_k^{n + d + 1}
\chi_{^c Q_k^*}}{|x - x_n|^{n + d + 1}} \right\|_{{{p(\cdot)}}}^{p_-} \\
\leq C \lambda \left\| M (\chi_{Q_k^*}) \right\|_{{{p(\cdot)}}\frac{n + d + 1}{n}}^{\frac{n + d + 1}{n} p_-}
\leq C \lambda \left\| \chi_{Q_k^*} \right\|_{{{p(\cdot)}}\frac{n + d +
1}{n}}^{ \frac{n + d + 1}{n}p_-}
\leq C \left\| {\mathcal{M}_N}f \ \chi_{Q_k^{\ast}} \right\|_{{{p(\cdot)}}}^{p_-}.\end{gathered}$$ This completes the proof of .
The atomic decomposition: $({{p(\cdot)}},\infty)$ atoms {#section:p-infty-atoms}
=======================================================
We begin with the definition of atoms.
Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, and $q$, $1 < q \leq \infty$, a function $a(\cdot)$ is a $({{p(\cdot)}}, q)$ atom if $\operatorname{supp}(a)\subset
B=B(x_0,r)$ and it satisfies
1. $\| a \|_{q} \leq |B|^{\frac{1}{q}}\| \chi_B \|_{{{p(\cdot)}}}^{-1} $,\
2. ${\displaystyle}\int a(x) x^{\alpha} dx = 0$ for all ${\displaystyle}|\alpha| \leq \lfloor n ( p_0^{-1} - 1)\rfloor$.
In (i) we interpret $1/\infty=0$. These two conditions are called the size and vanishing moments conditions of atoms.
If $p_0>1$ (which can happen if $p_->1$), then $\lfloor n ( p_0^{-1} -
1)\rfloor<0$, and we interpret this to mean that no vanishing moments are required.
In the remainder of this section we consider the case $q=\infty$.
\[Thm-Atomic\] Suppose ${{p(\cdot)}}\in {M\mathcal{P}_0}$. Then a distribution $f$ is in $H^{{{p(\cdot)}}}({\mathbb R}^n)$ if and only if there exists a collection $\{ a_j \}$ of $({{p(\cdot)}}, \infty)$ atoms supported on balls $\{ B_j \}$, and non-negative coefficients $\{\lambda_j\}$ such that $$f = \sum_j \lambda_j a_j,$$ where the series converges in $H^{{p(\cdot)}}({\mathbb R}^n)$. Moreover, $$\label{eqn:atomic1}
\| f \|_{H^{{{p(\cdot)}}}} \simeq
\inf\left\{ \bigg\| \sum_j \lambda_j \frac{\chi_{B_j}}
{\| \chi_{B_j} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}} : f = \sum_j \lambda_j a_j \right\}.$$
\[cor-Atom-Norm\] As an immediate corollary we get that $({{p(\cdot)}},\infty)$ atoms are uniformly bounded in $H^{{p(\cdot)}}$. However, as we will see, unlike the classical case we will not use this fact to prove the boundedness of operators.
Theorem \[Thm-Atomic\] follows from two lemmas whose proof we defer momentarily.
\[lemma:atoms-in-Hp\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, suppose $\{a_j\} $ is a sequence of $({{p(\cdot)}}, \infty)$ atoms, supported on $B_j
= B(x_j, r_j)$, and $\{\lambda_j\}$ is a non-negative sequence that satisfies $$\label{eqn:atomHp1}
\bigg\| \sum_j \lambda_j
\frac{\chi_{B_j}}{\| \chi_{B_j} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}} < \infty.$$ Then the series $f = \sum_j \lambda_j a_j$ converges in $H^{{{p(\cdot)}}}$, and $$\begin{aligned}
\label{Atom-Coeff1}
\| f \|_{H^{{{p(\cdot)}}}} \leq
C (n, {{p(\cdot)}}, p_0)\bigg\| \sum_j \lambda_j \frac{\chi_{B_j}}{\| \chi_{B_j} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}}.
\end{aligned}$$
\[lemma:Atom-Stein\] Let ${{p(\cdot)}}\in {M\mathcal{P}_0}$. If $f \in H^{{{p(\cdot)}}}$, then there exist $({{p(\cdot)}},\infty)$ atoms $\{ a_{k, j} \}$, supported on balls $B_{k, j}$, and non-negative coefficients $\{ \lambda_{k, j} \}$ such that $$\begin{aligned}
\label{Atom-Decomp}
f = \sum_{k, j} \lambda_{k, j} a_{k, j}.
\end{aligned}$$ Moreover, $$\begin{aligned}
\label{Atom-Coeff2}
\bigg\| \sum_{k, j} \lambda_{k, j} \frac{\chi_{B_{k, j}}}
{\| \chi_{B_{k, j}} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}} \leq C (n, {{p(\cdot)}}, p_0)\| f \|_{H^{{{p(\cdot)}}}}.
\end{aligned}$$
By Lemmas \[lemma:atoms-in-Hp\] and \[lemma:Atom-Stein\], $f\in
{H^{p(\cdot)}}$ if and only if it has the desired atomic decomposition. Therefore, it remains to show that holds. Given $f \in H^{{{p(\cdot)}}}$, there exists an atomic decomposition such that holds. This shows that the $H^{{p(\cdot)}}$ norm of $f$ dominates the infimum of the atomic decomposition norms. To see the opposite inequality, given any decomposition $f = \sum_j
\lambda_j a_j$, holds. Since this is true for all atomic decomposition, we have that $\| f \|_{H^{{{p(\cdot)}}}}$ is majorized by the infimum of the atomic decomposition norms.
Throughout the rest of this section, let $d = \lfloor
n(\frac{1}{p_0} - 1) \rfloor$ and $\gamma = (n + d + 1)/n$. Since ${{p(\cdot)}}\in {M\mathcal{P}_0}$, $M$ is also bounded on $L^{\gamma{{p(\cdot)}}}$. For by definition, $d > n(\frac{1}{p_0} - 1) - 1$, and this is equivalent to $\frac{n + d + 1}{n} > \frac{1}{p_0}$. Thus by Lemma \[lemma:max-up\] we get the boundedness of $M$.
Fix $\Phi\in {\mathcal{S}}$ such that $\int \Phi\,dx \neq 0$ and $\operatorname{supp}(\Phi)\subset B(0,1)$. Fix atoms $\{ a_j \}$ with support $\{
B_j \}$ and coefficients $\{\lambda_j\}$ such that holds. Given $B = B(x_0, r)$, let $2B =
B(x_0, 2r)$. We consider the case $p_-<1$; if $p_-\geq 1$ the proof is essentially the same, omitting the exponent $p_-$. By Lemma \[lemma:minkowski-low\], $$\begin{gathered}
\| {M_{\Phi, 0}}f \|_{{{p(\cdot)}}}^{p_-} \lesssim \| \sum_{j} \lambda_j {M_{\Phi, 0}}(a_j) \|_{{{p(\cdot)}}}^{p_-} \\
\leq \underbrace{\bigg\| \sum_{j} \lambda_j {M_{\Phi, 0}}(a_j)
\cdot \chi_{2B_j} \bigg\|_{{{p(\cdot)}}}^{p_-}}_{I_1}
+ \underbrace{\bigg\| \sum_{j} \lambda_j {M_{\Phi, 0}}(a_j)
\cdot \chi_{(2B_j)^c}\bigg \|_{{{p(\cdot)}}}^{p_-}}_{I_2}.
\end{gathered}$$ We first estimate $I_1$. By the size condition on $({{p(\cdot)}},\infty)$ atoms, we have that $$\label{eqn:atom-size}
{M_{\Phi, 0}}a_j (x) \leq \|a_j\|_\infty \|\Phi\|_1 \leq c\| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1}.$$ Define $g_j = (\| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1} \lambda_j)^{p_0}
\chi_{B_j}$. If $x \in \chi_{2B_j}$, then by the definition of the maximal operator, $$M g_j (x) \geq (\| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1}
\lambda_j)^{p_0} \frac{1}{|2B_j|} \int_{2B_j} \chi_{B_j} dx
= 2^{-n}(\| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1} \lambda_j)^{p_0}.$$ Then by Lemmas \[lemma:homog-exp\] and \[lemma:max-vector\], $$\begin{gathered}
I_1
\leq C\bigg\| \sum_j \| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1} \lambda_j
\ \chi_{2B_j} \bigg\|_{{{p(\cdot)}}}^{p_-} \leq C \bigg\| \sum_j M
(g_j)^{1/p_0}
\bigg\|_{{{p(\cdot)}}}^{p_-} \\
= C \bigg\| \bigg( \sum_j M (g_j)^{1/p_0} \bigg)^{p_0}
\bigg\|_{\frac{{{p(\cdot)}}}{p_0}}^{\frac{p_-}{p_0}}
\leq C \bigg\| \bigg(
\sum_j g_j^{1/p_0} \bigg)^{p_0}
\bigg\|_{\frac{{{p(\cdot)}}}{p_0}}^{\frac{p_-}{p_0}}
= C \bigg\| \sum_j\| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1}
\lambda_j \chi_{B_j} \bigg\|_{{{p(\cdot)}}}^{p_-}.\end{gathered}$$ To estimate $I_2$, let $a$ be an atom supported on $B=B(x_0,r)$. Then arguing as in [@stein93 p. 106] we have for $x \in (2B)^c$ the pointwise estimate $$\begin{gathered}
\label{Atom-pw}
{M_{\Phi, 0}}a(x) \leq c \left( \frac{r}{|x - x_0|} \right)^{n + 1 + d}
\avgint_B a(y)\,dy \\
\leq \left( \frac{r}{|x - x_0|} \right)^{n + 1 + d} \|a\|_\infty
\leq c \left( \frac{r}{|x - x_0|} \right)^{n \gamma} \|\chi_{B} \|_{{{p(\cdot)}}}^{-1}. \end{gathered}$$ Now arguing as we did in the proof inequality , we have for each $j$ that $$\label{eqn:atom-pw2}
{M_{\Phi, 0}}(a_j) (x)\leq c \left( \frac{r_j}{|x - x_j|} \right)^{n\gamma} \| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1}
\leq c \| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1} M (\chi_{B_j})^{\gamma}.$$ We can now estimate $I_2$: by Lemmas \[lemma:homog-exp\] and \[lemma:max-vector\], $$\begin{gathered}
I_2 \leq \bigg\| \sum_j \lambda_j {M_{\Phi, 0}}(a_j)
\cdot \chi_{(2B_j)^c} \bigg\|_{{{p(\cdot)}}}^{p_-} \\
\leq c \bigg\| \sum_j \frac{\lambda_j}{\| \chi_{B_j} \|_{{{p(\cdot)}}}}
M (\chi_{B_j})^{\gamma} \bigg\|_{{{p(\cdot)}}}^{p_-}
= \bigg\| \bigg( \sum_j M \bigg( \frac{\lambda_j^{1/\gamma}}
{\| \chi_{B_j} \|_{{{p(\cdot)}}}^{1/\gamma}}
\chi_{B_j} \bigg) ^{\gamma} \bigg)^{1/\gamma} \bigg\|_{\gamma {{p(\cdot)}}}^{\gamma p_-} \\
\leq C \bigg\| \bigg( \sum_j \| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1}
\lambda_j\chi_{B_j} \bigg)^{1/\gamma} \bigg\|_{\gamma {{p(\cdot)}}}^{\gamma p_-}
= C \bigg\| \sum_j \| \chi_{B_j} \|_{{{p(\cdot)}}}^{-1} \lambda_j \chi_{B_j} \bigg\|_{{{p(\cdot)}}}^{p_-}.\end{gathered}$$
We will prove this result assuming $f \in H^{{{p(\cdot)}}} \cap
L_{loc}^1$; then by Proposition \[Prop-Density\] and a density argument (cf. [@stein93 p. 109]) we get it for arbitrary $f\in H^{{p(\cdot)}}$.
Fix such an $f$ and for every $j \in {\mathbb Z}$, let $E_{j} = \{ x : {\mathcal{M}_N}f (x)
> 2^j \}$. By Theorem \[Thm-CZ\] we have that $f = g^{j} + b^{j}$, where $|g^{j} (x)| \leq c 2^j$ and $b^{j} = \sum_k b_k^{j}$, with each $b_k^{j}$ supported on a cube $Q_{k}^{j \ast}$. These cubes have bounded overlap $E_j = \bigcup_k Q_{k}^{j \ast}$. Moreover, we have that $$\label{bj-0}
\lim_{j \rightarrow \infty} \| b^{j} \|_{H^{{{p(\cdot)}}}} = 0.$$ To show this we proceed as in the proof of Lemma \[lemma:atoms-in-Hp\] (again only considering the case $p_-<1$): $$\| b^{j} \|_{H^{{{p(\cdot)}}}}^{p_-}\leq
\underbrace{\bigg\| \sum_k {M_{\Phi, 0}}(b_k^{j})
\cdot \chi_{Q_{k}^{j \ast}}\bigg \|_{{{p(\cdot)}}}^{p_-}}_{I_1}
+ \underbrace{\bigg\| \sum_k {M_{\Phi, 0}}(b_k^{j}) \cdot \chi_{(Q_{k}^{j
\ast})^c} \bigg\|_{{{p(\cdot)}}}^{p_-}}_{I_2}.$$ We first estimate $I_1$: by we have that $$I_1 \leq c\| \sum_k {\mathcal{M}_N}f \cdot \chi_{Q_{k}^{j \ast}}
\|_{{{p(\cdot)}}}^{p_-}
\leq c \| {\mathcal{M}_N}f \cdot \chi_{E_j} \|_{{{p(\cdot)}}}^{p_-}.$$ The last term tends to $0$ as $j \rightarrow 0$: this follows by Lemma \[lemma:norm-mod\] and the dominated convergence theorem.
To estimate $I_2$, let $x_{k, j}$ and $\ell_{k, j}$ be the center and side length of $Q_k^{j \ast}$. Then arguing as we did for inequality , if $x\in (Q_{k}^{j \ast})^c$, then $$M (\chi_{Q_k^{j \ast}}) (x) \geq c\frac{\ell_{k, j}^n}{|x -
x_k|^n}.$$ Then by inequality and Lemma \[lemma:max-vector\], $$\begin{gathered}
I_2
= \bigg\| \sum_k {M_{\Phi, 0}}(b_k^{j}) \cdot \chi_{Q_k^{j \ast}}
\bigg\|_{{{p(\cdot)}}}^{p_-}
\leq c \bigg\| \sum_k 2^j \frac{\ell_{k, j}^{n + 1 + d}}{|x - x_k|^{n + d + 1}} \cdot \chi_{Q_k^{j \ast}} \bigg\|_{{{p(\cdot)}}}^{p_-} \\
\leq c 2^{j p_-} \bigg\| \sum_k M (\chi_{Q_k^{j \ast}})^{\gamma}
\bigg\|_{{{p(\cdot)}}}^{p_-}
= c 2^{j p_-} \bigg\| \bigg( \sum_k (M \chi_{Q_k^{j \ast}})^{\gamma} \bigg)^{1/\gamma} \bigg\|_{\gamma {{p(\cdot)}}}^{\gamma p_-} \\
\leq c 2^{jp_-} \bigg\| \bigg( \sum_k (\chi_{Q_k^{j \ast}})^{\gamma}
\bigg)^{1/\gamma} \bigg\|_{\gamma {{p(\cdot)}}}^{\gamma p_-}
= c \bigg\| \sum_k 2^j \chi_{Q_k^{j \ast}} \bigg\|_{{{p(\cdot)}}}^{p_-} \\
\leq c\| 2^j \chi_{\{ x : {\mathcal{M}_N}f (x) > 2^j \}} \|_{{{p(\cdot)}}}^{p_-} \leq c\| {\mathcal{M}_N}f \chi_{E_j} \|_{{{p(\cdot)}}}^{p_-}.\end{gathered}$$ As before, the last term goes to $0$ as $j\rightarrow \infty$. This proves the limit .
As a consequence of we have that $g_j\rightarrow f$ in norm (and so in ${\mathcal{S}}'$) as $j\rightarrow \infty$. Further, since $g_j\rightarrow 0$ uniformly as $j\rightarrow -\infty$, we have that $$f = \sum_j (g^{j + 1} - g^{j}).$$ From the proof of Theorem \[Thm-CZ\], let $\{\eta_k^j\}$ be the partition of unity for $E_j$ with $\operatorname{supp}(\eta_k^j)\subset Q_k^{j*}$. Since $g^{j+1}-g^j = b^{j+1}-b^j$, $\operatorname{supp}(g^{j+1}-g^j)\subset
E_j$. Therefore, we have that $$f = \sum_{j,k} (g^{j + 1} - g^{j})\eta_k^j.$$ We now want to show that this expression can be rewritten as sum of atoms. Our argument follows Stein [@stein93 pp.108–9], and since many details are the same, we omit them here. Again as in the proof of Theorem \[Thm-CZ\], define the projections $\mathcal{P}_k^{j} : \mathcal{S}' \rightarrow \mathcal{H}_d$, where $\mathcal{H}_d$ is the space of polynomials of degree at most $d$, thought of as a subspace of the Hilbert space $L^2(Q_k^{j*},\tilde{\eta}_k^k\,dx)$. Define the polynomials $c_k^j =
\mathcal{P}_k^j (f)$ and $c_{\ell}^{j + 1} = \mathcal{P}_{\ell}^{j +
1} (f)$, and let $c_{k, \ell} = \mathcal{P}_{\ell}^{j + 1} [(f -
c_{\ell}^{j + 1}) \eta_k^j]$. For each $j$, we can then write $$g^{(j + 1)} - g^{j} = b^{j} - b^{(j + 1)} = \sum_k (f - c_k^j) \eta_k^j - \sum_{\ell} (f - c_{\ell}^{j + 1}) \eta_{\ell}^{j + 1} = \sum_k A_k^j,$$ where $$A_k^j = (f - c_k^j) \eta_k^j - \left( \sum_{\ell} (f - c_{\ell}^{j + 1}) \eta_{\ell}^{j + 1} \right) \eta_k^j + \sum_{\ell} c_{k, \ell} \eta_{\ell}^{j + 1}.$$ There exists a ball $B_{k, j}=B(x_{k,j},c\ell_{k,j})$ containing the cube $Q_k^{j \ast}$ such that $|B_{k,j}|\leq c|Q_k^{j*}|$. Moreover we have that $|A_k^j|\leq c2^j$ and $A_k^j$ satisfies the moment conditions for $({{p(\cdot)}},\infty)$ atoms. Therefore, if we define $$\label{eqn:patch2}
a_{k, j} = A_k^j c^{-1}2^{-j} \| \chi_{B_{k, j}} \|_{{{p(\cdot)}}}^{-1}, \qquad
\lambda_{k, j} = c 2^j \| \chi_{B_{k, j}} \|_{{{p(\cdot)}}},$$ the $a_{k,j}$ are $({{p(\cdot)}},\infty)$ atoms and we have the decomposition . It converges in ${\mathcal{S}}'$, and so, arguing as in the proof of Proposition \[Prop-Density\], it converges in $H^{{p(\cdot)}}({\mathbb R}^n)$.
Finally, we prove . We consider the case $p_-<1$; if $p_-\geq
1$, modify the following argument by replacing $1/p_0$ by $q>1$. Since $|B_{k,j}|\leq c|Q_k^{j*}|$, $M(\chi_{Q_k^{j*}}) \geq
c\chi_{B_{k,j}}=c\chi_{B_{k,j}}^{p_0}$. Therefore, by Lemmas \[lemma:homog-exp\] and \[lemma:max-vector\], $$\begin{gathered}
\bigg\| \sum_{k, j} \frac{\lambda_{k, j}}{\| \chi_{B_{k, j}} \|_{{{p(\cdot)}}}}
\ \chi_{B_{k, j}} \bigg\|_{{{p(\cdot)}}}
\leq C \bigg\| \sum_{k, j}\big(2^{jp_0} M (\chi_{Q_k^{j*}})\big)^{1/p_0} \bigg\|_{{{p(\cdot)}}}\\
= C \bigg\| \bigg(\sum_{k, j} M (2^{jp_0}\chi_{Q_k^{j*}})^{1/p_0}
\bigg)^{p_0} \bigg\|_{{{p(\cdot)}}/p_0}^{1/p_0}
\leq C \bigg\| \bigg(\sum_{k, j} 2^j \chi_{Q_k^{j\ast}}
\bigg)^{p_0}\bigg\|_{{{p(\cdot)}}/p_0}^{1/p_0} \\
= C \bigg\| \sum_{k, j} 2^j \chi_{Q_k^{j\ast}} \bigg\|_{{{p(\cdot)}}}
\leq C \bigg\| \sum_{ j} 2^j \chi_{E_j} \bigg\|_{{{p(\cdot)}}}.\end{gathered}$$ If $x \in {\mathbb R}^n$, there exists a unique $j_0 \in {\mathbb Z}$ such that $2^{j_0}
< {\mathcal{M}_N}f (x) \leq 2^{j_0 + 1}$. Hence, $$\sum_j 2^j \chi_{E_j} (x) = \sum_{j \leq j_0} 2^j = 2^{j_0 + 1} \leq 2 {\mathcal{M}_N}f (x).$$ If we combine this with the previous estimate, we get .
The atomic decomposition: $({{p(\cdot)}},q)$ atoms {#section:pq-atoms}
==================================================
In this section we consider the atomic decomposition when $q<\infty$. Our first main result is that when $q$ is sufficiently large, the analog of Theorem \[Thm-Atomic\] holds. Furthermore, we show that in this case we can give a *finite* atomic decomposition of $H^{{{p(\cdot)}}}$ (Theorem \[Var-Thm\] below). Lastly, by minor modifications to the proof of Theorem \[Var-Thm\], we give a finite atomic decomposition of the weighted Hardy space $H^{p_0} (w)$ (Theorem \[lemma:Weighted-FinDecomp\] below). We use this to prove the boundedness of singular integral operators on $H^{{{p(\cdot)}}}$ in Section 8.
Infinite atomic decomposition using $({{p(\cdot)}}, q)$ atoms
-------------------------------------------------------------
We extend Theorem \[Thm-Atomic\] by giving an atomic decomposition using $({{p(\cdot)}}, q)$ atoms.
\[thm:qatoms\] Suppose ${{p(\cdot)}}\in {M\mathcal{P}_0}$. Then a distribution $f$ is in $H^{{p(\cdot)}}$ if and only if for $q>1$ sufficiently large, there exists a collection $\{a_j\}$ of $({{p(\cdot)}},q)$ atoms supported on balls $\{B_j\}$, and non-negative coefficients $\{\lambda_j\}$ such that $$f =\sum_j \lambda_j a_j,$$ where the series converges in $H^{{p(\cdot)}}({\mathbb R}^n)$. Moreover, $$\label{eqn:qatom1}
\| f \|_{H^{{{p(\cdot)}}}} \simeq
\inf\left\{ \bigg\| \sum_j \lambda_j \frac{\chi_{B_j}}
{\| \chi_{B_j} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}} : f = \sum_j \lambda_j a_j \right\}.$$
Denote the norm of the maximal operator by $\|M\|_{({{p(\cdot)}}/p_0)'}$. Then it suffices to take $q> \max(1,p_+,p_0(1+2^{n+3}\|M\|_{({{p(\cdot)}}/p_0)'}))$.
One half of the proof of Theorem \[thm:qatoms\] is immediate: since for any $q$, $1<q<\infty$, $|B|^{1/q}\|a\|_q \leq \|a\|_\infty$, $({{p(\cdot)}},\infty)$ atoms are $({{p(\cdot)}},q)$ atoms. Therefore, by Lemma \[lemma:Atom-Stein\], every function $f\in {H^{p(\cdot)}}$ can be written as the sum of $({{p(\cdot)}},q)$ atoms and $\|f\|_{H^{{p(\cdot)}}}$ has the desired bound. Note that in this case there are no restrictions on $q$. The heart of the proof, therefore, is to prove the converse.
\[lemma:q-atoms-in-Hp\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, there exists $q=q({{p(\cdot)}},p_0,n)>\max(p_+,1)$ such that if $\{a_j\} $ is a sequence of $({{p(\cdot)}}, q)$ atoms supported on $B_j
= B(x_j, r_j)$, and $\{\lambda_j\}$ is a non-negative sequence that satisfies $$\label{eqn:qatomHp1}
\bigg\| \sum_j \lambda_j
\frac{\chi_{B_j}}{\| \chi_{B_j} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}} < \infty,$$ then the series $f = \sum_j \lambda_j a_j$ converges in $H^{{{p(\cdot)}}}({\mathbb R}^n)$, and $$\begin{aligned}
\label{eqn:qatomHp2}
\| f \|_{H^{{{p(\cdot)}}}} \leq
C (n, {{p(\cdot)}}, p_0,q)\bigg\| \sum_j \lambda_j \frac{\chi_{B_j}}{\| \chi_{B_j} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}}.
\end{aligned}$$
To prove Lemma \[lemma:q-atoms-in-Hp\] we will adapt the proof of Rubio de Francia extrapolation in the setting of variable Lebesgue spaces. This was first proved in [@MR2210118] (see also [@cruz-fiorenza-book; @cruz-martell-perezBook]). We need more careful control of the constants than was given in the original proof, and so we will reproduce the key steps.
To apply extrapolation we need a version of Lemma \[lemma:q-atoms-in-Hp\] for weighted $H^p$ spaces. To state it we introduce some definitions and preliminary results. For complete information on the theory of weights, see [@duoandikoetxea01; @garcia-cuerva-rubiodefrancia85]. By a weight $w$ we will always mean a non-negative, locally integrable function that is positive almost everywhere. We will say that $w\in
A_1$ if $$[w]_{A_1} = \operatorname*{ess\,sup}_{x\in {\mathbb R}^n} \frac{Mw(x)}{w(x)} < \infty.$$ Equivalently, $w\in A_1$ if given any ball $B$, $$\avgint_B w(y) \,dy \leq [w]_{A_1} \operatorname*{ess\,inf}_{x\in B} w(x).$$ A weight satisfies the reverse Hölder inequality with exponent $s>1$, denoted by $w\in RH_s$, if for every cube $B$, $$\left(\avgint_B w(x)^s\,dx\right)^{1/s} \leq C\avgint_Q
w(x)\,dx;$$ the best possible constant is denoted by $[w]_{RH_s}$. Note that if $w\in RH_s$, then by Hölder’s inequality, $w\in RH_t$ for all $t$, $1<t<s$, and $[w]_{RH_t}\leq [w]_{RH_s}$. If $w\in A_1$, then $w\in
RH_s$, and we have sharp control over the exponent $s$.
\[lemma:sharp-RH\] Given $w\in A_1$, then $w\in RH_s$, where $s=1+\big(2^{n+2}[w]_{A_1})^{-1}$.
\[remark:rh-const\] This result is proved in [@MR2427454] (see also [@cruz-fiorenza-book]), where everything is done in terms of cubes instead of balls. However, because $w\in A_1$ is doubling, the reverse Hölder inequality holds for balls with same exponent; in this case the constant $[w]_{RH_s}$ depends on $n$ and $[w]_{A_1}$.
Given a weight $w\in A_1$ and $p_0>0$, the weighted Hardy space $H^{p_0}(w)$ consists of all tempered distributions $f$ such that $$\|f\|_{H^{p_0}(w)}= \|{M_{\Phi, 0}}f\|_{L^{p_0}(w)} = \left(\int_{{{\mathbb R}^n}}{M_{\Phi, 0}}f(x)^{p_0}
w(x)\,dx\right)^{1/p_0} < \infty.$$ These spaces have an atomic decomposition: see Strömberg and Torchinsky [@MR1011673]. We state their result in the form we need to apply it; see Remark \[remark:atom-switch\] below.
\[lemma:ST-atoms\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$ and $q>\max(p_0,1)$, suppose $\{a_j\}$ is a sequence of $({{p(\cdot)}},q)$ atoms, $\{\lambda_j\}$ is a non-negative sequence, and $w\in A_1 \cap
RH_{(q/p_0)'}$. If $$\bigg\| \sum_j \lambda_j \frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg\|_{L^{p_0}(w)} <
\infty.$$ Then the series $$f = \sum_j \lambda_j a_j$$ converges in $H^{p_0}(w)$ and $$\|f\|_{H^{p_0}(w)} \leq C ({{p(\cdot)}},p_0,q,n, [w]_{A_1},[w]_{RH_{(q/p_0)'}})\bigg\| \sum_j \lambda_j
\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg\|_{L^{p_0}(w)}.$$
\[remark:atom-switch\] In [@MR1011673 Chapter VIII, Theorem 1] this result is stated for atoms $\bar{a}_j$ that (obviously) do not depend on a variable exponent ${{p(\cdot)}}$. To pass between the two kinds of atoms, it suffices to take $\bar{a}_j=\|\chi_{B_j}\|_{{p(\cdot)}}a_j$ and $\bar{\lambda}_j = \lambda_j\|\chi_{B_j}\|_{{p(\cdot)}}^{-1}$. The atoms $\bar{a}_j$ are required to have vanishing moments for $|\alpha| \leq \lfloor
d/p-n \rfloor$, where $d$ is a constant such that for all $t\geq 1$, $$w(B(x,tr)) \leq Kt^d w(B(x,r)).$$ If $w\in A_1$, then this is true with $d=n$: $$\begin{gathered}
w(B(x,tr)) \leq [w]_{A_1}|B(x,tr)| \operatorname*{ess\,inf}_{y\in B(x,tr)} w(y) \\
\leq [w]_{A_1} t^n |B(x,r)| \operatorname*{ess\,inf}_{y\in B(x,r)} w(y)
\leq [w]_{A_1} t^n w(B(x,r)).\end{gathered}$$
Fix ${{p(\cdot)}}\in {M\mathcal{P}_0}$; by Lemma \[lemma:diening\] the maximal operator is bounded on $L^{({{p(\cdot)}}/p_0)'}({\mathbb R}^n)$. Denote the norm of the maximal operator by $\|M\|_{({{p(\cdot)}}/p_0)'}$. Fix $q> \max(1,p_+,p_0(1+2^{n+3}\|M\|_{({{p(\cdot)}}/p_0)'}))$; the reason for this choice will be made clear below.
We will first show that if $a$ is a $({{p(\cdot)}},q)$ atom with support $B$, then $a\in H^{{p(\cdot)}}$. To do so we will show that $\|{M_{\Phi, 0}}a\|_{{p(\cdot)}}< \infty$. By Lemma \[lemma:minkowski-low\] (if $p_-<1$; the case $p_-\geq 1$ is handled similarly), $$\|{M_{\Phi, 0}}a\|_{{p(\cdot)}}^{p_-} \leq \|{M_{\Phi, 0}}(a)\chi_{2B}\|_{{p(\cdot)}}^{p_-}+
\|{M_{\Phi, 0}}( a)\chi_{(2B)^c}\|_{{p(\cdot)}}^{p_-} = I_1 + I_2.$$ By Lemma \[lemma:imbed\], since $q> \max(p_+,1)$ and ${M_{\Phi, 0}}$ is bounded on $L^q$, $$I_1= \|{M_{\Phi, 0}}(a)\chi_{2B}\|_{{p(\cdot)}}\leq (1+|2B|)\|{M_{\Phi, 0}}(a) \chi_{2B}\|_q
\leq C(1+|2B|)\|a\|_q < \infty.$$
To show that $I_2$ is finite, by inequality and the definition of $({{p(\cdot)}},q)$ atoms, and arguing as we did for , for $x\in (2B)^c$ we have that $$\begin{gathered}
{M_{\Phi, 0}}a(x) \leq c \left(\frac{r}{|x-x_0|}\right)^{n\gamma}
|B|^{-1/q}\|a\|_q \\
\leq \left(\frac{r}{|x-x_0|}\right)^{n\gamma}
\|\chi_B\|_{{p(\cdot)}}^{-1} \leq c \|\chi_B\|_{{p(\cdot)}}^{-1} M(\chi_B)(x)^\gamma,\end{gathered}$$ where $x_0$ is the center of $B$ and $\gamma= (n+d+1)/n$. As we noted in the proof of Theorem \[Thm-Atomic\], $M$ is bounded on $L^{\gamma{{p(\cdot)}}}$. Therefore, by Lemma \[lemma:homog-exp\], $$I_2= \|{M_{\Phi, 0}}a\chi_{(2B)^c}\|_{{p(\cdot)}}\leq
c\|\chi_B\|_{{p(\cdot)}}^{-1} \|M(\chi_B)\|_{\gamma {{p(\cdot)}}}^\gamma
\leq c\|\chi_B\|_{{p(\cdot)}}^{-1} \|\chi_B\|_{\gamma {{p(\cdot)}}}^\gamma < \infty.$$
To construct our weight $w$, form the Rubio de Francia iteration algorithm with respect to $L^{({{p(\cdot)}}/p_0)'}$. Given a function $h$, define $$\mathcal{R} h = \sum_{i=0}^\infty \frac{M^i f}{2^i \|M\|_{({{p(\cdot)}}/p_0)'}},$$ where $M^0 h = |h|$ and for $i\geq 1$, $M^i h = M\circ \cdots \circ M
h$ is $i$ iterates of the maximal operator. Three facts follow at once from this definition (cf. [@MR2210118; @cruz-martell-perezBook]):
1. $|h| \leq \mathcal{R} h$;
2. $\mathcal{R}$ is bounded on $L^{{{p(\cdot)}}/p_0)'}(\mathcal{R}^n)$ and $\|\mathcal{R}
h\|_{({{p(\cdot)}}/p_0)'} \leq 2\|h\|_{({{p(\cdot)}}/p_0)'}$;
3. $\mathcal{R} h \in A_1$ and $[\mathcal{R} h]_{A_1} \leq
2\|M\|_{({{p(\cdot)}}/p_0)'} = C({{p(\cdot)}},p_0,n)$.
By Lemma \[lemma:sharp-RH\] we have that $\mathcal{R} h\in RH_s$, where $s= 1+(2^{n+3}\|M\|_{({{p(\cdot)}}/p_0)'})^{-1}$. Therefore, since $q \geq p_0(1+2^{n+3}\|M\|_{({{p(\cdot)}}/p_0)'})$, we have that $\mathcal{R} h \in
RH_{(q/p_0)'}$ and $[\mathcal{R} h]_{Rh_{(q/p_0)'}}\leq C({{p(\cdot)}},p_0,n)$. We stress that all of these constants are independent of $h$.
Fix a sequence of atoms $\{a_j\}$ and constants $\{\lambda_j\}$ as in the hypotheses. Let $f =\sum \lambda_ja_j$; [*a priori*]{} we do not know that this series converges in $H^{{p(\cdot)}}$. To avoid this problem, define the functions $$f_k = \sum_{j=1}^k \lambda_j a_j.$$ Then $f_k \in H^{{p(\cdot)}}({\mathbb R}^n)$: since $a_j\in H^{{p(\cdot)}}$, by Lemma \[lemma:minkowski-low\] (if $p_-<1$, the case $p_-\geq 1$ is handled similarly) $$\|{M_{\Phi, 0}}f_k \|_{{p(\cdot)}}^{p_-} \leq \sum_{j=1}^k \lambda_j^{p_-}\|{M_{\Phi, 0}}a_j\|_{{p(\cdot)}}^{p_-}<\infty.$$ Furthermore, by Lemma \[lemma:ST-atoms\], given any function $h$, $f_k \in
H^{p_0}(\mathcal{R} h)$, and $$\label{eqn:ST-est}
\|f\|_{H^{p_0}(\mathcal{R} h)} \leq
C({{p(\cdot)}},p_0,n)\bigg\|\sum_{j=1}^k \lambda_j\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}
\bigg\|_{L^{p_0}(\mathcal{R} h)}.$$
We will now show that holds for each $f_k$ with a constant independent of $k$. By Lemmas \[lemma:holder\] and \[lemma:homog-exp\], $$\|{M_{\Phi, 0}}f_k \|_{{p(\cdot)}}^{p_0} = \|({M_{\Phi, 0}}f_k)^{p_0} \|_{{{p(\cdot)}}/p_0}
\leq C({{p(\cdot)}},p_0)\sup \int_{{{\mathbb R}^n}}{M_{\Phi, 0}}f_k(x)^{p_0} h(x)\,dx,$$ where the supremum is taken over all $h\in L^{({{p(\cdot)}}/p_0)'}$ with $\|h\|_{({{p(\cdot)}}/p_0)'}\leq 1$. (We may assume that $h$ is non-negative.) Fix such a function $h$; we will estimate the integral on the right-hand side with a constant independent of $h$. By the properties of the Rubio de Francia iteration algorithm, and Lemmas \[lemma:holder\] and \[lemma:homog-exp\], $$\begin{aligned}
\int_{{{\mathbb R}^n}}{M_{\Phi, 0}}f_k(x)^{p_0} h(x)\,dx
& \leq \int_{{{\mathbb R}^n}}{M_{\Phi, 0}}f_k(x)^{p_0} \mathcal{R} h(x)\,dx \\
& \leq C\int_{{{\mathbb R}^n}}\bigg(\sum_{j=1}^k
\lambda_j\frac{\chi_{B_j}(x)}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg)^{p_0} \mathcal{R} h(x)\,dx \\
& \leq C\bigg\|\bigg(\sum_{j=1}^k
\lambda_j\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg)^{p_0}\bigg\|_{{{p(\cdot)}}/p_0}\|\mathcal{R}
h\|_{({{p(\cdot)}}/p_0)'} \\
& \leq C\bigg\|\sum_{j=1}^k
\lambda_j\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg\|_{{{p(\cdot)}}}^{p_0}.\end{aligned}$$ Inequality for $f_k$ now follows and the constant depends only on ${{p(\cdot)}}$, $p_0$ and $n$.
To complete the proof we need to show that holds for $f$. But the same argument that proved this inequality for $f_k$ shows that if $l>k$, $$\label{eqn:cauchy}
\|f_l -f_k\|_{H^{{p(\cdot)}}({\mathbb R}^n)} \leq
C\bigg\|\sum_{j=k+1}^l
\lambda_j\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg\|_{{{p(\cdot)}}}.$$ However, by hypothesis we have that $$\bigg\|\sum_{j=1}^\infty
\lambda_j\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg\|_{{{p(\cdot)}}}<\infty$$ and therefore the partial sums of this series are Cauchy in $L^{{p(\cdot)}}({\mathbb R}^n)$. Hence, as $k,\,l \rightarrow \infty$, the right-hand side of tends to $0$. Therefore, the sequence $\{f_k\}$ is Cauchy in $H^{{p(\cdot)}}$ and so by Proposition \[prop:complete\] converges to $f$ in $H^{{p(\cdot)}}$. Therefore, by the monotone convergence theorem in variable Lebesgue spaces (Lemma \[lemma:monotone\]) we have that $$\begin{gathered}
\|f\|_{H^{{p(\cdot)}}({\mathbb R}^n)} = \lim_{k\rightarrow \infty} \|f_k
\|_{H^{{p(\cdot)}}({\mathbb R}^n)} \\
\leq C \lim_{k\rightarrow \infty}
\bigg\|\sum_{j=1}^k
\frac{\lambda_j}{\|\chi_{B_j}\|_{{p(\cdot)}}}\chi_{B_j}\bigg\|_{{{p(\cdot)}}}
=C \bigg\|\sum_{j}
\frac{\lambda_j}{\|\chi_{B_j}\|_{{p(\cdot)}}}\chi_{B_j}\bigg\|_{{{p(\cdot)}}}. \end{gathered}$$
Finite atomic decompositions
----------------------------
Given $q<\infty$, let $H^{{{p(\cdot)}},q}_{fin}$ be the subspace of $H^{{p(\cdot)}}$ consisting of all $f$ that have decompositions as finite sums of $({{p(\cdot)}},q)$ atoms. By Theorem \[Thm-Atomic\], if $q$ is sufficiently large, $H^{{{p(\cdot)}},q}_{fin}$ is dense in $H^{{p(\cdot)}}$. Our next result shows that on this subspace the atomic decomposition norm, restricted to finite decompositions, is equivalent to the $H^{{{p(\cdot)}}}$ norm. This extends a result from [@MR2399059] to the variable setting.
\[Var-Thm\] Let ${{p(\cdot)}}\in {M\mathcal{P}_0}$ and fix $q$ as in Theorem \[thm:qatoms\]. For $f \in H_{fin}^{{{p(\cdot)}}, q} ({\mathbb R}^n)$, define $$\label{eqn:VT1}
\|f\|_{H_{fin}^{{{p(\cdot)}}, q}} = \inf\bigg\{ \bigg\|\sum_{j=1}^k
\lambda_j\frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}}\bigg\|_{{p(\cdot)}}: f = \sum_{j=1}^k \lambda_j a_j \bigg\},$$ where infimum is taken over all finite decompositions of $f$ using $({{p(\cdot)}}, q)$ atoms $a_j$, supported on balls $B_j$. Then $$\| f \|_{H^{{{p(\cdot)}}}} \simeq \| f \|_{H_{fin}^{{{p(\cdot)}}, q}}.$$
Our argument is based on the proof of [@MR2492226 Theorem 6.2]. It requires two lemmas. The first introduces a non-tangential variant of the grand maximal operator. A proof can be found in Bownik [@MR1982689 Prop. 3.10].
\[lemma:bownik-NT\] Define the non-tangential grand maximal function $\mathcal{M}_{N,1}$ by $$\mathcal{M}_{N,1} f(x) \sup_{\Phi \in {\mathcal{S}}_N} \sup_{|y-x|<t}
|\Phi_t*f(x)|.$$ Then for all $x\in {\mathbb R}^n$ and tempered distributions $f$, $$\mathcal{M}_{N,1} f(x) \approx {\mathcal{M}_N}f(x),$$ where the constants depend only on $N$.
The second lemma is a decay estimate for the grand maximal operator.
\[lemma:decay-Mg\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, suppose $f\in H^{{p(\cdot)}}$ is such that $\operatorname{supp}(f)\subset B(0,R)$ for some $R>1$. Then for all $x\in
B(0,4R)^c$, $${\mathcal{M}_N}f(x) \leq C(N,{{p(\cdot)}},p_0)\|\chi_{B(0,R)}\|_{{p(\cdot)}}^{-1}.$$
To prove the desired estimate, it will suffice to show that for any $\Phi \in \mathcal{S}_N$, $x \in B(0, 4R)^c$, and $t > 0$, $$|f*\Phi_t(x)| \leq C\| \chi_{B(0, R)} \|_{{{p(\cdot)}}}^{-1},$$ where the constant $C$ is independent of $f$, $\Phi$, $x$ and $t$. We consider two cases, depending on the size of $t$.
[**Case 1: $t \geq R$.**]{} Given $x \in B(0, 4R)^c$ and $t \geq R$, we claim that there exists $\Psi \in \mathcal{S}$ so that $f \ast \Phi_t (x) = f \ast
\Psi_R (0)$. Let $\theta \in C_c^{\infty}$ be such that $\operatorname{supp}(\theta) \subset B(0, 2)$ and $\theta = 1$ on $B(0, 1)$, and define $\Psi(z) =
\Phi(\frac{x}{t} + \frac{Rz}{t}) \theta(z) (R/t)^n$. Then $$\begin{gathered}
f \ast \Phi_t (x)
= \int f(y) t^{-n} \Phi\left(\frac{x - y}{t}\right) dy \\
= \int f(y) t^{-n} \ \underbrace{t^{-n} \Phi \left(\frac{x - y}{t}\right)
\theta\left(\frac{y}{R}\right)}_{\Psi_R(0 - y)} dy
= f \ast \Psi_R (0).\end{gathered}$$ We actually have that $c\Psi \in \mathcal{S}_N$, where $c=c(\theta,N)$. To see this, recall that since $\Phi\in {\mathcal{S}}_N$, $\|\partial^\beta \Phi\|_\infty \leq c$ for all $|\beta| \leq N$. Fix $z \in \operatorname{supp}(\Psi) = B(0, 2)$. Then for any multi-index $|\beta| \leq N$, $$|{\partial}^{\beta} \Psi(z)|
\leq \left( \frac{R}{t} \right)^{n }
\sum_{\gamma \leq \beta} {\beta \choose \gamma}
\bigg| {\partial}^{\gamma} \Phi\left(\frac{x + Rz}{t}\right ) \left( \frac{R}{t} \right)^{\gamma }
{\partial}^{\beta - \gamma} \theta\left(\frac{y}{R}\right) R^{-|\beta| + |\gamma|} \bigg|.$$ Since $t\geq R>1$, we see that $|{\partial}^{\alpha} \psi(z)| \leq C(\theta,N)$. Hence, $$\sup_{|\alpha|,|\beta|\leq N} \| \Psi \|_{\alpha,\beta}
= \sup_{|\alpha|,|\beta| \leq N} \sup_{z \in B(0, 2)} |z^\alpha {\partial}^{\beta} \Psi(z)|
\leq C(\theta, N).$$ Since $c(N,\theta)\Psi \in {\mathcal{S}}_N$, by Lemma \[lemma:bownik-NT\] we have the pointwise bound $$\begin{gathered}
|f \ast \Phi_t (x)|
= |f \ast \Psi_R (0)|
\leq \inf_{z \in B(0, R)} M_{\Psi, 1} f(z) \\
\leq C(N,\theta)\inf_{z \in B(0, R)} \mathcal{M}_{N,1} f(z)
\leq C(N,\theta) \inf_{z \in B(0, R)} {\mathcal{M}_N}f(z).
\end{gathered}$$ Therefore, by Lemmas \[lemma:holder\], \[lemma:homog-exp\] and \[lemma:kopaliani\], $$\begin{gathered}
|f \ast \Phi_t (x)|^{p_0}
\leq C \avgint_{B(0, R)} {\mathcal{M}_N}f(z)^{p_0} dz \\
\leq C|B(0,R)|^{-1}\|({\mathcal{M}_N}f)^{p_0}\|_{{{p(\cdot)}}/p_0}\|\chi_{B(0,R)}\|_{({{p(\cdot)}}/p_0)'} \\
\leq C\|{\mathcal{M}_N}f\|_{{p(\cdot)}}^{p_0}\|\chi_{B(0,R)}\|_{{{p(\cdot)}}/p_0}^{-1}
\leq C \| f \|_{H^{{{p(\cdot)}}}}^{p_0} \| \chi_{B(0, R)} \|_{{{p(\cdot)}}}^{-p_0},\end{gathered}$$ where $C=C(N,{{p(\cdot)}},p_0,n)$.
[**Case 2: $t < R$.**]{} This case is similar to the previous one but we need to construct $\Psi$ differently as we need our estimate to hold at more points than the origin. Fix $z\in B(0,R/2)$ and choose $u\in B(0,R/2)$ such that $|z-u|<t$. We claim there exists $\Psi$ (depending on $u$, $t$ and $R$) such that $f \ast \Phi_t (x) = f \ast
\Psi_t (u)$. As before, let $\theta \in C_c^{\infty}$ be supported on $B(0, 2)$ and $\theta = 1$ on $B(0, 1)$. Define $\Psi$ by $$\Psi(z) = \Phi\left( \frac{x - u}{t} + z \right)
\theta\left( \frac{u}{R} - \frac{t}{R} z\right ).$$ Then we have that $$f \ast \Phi_t (x)
= \int f(y) \Phi_t (x - y) dy
= \int f(y) \underbrace{\Phi_t (x - y) \theta(y/R)}_{\Psi_t (u - y)} dy
= f \ast \Psi_t (u).$$ Assume for the moment that $c(\theta,N)\Psi \in {\mathcal{S}}_N$. Then by Lemma \[lemma:bownik-NT\], $$|f*\Psi_t(u)| \leq M_{\Psi,1} f(z) \leq C(\theta,N)
\mathcal{M}_{N,1}f(z) \leq C(\theta,N){\mathcal{M}_N}f(z).$$ Since this holds for every $z\in B(0,R/2)$, we have that $$|f \ast \Phi_t (x)| \leq C(\theta,N)\inf_{z\in B(0,R/2)} {\mathcal{M}_N}f(z),$$ and we can repeat the above argument to get the desired estimate.
It remains to show that $c(\theta,N)\Psi \in \mathcal{S}_N$; it will suffice to show that for all $\beta$ such that $|\beta| \leq N$, $${\displaystyle}\sup_{z \in {\mathbb R}^n} |{\partial}^{\beta} \psi(z)| (1 + |z|)^N \leq C(\theta,N).$$ Since $\Phi\in {\mathcal{S}}_N$, for all $|\beta|\leq N$, $(1+|y|)^N|\partial^\beta \Phi(y)| \leq c(N)$. Therefore, by the product rule, since $t<R$, $$\begin{gathered}
|{\partial}^{\beta} \Psi(z)|
\leq \sum_{\gamma\leq\beta}{\beta \choose \gamma}
\bigg| \partial^\beta \Phi\left(\frac{x-u}{t}+z\right)\bigg|
\left(\frac{t}{R}\right)^{|\beta|-|\gamma|}
\bigg|\partial^{\beta-\gamma}\theta\left(\frac{u}{R}-\frac{t}{R}z\right)\bigg|\\
\leq \frac{C(\theta,N)}{(1+|\frac{x-u}{t}+z|)^N}.\end{gathered}$$
To estimate the last term, note first that since $x \not\in B(0, 4R)$ and $u \in B(0, R/2)$, $$\frac{|x - u|}{t} > \frac{7}{2}\frac{R}{t}.$$ Second, since the $\theta$ term is non-zero only if $|\frac{u -
zt}{R}| \leq 2$, we must have that $|\frac{u}{t} - z| <
\frac{2R}{t}$, which implies $|z| < |\frac{u}{t}| + 2\frac{R}{t} =
\frac{5}{2}\frac{R}{t}$. Together these two estimates show that $|\frac{x - u}{t} + z| > \frac{7}{2}\frac{R}{t} -
\frac{5}{2}\frac{R}{t}
= \frac{R}{t}$. Therefore, for $z \in \operatorname{supp}(\Psi)$, $$|{\partial}^{\beta} \psi(z)| (1 + |z|)^N
\leq C \frac{(1 + |z|)^N}{(1 + |\frac{x - u}{t} + z|)^N} \leq C
\left( \frac{1 + 3R/t}{1 + R/t} \right)^N \leq C(\theta, N).$$ This completes the proof.
Since the infimum over finite sums in is larger than the infimum when taken over all possible atomic decompositions, by Theorem \[thm:qatoms\] we have that $\| f \|_{H^{{{p(\cdot)}}}}
\leq C\| f \|_{H_{fin}^{{{p(\cdot)}}, q}}$.
To prove the reverse inequality, fix $f \in H_{fin}^{{{p(\cdot)}}, q}$. By homogeneity we may assume that $\| f \|_{H^{{{p(\cdot)}}}} = 1$; we will show that $\| f \|_{H_{fin}^{{{p(\cdot)}}, q}} \leq C(N,{{p(\cdot)}},p_0,q,n)$. Since $f$ has a finite atomic decomposition, there exists $R>1$ such that $\operatorname{supp}(f)\subset B(0,R)$. By Lemma \[lemma:decay-Mg\], $$\label{Key-PW}
{\mathcal{M}_N}f (x) \leq c \| \chi_{B(0, R)} \|_{{{p(\cdot)}}}^{-1}.$$ Let $\Omega_j = \{ x : {\mathcal{M}_N}f(x) > 2^j \}$; define $j' = j'(f,
{{p(\cdot)}})$ to be the smallest integer such that for all $j> j'$, $\overline{\Omega}_j \subset B(0, 4R)$. By it suffices to take $j'$ to be the largest integer such that $2^{j'} < c \| \chi_{B(0, R)} \|_{{{p(\cdot)}}}^{-1}$. By Lemma \[lemma:Atom-Stein\] we can form the “canonical” decomposition of $f$ in terms of $({{p(\cdot)}},\infty)$ atoms: $$f = \sum_{j} \sum_k \lambda_{k,j} a_{k,j} = \sum_{j \leq j'} \sum_k
\lambda_{k,j} a_{k,j} +
\sum_{j> j'} \sum_k \lambda_{k,j} a_{k,j} = h + \ell.$$ We will rewrite the sum $h+\ell$ as a finite atomic decomposition in terms of $({{p(\cdot)}},q)$ atoms. To do so, we will use the finer properties of the atoms $a_{k,j}$ that are implicit in the proof of Lemma \[lemma:Atom-Stein\]. First, $\operatorname{supp}(h),\,\operatorname{supp}(\ell)\subset B(0,4R)$. The atoms $a_{k,j}$ are supported in $\Omega_j$, so by our choice of $j'$, $\operatorname{supp}(\ell)\subset B(0,4R)$. Since $\operatorname{supp}(f)\subset B(0,R)$, we also have that $\operatorname{supp}(h)\subset
B(0,4R)$.
Second, $h,\,\ell \in L^q$. Since $f$ has a finite $({{p(\cdot)}},q)$ atomic decomposition it is in $L^q$; since $q>1$ we also have that ${\mathcal{M}_N}f\in
L^q$. If we fix $x\in \operatorname{supp}(\ell)$, then there exists $s>j'$ such that $x \in \Omega_{s} \backslash \Omega_{s + 1}$. By construction (see ) the sets $\operatorname{supp}(a_{k,j})$ have bounded overlap and $|\lambda_{k,j} a_{k,j}|\leq c2^j$. Hence, $$\label{eqn:patch}
\sum_{j >j'} \sum_k |\lambda_{k,j} a_{k,j} (x)|
= \sum_{j'<j \leq s} \sum_k|\lambda_{k,j} a_{k,j} (x)|
\leq c \sum_{j \leq s} 2^j =c2^{s+1} \leq c {\mathcal{M}_N}f(x).$$ Thus $\ell \in L^q$, and so $h = f - \ell \in L^q$ as well.
Third, $h,\,\ell$ satisfy the vanishing moment condition for all $|\alpha|\leq \lfloor n(1/p_0-1) \rfloor$. Since $f$ is a finite sum of $({{p(\cdot)}},q)$ atoms, it has vanishing moments for these $\alpha$. Since $\operatorname{supp}(\ell)\subset B(0,4R)$, by Hölder’s inequality, $\ell
\in L^1$. Moreover, given any monomial $x^\alpha$, by $$\bigg\| \sum_{j> j'} \sum_k |x^\alpha||\lambda_{k,j} a_{k,j}| \bigg\|_{L^1}
\leq (4R)^{|\alpha|}\bigg\| \sum_{j > j'} \sum_k |\lambda_{k,j} a_{k,j}|
\bigg\|_{L^q}
\cdot |B(0, 4R)|^{1/q'} < \infty.$$ Thus the sum on the left-hand side converges absolutely in $L^1$ and so we can exchange sum and integral to get that $\ell$ has the same vanishing moments as each $a_{k,j}$. Finally, since $h = f - \ell$, $h$ also has the same vanishing moments.
Fourth, there exists a constant $c$ such that $ch$ is a $({{p(\cdot)}},
\infty)$ atom supported on $B(0, 4R)$. To show this we only need to check the size condition. Fix $x \in {\mathbb R}^n$; then by the same estimates for $a_{k,j}$ we used above, $$|h(x)| \leq \sum_{j \leq j'} \sum_k |\lambda_{k,j}
a_{k,j}(x)| \leq c \sum_{j \leq j'} 2^j \leq c2^{j'} \leq c \| \chi_{B(0, R)} \|_{{{p(\cdot)}}}^{-1},$$ where the last follows by our choice of $j'$.
Finally, we show that $\ell$ can be rewritten as a finite sum of $({{p(\cdot)}},q)$ atoms. Let $F_i = \{ (j, k) : |j| + |k| \leq i \}$ and define the finite sum $\ell_i$ by $$\ell_i= \sum_{F_i} \lambda_{k,j} a_{k,j}.$$ Since the sum for $\ell$ converges absolutely in $L^q$, we can find $i$ such that $\|\ell-\ell_i\|_q$ is as small as desired. In particular, we can find $i$ such that $\ell -
\ell_i$ is a $({{p(\cdot)}}, q)$ atom.
Therefore, $$f = c(h/c) + (\ell - \ell_i) + \sum_{(j, k) \in F_i} \lambda_{k,j} a_{k,j}$$ is a finite decomposition of $f$ as $({{p(\cdot)}},q)$ atoms. To complete the proof we will use Lemma \[lemma:Atom-Stein\] to get the desired estimate on $\|f\|_{H^{{{p(\cdot)}},q}_{fin}}$. Let $\tilde{B} = B(0, 4R)$. By the definition of the finite atomic norm and Lemma \[lemma:minkowski-low\] (if $p_-<1$), $$\begin{gathered}
\| f \|_{H_{fin}^{{{p(\cdot)}}, q}}^{p_-}
\leq \bigg\| \frac{c \chi_{\tilde{B}}}{\| \chi_{\tilde{B}}
\|_{{{p(\cdot)}}}^{p_-}}
+ \frac{\chi_{\tilde{B}}}{\| \chi_{\tilde{B}} \|_{{{p(\cdot)}}}}
+ \sum_{(j, k) \in F_i} \frac{\lambda_{k,j} \chi_{B_{k, j}}}{\|
\chi_{B_{k, j}} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}}^{p_-} \\
\leq c^{p_-}+1+\bigg\|\sum_{(j, k) \in F_i} \frac{\lambda_{k,j} \chi_{B_{k, j}}}{\|
\chi_{B_{k, j}} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}}^{p_-}
\leq C+\bigg\|\sum_{j, k} \frac{\lambda_{k,j} \chi_{B_{k, j}}}{\|
\chi_{B_{k, j}} \|_{{{p(\cdot)}}}} \bigg\|_{{{p(\cdot)}}}^{p_-}
\leq C+C\|f\|_{H^{{p(\cdot)}}({\mathbb R}^n)}
\leq C.\end{gathered}$$ This completes the proof.
Finite atomic decompositions for weighted Hardy spaces
------------------------------------------------------
We end this section by showing that a version of Theorem \[Var-Thm\] holds for the weighted Hardy spaces. This result is of interest in its own right, but we give it primarily because we will need it in the next section to prove the boundedness of singular integrals on $H^{{p(\cdot)}}$. For this reason we only prove one particular case; we leave it to the interested reader to prove the more general result implicit in our work.
Let ${{p(\cdot)}}\in {M\mathcal{P}_0}$, and let $q>1$. Given $w\in A_1$, define $H_{fin}^{p_0, q} (w)$ to be the set of all finite sums of $({{p(\cdot)}},q)$ atoms. By the proof of Lemma \[lemma:q-atoms-in-Hp\] we have that for $q$ sufficiently large, $H_{fin}^{{{p(\cdot)}}, q} ({\mathbb R}^n) =
H_{fin}^{p_0, q} (w)$ as sets. Given $f \in
H_{fin}^{p_0, q} (w)$, define a weighted atomic decomposition norm on $H_{fin}^{p_0, q} (w)$ by $$\| f \|_{H_{fin}^{p_0, q} (w)} = \inf \bigg\{ \bigg\|
\sum_{j = 1}^k \lambda_j^{p_0} \frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}^{p_0}}
\bigg\|_{L^1 (w)}^{1/p_0} :
f = \sum_{j = 1}^k \lambda_j a_j \bigg\},$$ where the infimum is taken over all decompositions of f as a finite sum of $({{p(\cdot)}},q)$ atoms.
\[lemma:Weighted-FinDecomp\] Given ${{p(\cdot)}}\in {M\mathcal{P}_0}$, fix $q$ as in Theorem \[thm:qatoms\] and let $w \in A_1 \cap L^{({{p(\cdot)}}/p_0)'} ({\mathbb R}^n)$. Then there exists $C = C({{p(\cdot)}}, p_0, [w]_{A_1}, \| w \|_{({{p(\cdot)}}/p_0)'})$ such that $$\| f \|_{H_{fin}^{p_0, q} (w)} \leq C \| f \|_{H^{p_0} (w)}.$$
We note in passing that Lemma \[lemma:Weighted-FinDecomp\] is not the same as [@MR2492226 Theorem 6.2] because the the atoms given there are defined using the weighted $L^q$-norm, and we cannot pass between these two types of atoms simply by multiplying by a constant.
The proof is very similar to the proof of Theorem \[Var-Thm\]; here we sketch the changes required. Fix $f \in H_{fin}^{p_0, q}
(w)$; then $f \in H_{fin}^{{{p(\cdot)}}, q}({\mathbb R}^n)$, and is supported on a ball $B=B(0, R)$ for some $R > 1$. Let $\tilde{B} = B(0, 4R)$. By Lemma \[lemma:decay-Mg\], for $x
\not\in \tilde{B}$, we have ${\mathcal{M}_N}f(x) \leq c \| \chi_{B}
\|_{{{p(\cdot)}}}^{-1}$.
Assume that $\| f\|_{H^{p_0} (w)} = 1$; we will show that $\| f \|_{H_{fin}^{p_0, q}
(w)} \leq C$. By the proof of [@MR1011673 Chapter 8, Theorem 1] we have that $$f = \sum_{k, j} \lambda_{k, j} a_{k, j}$$ where $\{a_{k,j}\}$ are $({{p(\cdot)}},\infty)$ atoms supported on balls $B_{k,j}$, $\{\lambda_{k,j}\}$ are non-negative, and $$\label{eqn:wtd-atom}
\bigg\| \sum_{k, j} \lambda_{k, j}^{p_0} \frac{\chi_{B_{k,j}}}{\|\chi_{B_{k, j}}\|_{{p(\cdot)}}^{p_0}}
\bigg\|_{L^1 (w)} \leq C \| f \|_{H^{p_0} (w)}^{p_0}.$$ (As we noted in Remark \[remark:atom-switch\], this is a restatement of the results from [@MR1011673] to our setting.) This decomposition is constructed in a fashion very similar to that of Lemma \[lemma:Atom-Stein\] and the atoms and coefficients have much the same properties. Therefore, if we let $j'$ be the smallest integer such that $2^{j'} \leq \|\chi_B\|_{{p(\cdot)}}^{-1}$ and regroup the sum as $$f = \sum_{j \leq j'} \sum_k {\lambda}_{k, j} {a}_{k, j}
+ \sum_{j > j'} \sum_k {\lambda}_{k, j} {a}_{k, j} = h + \ell,$$ the argument proceeds exactly as before. This allows us to write $$f = c (h/c) + (\ell - \ell_i) + \sum_{F_i} {\lambda}_{k, j} {a}_{k, j},$$ where $h$ is a $({{p(\cdot)}},\infty)$ atom, $i$ is chosen large enough that $(\ell-\ell_i)$ is $({{p(\cdot)}},q)$ atom, and the sum is a finite sum of $({{p(\cdot)}},\infty)$ atoms. Moreover, we have that $$\begin{gathered}
\| f \|_{H_{fin}^{p_0, q} (w)}^{p_0}
\leq \bigg\| c^{p_0}\frac{\chi_{\tilde{B}}}{\|\chi_{\tilde{B}}\|_{{p(\cdot)}}^{p_0}}
+ \frac{\chi_{\tilde{B}}}{\|\chi_{\tilde{B}}\|_{{p(\cdot)}}^{p_0}}
+ \sum_{k,j} {\lambda}_{k, j}^{p_0} \frac{\chi_{B_{k,
j}}}{\|\chi_{B_{k, j}}\|_{{p(\cdot)}}} \bigg\|_{L^1 (w)} \\
\leq C\frac{w(\tilde{B})}{\|\chi_{\tilde{B}}\|_{{p(\cdot)}}^{p_0}} +
\bigg \| \sum_{k, j} {\lambda}_{k, j}^{p_0} \chi_{B_{k, j}} \bigg\|_{L^1 (w)}.\end{gathered}$$ By , since the $\lambda_{k,j}$ are non-negative, the last term is bounded by $C\|f\|_{H^{p_0}(w)}^{p_0} =C$. To bound the first term, note that by Lemmas \[lemma:holder\] and \[lemma:homog-exp\], $$\begin{gathered}
w (\tilde{B}) = \int_{\tilde{B}} w (x) dx
\leq C({{p(\cdot)}},p_0)\| \chi_{\tilde{B}} \|_{{{p(\cdot)}}/p_0} \| w \|_{({{p(\cdot)}}/p_0)'} \\
\leq C ({{p(\cdot)}},p_0,\| w \|_{({{p(\cdot)}}/p_0)'} )\| \chi_{\tilde{B}} \|_{{{p(\cdot)}}}^{p_0}. \end{gathered}$$ Hence, $\| f \|_{H_{fin}^{p_0, q} (w)} \leq C({{p(\cdot)}}, p_0, [w]_{A_1}, \|
w \|_{({{p(\cdot)}}/p_0)'})$ and our proof is complete.
Boundedness of operators on $H^{{{p(\cdot)}}}$ {#section:sio}
==============================================
In this section we show that convolution type Calderón-Zygmund singular integrals with sufficient regularity are bounded on $H^{{p(\cdot)}}$. Our two main techniques are the finite atomic decomposition from Section \[section:pq-atoms\] and weighted norm inequalities. First we define the class of singular integrals we are interested in.
\[eqn:kregular\] Let $K \in \mathcal{S}'$. We say $Tf = K \ast f$ is a convolution-type singular integral operator with regularity of order $k$ if the distribution $K$ coincides with a function on ${\mathbb R}^n\setminus \{0\}$ and has the following properties:
1. $\hat{K} \in L^{\infty}$;
2. for all multi-indices $0 \leq |\beta| \leq k + 1$ and $x \neq 0$, $ {\displaystyle}|{\partial}^{\beta} K(x)| \leq \frac{C}{ |x|^{n+ |\beta|}}$.
Singular integrals that satisfy this definition are bounded on $L^p$, $1<p<\infty$. More importantly, the pointwise smoothness conditions guarantee that they satisfy weighted norm inequalities. In particular, we have the following weighted Kolmogorov inequality; for a proof, see [@duoandikoetxea01; @garcia-cuerva-rubiodefrancia85].
\[lemma:kolmogorov\] Let $T$ be a convolution-type singular integral operator as defined above. Given $w\in A_1$ and $0<p<1$, then for every ball $B$, $$\int_{B} |Tf(x)|^{p}w(x)\,dx
\leq C(T,n,p,[w]_{A_1})w(B)^{1-p}\left(\int_{{{\mathbb R}^n}}|f(x)|w(x)\,dx\right)^p.$$
Our main results in this section are the following two theorems.
\[thm:bdd-Lp\] Given ${{p(\cdot)}}\in M\mathcal{P}_0$ and $q>1$ sufficiently large (as in Theorem ), let $T$ be a singular integral operator that has regularity of order $k \geq \lfloor n (\frac{1}{p_0} - 1) \rfloor $. Then $$\|Tf\|_{{p(\cdot)}}\leq C(T,{{p(\cdot)}},p_0,q,n)\|f\|_{H^{{p(\cdot)}}}.$$
\[thm:bdd-Hp\] Given ${{p(\cdot)}}\in M\mathcal{P}_0$ and $q>1$ sufficiently large (as in Theorem ), let $T$ be a singular integral operator that has regularity of order $k \geq \lfloor n (\frac{1}{p_0} - 1) \rfloor $. Then $$\|Tf\|_{H^{{p(\cdot)}}} \leq C(T,{{p(\cdot)}},p_0,q,n)\|f\|_{H^{{p(\cdot)}}}.$$
We will prove both theorems as a consequence of a more general result for sublinear operators.
\[thm:gen-bdd\] Given ${{p(\cdot)}}\in M\mathcal{P}_0$ with $0<p_0<1$, and $q>1$ sufficiently large (as in Theorem ), suppose that $T$ is a sublinear operator that is defined on $({{p(\cdot)}},q)$ atoms. Then:
1. If for all $w \in A_1\cap RH_{(q/p_0)'}$ and every $({{p(\cdot)}},
q/p_0)$ atom $a(\cdot)$ with support $B$, $$\label{Condition:Hp-Lp}
\| Ta \|_{L^{p_0} (w)} \leq C(T,{{p(\cdot)}},p_0,q,n,[w]_{A_1},[w]_{RH_{(q/p_0)'}}) \frac{w(B)^{1/p_0}}{\| \chi_{B} \|_{{{p(\cdot)}}}},$$ then $T$ has a unique, bounded extension $\tilde{T} : H^{{{p(\cdot)}}}
\rightarrow L^{{{p(\cdot)}}}$.
2. If for all $w \in A_1\cap RH_{(q/p_0)'}$ and every $({{p(\cdot)}},
q/p_0)$ atom $a(\cdot)$ with support $B$, $$\label{Condition:Hp-Hp}
\| Ta \|_{H^{p_0} (w)} \leq C(T,{{p(\cdot)}},p_0,q,n,[w]_{A_1},[w]_{RH_{(q/p_0)'}}) \frac{w(B)^{1/p_0}}{\| \chi_{B} \|_{{{p(\cdot)}}}},$$ then $T$ has a unique, bounded extension $\tilde{T} : H^{{{p(\cdot)}}}
\rightarrow H^{{{p(\cdot)}}}$.
The additional hypothesis that $0<p_0<1$ is not a real restriction, since by Lemma \[lemma:max-up\] we may take $p_0$ as small as desired.
Note that when ${{p(\cdot)}}$ is constant and $w\equiv 1$, then conditions and reduce to showing that $T$ is uniformly bounded on atoms, which is the condition used to prove singular integrals are bounded on classical Hardy spaces.
First suppose that holds. Fix $f \in
H_{fin}^{{{p(\cdot)}}, q/p_0}$; by Theorem \[thm:qatoms\] this set is dense in $H^{{p(\cdot)}}$. Since $T$ is well-defined on the elements of $H_{fin}^{{{p(\cdot)}}, q/p_0}$, it will suffice to prove that $$\label{eqn:bdd-on-atoms}
\|Tf\|_{{L^{p(\cdot)}}} \leq C(T,{{p(\cdot)}},p_0,q,n)\|f\|_{H^{{p(\cdot)}}}.$$ For in this case by a standard density argument there exists a unique bounded extension $\tilde{T}$ such that $\tilde{T} : H^{{{p(\cdot)}}}
\rightarrow L^{{{p(\cdot)}}}$.
To prove we will use the extrapolation argument in Lemma \[lemma:ST-atoms\] to reduce the variable norm estimate to a weighted norm estimate. Arguing as we did in that proof, we have that $$\| Tf \|_{L^{{{p(\cdot)}}}}^{p_0} \leq \sup\int |Tf(x)|^{p_0} \mathcal{R}g(x) dx,$$ with the supremum taken over all $g \in L^{({{p(\cdot)}}/p_0)'}$ with $\| g
\|_{({{p(\cdot)}}/p_0)'} \leq 1$. Suppose for the moment that we can prove that for all $f\in H_{fin}^{{{p(\cdot)}}, q/p_0}$, $$\label{Base:Hp-Lp}
\|Tf\|_{L^{p_0}(\mathcal{R} g)} \leq
C(T,{{p(\cdot)}},p_0,q,n)\|f\|_{H^{p_0}(\mathcal{R} g)}.$$ (In particular, the constant is independent of $g$.) Then we can continue the argument as in the proof of Lemma \[lemma:ST-atoms\] to get $$\begin{gathered}
\| Tf \|_{L^{p_0} (\mathcal{R}g)}^{p_0}
\leq C \| f \|_{H^{p_0} (\mathcal{R}g)}^{p_0} \leq C\int {\mathcal{M}_N}f(x)^{p_0} \mathcal{R}g(x) dx \\
\leq C \| ({\mathcal{M}_N}f)^{p_0} \|_{{{p(\cdot)}}/p_0} \ \| \mathcal{R}g \|_{({{p(\cdot)}}/p_0)'}
\leq C \| {\mathcal{M}_N}f \|_{{{p(\cdot)}}}^{p_0} \leq C\| f \|_{H^{{{p(\cdot)}}}}^{p_0}. \end{gathered}$$ This gives us .
To complete the proof we will show . Recall that as sets, $H_{fin}^{p_0, q/p_0} (\mathcal{R}g) = H_{fin}^{{{p(\cdot)}}, q/p_0}$. Therefore, let $$f = \sum_{j = 1}^k {\lambda}_j {a}_j$$ be an arbitrary finite decomposition of $f$ in terms of $({{p(\cdot)}},q/p_0)$ atoms. Since, $0 < p_0 < 1$, by the sublinearity of $T$, convexity and , $$\begin{gathered}
\| Tf \|_{L^{p_0} (\mathcal{R}g)}^{p_0} = \int |Tf(x)|^{p_0} \mathcal{R}g (x) dx
\leq \sum_{j = 1}^k {\lambda}_j^{p_0} \int_{B_j} |T{a}_j (x)|^{p_0} \mathcal{R}g(x) dx \\
\leq C \sum_{j = 1}^k {\lambda}_j^{p_0} \frac{\mathcal{R}g (B_j)}{\| \chi_{B_j} \|_{{{p(\cdot)}}}^{p_0}}
= C\bigg\| \sum_{j = 1}^k {\lambda}_j^{p_0} \frac{\chi_{B_j}}{\|\chi_{B_j}\|_{{p(\cdot)}}^{p_0}} \bigg\|_{L^1 (\mathcal{R}g)}.
\end{gathered}$$ This is true for any such decomposition of $f$. Therefore, since $\mathcal{R}g \in A_1 \cap L^{({{p(\cdot)}}/p_0)'}$ by construction, by Lemma \[lemma:Weighted-FinDecomp\] we can take the infimum over all such decompositions to get $\| Tf
\|_{L^{p_0} (\mathcal{R}g)} \leq C \| f \|_{H^{p_0} (\mathcal{R}g)}$, where $C=C(T,{{p(\cdot)}}, p_0, q,n)$. This proves for all $f\in H_{fin}^{{{p(\cdot)}}, q/p_0}$.
We now consider the case when condition holds. The proof is essentially the same as before, except instead of proving , we need to prove that for all $f\in H_{fin}^{{{p(\cdot)}}, q/p_0}$, $$\label{Base:Hp-Hp}
\|Tf\|_{H^{p_0}(\mathcal{R} g)} \leq
C(T,{{p(\cdot)}},p_0,q,n)\|f\|_{H^{p_0}(\mathcal{R} g)}.$$ Given this, we can then repeat the extrapolation argument as before. To prove we use the same argument used to prove , replacing $Tf$ with ${M_{\Phi, 0}}(Tf)$ where $\Phi\in {\mathcal{S}}$ with $\int \Phi\,dx =1$, and using instead of .
By Theorem \[thm:gen-bdd\] it will suffice to show that condition holds for all $({{p(\cdot)}},q/p_0)$ atoms and all $w\in A_1\cap RH_{(q/p_0)'}$.
Fix such an atom $a(\cdot)$ with support $B=B(x_0,r)$. Let $2B =
B(x_0, 2r)$ and write $$\begin{gathered}
\|Ta\|_{L^{p_0}(w)}^{p_0}
= \int |Ta (x)|^{p_0} w (x) dx \\
= \int_{2B} |Ta (x)|^{p_0} w(x) dx + \int_{(2B)^c} |Ta (x)|^{p_0}
w(x) dx
= I_1 + I_2.\end{gathered}$$ We first estimate $I_1$. By Lemma \[lemma:kolmogorov\] there exists a constant $C=C(T,n,p_0,[w]_{A_1})$ such that $$\begin{gathered}
\int_{2B} |Ta (x)|^{p_0} w(x) dx
\leq C w(B)^{1 - p_0} \left( \int_{{{\mathbb R}^n}}|a(x)| w(x) dx \right)^{p_0} \\
\leq C w(B)^{1 - p_0} |B|^{p_0} \left( \avgint_B |a|^{q/p_0} dx \right)^{1/q}
\left( \avgint_B w(x)^{(q/p_0)'} dx \right)^{p_0/(q/p_0)'}.
\end{gathered}$$ Since $a(\cdot)$ is a $({{p(\cdot)}},q/p_0)$ atom and $w\in RH_{(q/p_0)'}$, we get that $$I_1 \leq C [w]_{RH_{(q/p_0)'}}^{p_0} \ w(B)^{1 - p_0} |B|^{p_0} \|
\chi_B \|_{L^{{{p(\cdot)}}}}^{-p_0} |B|^{-p_0} w(B)^{p_0} = C
[w]_{RH_{(q/p_0)'}}^{p_0} \ w(B) \| \chi_B \|_{L^{{{p(\cdot)}}}}^{-p_0}.$$ To estimate $I_2$, we start with a pointwise estimate. Let $d = \lfloor n(\frac{1}{p_0} - 1) \rfloor$. We claim that there exists a constant $C = C(T, n)$ such that for all $x \in (2B)^c$, $$\label{Ta-pointwise}
|Ta(x)| \leq C \frac{|B|^{1 + \frac{d + 1}{n}}}{\| \chi_B \|_{{{p(\cdot)}}}} \cdot \frac{1}{|x - x_0|^{n + d + 1}}.$$ To prove this, let $P_d$ be the Taylor polynomial of $K$ of degree $d$ centered at $x - x_0$. By our definition of $d$ and our assumption on $k$, $d + 1 \leq k + 1$. Therefore, the remainder $|K(x - y)| - P_d (y)|$ can be estimated by Condition (2) in Definition \[eqn:kregular\]. Hence, by the vanishing moment and size conditions on $a(\cdot)$ and Hölder’s inequality, $$\begin{aligned}
|Ta (x)|
&\leq \int |K(x - y) - P_d (y)| |a(y)| dy \\
& \leq \frac{C}{|x - x_0|^{n + d + 1}} \int_{B(x_0, r)} |y - x_0|^{d + 1} |a(y)| dy \\
&\leq C \frac{r^{d + 1} |B| }{|x - x_0|^{n + d + 1}}\avgint_B a(y)\,dy \\
& \leq C \frac{|B|^{\frac{n + d + 1}{n}} |B|^{-p_0/q}\|a\|_{q/p_0}}{|x
- x_0|^{n + d + 1}} \\
& \leq C \frac{|B|^{1 + \frac{d + 1}{n}}}{\| \chi_B \|_{{{p(\cdot)}}}} \cdot \frac{1}{|x - x_0|^{n + d + 1}}.
\end{aligned}$$ Given we have that $$\int_{(2B)^c} |Ta(x)|^{p_0} w(x) dx
\leq C \frac{|B|^{p_0 (\frac{n + d + 1}{n})}}
{\| \chi_{B} \|_{{{p(\cdot)}}}^{p_0}} \underbrace{\int_{(2B)^c} \frac{w(x)}{|x - x_0|^{p_0 (n + d + 1)}} dx}_{J}.$$ To complete the proof we will show that there exists a constant $C = C(n, p_0)$ such that $$\label{Est:J}
J \leq C\frac{[w]_{A_1} w(B)}{|B|^{p_0 (\frac{n + d + 1}{n})}}.$$ The proof of this is standard; for the convenience of the reader we sketch the details. Write $$(2B)^c = \bigcup_{i =1}^{\infty} (2^{i + 1} B \backslash 2^i B);$$ then for $x \in 2^{i + 1} B \backslash 2^i B$, we have $|x - x_0|
\simeq 2^i r \simeq 2^i |B|^{1/n}$. Since $w \in A_1$ and $p_0 (n + d
+1) > n$, we can estimate as follows: $$\begin{aligned}
J
&= \sum_{i = 1}^{\infty} \int_{2^{i + 1} B \backslash 2^i B} \frac{w(x)}{|x - x_0|^{p_0 (n + d + 1)}} dx \\
&\leq \frac{C}{|B|^{p_0 (\frac{n + d + 1}{n})}}
\sum_{i = 1}^{\infty} \frac{1}{2^{ip_0 (n + d + 1)}} \int_{2^{i + 1} B \backslash 2^i B} w(x) dx \\
&= \frac{C}{|B|^{p_0 (\frac{n + d + 1}{n})}}
\sum_{i = 1}^{\infty} \frac{2^{n(i + 1)}|B| }{2^{ip_0 (n + d + 1)}} \ \avgint_{2^{i + 1} B} w(x) dx \\
&\leq \frac{C 2^n[w]_{A_1}}{|B|^{p_0 (\frac{n + d + 1}{n})}}
\sum_{i = 1}^{\infty} \frac{1}{2^{ip_0 (n + d + 1)-in}} \left( |B| \ \operatorname*{ess\,inf}_{x \in B} w(x) \right) \\
& = \frac{C [w]_{A_1} w(B)}{|B|^{p_0 (\frac{n + d + 1}{n})}}.
\end{aligned}$$ This gives us and so completes the proof.
Our argument is similar to the proof of Theorem \[thm:bdd-Lp\]. By Theorem \[thm:gen-bdd\] it will suffice to show that condition holds for an arbitrary $({{p(\cdot)}},q/p_0)$ atom $a(\cdot)$ with support $B=B(x_0,r)$, and all $w\in A_1\cap RH_{(q/p_0)'}$. Fix $\Phi \in \mathcal{S}$ with $\int \Phi = 1$; then we can estimate $\| Ta \|_{H^{p_0} (w)}$ as follows: $$\begin{aligned}
\| Ta \|_{H^{p_0}(w)}^{p_0}
&\lesssim \int_{2B} {M_{\Phi, 0}}(Ta) (x)^{p_0} w(x) dx + \int_{(2B)^c} {M_{\Phi, 0}}(Ta) (x)^{p_0} w(x) dx = J_1 + J_2.
\end{aligned}$$ To estimate the $J_1$ we first use the fact that ${M_{\Phi, 0}}(Ta) \leq c
M (Ta)$. Moreover, we have that since $w\in A_1$, the Hardy-Littlewood maximal operator also satisfies Kolmogorov’s inequality (see [@duoandikoetxea01; @garcia-cuerva-rubiodefrancia85]): $$\begin{aligned}
J_1 \leq C w(2B)^{1 - p_0} \big( \underbrace{\int_{{{\mathbb R}^n}}|Ta(x)| w(x) dx}_L \big)^{p_0}.
\end{aligned}$$ To get the desired estimate for $J_1$ it will suffice to show that $$L = \int_{{{\mathbb R}^n}}|Ta(x)| w(x) dx \leq \frac{w (B)}{\| \chi_{B} \|_{{{p(\cdot)}}}}.$$ To prove this, we again split the integral: $$L = \int_{2B} |Ta(x)| w(x) dx + \int_{(2B)^c} |Ta(x)| w(x) dx = L_1 + L_2.$$ To estimate $L_1$ we apply Hölder’s inequality, the boundedness of $T$ on $L^{q/p_0}$, and the fact that $w\in RH_{(q/p_o)'}$ to get $$\begin{gathered}
L_1 = \int_{2B} |Ta(x)| w(x) dx
\leq \left( \int_{2B} |Ta(x)|^{q/p_0} dx \right)^{p_0/q}
\left( \int_{2B} w(x)^{(q/p_0)'} dx \right)^{1/(q/p_0)'} \\
\leq \| a \|_{L^{q/p_0}} \cdot |2B|^{1/(q/p_0)'} \left( \avgint_{2B} w(x)^{(q/p_0)'}
dx \right)^{1/(q/p_0)'}
\leq C(n,[w]_{A_1},[w]_{RH_{(q/p_0)'}}) \frac{w(B)}{\| \chi_B \|_{{{p(\cdot)}}}}.
\end{gathered}$$ To estimate $L_2$ we repeat the argument we used to estimate $I_2$ in the proof of Theorem \[thm:bdd-Lp\], replacing the exponent $p_0$ by $1$. Then using the pointwise estimate for $Ta$ and the decomposition argument, we have that $$\begin{gathered}
L_2 \leq C \frac{|B|^{\frac{n + d + 1}{n}}}{\| \chi_B \|_{{{p(\cdot)}}}} \left( \int_{(2B)^c} \frac{w(x)}{|x - x_0|^{n + d + 1}} dx \right) \\
\leq C \frac{|B|^{\frac{n + d + 1}{n}}}{\| \chi_B \|_{{{p(\cdot)}}}} \cdot \frac{w(B) [w]_{A_1}}{|B|^{\frac{n + d + 1}{n}}} \cdot \left( \sum_{i = 0}^{\infty} \frac{2^{ni}}{2^{i(n + d + 1)}} \right) \leq C \frac{w(B)}{\| \chi_B \|_{{{p(\cdot)}}}}.
\end{gathered}$$ To estimate $J_2$, we will prove a pointwise bound for ${M_{\Phi, 0}}(Ta_j) (x)$ for $x \in (2B_j)^c$ similar to . Define $K^{(t)} = K
\ast \Phi_t$; then $K^{(t)}$ satisfies condition (3) of Definition \[eqn:kregular\] uniformly for all $t>0$. Moreover, for $x\in (2B)^c$, the integral for $K*a(x)$ converges absolutely, so $|\Phi_t*(K*a)(x)|=|K^{(t)}*a(x)|$.
Again let $d = \lfloor n(\frac{1}{p_0} - 1) \rfloor$ and fix $t>0$. If $P_d$ is the Taylor polynomial of $K^{(t)}$ centered at $x - x_0$, we can argue exactly as we did to prove to get $$\begin{aligned}
|K^{(t)} \ast a (x)|
& = \left|\int [K^{(t)}(x - y) - P_d (y) ] a (y)\, dy\right| \\
& \leq \frac{C}{ |x - x_0|^{n+ d + 1}} \int_{B(x_0,r)} |y-x_0|^{d + 1}|a(y)| dy \\
& \leq C\frac{|B|^{\frac{n+d+1}{n}}|B|^{-p_0/q}}{|x - x_0|^{n+ d + 1}}
\|a\|_{L^{q/p_0}} \\
& \leq C\frac{|B|^{1+\frac{d + 1}{n}}}{\| \chi_{B} \|_{{p(\cdot)}}}\frac{1}{|x - x_0|^{n + d + 1}}.\end{aligned}$$ The final constant is independent of $t$, an so we can take the supremum over all $t$ to $${M_{\Phi, 0}}(Ta)(x) \leq C\frac{ |B|^{1+\frac{d + 1}{n}}}{\| \chi_{B} \|_{{{p(\cdot)}}}}\frac{1}{|x - x_0|^{n + d + 1}}.$$ Then arguing as we did before, by we have that $J_2 \leq w(B)/ \| \chi_B
\|_{{{p(\cdot)}}}^{p_0}$. This completes the proof.
[^1]: The first author is supported by the Stewart-Dorwart faculty development fund at Trinity College and by grant MTM2012-30748 from the Spanish Ministry of Science and Innovation
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper proposes a novel cost-reflective and computationally efficient method for allocating distribution network costs to residential customers. First, the method estimates the growth in peak demand with a 50% probability of exceedance (50POE) and the associated network augmentation costs using a probabilistic long-run marginal cost computation based on the Turvey perturbation method. Second, it allocates these costs to customers on a cost-causal basis using the Shapley value solution concept. To overcome the intractability of the exact Shapley value computation for real-world applications, we implement a fast, scalable and efficient clustering technique based on customers’ peak demand contribution, which drastically reduces the Shapley value computation time. Using customer load traces from an Australian smart grid trial (*Solar Home Electricity Data*), we demonstrate the efficacy of our method by comparing it with established energy- and peak demand-based cost allocation approaches.'
author:
- 'Donald Azuatalam, Archie C. Chapman, and Gregor Verbič, [^1]'
title: 'A Turvey-Shapley Value Method for Distribution Network Cost Allocation'
---
Turvey perturbation method, long-run marginal cost, Shapley value, k-means clustering, cost-causality, demand-based tariffs, cost-reflective network pricing
Introduction {#section1}
============
The rapid rise in the penetration levels of distributed energy resources in low-voltage distribution networks necessitates the design of network tariffs that allocate associated network costs in an efficient, fair and equitable manner to network users. Hence, distribution network service providers (DNSPs) and regulators in most jurisdictions are challenged with the tasks of designing efficient tariffs that are reflective of network cost drivers [@picciariello2015distributed]. Recent studies in this area have explored different methods for distribution network pricing, including transmission network pricing methodologies, such as *Locational Marginal Pricing* (LMP), *Postage Stamp* (PS), *MW-Mile*, *MVA-Mile*, *Average Participation* (AP), and *Marginal Participation* (MP) [@rubio2000marginal; @zolezzi2001review; @sotkiewicz2006allocation; @li2008cost; @brown2015efficient], *Ramsey-Boiteux pricing*, *cooperative game theory* or other extemporaneous cost allocation methodologies. However, in order to establish the performance of these methods with respect to established tariff design principles [@bonbright1961principles], which includes *cost reflectiveness*, *efficiency*, *stability* and *fairness*, we need to define a measure (benchmark) of overall performance, with which to compare existing methods. To this end, the purpose of our study is to use a principled cost allocation method as a performance benchmark for other allocation methodologies.
Distribution network costs typically comprises three major cost components–energy costs, long-run marginal cost (LRMC), and residual (such as retail charges) costs. In order to adequately recover these cost components, network tariffs should be structured such that the fixed, energy and/or demand charges efficiently send the right price signals to customers to respond appropriately. For example, [@abdelmotteleb2018designing] used a three-part tariff for distribution network cost allocation. Here, the residual costs (\$/day) were recovered through *Ramsey pricing*, while the LRMC (\$/kW) and energy costs (\$/kWh) were recovered through *coincident peak pricing* and *distribution locational marginal pricing* (DLMP) respectively.
Nevertheless, network tariffs historically only consisted of energy-based (volumetric) and residual charges due to two reasons: (i) there was little need to signal the LRMCs because loads per feeder were relatively flat and (ii) pricing mechanisms available to utilities were severely constrained by metering technology. However, with the introduction of smart meters, it is possible to implement tariffs which reflect congestion costs that drives network investments. As such, network tariffs should be based on customers demand at network peak [@lewis1941two; @boiteux1952determination; @nijhuis2017analysis; @passey2017designing; @abdelmotteleb2018designing; @nelson2013new]. It should be time- and location-specific and should account for network losses and actual customer energy use. Additionally, a fair and equitable tariff should also eliminate or reduce inter-customer subsidies created due to PV owners, while safeguarding vulnerable customers [@picciariello2015electricity; @simshauser2014network].
Unlike volumetric tariffs, peak demand-based tariffs are robust to technological changes (such as solar PV, EV or battery storage) which reshapes customers’ demand profiles while effectively signalling peak demand costs to customers [@simshauser2014network; @pimm2018time]. Thus, residual and/or LRMC costs can be recovered partly through demand-charges instead of constantly increasing fixed or energy charges for all customers [@brown2015efficient; @ahmad2018pricing]. So far, *coincident peak pricing* (CPP) and *critical peak pricing* have been proposed to mitigate the impacts of DER on the equity of network cost allocations. Although peak demand-based tariffs are more complex than energy-based tariffs, they can better allocate network costs on a cost-causal basis and ensure a stable revenue for network companies [@simshauser2014network; @ahmad2018pricing]. Furthermore, [@ahmad2018rate] showed that customer bill volatility reduces with demand-based tariffs compared to *real time pricing* and *time-of-use tariffs*. Contrarily, [@passey2017designing] tested [demand-based tariffs]{} proposed by the Australian Energy Regulator (AER) on households in Sydney. It was concluded that without due adjustments made, these tariffs are low in cost-reflectivity. Generally, the suitability of network tariffs in terms of fairness and cost-reflectivity depends on the assumptions made in the tariff design and on customers’ price response [@stenner2015australian]. This further highlights the need for a principled cost allocation benchmark.
Further still, the distribution network tariff design problem encompasses more than the aforementioned requirements. Beyond these, there are other three salient tariff design questions that need to be answered in order to achieve a cost-reflective pricing and appropriate customer response: (i) What LRMC calculation method should be used – the *Turvey perturbation method* or the *Average Incremental Cost method*? (ii) What peak demand should be the basis for charging customers – individual customer peak, distribution network coincident peak or zone-substation peak? and, (iii) What is the optimal frequency of peak demand measurement – monthly or yearly basis? [@ahmad2018pricing]. The answers to these questions form the basis for practically implementing tariffs that better recuperate forward-looking network costs. However, there are no clear-cut answers to these questions, because choices for network companies depend on other factors, such as customer socio-demographics, customer class, and the availability of smart meters and energy management systems. Nonetheless, recent research in the area argues that demand-based tariffs should be based on network coincident peak since it better signals LRMC. However, in practice, customers’ coincident peak demand is hard to measure and thus CPP is difficult to implement. Thus, demand charges are a step-forward to attaining optimal network tariffs.
In this paper, we seek, based on established economic principles, to make a further step towards the design of equitable network pricing. Our focus is to provide a measure for the fair and efficient allocation of costs that signal the drivers of future network investment. To achieve this, we develop a novel method to apportion the LRMC, using a probabilistic approach to the *Turvey perturbation method* ([@turvey1969marginal]) linked via the *characteristic function* of a cooperative game[^2] to the *Shapley value* (SV) cost allocation rule [@Shapley1953; @ChalkiadakisEtal2011].
In more detail, the [Turvey perturbation method]{} is a forward-looking and more time- and location-specific method for LRMC estimation, compared to the simpler [average incremental cost]{} methods widely used by network companies [@ahmad2018pricing]. Furthermore, [@biggardarr; @neraconsult] argue that the [Turvey perturbation]{} method is the preferred option since it better aligns with the underpinning principles governing LRMC. However, research in [@toothrichard] concluded that both methods can be equally used for LRMC calculations.
At the same time, the SV gives a vector-valued solution to a cooperative *transferable utility* (TU) game, where the total cost or worth of a coalition is defined by a single-valued *characteristic function*. In our method, the characteristic function is the probabilistic LRMC defined using the Turvey perturbation method. The SV has found several applications including consumer demand response compensation [@o2015shapley; @bakr2015using; @Chapman_IREP2017], transmission network cost allocation [@zolezzi2002transmission; @tan2002application; @khare2015shapley], distribution network loss allocation [@sharma2017loss], and other cost allocation problems [@ghassemi2008cooperative; @stanojevic2010economic; @byun2009fair; @o2015shapley; @bakr2015using]. However, due to the computational complexity of computing the *exact* SV for a large number of players, it’s application is usually limited to small problems. Recent research in this area, nonetheless, has seen developments of approximate methods of calculating the SV in polynomial time [@fatima2008linear; @fatima2007randomized; @stanojevic2010economic; @david2005shapley; @castro2009polynomial; @byun2009fair; @bakr2015using]. In [@bakr2015using], a comparison of the accuracy and scalability of two approximate SV computation algorithms was made, namely *linear-time approximation* [@fatima2008linear] and *stratified sampling* [@castro2009polynomial; @maleki2013bounding] techniques. While the *stratified sampling* approach was more accurate, the *linear-time approximation* required less memory and computation time as the number of players increased. This is a general finding, so with these methods, there is always a trade-off between accuracy and computational complexity. Moreover, some of these methods are only suited to *weighted voting games*, which does not match our cost-allocation problem. In a different research direction, a clustering approach was adopted in [@david2005shapley], where customers are segmented into major classes. However, customers in the same class are assigned the same SV. Conversely, for distribution network cost allocation, this is not the case, as the SV should be different for all customers.
In light of these shortcomings, we derive a computationally-efficient clustering algorithm, to allocate network costs based on the Turvey-Shapley value method. The SV is computed at the level of clusters, and individual customers are allotted a portion of this SV based on their average coincident peak demand contribution to each coalition of their representative cluster. This approximation approach is validated by comparison to the exact SV calculation, for which the SV estimation error is shown to be small and reduces as the number of customers increases (i.e. when approximation become computationally necessary). Furthermore, as the SV method best allocates network costs in a principled, fair and stable manner, we used it as a benchmark to measure the cost-reflectivity of other cost allocation methodologies. In summary, the analysis in this paper extends the preliminary results in our earlier conference paper [@azuatalam2019shapley] in the following ways:
- We propose a probabilistic approach to the [Turvey]{} LRMC computation via a Weibull distribution, which gives an unbiased estimate of forward-looking network costs.
- We propose a peak load contribution clustering technique interleaved with the Turvey-Shapley value method to compute the Shapley value for large number of customers with low computation time and estimation error.
- We demonstrate the effectiveness of our methodology using real customer load traces from the *Solar Home Electricity Data*[^3]. Our results show that the proposed allocation method is the most reflective of network capacity costs compared to other cost allocation methodologies.
The remainder of this paper is structured as follows. Section \[prelim\] provides preliminaries on cooperative game theory and the Shapley value. In Section \[method\], we describe the methodology, including the Turvey pricing method. The results are presented and discussed in Section \[results\]. Section \[conclusions\] concludes.
Preliminaries {#prelim}
=============
In this section, we provide a background to cooperative games, the *Shapley value* and its characteristics, and the *Turvey perturbation method*.
Cooperative Games
-----------------
Formally, we consider the class of *transferable utility* (TU) games, which are cooperative games that allow the transfer of worths between players. If the players in a cooperative game agree to work together, they form a *coalition*. If all $n$ players form a coalition, it is called the *grand coalition*. Each player incurs some private *cost* in completing its component of the joint action, while collectively, the joint action has some *worth* associated with it.
A TU game is given by $\Gamma = \langle \mathcal{N},w\rangle$ where:
- $\mathcal{N}$ is a set of $n$ *players*, and
- $w(\mathcal{S})$ is a *characteristic function*, $w:2^{n}\to \Re_+$ with $w(\emptyset) = 0$, that maps from each possible coalition $\mathcal{S}\subseteq \mathcal{N}$ to the *worth* of $\mathcal{S}$.
Before defining the Shapley value, we first formally define some important characteristics of any solution to a TU game.
Given a $\Gamma$, a solution concept defines a *worth* to each player, which is a vector of transfers (worths), $\phi = (\phi_1,\ldots, \phi_i,\ldots,\phi_n) \in \Re^n$.
We denote the sum of worths as $\sum_{i\in \mathcal{S}} \phi_i = \phi(\mathcal{S})$. Some desirable properties of solutions concepts include the following; a solution is:
- *Efficient* if $\phi(\mathcal{S}) = w(\mathcal{S})$, so that the worth vector exactly divides the coalitions worth,
- *Symmetric* if $\phi_i = \phi_j$ if $w(\mathcal{S}\cup \{i\}) = w(\mathcal{S}\cup \{j\}),\ \forall ~\mathcal{S}\subseteq N\setminus \{i,j\}$. This means that equal worths are made to symmetric players, where symmetry means that we can exchange one player for the other in any coalition that contains only one of the players and not change the coalition’s worth.
- *Additive* if for any two additive games the solution can be given by $\phi_i(v_1+v_2) = \phi_i(v_1) + \phi_i(v_2)$ for all players. That is, an additive solution assigns worths to the players in the combined game that are the sum of their worths in the two individual games.
- *Zero worth to a null player* if a player $i$ that contributes nothing to any coalition, such that $w(\mathcal{S}\cup \{i\})=w(\mathcal{S})$ for all $\mathcal{S}$, then the player receives a worth of 0.
The Shapley Value (SV) {#section_shapley_value}
----------------------
Solution concepts in cooperative game theory define divisions of the group reward among players, while considering the rewards available to each alternative coalition of players. The SV is one of such solution concepts which also satisfies the desirable properties listed above, by virtue of its definition.
The SV allocates to player $i$ in a coalitional game $\langle w,\mathcal{N}\rangle$ the worth: $$\label{defShapleyValue1}
\phi _{i}(w)
= \frac {1}{n}\sum_{\mathcal{S}\subseteq \mathcal{N}\setminus i}
\binom{n-1}{|\mathcal{S}|}^{-1}\left(w(\mathcal{S}\cup \{i\})-w(\mathcal{S})\right)$$
Here, the vector-valued function $\phi$ has the following intuitive interpretation: consider a coalition being formed by adding one player at a time. When $i$ joins the coalition $\mathcal{S}$, its *marginal worth* is given by $w(\mathcal{S}\cup \{i\}) - w(\mathcal{S})$. Then, for each player, its SV worth is the average of its marginal contributions over the possible different orders in which the coalition can be formed.
The expression in can also be interpreted as a player’s contribution to all subsets of $\mathcal{N}$ that do not contain it, where the binomial term is the number of coalitions of size $|\mathcal{S}|$. We can further expand this expression to identify a useful approximation. Specifically, the summation in can be expressed in terms of the size of the coalitions that $i$ is added to, as follows: $$\label{eqn:SV_expanded}
\phi_{i}(w)
= \frac {1}{n}\sum_{k=0}^{n-1}
\left( \binom{n-1}{k}^{-1}
\sum_{\mathcal{S}\in \mathcal{S}^k} \left(w(\mathcal{S}\cup \{i\}) - w(\mathcal{S})\right) \right)$$ where: $\mathcal{S}^k = \{\mathcal{S}\subseteq \mathcal{N}\setminus i \,:\, |\mathcal{S}| = k \}$ is the set of coalitions of size $k$ that exclude $i$. We can now approach the SV by approximating the inner term for each size $k \in \{0,\ldots,n-1\}$.
One approach is to statistically estimate the term: $$\label{eqn:SV_randomised}
\chi^k = \binom{n-1}{k}^{-1} \sum_{\mathcal{S}\in \mathcal{S}^k} \left(w(\mathcal{S}\cup \{i\}) - w(\mathcal{S})\right)$$ using a sample-based approach. This is a randomised sampling algorithm described in detail in Section \[section\_randomised\_sampling\].
$\mathcal{N}$: set of players/customers, $\mathcal{N} = \{1,..,n\}$\
$\mathcal{L}$: set of all coalitions, $\mathcal{L} = \{\mathcal{S}^1,..,\mathcal{S}^l,..., \mathcal{S}^{2^n - 1}$}
$\mu$ = 50POE yearly peak demand of $\mathcal{S}$ **Find** $X\sim\mathcal{W}(\alpha,\beta)$ of $\mu$ exceeding line limit $x$
$\alpha = \mu/ \Gamma(1+1/\beta)$ with $\beta=1.5$ $w(\mathcal{S}) = P(X \geq x)Y$ $w(\mathcal{S}) = 0$ **Compute** $M_{\mathcal{S}_i}=w(\mathcal{S}) - w(\mathcal{S} - \{i\})$ **Compute** $K_{\mathcal{S}_i}=(1/n!)(|\mathcal{S}|-1)!\;(n-|\mathcal{S}|)!$ **Return** SV of $i$, $\phi_{i}(w) = \sum_{\substack{\mathcal{S}\in \mathcal{L}}} K_{\mathcal{S}_i} M_{\mathcal{S}_i}$
The Turvey Perturbation Method {#turvey_method_section}
------------------------------
The *Turvey perturbation method* [@turvey1969marginal] is one technique used to estimate the LRMC of capacity-based investments. It quantifies the effects of a (small) permanent change in demand $Q$ on future capital costs $C$. It is defined in [@toothrichard] as:
$$\label{eqn:turvey}
\textrm{Turvey LRMC} = \frac{PV(\Delta C)}{PV(\Delta Q)}$$
The expression in translates to–the ratio of the present value of change in costs (due to a permanent change in demand) to the present value of the permanent change in demand. The Turvey perturbation method therefore involves forecasting demand over the estimation period, with a certain confidence level. For this, we assume a small growth in yearly peak demand with a 50% probability of exceedance.
As explained in detail in the next section, in our methodology, the probability that this value exceeds the network line limit informs the LRMC. This probabilistic approach to the Turvey perturbation method, achieved via a Weibull distribution, is used to construct the *characteristic function* for the SV computation.
Methodology {#method}
===========
In this section, we detail the steps taken to assess the cost-reflectivity in the allocation of distribution network tariffs to LV residential customers, with the SV allocation being the benchmark. First, we explain the *Turvey-Shapley value* LRMC estimation and allocation methodology. Second, we describe algorithms to determine the exact SV for a set of customers $\mathcal{N}$, and its approximation. The approximation algorithms are required to compute the SV for $n>25$ players, with lower computational burden and minimal loss in accuracy.
Turvey-Shapley Value LRMC Methodology
-------------------------------------
The proposed Turvey-Shapley value LRMC methodology involves interleaving of a novel probabilistic approach to the Turvey perturbation method with the SV *characteristic function*, and is illustrated in Algorithm \[ShapleyvalueExact1\]. In this section, we explain the steps for the Turvey LRMC estimation and SV cost allocation.
First, we calculate the line limit $x$ of the given network with line augmentation cost $Y$. This is given as the yearly peak demand of the network (i.e. grand coalition of customers) multiplied by a factor of 1.5 (to account for distribution line emergency limit). We have assumed that the probability of a coalition’s 50 POE peak demand exceeding the line limit follows a Weibull distribution, $X\sim\mathcal{W}(\alpha,\beta)$[^4]. The two-parameter Weibull distribution function is defined as: $$F_{\alpha,\beta}(x) = 1 - \exp \Big[-\Big(\frac{x}{\alpha}\Big)^{\beta}\Big] \quad \mathrm{for} \quad x \geq 0$$ where $\alpha>0$ is the *scale* parameter and $\beta>0$ is the *shape* parameter. In order to obtain the standard fat-tailed Weibull distribution that is required in this study, $\beta$ is taken as 1.5.
Then, for each coalition, we assume a yearly peak demand growth rate of 1% as the 50 POE value, which is taken as the mean $\mu$ of the Weibull distribution. Given the mean and shape factor, we calculate the scale factor of the distribution. If the tail probability $P(X \geq x)$ of a coalition’s 50 POE peak demand exceeding the line limit is less than 0.001, we neglect the coalition cost (set as zero) in the incremental cost (IC) calculations for a particular customer in each coalition size. For example, if $Y$ is \$1M, then we neglect coalitions with cost less than \$1k, which improves the incremental cost computation time of each customer. Otherwise, the coalition cost is given as $P(X \geq x) Y$, that is, the coalition’s expected LRMC under the corresponding Weibull distribution. The SV for each customer $i \in \mathcal{N}$ is calculated as the average of its marginal contributions to all coalitions containing $i$, as in .
In the next three subsections, we describe the exact SV algorithm (*Exact*), and two algorithms (*Sampling* and *Clustering*) to compute the approximate SV for up to 25 customers.
Direct Enumeration
------------------
The exact algorithm (*Exact*), also known as *direct enumeration*, is based on and is described in Algorithm \[ShapleyvalueExact1\]. In terms of computational speed, it performs well with $n<20$ players. But with $n>25$, its performance (w.r.t. speed and memory requirements) deteriorates because of the time taken and memory required to compute the large ($2^{n}-1$) $\times$ $n$ coalition matrix. Note that Line 2 in Algorithm \[ShapleyvalueExact1\] can be broken down into $n$ coalition sizes according to .
Randomised Sampling {#section_randomised_sampling}
-------------------
As explained in Section \[section\_shapley\_value\], we use a sample-based randomised algorithm which statistically estimates , to provide approximate SV calculations based on . With this, we do not perform all the incremental cost calculations ($2^{n-1}$) required to compute the exact SV for each customer. Instead, we do this only for coalition sets that contain more than 10,000 possible coalitions of the same size. In Algorithm \[ShapleyvalueApprox\] (*Sampling*), we first select randomly a pilot sample ($|\mathcal{P}| = K$) from such large coalitions, where $K$ is determined by trial and error. Then, the standard deviation $\sigma_{\mathcal{P}}$ of the marginal contribution of customer $i$ to the sampled coalition is computed, followed by the calculation of the optimal sample size using: $$\label{eqn6}
|\mathcal{P}'| = \Big(\frac{Z\sigma_{\mathcal{P}}}{d}\Big)^2$$ where $Z =1.96$ is the z-score of 95% confidence in a Gaussian distribution and $d =0.2$ is taken to be the margin of error for the sampling estimation.
$\mathcal{N}$: set of players/customers, $\mathcal{N} = \{1,...,n\}$\
$\mathcal{K}$: partition of set of $k$-sized coalitions, $\mathcal{K} = \{\mathcal{S}^1,..,\mathcal{S}^k,..,\mathcal{S}^{n}\}$
**Find** $\mathcal{G}\subset \mathcal{S}^k: \forall\ \mathcal{Z} \in \mathcal{G}, i \in \mathcal{Z}$ $\theta_g = \frac{(k-1)!\;(n-k)!}{n!}\left(w(g) - w(g - \{i\})\right)$ $\vartheta_{\mathcal{S}^k} = \sum_{g\in \mathcal{G}}\theta_g$ **Sample** $|\mathcal{P}|$ coalitions from $|\mathcal{G}|:|\mathcal{P}| = K$ $\theta_p = \left(w(p) - w(p - \{i\})\right)$ **Compute** standard deviation $\sigma_{\mathcal{P}}$ of all $\theta_{p \in \mathcal{P}}$ **Compute** optimal sample size, $|\mathcal{P}'|$ using **Sample** $|\mathcal{P}'|$ coalitions from $|\mathcal{G}|$ $\theta_{p'} = \frac{(k-1)!\;(n-k)!}{n!}\left(w(p') - w(p' - \{i\})\right)$ $\vartheta_{\mathcal{S}^k} = \sum_{p'\in \mathcal{P}'}\theta_{p'}$ $\phi_i = \sum_{\mathcal{S}^k \in \mathcal{K}}\vartheta_{\mathcal{S}^k}$
Clustering Method
-----------------
The clustering method illustrated in Algorithm \[ShapleyvalueClustering\] is also based on . In this method, we first cluster $|\mathcal{H}^m| = 125$ customers from the Ausgrid *Solar Home Electricity Data* set into $n=5$ representative load profile clusters (with minimum customer set $|\mathcal{H}^1| = 25$). Here, end-users are clustered based on their half-hourly average daily consumption pattern for a year using the k-means clustering algorithm. Next, for each set of network users $\mathcal{H}^h \in \mathcal{H}$, we find the yearly demand (with 30-minute resolution) of each cluster $i\in \mathcal{N}$, by summing the half-hourly demand of all customers belonging to cluster $i$. Then, we find the average contribution $\boldsymbol{V}_g^{\mathcal{H}^h_i}$ of member customers to the yearly peak demand of cluster $i$ over all coalitions $g\in\mathcal{G}$. After computing the SV of each cluster, the cluster cost is then apportioned to its member customers according to their contribution. It is worth noting that the algorithm can be scaled to compute the SV for $|\mathcal{H}^m| > 125$ network users in our dataset, with just an insignificant clustering overhead computation cost for allocating customers into 5 clusters; and moreover, it would scale to settings with up to 25 clusters irrespective of the total number of customers.
.
$\mathcal{H}$: partition of set of network users, $\mathcal{H}:= \{\mathcal{H}^1,..,\mathcal{H}^h,..,\mathcal{H}^m\}$\
$\mathcal{N}$: set of players/clusters, $\mathcal{N} = \{1,...,n\}$\
$\mathcal{K}$: partition of set of $k$-sized coalitions, $\mathcal{K} = \{\mathcal{S}^1,..,\mathcal{S}^k,..,\mathcal{S}^{n}\}$
**Cluster** $|\mathcal{H}^m|$ customers into $n$ clusters **Sample** $|\mathcal{H}^h|$ customers uniformly from $|\mathcal{H}^m|$ **Find** $\mathcal{G}\subset \mathcal{S}^k: \forall\ \mathcal{Z} \in \mathcal{G}, i \in \mathcal{Z}$ **Find** $\mathcal{H}^h_i\subset \mathcal{H}^h$ $\boldsymbol{V}_g^{\mathcal{H}^h_i} = [v_g^1,...,v_g^{|\mathcal{H}^h_i|}]$ $\theta_g = \frac{(k-1)!\;(n-k)!}{n!}\left(w(g) - w(g - \{i\})\right)$ $\vartheta_{\mathcal{S}^k} = \sum_{g\in \mathcal{G}}\theta_g$ $\phi_i = \sum_{\mathcal{S}^k \in \mathcal{K}}\vartheta_{\mathcal{S}^k}$ $\boldsymbol{\phi_t^{\mathcal{H}^h_i}} = (\phi_i/|\mathcal{G}|)\sum_{g \in \mathcal{G}} \boldsymbol{V}_g^{\mathcal{H}^h_i}$
Case Study, Results and discussion {#results}
==================================
In this section, we assess the computational performance and accuracy of the SV approximation methods, the correlation of SV with three peak demand indicators commonly used to define tariffs, and compare alternative pricing methods with the SV cost allocation.
To begin, we define the following three peak demand indicators as follows:
- *Coincident* peak demand (CPD): This refers to a customer’s coincident peak demand at the time of the network’s yearly peak load.
- *Individual* peak demand (IPD): This refers to a customer’s yearly peak demand.
- *Total* peak demand (TPD): This refers to the sum total of a customer’s monthly peak demand values in a year.
As case study, the net load traces (solar PV and demand) used in this work were sourced from the Ausgrid (DNSP in NSW) *Solar Home Electricity Data*. The dataset comprises three years of half-hourly resolution smart meter data for the period between July 2010 to June 2013, for 300 residential customers in the Sydney region of Australia. However, we could only extract 125 customers from this dataset with complete solar PV and demand data, for the period between July 2012 to June 2013. This information is used to obtain the above defined peak demand indicators for each customer.
We also employ network tariff data from Ausgrid[^5], given in Table \[table1\][^6], which enables us make a rough estimate of the revenue obtained for these customers under the flat and ToU energy network prices.
\[table1\]
[c@c@c@c@c@c@c@]{} &
[@c@]{}Fixed\
charge\
&
[@c@]{}Anytime\
Energy\
&
[@c@]{}Off peak\
Energy\
&
[@c@]{}Shoulder\
Energy\
&
[@c@]{}Peak\
Energy\
&\
*Flat* & 40.097 & 11.163 & - & - & -\
*ToU* & 40.097 & - & 2.805 & 7.086 & 27.335\
SV Computation Time and Accuracy
--------------------------------
Here, we compare the computational performance and accuracy of the different SV calculation algorithms. For this first set of computations, we have assumed that all customers possess solar PV, so the net load is used to compute their monthly peak demand. Fig. \[sv\_time\] shows the SV computation time in minutes for all customer sizes from 5 up to 25 users. A related point to consider is that the exact SV computation for more than 25 users is not computationally feasible. The *Exact* algorithm performs best for $n < 15$ customers, but with $n \geq 15$, this is not the case. It takes 418 minutes to compute the SV for $n=25$ customers, due to the time and memory consuming coalition matrix generation and the corresponding coalition cost function calculations.
Conversely, there is a significant improvement in computational performance with the first approximate algorithm compared to *Exact*. *Sampling* takes only 98 minutes to compute the SV for $n=25$ (about a quarter of the time taken for *Exact*). This reduction in computation time is as a result of performing IC calculations for a select (optimal) sample, using a constant number as a pilot sampling size, instead of performing the IC calculations for all coalitions (in *Exact*) at the same time. Furthermore, splitting the total coalitions into $n$ coalition sizes in *Sampling* overcomes the memory limitations of *Exact*.
On the other hand, the clustering algorithm takes the least time to compute the SV for $n \geq 15$ customers. This is because the SV calculation is done for only 5 players (or clusters), with a little overhead computation cost for clustering.
To evaluate the accuracy of the sampling and the clustering technique, we find the root mean square error (RMSE) in the SV estimation. Fig. \[sv\_error\] shows RMSE values between $10^{-2}$ and $10^{-5}$ relative to mean values between $0.04$ and $0.2$ for $n\leq25$ customers. Although the sampling approach is more accurate than the clustering technique for $5<n\leq25$, it cannot be up-scaled to $n>25$ players, without a significant increase in the computation time. Besides, with the clustering technique, the estimation error reduces as more customers are added to make the clusters more representative.
-------------- ------------ -------- -------- -------- -------- --------
25 50 75 100 125
Coincident 0.8773 0.9134 0.9114 0.9272 0.9478
Individual 0.6668 0.5891 0.5644 0.5321 0.5065
Total 0.6626 0.5833 0.5541 0.5140 0.4848
Coincident 0.7967 0.8380 0.8780 0.9098 0.9473
Individual 0.4113 0.3625 0.3291 0.2918 0.2493
Total 0.5700 0.5512 0.5277 0.5016 0.4685
\[sv\_corr\]
-------------- ------------ -------- -------- -------- -------- --------
: Mean SV correlation with Peak demand Indicators
SV Linear Correlation with Peak demand Indicators
-------------------------------------------------
This section shows the results obtained by finding the linear correlation between the SV computed using the *clustering* technique and the peak demand indicators, for two scenarios (i) all customers without PV and (ii) all customers with PV. Since the SV is computed for 100 Monte Carlo runs based on uniform random sampling, the Pearson’s correlation coefficients (R-value) are presented as box plots in Fig. \[sv\_correlation\] while Table \[sv\_corr\] shows the mean values, for $\mathcal{H}$ of size 25, 50, 75, 100, and 125.
For Scenario 1 (Fig. \[sv\_correlation\_nopv\]), the SV correlates more with *Individual* peak demand than with *Total* peak demand but for Scenario 2 (Fig. \[sv\_correlation\_pv\]), the converse is the case. It is worth noting that *Individual* and *Total* corresponds to charging customers based on their yearly peak load and monthly peak load respectively. Without PV, a customer’s true demand is revealed, which is less sensitive to weather, so individual peak demand dominates. However, with PV, a customer’s demand profile is modified with PV generation which is season-dependent, and as such it’s better to charge customers on a monthly basis. Nevertheless, for both scenarios, the SV correlates most with *Coincident* peak demand, because it drives augmentation cost the most.
Figs. \[sv\_correlation\_peak\_nopv\] and \[sv\_correlation\_peak\_pv\] show the scatter plot for SV correlation with the peak demand indicators for $n=125$ customers without PV and with PV respectively. For Scenario 1, the mean R-values are 0.948, 0.507 and 0.485, for *Coincident*, *Individual*, and *Total* peak demand respectively while the R-values for Scenario 2 are 0.947, 0.249 and 0.468, for *Coincident*, *Individual*, and *Total* peak demand respectively. While CPD and TPD have similar values in both scenarios, IPD is considerably different. This is because a customer’s (individual) yearly peak demand changes significantly with the addition of PV.
Comparison of Cost Allocation Methodologies
-------------------------------------------
Here, we evaluate how different energy-based and peak demand-based cost allocation methodologies compare to the SV, by measuring the correlation between normalised customer cost allocation and specifically-defined peak demand indicators. First, we perform the SV computation for $n=125$ customers, for both scenarios. Then, we estimate the revenue obtained by the DNSP for these customers under the two tariffs (Flat- and ToU-based) described in Section \[method\]. For this, we have have neglected the feed-in-tariff (FiT) as it is administered through a different mechanism and handled by retailers in the Australian electricity system’s regulatory and billing arrangements. Therefore, we use only the power import from the grid for our calculations.
We consider the following cost allocation methods in our analysis:
- Energy-based (EB - Flat or ToU)
- Coincident peak load (CP - *Coincident* peak)
- Yearly peak load (YP - *Individual* peak)
- Monthly peak load (MP - *Total* peak) and
- Shapley value cost allocation (SV)
In the first method, cost allocation is done according to the revenue calculations under Flat and ToU tariff for $n=125$ customers, using tariff values in Table \[table1\]. For the rest, we split the total revenue obtained using the energy-based tariffs, according to the normalised SV and the normalised *Coincident*, *Individual*, and *Total* peak demand values for each customer.
The top part of Fig. \[sv\_corr\_error\] shows the correlation between the different cost allocation methodologies and the peak demand indicators. From these, we can deduce that the SV provides a fine balance between coincident, individual and aggregate peak demand, since it properly accounts for network usage at times other than the coincident network peak. This is by virtue of the way it is being computed, by evaluating the marginal contribution of customers to all possible coalitions. Although, the major cost driver for distribution networks is the coincident peak demand, it is necessary for a cost-reflective tariff design to appropriately account for aggregate and individual customer peak demand. Furthermore, the yearly and monthly peak demand allocation are also not cost-reflective, since they have a lower correlation with coincident peak demand compared with the SV allocation. Energy-based allocation methods perform worst as they show much lower correlation with the peak demand indicators. However, the results for Scenario 2 show that ToU-energy based allocation is better than that of Flat-energy based allocation as it shows comparatively higher R-values. Moreover, these results implicitly show that inter-customer subsidies will be reduced since customers would be paying their fair share, given the similar R-value for *Coincident* peak demand in both scenarios.
At the bottom part of Fig. \[sv\_corr\_error\], we show the error (RMSE) in actual cost values that will arise when a DNSP allocates cost to customers using less optimal cost allocation methodologies. This translates to inaccurate wealth transfer amongst customers in a distribution network. As expected, there are higher RMSE values resulting from energy-based cost allocation methods. This implies that they are least cost-reflective for both scenarios, both in terms of cost-causality and equity in cost allocation. Also, for both scenarios, the coincident peak load allocation results in the least error because the main cost driver for networks is the coincident peak. While the monthly peak allocation, and the yearly peak demand allocation method results in similar RMSE values for Scenario 1, this is not the case for the second scenario. When all customers possess PV, the monthly peak allocation results in slightly lower RMSE compared to the yearly peak demand allocation. This shows that for residential customers with PV, using a monthly peak demand network tariff is more cost reflective than a yearly peak demand tariff. It will be unfair to charge customers based on their sole highest yearly peak demand, which occurs in just one month of a year, and does not account for the seasonality in energy consumption which comes with PV generation.
Conclusions and Future Work {#conclusions}
===========================
In this work, we showed the efficacy of the *Turvey-Shapley value* method in calculating and apportioning LRMC to network users in a cost-reflective way and with low computational burden using the proposed clustering technique.
We have demonstrated that the SV is a cost-reflective cost allocation method for networks with a large number of customers, regardless of PV adoption. This is because the SV is a principled technique, which provides a proper balance between the peak demand indicators (network cost drivers). This makes for a fair cost allocation, which further reduces inter-customer subsidies. Other peak demand-based cost allocation approaches perform well up to the extent to which they appropriately balance the peak demand indicators, but with a greater emphasis on coincident peak demand. Furthermore, our results show that energy-based cost allocation methodologies are least cost-reflective as they least correlate with the peak demand indicators.
For future work, we will consider LRMC allocation for customers with both PV and batteries. In this case, an optimisation would have to be solved for each cost function computation.
[^1]: Donald Azuatalam, Archie C. Chapman and Gregor Verbič, are with the School of Electrical and Information Engineering, The University of Sydney, Sydney, New South Wales, Australia. E-mail: [email protected], [email protected], [email protected].
[^2]: A cooperative game models a game where a group of players cooperate to earn a joint reward, which has to be shared among the players in a fair and stable way.
[^3]: Dataset is available at https://data.nsw.gov.au/
[^4]: This choice is not essential to the method; any other fat-tailed distribution could be used as well.
[^5]: Ausgrid Network Price List. Available at https://www.ausgrid.com.au/Industry/Regulation/Network-prices.
[^6]: Peak: summer weekdays (Nov. to Mar.) between 2pm to 8pm, winter weekdays (Jun. to Aug.) between 5pm to 9pm; Shoulder: weekdays year round, between 7am to 10pm (exc. Peak periods); Off-peak: all other times.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Underground online forums are platforms that enable trades of illicit services and stolen goods. Carding forums, in particular, are known for being focused on trading financial information. However, little evidence exists about the sellers that are present on active carding forums, the precise types of products they advertise, and the prices buyers pay. Existing literature mainly focuses on the organisation and structure of the forums. Furthermore, studies on carding forums are usually based on literature review, expert interviews, or data from forums that have already been shut down. This paper provides first-of-its-kind empirical evidence on active forums where stolen financial data is traded. We monitored five out of 25 discovered forums, collected posts from the forums over a three-month period, and analysed them quantitatively and qualitatively. We focused our analyses on products, prices, seller prolificacy, seller specialisation, and seller reputation, and present a detailed discussion on our findings.
carding forums $\cdot$ underground forums $\cdot$ CVV
author:
- |
Andreas Haslebacher, Jeremiah Onaolapo, and Gianluca Stringhini\
University College London\
`[email protected]`\
`{j.onaolapo,g.stringhini}@cs.ucl.ac.uk`
bibliography:
- 'biblio.bib'
title: |
All Your Cards Are Belong To Us:\
Understanding Online Carding Forums
---
Introduction
============
£479 million of fraud losses on UK issued credit and debit cards were recorded in 2014 [@ffauk2015]. Almost 70% of these losses originate from “remote purchase fraud.” This category of fraud denotes that card details obtained through illicit methods such as phishing, skimming or hacking are used for fraudulent online transactions. Since consumers have mostly shifted from cash to transactions via payment cards and have become accustomed to online payments, opportunities for theft of payment card details have soared and they have attracted the attention of cybercriminals.
Theft of card information is usually the first step in the chain of credit card fraud. Further stages are resale, validation and monetization of the stolen data. These deals and activities take place in a massive underground economy, usually aided by underground online forums. These forums are popular platforms where card details are traded, thus generating huge revenues for cybercriminals. On these forums, fraudsters typically open a thread and write an advertisement for their products as a first posting. Potential buyers either reply within that thread asking to contact them or they contact the seller themselves using private message services or instant messaging services like ICQ.
The sales volumes thus generated appear to be substantial. It is estimated, for example, that the closure of several credit card related forums in 2012 prevented international fraud to the tune of £500 million [@gold2013identity]. It is therefore important to understand the characteristics of these online forums and the activity of cybercriminals using them.
The body of research into underground forums is growing but still limited. In particular, there are only a few studies available about credit card related forums. These studies mainly focus on the organisation and the structure of the forums but less on the content itself, that is, the products traded and the activity of the traders on these forums [@allodi2015then; @afroz2013honor; @yip2013forums]. In addition, existing studies are usually based on either expert interviews or examinations of forums that have been shut down by law enforcement agents. Active forums are rarely analysed. The examination of closed forums may be problematic since they may differ from those still existing, especially when this difference is the reason why they are closed. Moreover, cybercrime evolves rapidly and tackling this type of crime requires an understanding of the current situation and activities.
In this paper, we collected data directly from the discussions on the underground forums that we studied, with emphasis on product offers and advertisements posted by potential sellers on the forums. Since sellers come earlier in the fraud chain than the other actors, it might be more efficient to tackle credit card fraud by stopping sellers than buyers, for example. As a result, we excluded buyers and money mules from this study (we will study them in future work).
[**Research question.**]{} In order to overcome the literature gaps and methodological concerns we highlighted earlier, this paper aims to shed light on the current situation of underground online forums by analysing real data collected from active forums. We set out to answer the following research question: What does the current situation of underground credit card forums look like? That is, what are the typical features of the forums, what products are sold and how can the activity of the traders be characterised?
[**Hypotheses.**]{} After studying existing research literature, we defined the following hypotheses to guide our analyses:
- [Hypothesis 1]{} ([$H_{1}$]{}). Prices of credit card numbers combined with additional information about the cardholder are higher than prices of credit card numbers alone.
- [Hypothesis 2]{} ([$H_{2}$]{}). On active carding forums, a small number of traders are responsible for a large proportion of traffic.
- [Hypothesis 3]{} ([$H_{3}$]{}). Specialisation is discernible on carding forums, that is, most of the traders sell only one product type.
- [Hypothesis 4]{} ([$H_{4}$]{}). Specialised traders sell their products at lower prices than unspecialised traders.
- [Hypothesis 5]{} ([$H_{5}$]{}). The carding forums under analysis have working reputation systems that are sophisticated as those of legal marketplaces.
- [Hypothesis 6]{} ([$H_{6}$]{}). The vast majority of actors are not operating on more than one forum.
Our analyses confirmed [$H_{1}$]{}, [$H_{2}$]{} and [$H_{6}$]{}. [$H_{4}$]{} was partially rejected, while [$H_{3}$]{} and [$H_{5}$]{} were completely rejected. Details of the analyses are in Section \[sec:analysis\].
[**Contributions.**]{} In this paper, we established an outline of active forums and their defining features. To this end, we analyzed five carefully selected active forums in detail regarding their traders and products. This investigation thus provides insights into current underground online forums, unlike previous work that studied forums that were shut down. This paper provides some insights into product proportions on carding forums, since existing literature does not provide any sound information on that. Overall, we present a comprehensive overview of carding forums. We identified various product types with different prices. Our findings suggest that a small number of traders are responsible for the majority of the traffic observed on the underground forums. A distinct pattern of seller specialisation is not yet discernible from our findings.
Background and Related Work
===========================
In this section, we explore products sold on underground forums, their prices, seller prolificacy, seller specialization, and seller reputation, as presented in previous work. In addition, we develop some hypotheses from the general findings presented in existing literature. These hypotheses form the basis of our analysis, presented later in this paper.
Products and prices
-------------------
Products and services traded on carding forums can generally be classified as credit card information, bank account information, credentials or online payment services [@ablon2014markets]. In 2011, Motoyama et al. [@motoyama2011analysis] analysed the records of six closed forums and identified online payments, game-related accounts, credit cards and financial accounts as being the items most traded. Stone-Gross et al. analysed the Spamdot forum, studying the tightly connected community of buyers and sellers that were active on it [@stone2011underground]. Onaolapo et al. showed that cybercriminals active on such forums actively look for free samples of stolen credentials, and assess their quality before making a purchase [@onaolapo2016happens].
Credit card information is generally divided into three groups: *credit card numbers*, *dumps*, and *fullz* [@holt2010exploring]. Credit card numbers (also known as *“CVV”*) include at least the information printed on the card, that is, actual credit card number, cardholder name, expiration date and security code *CCV2* on the back of the card (not to be confused with CVV), and sometimes, the billing address and phone number. Dumps denote information from the tracks on the magnetic stripe of a card. These data are required to clone physical credit cards. Fullz provide further information on the cardholder including, for example, date of birth or social security number [@holt2010exploring].
On the prices sought for products on carding forums, Shulman [@shulman2010underground] states that, in 2006, credit card numbers were traded for \$1-\$25 each. Only two years later, credit card numbers were available for \$0.06. Shulman [@shulman2010underground] mentions three reasons that account for this decline: CVVs are becoming a commodity, monetizing information is more difficult and credit cards are beset by stolen online credentials. In the April 2015 report on Internet security, Symantec [@symantec2015] indicates a price range of \$0.50-\$20 for CVVs. These rates, on the one hand, confirm that there are details sold at low prices and, on the other hand, show that there are still cards sold for \$20.
Sood and Enbody [@sood2013crimeware] provide a more detailed estimation of rates charged per credit card number. Numbers from the USA cost \$4-\$10 on average, from Canada \$5-\$7 and from the UK \$6-\$8. Classified according to credit card types, a classic or standard credit card number from the USA or Canada costs \$8-\$10, a gold card \$15-\$20 and an Amex \$6-\$10. Classic and Amex cards are the cheapest in the listing of Sood and Enbody [@sood2013crimeware]. Nevertheless, they are still more expensive than the lower limits of their quoted price range (\$4/\$5). However, it is not ideal that these rates have not been observed but estimated and it remains unclear on which basis they have been calculated.
Reasons for price differences include the types of cards and countries of origin, as already mentioned, in addition to the rarity and the quantity of the products to be purchased [@hutchings2014crime]. Discounts on purchases of large card quantities lower the price per item. Furthermore, cards with more personal information available, with high balances and extended expiration dates and freshly acquired cards tend to be more expensive [@ablon2014markets].
Since there is no sound information available on the product proportions on carding forums so far, this paper provides some insights into that. Existing literature does not state reliable prices for dumps or fullz. As cards containing more information tend to be more expensive than those with less information, however, we infer and hypothesise ([$H_{1}$]{}) that prices for fullz are higher than those for credit card numbers. For dumps, no hypothesis can be formulated derived from existing research literature. Due to the effort needed to monetise the information, low prices would be expected. However, once copied and successfully used to conduct a transaction, such a clone might be a lucrative means of payment. Also, the efforts necessary to steal the data (e.g. skimming dumps in a restaurant or collecting data on fake Internet sites) does not give a clear indication of the expected price differences between dumps and CVVs.
Seller prolificacy
------------------
Generally, there are several types of participants on the forums: sellers, buyers, intermediaries, mules, administrators, and others. These roles are not mutually exclusive; sellers may simultaneously be buyers. Although the total number of participants is in unclear, Ablon et al. [@ablon2014markets] argue that, based on expert interviews and literature review, the total number of participants on the forums is likely to rise. The increasing spread of different marketplaces and forums would facilitate access to one of them. At least from a historic perspective, Christin [@christin2013traveling] confirmed this growth of participants on underground platforms as he observed a linear increase of sellers during his half-year analysis of Silk Road, a large underground marketplace. In the aftermath of Silk Road’s take-down in 2013, the number of sellers on competitor[-]{} and newcomer-marketplaces has substantially increased, surpassing the original number of sellers on Silk Road [@soska2015measuring].
In terms of geographic location, forum users come from all over the world. Regarding the sellers, De Carbonnel [@decarbonnel2013] claims that Russian participants deliver the best quality, while participants from China, Latin America and Eastern Europe are the leaders with respect to quantity. These geographic patterns, however, vary depending on the types of forums and the services provided. An analysis of a marketplace offering SEO services locates the sellers mainly in India, Bangladesh, and the USA [@farooqi2015characterizing].
Examinations in relation to sales quantity reveal substantial differences in seller prolificacy. Farooqi et al. [@farooqi2015characterizing] identified an “insider ring” composed of several top sellers. This means that a small number of traders account for a large proportion of traffic on the marketplace. One common characteristic is that they joined the community very early and are frequent visitors to the pages. Christin [@christin2013traveling] agrees on the existence of several long-time sellers but also reports on a continual “come and go” of sellers. It is unclear, however, whether they leave the community after having made sales or due to unsuccessful attempts.
In terms of seller prolificacy, Motoyama et al. [@motoyama2011analysis] analysed the records of 6 closed forums and concluded that 10% of the sellers are responsible for 40%-50% of the goods traded. More generally expressed, D[é]{}cary-H[é]{}tu and Lepp[ä]{}nen [@decary2013criminals] reason that some sellers are more effective than others. Their conclusion is based on counting of advertisements of sellers on one underground forum. However, it is doubtful that counting ads is the right approach of quantifying success. Moreover, the analysis of several forums instead of one might have produced more reliable results.
It is common ground among crime scientists that crime is distributed neither randomly nor evenly [@felson2010crime]. That implies a small group accounts for more offenses than its expected share would be. As earlier stated, studies on marketplaces suggest the presence of some highly prolific users. We hypothesise ([$H_{2}$]{}) that on active carding forums, a small number of traders are responsible for a large proportion of traffic.
Seller specialisation
---------------------
Looking at the products sold per seller, several studies found evidence of specialisation amongst sellers. Derived from literature review and expert interviews, Kraemer-Mbula et al. [@kraemer2013cybercrime], for example, promote an ecosystem perspective to understand the actions of underground traders. Comparable to the legitimate business community, underground ecosystems includes actors that compete against each other, targeting competitive advantage. They try to reach this advantage by specialising in a particular type of product [@kraemer2013cybercrime].
By applying a framework of social organisation, Holt [@holt2013exploring] identified specialisation on underground forums too. While one third of sellers offered various products, two thirds focused on only one product category. As the Symantec report [@symantec2015] illustrates, there are perpetrators specialising in writing viruses, in distributing malware or in monetising credit cards, for example. In recent years, Symantec has observed an increasing professionalisation in all aspects in the underground economy. Their findings are supported by research literature. Sood and Enbody [@sood2013crimeware] also identified specialisation as a trend in underground markets. They argue that these markets are increasingly accessible to people with various technical skills. Hence, there is a division of labour due to differing skills. While analysing seller characteristics on black marketplaces, Soska and Christin [@soska2015measuring] discovered numerous specialised sellers, though there was a notable number of vendors selling different products as well.
These findings indicate that specialisation is present in the underground ecosystem as in the legitimate business world. Hence, we hypothesise ([$H_{3}$]{}) that specialisation is also discernible on carding forums, that is, most of the traders sell only one product type.
What does that mean in terms of product prices? We did not find any association between specialisation and product prices in existing literature. Resorting to economic theories [@smith1937inquiry], there are long-established economic “laws” that basically state that concentrating on one production task leads to a higher efficiency at that particular task. This efficiency enables an increase in production compared to unspecialised suppliers. Due to such economies of scale, products and services can be offered at significantly lower costs and prices can be cut. Applying this to traders on carding forums, we expect cost[-]{} and price-reducing effects when sellers specialise in trading of a single product category (due to the economies of scale). Thus, we hypothesise ([$H_{4}$]{}) that specialised traders sell their products at lower prices than unspecialised traders.
Seller reputation
-----------------
One key aspect in the underground economy is reputation [@ablon2014markets; @decary2013criminals; @motoyama2011analysis]. A reputable seller is more likely to be trusted and thus more likely to engage in trades and to complete transactions. On forums, reputation is usually established by positive customer feedback. Buyers may rate their sellers by giving positive ratings if the ordered products have been successfully delivered, and negative ratings if the seller has not delivered and was rather a ripper. Consequently, a seller’s positive reputation score presents his/her threads in a more credible light, and these sellers have a higher chance of acquiring multiple customers [@holt2013exploring].
However, the effort to establish baseline reputation appears to be laborious. Before half of traders receive their first positive feedback, for example, they write approximately sixty posts [@motoyama2011analysis]. In this case, the reputation process is intrinsically peer-driven. Sellers are dependent on recommendations by buyers. Sometimes, forum administrators provide a vetting process, often in addition to the peer-driven process and often with intransparent criteria. In those cases, entry costs are relatively high and access to higher tiers is tight [@ablon2014markets].
The emphasis on reputation and trust means that it is indispensable for competitive forums to have a well-functioning reputation system. Again, since the above-mentioned findings are widely based on expert interviews and therefore remain relatively vague, the actual status regarding currently running carding forums is not known. Since trades on carding forums depend on relationships between mutually distrustful parties, we argue that trust is even more important than in legitimate trades. In the event of an unsuccessful deal, the parties hardly have any legal remedies and countermeasures available, except for a negative reputation rating. We thus hypothesise ([$H_{5}$]{}) that the carding forums to be analysed have working reputation systems that are at least as sophisticated as those of legal marketplaces, for instance eBay. This expectation applies only to open forums where everybody can participate.
As discussed, the efforts needed for gaining trust are extensive. A consequence might be that sellers concentrate on establishing reputation on one specific forum instead of several forums. It is therefore not expected that sellers are present on multiple forums. By assuming this, we support Motoyama’s et al. [@motoyama2011analysis] expectation of non-existing multiple accounts. In contrast, we disagree with Ablon et al. [@ablon2014markets] who argue without providing any reasons that sellers would advertise on multiple marketplaces. We hypothesise ([$H_{6}$]{}), therefore, that the vast majority of actors are not operating on more than one forum.
Methodology {#sec:methodology}
===========
In this section, we describe our data collection approach. We collected names of underground forums from various sources, and selected 5 forums that matched our selection criteria for examination. Data spanning a period of three months was collected from the forums, and we tested our hypotheses on the data.
Forum search
------------
The first step of the examination is the forum search. We took the following steps to find carding forums: First, we collected names of forums that were mentioned by research literature. Second, we carried out searches via Google. Third, we used other search engines and information pages, namely `Onion.city` search via Tor network, `webstatsdomain.org` and “The Hidden Wiki.” Finally, we searched forums that we already found for references to other forums. In the latter case, we adopted the method of snowball sampling [@biernacki1981snowball]. The only selection criterion at this point was that, due to the authors’ language abilities, the forums had to be at least partly in English or in German.
By this means, we found 25 forums, 15 of them via Google. The 25 forums are listed in Table \[tab:discoveredforums\]. The forum names mentioned in existing literature research were of little use since all the mentioned forums had already been shut down. We found five forums through other forums, and five from listings and other search pages. Although numerous forums were listed, most of them did not exist anymore. Two of the forums discovered during the first search in February 2015 were shut down at the beginning of the analysis in June 2015. Notwithstanding, the carding underworld seems to be dynamic. We found one active forum containing posts dating back to 2008.
**Forum name** **Forum address (http://...)**
------------------ --------------------------------------------------------------
Agoraforum [lacbzxobeprssrfx.onion/](lacbzxobeprssrfx.onion/)
**Altenen** [www.altenen.com](www.altenen.com)
**Crdpro** [crdpro.su](crdpro.su)
**Crimenetwork** [crimenc5wxi63f4r.onion](crimenc5wxi63f4r.onion)
Cardingforum [www.cardingforum.org](www.cardingforum.org)
Hackingforum [hackingforum.ru](hackingforum.ru)
Unixorder [www.unixorder.com](www.unixorder.com)
Crdclub [crdclub.ws](crdclub.ws)
Carderscave [www.carderscave.ru](www.carderscave.ru)
Infraud [infraud.cc](infraud.cc)
Lampeduza [lampeduza.so](lampeduza.so)
Blackstuff [www.blackstuff.net/forum.php](www.blackstuff.net/forum.php)
Bus1Nezz [bus1nezz.biz](bus1nezz.biz)
Cardingmafia [www.cardingmafia.ws](www.cardingmafia.ws)
**Bpcsquad** [www.bpcsquad.com](www.bpcsquad.com)
Procarder [www.procarder.ru](www.procarder.ru)
Cardersforum [www.cardersforum.se/](www.cardersforum.se/)
Crimes [crimes.ws/](crimes.ws/)
Carderbase [carderbase.su](carderbase.su)
Carder [carder.me](carder.me)
Darkstuff [www.darkstuff.net](www.darkstuff.net)
Coinodeal [coinodeal.com](coinodeal.com)
**Tuxedocrew** [www.tuxedocrew.biz](www.tuxedocrew.biz)
Privatemarket [privatemarket.us](privatemarket.us)
Omerta [omerta.cm](omerta.cm)
: Names and web addresses of discovered forums. The ones we focused on are highlighted in boldface text.[]{data-label="tab:discoveredforums"}
Besides forums, we discovered more than two dozen stores (e.g. *Globalcards* and *Dexter*, offering mainly credit card numbers). We did not include these single-vendor marketplaces in our analysis since they differ significantly from forums. They do not gather multiple sellers, they have no reputation systems, and users normally do not communicate with each other. Hence, they do not meet our interest in interactions between forum members.
To narrow down the analyses, we chose five out of the 25 forums (see Table \[tab:threadsandproductsperforum\]) for detailed examination: *Altenen*, *Crdpro*, *Crimenetwork*, *Bpcsquad*, and *Tuxedocrew*. The first three are the largest forums we found (as measured by number of posts) and should thus be the most fruitful ones. We excluded *Agoraforum* despite possessing the greatest number of posts, because 99% of its posts are requests for referral links for registration on Agora Marketplace. Tuxedocrew is included as it is one of the smallest forums and has existed for around two years. Thus, it is not entirely new and it might provide interesting insights when its content is compared to that of larger forums. Finally, we chose Bpcsquad since it is a medium-sized forum. It is remarkable to note that it is the largest one of the very new forums. To sum up, our selection criteria are forum size, founding date, and, to a lesser degree, content. These criteria should ensure a good mix. Altenen, Bpcsquad and Tuxedocrew are mainly or exclusively in English, Crimenetwork in German and Crdpro half in English, half in Russian.
Temporal sampling
-----------------
In order to have a comparable time-coverage of all five forums, we monitored activity on them over a period of three months, specifically from April to June 2015. This means that a snapshot was made by the end of June and data of the previous three months was collected. This three-month period was determined by the largest forum, Altenen.
This limitation to three months meant that no full activity-record could be recorded. Furthermore, it introduced the risk of catching three “special” months instead of a whole year’s coverage. However, we argue that the current situation is of interest and not the past, and that three months are still more advantageous than shorter periods. Moreover, the collected data showed that a substantial volume of posts can be captured in three months, especially from the larger forums. Admittedly, a longer period would be beneficial for the smaller forums. In summary, a consistent and thus comparable time period is favoured over a larger number of posts from small forums.
Data collection
---------------
We collected information on the numbers of members and posts, content, forum-accessibility, languages, and founding dates. For the selected forums, threads where potential sellers advertise their products were collected. A systematic review of the entirety of these forums was not possible. No activity records, copies of databases or web crawler services were available. The analysis was effected from the user’s perspective. For instance, we did not analyse private messages used to arrange and complete trades. Nevertheless, this method provides an enlightening snapshot of the current carding situation. Where necessary and possible, we set up login credentials to gain wider access to the forums.
Threads published between April and June 2015 were collected for further analysis. Ads that were created before the three-month observation period were not collected. However, it is likely that older threads were still successfully promoting products and generating sales. Therefore, we captured older threads in cases where an activity in the form of answer postings or vouchings during the three months was registered. Such activity suggests that deals had taken place. Indeed, it was crucial to consider such older threads since it was expected that long-established insider rings existed on the forums, as pointed out by Farooqi et al. [@farooqi2015characterizing].
The threads usually describe the advertised products and their prices. Whenever an unspecified price range was indicated in an ad, we chose the lowest price for analysis. Calculating the mean value may distort the picture presuming that, for example, if only one high-priced gold card is offered in addition to many low-priced standard cards. In cases where various products were advertised in a single thread, each entry was considered equally.
To keep the focus on carding, we limited the spectrum of investigation to typical financial cybercrime related data: credit card numbers (CVVs), dumps, fullz, PayPal-credentials and Western Union (WU) payment transfers. We excluded other carding-related services such as ordinary online store credentials or monetisation-services.
In order to operationalise “traffic” on the forums, as necessary for hypothesis [$H_{2}$]{}, D[é]{}cary-H[é]{}tu and Lepp[ä]{}nen [@decary2013criminals] counted advertisements as indicators. In our view, however, the consideration of vouchings would be more promising to obtain an accurate impression of a seller’s “performance.” Vouchings are evident signs that successful transactions have been made. Yet, since probably not every buyer vouches for the seller, counting the number of vouchings tends to underestimate the traffic. Conversely, there might be rippers vouching for each other without having made any transaction. Since Farooqi et al. [@farooqi2015characterizing] and Christin [@christin2013traveling] also relied on vouchings and member feedbacks, using them to calculate revenues, counting vouchings seems to be an appropriate method.
To determine whether users are specialised in one product category, as a prerequisite to be able to test [$H_{3}$]{} and [$H_{4}$]{}, we checked their personal profile sites. These pages display complete lists of all threads and posts written by the corresponding users. This method enabled us to see whether multiple products were advertised. The definition of specialisation is relatively strict. For instance, if users sold credit card numbers and PayPal-credentials, we did not consider them as being specialised. Only very narrowly related categories, for instance credit card numbers and fullz, were treated as identical product categories in this respect.
Hypothesis [$H_{6}$]{} requires us to determine whether the same users are present on several forums. We carried out searches for users throughout the selected forums, and compared their identity details. Where applicable, these are username, email address, ICQ-number and Yahoo-ID. These details were collected from the postings and the users’ profile pages.
Finally, we gathered information on the reputation system of each forum from various sources. Depending on the forum, these are the FAQs, specially installed forum threads, terms and conditions or customer information sites. Also own observation and interpretation were employed to grasp how the reputation systems functioned.
In total, we collected 388 threads. They advertised 987 individual products in total, that is, on average, each thread promoted 2.5 individual products (e.g. CVV USA Classic). The figures for the individual forums are in Table \[tab:threadsandproductsperforum\].
**Forum** **Threads** **Individual products**
-------------- ------------- -------------------------
Altenen 206 431
Crdpro 57 270
Crimenetwork 96 136
Bpcsquad 25 130
Tuxedocrew 4 20
: Threads and individual products per forum[]{data-label="tab:threadsandproductsperforum"}
Analytical strategy
-------------------
Monitoring of the forums required a combined methodical approach. We analysed the content we collected both qualitatively and quantitatively. Qualitative analysis was applied for content categorisation and analysis of reputation, while quantitative analysis was applied for comparisons of products and prices and determination of traffic per seller.
To prepare the data for analysis, a categorisation of thread content was necessary. This procedure required a qualitative research approach and was thus done manually in Excel using content analysis method. We stuck to clear coding rules in order to avoid subjective and inconsistent categorisations. Some approximate categories were already provided by existing research literature (e.g. “credit cards”). However, these categories were somewhat too coarse and further sub-categories had to be created (e.g. “CVV” or “dumps”). Hence, The categorisation process is a combined product of deductive practice (assigning content to given categories) and inductive practice (building new categories based on content) [@kluge1999empirisch]. Yet, it is important to ensure that the categories do not become too small and thus render subsequent quantitative calculations impossible. For example, dumps are sometimes advertised divided into *track 1* or *track 2* dumps. Since the aim of the categorisation is to obtain meaningful product categories, we avoided such fine distinctions. Therefore, credit card data was coded according to product category, country and product type (e.g. CVV, UK, gold). Visa and Mastercard details were not explicitly differentiated since they are usually treated interchangeably by the traders.
Next, we imported the data into SPSS software for statistical analyses. At first, we ran general frequency calculations. In order to test [Hypothesis 2]{}, the number of traders in relation to the generated traffic was quantitatively computed in the shape of a Lorenz curve. For [$H_{3}$]{}, the frequencies of the specialised users were compared to the unspecialised ones. [$H_{1}$]{} and [$H_{4}$]{} required the application of inference statistics. Since their price distributions resembled Poisson rather than a normal distribution, we performed Mann-Whitney-U tests to test whether there were significant differences between the values. Regarding [$H_{5}$]{}, the reputation systems of the forums were evaluated qualitatively. To assess the degree of sophistication in relation to legitimate marketplaces, we compared them to eBay’s system. Strictly speaking, eBay is not a forum and thus not fully comparable. However, we found no suitable legitimate large-scale forum set up to enable trading. Furthermore, eBay comes close to the system of advertising, trading, and buying exercised on forums. Finally, in order to examine whether sellers operated on more than one forum ([$H_{6}$]{}), we reproduced and interpreted the proportions of multiple representations across all forums.
We carried out all these calculations both for the entirety of the threads and for each forum separately. Therefore, the results (Section \[sec:analysis\]) are reported in aggregated form, and where applicable, and if enough cases are available, for every individual forum.
Data Analysis {#sec:analysis}
=============
In this section, we describe our analyses of the five selected forums, and our findings.
Overview
--------
We discovered 25 forums, out of which we selected five forums for analyses. In this section, we describe general attributes of the 25 discovered forums. The attributes are name, members, total posts, accessibility, languages, and founding date.
[**Name.**]{} The names and full website addresses of the discovered forums are listed in Table \[tab:discoveredforums\]. They have a wide array of top-level domains, for example [.com](.com), [.ws](.ws) (Samoa), and [.so](.so) (Somalia). The locations that the forums really operate from are usually unclear, and apart from two German-speaking platforms, cannot be derived from the forum content.
[**Members.**]{} As of June 2015, the smallest forum had 1,100 members, while the largest had 148,800 members. A comparison between the first search in February 2015 and the second in June 2015 revealed some substantial increases in members. Cardersforum, for example, grew from 44,200 to 45,700 members (3.4% increment), and Cardingmafia grew from 98,700 to 121,500 members (23.1% increment). Altenen, already a large forum in February, was more than twice as large four months later (from 60,700 to 148,800 members, 145.1% increment).
It is not clear how many of the members on each forum were actually contributing. Two forums indicated in their forum statistics that only a fraction of members were really active. On Altenen, these were 38,300 of its 148,800 members (25.7%); on Carder, these were 1,500 of 10,200 (14.7%). Neither forum disclosed what “active members” actually meant.
[**Total posts.**]{} The number of posts varied between 150 and 15,778,599. Besides advertisements, the posts comprised mainly answers to advertisements or contributions to discussions. Typical answers to offers for sale are “interested please contact me” or “made deal and worked.” In line with the increase in number of members, the number of posts also increased between the two searches (e.g. Cardingmafia from 31,900 to 37,600 posts, 17.9% increment). On Altenen, the number of posts doubled (from 607,100 to 1,265,500 posts, 108.5% increment).
[**Accessibility.**]{} Nine forums were completely open. This means that everybody could access them for free or even without registration. More than half of the forums had private VIP areas that required special registration to join. Access to these areas usually required a recommendation or an invitation by other members. Three forums charged registration fees, \$50 in the case of Lampeduza, and \$100 in the cases of Infraud and Omerta respectively.
[**Languages.**]{} The forums were mostly in English and Russian, with two in German. Some forums contained international sections in various languages. However, the number of posts in such sections were consistently small.
[**Founding date.**]{} We estimated the founding dates of the forums from the oldest posts found on them, mostly in the introduction or announcement sections. These are not necessarily the founding dates as older posts might have been deleted in the meantime. In addition, forums might have been shut down and reopened under another name (e.g. Crdpro was formerly Carderpro). We estimated founding dates between 2008 and 2014. A large number of the forums were launched in 2013.
In terms of size, the median number of members was 28,850, while the median number of posts was 58,150. Excluding the special case of Agoraforum, the number of posts per member varied between 0.1 (Privatemarket) and 18.6 (Crimenetwork). There are thus forums where only a fraction of the members post messages and there are some where members post numerous messages on average. However, these results have to be treated with caution as posts or members might have been deleted during the existence of the forums.
Detailed analysis
-----------------
In this section, we present the results of hypotheses testing for the five selected forums.
[**Products and prices.**]{} Table \[tab:productsandprices\] presents the numbers, proportions (in %) and prices (in US\$) per product category. CVVs are further divided by product type. Credit card numbers cost on average \$10, dumps and fullz more than \$30. PayPal credentials are advertised for \$3. Western Union payments of \$100 are sold for \$15. As hypothesised ([$H_{1}$]{}), prices for fullz are higher than those for credit card numbers (CVV: mean = 10.08, median = 10.00; fullz: mean = 31.82, median = 30.00). The difference is statistically significant (Mann-Whitney-U = 4011, z = -12.86, p < 0.01).
**Products** **Number** **Proportion (%)** **Price (\$)**
---------------- ------------ -------------------- ----------------
**CVVs** 465 47.1 10.08
*Classic* *98* *9.9* *9.93*
*Gold* *14* *1.4* *16.86*
*Amex* *66* *6.7* *12.34*
*others* *16* *1.6* *13.00*
*unspecified* *271* *27.5* *9.06*
**Dumps** 234 23.7 34.52
**Fullz** 140 14.2 31.82
**PayPal** 133 13.5 3.01
**WU (\$100)** 15 1.5 15.00
**Total** 987 100.0
: Products and prices (mean) in total.[]{data-label="tab:productsandprices"}
Examined per individual forum, prices of CVVs do not vary substantially, those for the other products show considerable variation. Besides price differences, the product proportions alsi vary per forum. On Altenen, for example, dumps have a share of 8% of the products analysed. On Crdpro, their proportion is 52%. Yet the absolute numbers are partly very low and the values may thus lack reliability. The prices per product category depend widely on the effective composition of these categories, that is, on the relative frequencies per product type and country of origin. Since the absolute numbers are too low to display these values for each forum, Table \[tab:productsandpricespertype\] shows them summarised across all forums. Only product types consisting of at least 10 cases are considered. There is indeed substantial variation between different product types and countries. Amex and Gold cards are consistently more expensive than classic VISA or Mastercard cards. US products are the cheapest, while European products tend to be more expensive.
**Number** **Price (\$)**
----------- --------------- ---------------- ------- -------
**CVV** **Australia** **Classic** 12 12.75
**** **Canada** **Amex** 10 14.20
**Classic** 14 10.92
**** **UK** **Amex** 10 15.10
**Classic** 17 11.94
**** **USA** **Amex** 25 7.02
**Classic** 31 5.46
**Dumps** **Canada** **Classic** 14 31.43
**Gold** 12 45.25
**** **EU** **Classic** 20 41.75
**Gold** 20 58.50
**** **USA** **Classic** 29 19.17
**Gold** 27 30.93
: Products and prices (mean) per product type with at least 10 cases.[]{data-label="tab:productsandpricespertype"}
[**Seller prolificacy.**]{} The products are not sold evenly throughout all sellers on the forums. The Lorenz curve in Figure \[fig:trafficpertradersall\] shows that around 70% do not generate any obvious traffic as seller, whereas a single user generates 44% of all traffic. This user joined Altenen in summer 2014 and sells CVVs of various countries.
Disentangling the individual forums from the total results in similar pictures (Figures \[fig:trafficpertradersaltenen\], \[fig:trafficpertraderscrdpro\] and \[fig:trafficpertraderscrimenet\]). Altenen has the most unequal distribution. Crimenetwork’s distribution is not as extreme but still far from being equal. Bpcsquad and Tuxedocrew have too few vouchings to calculate a Lorenz curve. However, Tuxedocrew has posts dating back to 2013 that still receive vouchings but only in small numbers. Crdpro does not diverge from these distributions.
Considering the total amount of traffic, however, it is striking that the figures are low, both in the English and in the Russian speaking part. Although Crdpro has 17 times as many users as Bpcsquad, for example, it produced only twice as many advertising threads during the time of observation. Detailed analysis revealed that a substantial number of these threads contain links to shops. Furthermore, there have been no new entries in the two VIP areas since 2013 and the forum appeared to have been disconnected during some summer months in 2013.
In terms of products, the top three sellers on Altenen sell CVVs and WU payments, on Crdpro they sell dumps and on Crimenetwork again CVVs. As far as reputation is concerned, the high-profile sellers have usually high reputation ratings. To conclude this section, the hypothesis ([$H_{2}$]{}) that a small number of traders are responsible for a large proportion of traffic is accepted. However, we could not confirm the presence of an insider ring, as proposed by Farooqi et al. [@farooqi2015characterizing]. Overall, only five out of the twenty most prolific users registered in the founding year of the according forum. The others joined later. However, it is possible that some sellers have more than one account and have thus several “joining-dates.” This possibility does not seem to be very likely, mainly due to the expected effort needed for establishing reputation for each account.
[**Seller specialisation.**]{} In total, the majority of the users on the forums are not specialised (see Table \[tab:specandunspecusers\]), that is, most users sell more than one type of product and [Hypothesis 3]{} has to be rejected.
------------------ ------------ -------------------- ------------ --------------------
**Number** **Proportion (%)** **Number** **Proportion (%)**
**Altenen** 64 31.1 142 68.9
**Crdpro** 41 73.2 15 26.8
**Crimenetwork** 24 25.0 72 75.0
**Bpcsquad** 15 60.0 10 40.0
**Tuxedocrew** 3 100.0 0 0.0
**Total** 147 38.1 239 61.9
------------------ ------------ -------------------- ------------ --------------------
Regarding Crdpro, Bpcsquad and Tuxedocrew, the hypothesis would be true. However, a closer look on Crdpro reveals that its users generally sell a large variety of the same product category, instead, for instance credit cards from many different countries. This pattern is exactly the opposite of Crimenetwork’s. Users on Crimenetwork usually sell different product categories but not various types within the same product category, for example only CVVs from Germany.
The results for [$H_{3}$]{} raise the question about differences between specialised and unspecialised sellers. Building upon [$H_{2}$]{} and taking into account the number of vouchings these users receive, no major differences are discernible. Regarding Altenen, for example, seven out of the twenty users with the most vouchings are specialised. This equals approximately the calculated specialisation rate of 31%.
In terms of prices per product, there are some differences. CVVs, dumps, and PayPal-credentials advertised by specialised users are cheaper than those of unspecialised users; fullz are more expensive (see Table \[tab:productcategories\]). However, only the price difference for dumps is statistically significant at the 95% confidence level (Mann-Whitney-U = 1643, z = -4.83, p < 0.01). [Hypothesis 4]{} thus has to be partially rejected.
That said, it is delicate to summarise product types because every type and country has its own price. A careful comparison would only contain a single product type. Hence, we made such a price comparison for US Classic CVVs, the most prevalent product type. The result shows no significant difference (Mann-Whitney-U = 54, z = -0.63, n.s.).
A striking aspect that Table \[tab:productcategories\] reveals is the distribution of advertised product categories among specialised and unspecialised users. Unspecialised users hardly advertise any dumps (7.5% of all products). In contrast, dumps are almost half of the products (45.4%) that specialised users advertise.
------------ ------------ -------------------- ---------------- ------------ -------------------- ----------------
**Number** **Proportion (%)** **Price (\$)** **Number** **Proportion (%)** **Price (\$)**
**CVV** 166 39.1 9.28 299 54.7 10.46
**Dumps** 193 45.4 32.56 41 7.5 42.61
**Fullz** 45 10.6 35.86 95 17.4 30.20
**Paypal** 21 4.9 1.99 112 20.5 3.17
**Total** 425 100.0 547 100.0
------------ ------------ -------------------- ---------------- ------------ -------------------- ----------------
[**Seller reputation.**]{} We hypothesised ([$H_{5}$]{}) that carding forums have working reputation systems at least as sophisticated as those of legal marketplaces. The reputation systems are as follows:
*eBay:* Buyers on eBay can leave feedback for a seller after a purchase and transaction ends. These ratings determine the “feedback score.” Positive feedback gives one point, neutral feedback does not change the score, and negative feedback subtracts one point [@ebayfeedback]. As an additional protection measure, eBay refunds the purchase price in the event of a non-delivery [@ebaymoneyback]. Furthermore, a seller protection service identifies high-risk buyers in order to avoid non-payments [@ebaysellerprotection].
*Altenen:* Altenen’s basic reputation system works in the same way as eBay’s. A user’s “reputation power” consists of the number of positive minus the number of negative feedback points. The median score of the observed sellers was 1. In addition, users are allowed to rate threads. Sellers also have the opportunity of paying \$50 to the “Altenen buyer protection reserve” that is used as backup payment service. In case of non-delivery, buyers receive their money back out of this fund. Altenen also offers an escrow service that protects buyers from non-delivery and sellers from non-payment. In using that service, a buyer pays the money plus a transaction fee of \$5-\$30 to Altenen. Once the requested products are delivered, the amount is released to the seller.
*Crdpro:* In theory, Crdpro has a feedback system identical to Altenen’s. However, the system was disabled during our observation. As a consequence, apart from some long-standing members, all other members had neutral feedback scores. An escrow service was not provided.
*Crimenetwork:* Crimenetwork’s reputation system was based on “likes,” similar to Facebook. Members may “like” other members. Like-scores between 0 and 857 were recorded with a median of 36 likes. Crimenetwork’s escrow service is comparable to Altenen’s. A fee of 4% of the purchase price is charged for successful transactions.
*Bpcsquad:* As seen with Altenen and Crdpro, members may rate other members on Bpcsquad by giving positive, neutral or negative feedback. The scores ranged from 0 to 80, with a median of 0. Furthermore, there is a thread rating possibility. Bpcsquad also activated an escrow service but did not provide information on transaction fees.
*Tuxedocrew:* Tuxedocrew’s reputation system differed from those seen so far. Users could only rate threads but could not give any feedback for other users. Instead, the forum administrator could award users with special titles. The criteria that had to be met to receive these titles were not published on the forum. Tuxedocrew also offered an escrow service and charged a 15% fee.
Overall, only Altenen’s system appeared to be similarly elaborated as eBay’s. However, it had an amateurish touch, especially the \$50 buyer protection reserve which is not able to cover substantial amounts. The other forums had fewer features than eBay and even those were not always working. [$H_{5}$]{} is thus rejected.
[**Presence of sellers on multiple forums.**]{} Finally, we examined whether users were present on several forums. This was done for all kind of sellers including high[-]{} and low-profile traders. In total, only six sellers were found trading on more than one forum, namely two on Altenen and Crdpro, two on Altenen and Bpcsquad, and two on Bpcsquad and Crdpro. A detailed analysis of these sellers showed that they were not high-profile but rather low-profile unsuccessful traders trying their luck on several platforms. [Hypothesis 6]{} is thus confirmed; concentration on a single forum was expected.
Discussion
==========
In this section, we summarise our findings on the carding forums that we studied. The apparent lack of specialisation of forum users is also described. Finally, we highlight limitations of the study.
[**Summary of our findings.**]{} The prices sought for the products offered on the forums lie within the range given by the reviewed literature. Dumps and fullz are relatively expensive; they are more than three times as expensive as credit card numbers (CVVs). This may be due to the effort needed to gain or monetise the data, the amount of information available, the higher rewarding potential, and differing demand and supply. Brison [@brison2015] argues, for example, that dumps generally promise a higher payoff than CVVs. In contrast, CVVs are well-represented on the forums and thus seem to be available in abundance, which might push prices downwards. However, contrary to Shulman’s assumption [@shulman2010underground], the prices of CVVs are still solid. Taking into account the large proportion of CVVs on the investigated forums, trading credit card numbers is presumably still a lucrative business. PayPal-credentials are well-represented on some forums as well, but so far do not seem to replace credit cards as the most attractive trading goods. Western Union money transfer services play only a marginal role on most of the forums.
The products are advertised by sellers with varied success. Even though some users complete hundreds of transactions, most users do not sell anything at all. This means that the trading sections of the forums are profitable distribution channels for high-profile actors. This domination by a few traders implies that the forums are not typical forums characterised by mutual exchanging and participating users. In the carding world, there is somewhat a clash of prolificacy and -– arguably –- professionalism observable.
Referring back to the methodology part, counting of vouchings instead of ads, the latter [@decary2013criminals] was probably more suitable to determine criminal performance. Some prolific sellers had only one ad but received dozens of vouchings. Counting of ads would have overlooked that.
Specialisation is not a key characteristic of sellers, even not of high-profile traders. Specialisation was observed mostly on Crdpro. This might be due to the high proportion of dumps sold on this forum. Dumps constitute almost half of the products sold by specialised users on Crdpro. Dealing with dumps appears to demand a higher degree of specialisation than dealing with only electronically obtainable products. Unlike CVVs or credentials, the acquisition of dumps requires a connection to the physical world. Therefore, perpetrators cannot stay in the underground cyberworld only. As a result, it might be costlier for unspecialised users to acquire dumps, thus forcing them to sell dumps at higher prices, which would confirm Smith’s economic theory [@smith1937inquiry].
Yet the majority of sellers are not specialised. It could be argued that if they are apt or have valuable data sources, they know and distribute other types of illicit products and services. On the contrary, unsuccessful sellers try their luck with another product if it does not work with the first. These users, though, might as well be rippers. Advertising a large array of products might be done to give the impression of a prosperous seller, or they just try various products in the hope that somebody would engage in a trade eventually.
Overall, it is possible that the scope of analysis regarding specialisation was too narrow. There may be a specialisation in the underground world in larger terms where carding itself is already a specialisation. Another reason might be that carding is not as complicated as other cybercrimes like DDoS-attacks or large-scale spam campaigns. Regarding DDoS-attacks, for example, taking advantage of security vulnerabilities and manipulating compromised machines to send huge amounts of data may require more time and skills than stealing and trading credit card data. Therefore, it makes more sense to be specialised in those domains.
At least on the investigated forums, and given the available details, users are not present on more than one forum. It might thus be true that the effort needed to reach a certain reputation level deters users from establishing themselves on multiple forums, as Motoyama et al. [@motoyama2011analysis] proposed. This effort could also be the reason why most users do not have any ratings at all, as the analysis showed. Another reason for this, however, might be the presence of rippers. Regarding users with high reputation and many vouchings, it is highly unlikely to find any rippers among them. Among users without any reputation scores and vouchings, the proportion of rippers could be large. It is interesting to note that an expert interviewed by Ablon et al. [@ablon2014markets] estimates that around 30% of all sellers are rippers.
In general, and if not stated otherwise, all our findings apply to all five examined forums. However, there are some differences. Sales on Crimenetwork are not distributed as extremely unevenly as on other forums, neither are there numerous specialised users present. Various people sell various goods. Crimenetwork is thus more forum-like in terms of mutual exchange and participation than the other forums. The high number of posts per member confirms this perception. In addition, the forum gives the impression of being well-maintained. It has a myriad of banned users and the administrators comment rigorously if users do not stick to the rules (e.g. in case a post does not fit into a thread).
Crdpro is the obvious opposite. Its best times were probably in the past. It appears to be in decline. It does not seem to be monitored by the administrators, there is no escrow service, the reputation system does not work, and there is in general not much traffic. It might be a question of time until the entire forum will be closed.
On Tuxedocrew, the smallest forum, there are only four recorded threads receiving any vouches. However, the number of vouchings are low and it is thus questionable how fruitful the business really is. What might be possible is that some recurring customers buy a lot and do not always vouch. A reason for the small size of this forum might be the high charges for the escrow service or, even more likely, the lack of a user-based reputation system. Only the administrator is able to assess other users, based on intransparent criteria. This might be too little to build trust among the users and to boost trade.
Bpcsquad and Altenen do not particularly diverge from the general findings. Bpcsquad is relatively small and the low number of ratings may denote unsuccessful deals. It is thus uncertain how strongly this forum will grow in the future. In contrast, the enormous increase in members on Altenen is impressive. Apparently, it is attractive to be part of this large community. Forums with numerous users usually have diverse products, and a multitude of potential buyers, that is, high supply and demand. Both platforms have reasonable and working basic reputation systems. Altenen provides an additional, arguably pseudo-protection measure.
[**Limitations.**]{} We encountered a number of limitations during the study, and they are mentioned in this section. Firstly, the three-month period does not allow long-term conclusions. After all, due to the technique of considering the vouchings of this time period, older and often very profitable ads were included in the analysis. Secondly, the examination was carried out from a user’s perspective. That is, no private messages could be studied. In addition, VIP sections on the forums had to be ignored. Thus, the findings of this study do not give a complete picture of the forums. Nevertheless, we gathered substantial amounts of data that allowed some analyses and conclusions. The third limitation concerns the internal validity of the data. It cannot be excluded that other investigators, for instance, law enforcement agents engage in trades on the forums for research and investigative purposes. This might bias the data. However, we do not consider this possibility a substantial threat.
Another threat to internal validity is the recorded product prices. The prices advertised are not necessarily the prices that buyers eventually paid. No post was found where the possibility of price negotiations was mentioned. Nevertheless, it cannot be excluded that users pay other, probably lower prices than those advertised.
Finally, given that the examined forums trade different goods or attract certain types of users, the findings are an artefact of the forums in question and do not represent the entire carding underworld. The results are only valid for the five forums we analysed. The selection of the forums is thus a threat to external validity. This limitation concerns all hypotheses but especially [$H_{6}$]{} where an explicit cross-forum comparison was executed. In principle, this limitation was overcome by selecting five different forums based on various selection criteria.
[**Future work.**]{} Sellers were focus of this study. Future research should also consider their counterparts, the buyers. The reviewed literature did not cover buyers and they were also neglected in this paper. It might be useful to examine whether there are high-profile buyers and observe what they buy, and whether they resell the products, and to whom, if they resell. Regarding research design and methodology, a long-term or a follow-up study might be able to identify trends or confirm the patterns found in this study, respectively. Researchers could also consider engaging in trades and getting in touch with the traders, subject to ethical considerations. This method would allow researchers to collect more information on traders and gain better understanding of their roles within the fraud chain. These findings would help to shed more light on how to counter financial cybercrime in the future.
Conclusion
==========
This paper presented an overview of 25 existing online carding forums and an in-depth analysis of five of these forums, covering a three-month period of monitoring. What differentiates this study from others is, first, we investigated real data instead of drawing conclusions based solely on existing literature or expert opinion, second, we examined active forums instead of closed forums, and third, we applied a low-level focus on products, prices and sellers. Our findings suggest that the market of carding forums is dynamic. However, it is not clear how promising the future of carding forums is, especially with the emergence of single-vendor stores which could imply that high-profile sellers would leave existing carding forums to open their own single-vendor stores.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The estimation of cosmological parameters from a given data set requires a construction of a likelihood function which, in general, has a complicated functional form. We adopt a Gaussian copula and constructed a copula likelihood function for the convergence power spectrum from a weak lensing survey. We show that the parameter estimation based on the Gaussian likelihood erroneously introduces a systematic shift in the confidence region, in particular for a parameter of the dark energy equation of state $w$. Thus, the copula likelihood should be used in future cosmological observations.'
author:
- 'Masanori Sato$^{1,2,}$[^1], Kiyotomo Ichiki$^{1}$, Tsutomu T. Takeuchi$^{3}$'
bibliography:
- 'ms.bib'
title: Precise Estimation of Cosmological Parameters Using a More Accurate Likelihood Function
---
[**Introduction**]{}. Currently, a number of wide-field weak lensing (WL) surveys are planned, such as Subaru Hyper Suprime Cam Survey [@2006SPIE.6269E...9M], the Panoramic Survey Telescope & Rapid Response System (Pan-STARRS[^2]), the Dark Energy Survey (DES[^3]), the Large Synoptic Survey Telescope (LSST[^4]), the Joint Dark Energy Mission (JDEM[^5]), and Euclid [@2010arXiv1001.0061R], to address questions about the nature of dark energy and/or the properties of gravity on cosmological scales. To achieve the full potential of these planned future surveys, it is of great importance to employ adequate statistical measures and methods of weak lensing for estimating cosmological parameters.
For the cosmological parameter estimation, almost all previous authors have used the $\chi^2$ method in weak lensing analysis . However, the probability distribution function (PDF) of the weak lensing power spectrum was found to be well approximated by a $\chi^2$ distribution with a heavier positive tail than expected from the normal $\chi^2$ distribution [@2009ApJ...701..945S]. The $\chi^2$ distribution deviates from Gaussian distribution on large scales because the number of modes corresponding to the degrees of freedom are very small. Meanwhile, the $\chi^2$ distribution converges Gaussian distribution at high $\ell$ because of the central limit theorem. We have to make maximal use of these pieces of information accurately to constrain the cosmological parameters. If such information is not taken into account the likelihood function, the derived cosmological parameters can be systematically biased .
In a companion paper [@sit], we constructed a more properly accurate likelihood function using the Gaussian copula (hereafter “copula likelihood”) for the cosmic shear power spectrum, rather than the multivariate Gaussian distribution. We show that the copula likelihood well reproduces the $n$-dimensional probability distribution of the cosmic shear power spectrum estimated from 1000 realizations obtained from ray-tracing simulations performed by [@2009ApJ...701..945S].
In this Letter, we estimate the cosmological parameters using both the copula and Gaussian likelihoods in order to evaluate how the difference between the two likelihoods affects the parameter estimation. Using the Markov Chain Monte Carlo method, we also examine the impact on both ongoing and future surveys.
The cosmological parameters employed for our ray-tracing simulations are consistent with the WMAP 3-yr results (WMAP3) [@2007ApJS..170..377S]. Detailed descriptions of our ray-tracing simulations are summarized in [@2009ApJ...701..945S] (see also, [@2010arXiv1009.2558S]).
. The likelihood function plays a central role in various statistical analyses. In the companion paper [@sit], we derive the copula likelihood for the cosmic shear power spectrum. In this section, we briefly summarize the method.
From the Sklar’s theorem [@sklar1959fonctions], one can relate any $n$-point cumulative probability distribution (CDF) to one-point CDFs as $$\begin{aligned}
&{\rm Prob}(x_{1}\le\hat{x}_{1},x_{2}\le\hat{x}_{2},\dots,x_{n}\le\hat{x}_{n})\equiv
F(\hat{x}_{1},\hat{x}_{2},\dots,\hat{x}_{n})\nonumber\\
&=C(F_1(\hat{x}_{1}),F_2(\hat{x}_{2}),\dots,
F_n(\hat{x}_{n})).
\label{sklars_theorem}\end{aligned}$$ Here $\hat{x}_i$ ($i=1,2,\dots,n$) are independent and identically-distributed observed variables (in this work, the shear power spectrum, $P_\kappa(\ell)$), $C$ denotes the function called a copula, $F$ denotes the $n$-point CDF, and $F_{i}$ denotes the one-point CDFs (in this work, the cumulative $\chi^2$ distributions). Thus, the copula describes how one-point CDFs are joined together to give an $n$-point CDF. A comprehensive proof of Sklar’s theorem and rigorous definition of a copula are found in [@nelsen2006introduction; @2010ApJ...708L...9S; @2010MNRAS.406.1830T].
The multivariate Gaussian copula is a copula of $n$-dimensional random vector that is multivariate normally distributed. This copula is expressed as $$C(u_1,u_2,\dots,u_n)\equiv\Phi\left(\Phi_1^{-1}(u_1),\Phi_1^{-1}(u_2),\dots,\Phi_1^{-1}(u_n)\right),$$ where $\Phi_{1}^{-1}$ is the inverse function of a one-point Gaussian CDF with mean $\mu_i$ and standard deviation $\sigma_i$, and $u_i\equiv
F_i(\hat{x}_i)$. Here $\Phi$ is an $n$-point Gaussian CDF defined by $$\begin{aligned}
&\Phi(\hat{x}_1,\hat{x}_2,\dots,\hat{x}_n)=\int_{-\infty}^{\hat{x}_1}\int_{-\infty}^{\hat{x}_2}\dots\int_{-\infty}^{\hat{x}_n}\frac{1}{\sqrt{(2\pi)^n
{\rm
det(Cov)}}}\nonumber\\
&\times\exp\left(-\frac{1}{2}(\boldsymbol{x}-\boldsymbol{\mu})^{\rm T}{\rm
Cov}^{-1}(\boldsymbol{x}-\boldsymbol{\mu})\right)\pd x_1\pd x_2\dots
\pd x_n,\end{aligned}$$ with mean $\boldsymbol{\mu}$ and $n\times{n}$ covariance matrix. ${\rm Cov}^{-1}$ shows the inverse covariance matrix. Hereafter, $\boldsymbol{\mu}\equiv (\mu_1,\mu_2,\dots,\mu_n)$, $\boldsymbol{x}\equiv (x_1,x_2,\dots,x_n)$, and superscript ’T’ stands for the transpose of vector.
Using this copula, a log-likelihood $\ln\mathcal{L}$ is derived as [@sit] $$\begin{aligned}
-2\ln\mathcal{L}_{c}(\hat{x}_1,\hat{x}_2,\dots,\hat{x}_n)&=
\sum_{i=1}^{n}\sum_{j=1}^n(q_i-\mu_i){\rm
Cov}^{-1}(q_j-\mu_j)\nonumber\\
&-\sum_{i=1}^n\frac{(q_i-\mu_i)^2}{\sigma_i^2}-2\sum_{i=1}^n\ln f_i(\hat{x}_i),
\label{gau_copula_like}\end{aligned}$$ for a general probability distribution, while for Gaussian probability distribution it reduces $$-2\ln\mathcal{L}_{g}(\hat{x}_1,\hat{x}_2,\dots,\hat{x}_n)=
\sum_{i=1}^{n}\sum_{j=1}^n(\hat{x}_i-\mu_i){\rm
Cov}^{-1}(\hat{x}_j-\mu_j) .
\label{gau_like}$$ Here, functions $f_i$ represent the derivatives of the CDFs $F_i$, and we have omitted an irrelevant constant term in the above two equations. The relation between $u_i$ and $q_i$ is $$q_i=\sigma_i\Psi_1^{-1}(u_i)+\mu_i,$$ where $\Psi_1$ is a cumulative standard normal distribution. In what follows, we use Eqs. (\[gau\_copula\_like\]) and Eq. (\[gau\_like\]) to constrain the cosmological parameters and investigate how the different likelihood functions cause a difference in the posterior distribution function of cosmological parameters.
. In the case considered in this Letter, the observed variables $\hat{x}_i$ are binned values of the nonlinear convergence power spectrum $\hat{P}_{\kappa}(\ell_i)$, which are computed using the prescription of [@2003MNRAS.341.1311S] for the nonlinear effects. Note that we assume a bin width $\Delta\ln \ell=0.3$, single source redshift distribution, i.e. all lensed galaxies lie at $z_s=1.0$, and we do not consider intrinsic ellipticity dispersion $\sigma_{\epsilon}$ throughout this Letter. Instead, we use the information up to multipole $\ell_{\rm
max}=1071$, because the weak lensing power spectrum estimated from a realistic survey on smaller scales than $\ell\sim 1000$ is expected to be contaminated by the intrinsic ellipticity noise [@2009MNRAS.395.2065T]. When constructing the copula likelihood function, $\mu_i$ and ${\rm Cov}^{-1}$ are estimated from ray-tracing simulations performed by [@2009ApJ...701..945S]. It is known to be appropriate to adopt a $\chi^2$ distribution for the general probability distribution $f_i$, because the one-point PDF of the convergence power spectrum is fairly well described by $\chi^2$ distribution with mean and variance of $P_{\kappa}(\ell_i)=\langle{\hat{P}_{\kappa}(\ell_i)}\rangle$ and $\sigma^2(\ell_i)=\langle\hat{P}_{\kappa}(\ell_i)^2\rangle -
P_{\kappa}(\ell_i)^2$, respectively [see, @2009ApJ...701..945S]. From [@2009ApJ...700..479T], this $\chi^2$ distribution is written as $$f_{\chi^2}(\hat{P}_{\kappa}(\ell_i))=\frac{\hat{P}_{\kappa}(\ell_i)^{\Upsilon -1}}{\Gamma(\Upsilon)}\left(\Upsilon\frac{e^{-\hat{P}_{\kappa}(\ell_i)/P_{\kappa}(\ell_i)}}{P_{\kappa}(\ell_i)}\right)^{\Upsilon},
\label{eq:chi_square}$$ for $\hat{P}_{\kappa}(\ell_i)>0$ and $f_{\chi^2}=0$ for $\hat{P}_{\kappa}(\ell_i)\le 0$. Here $\Gamma(x)$ is the gamma function and we define $\Upsilon\equiv{P_{\kappa}(\ell_i)^2/\sigma^2(\ell_i)}$ which corresponds to the number of independent modes.
For simplicity, we work with two cosmological parameters in weak lensing given as $$\boldsymbol{p}=(\Omega_{\rm dm}, \ln 10^{10}\Delta_{\rm R}^2)$$ where $\Omega_{\rm dm}$ is the dark matter density today and $\Delta_{\rm R}^2$ is the amplitude of primordial curvature perturbations defined at $k=0.002\,{\rm Mpc}^{-1}$ [@2007ApJS..170..377S], whose fiducial values in our analysis are 0.196 and $20.45\times 10^{-10}$, respectively. The other cosmological parameters are fixed at the fiducial values of the WMAP 3-yr cosmology. When we estimate the equation of state parameter $w$ in the dark energy equation, combined with another probe of WMAP3 data, we work with three cosmological parameters given as $$\boldsymbol{p}=(\Omega_{\rm dm}, w, \ln 10^{10}\Delta_{\rm R}^2).$$ The parameter space we explore is as follows: $\Omega_{\rm
dm}=[0.002,1.0]$, $w=[-6.0,0]$, $\ln 10^{10}\Delta_{\rm R}^2=[-2.5,8.0]$ and the fiducial values of these parameters are taken as $0.196$, $-1$ and $3.02$, respectively. Assuming flat priors for the cosmological parameters, we employ the Markov Chain Monte Carlo method [@2002PhRvD..66j3511L] to constrain cosmological parameters given the cosmological observables. Sixteen parallel chains were computed and the convergence test is made based on the Gelman and Rubin statistics called “$R-1$” statistics [@gelman1992inference]. In this work, each chain typically has 300,000 points and $R-1<0.02$ for both of the two models.
. In this section, we constrain the cosmological parameters using both the copula and Gaussian likelihoods and evaluate how the difference between the two likelihoods affects the estimation. We consider two cases: ongoing and future surveys.
[*Impact on ongoing weak lensing surveys.*]{} In our ray-tracing simulations, the survey area is set as $\Omega_{\rm s}=25$ deg$^2$. Therefore the fundamental mode of ray-tracing simulations is $\ell_f=72$. Figure \[fig:om\_sigz\] shows 1$\sigma$ and 2$\sigma$ confidence level constraints on the $\Omega_{\rm m}$–$\sigma_8$ plane. Note that $\Omega_{\rm m}$ denotes the present-day total matter density, i.e., the sum of the current dark matter and baryon densities. The red and blue contours show the marginalized constraints obtained by the copula likelihood (Eq. \[gau\_copula\_like\]) and Gaussian likelihood (Eq. \[gau\_like\]), respectively. Clearly the results from the copula likelihood (red contours) and Gaussian likelihood (blue contours) are significantly different. The contours obtained with the copula likelihood are shifted toward a parameter region that gives a lower convergence power compared to that from the Gaussian likelihood. This result is attributed to the fact that the median of the $\chi^2$ distribution for the convergence power spectrum is smaller than that of the Gaussian distribution.
Figure \[fig:om\_wz20\] shows the constraints on the $\Omega_{\rm m}$–$w$ plane obtained when the weak lensing information is combined with data from WMAP3. Also in this case we see the difference between the results from the two likelihoods. Similarly, the contours obtained by the copula likelihood are shifted toward the parameter region that gives the lower power, i.e. smaller $\Omega_{\rm m}$ and larger $w$, than the Gaussian likelihood, as expected. Consequently, if one uses the (approximate) multivariate Gaussian likelihood, one will overestimate the lower bound and underestimate the upper one for the constraint on $w$. Specifically, we found an allowed range of the equation of state parameter $w$ for the Gaussian likelihood model as $-1.055 < w < -0.956~,$ while for the copula likelihood model as $-1.038 < w < -0.941~$ at $95$% confidence level. Therefore, a few percent of systematic error potentially exists in the parameter estimation if one uses approximate Gaussian likelihood.
![Two-dimensional constraints (1$\sigma$ and 2$\sigma$ C.L.) on $\Omega_{\rm m}$ and $\sigma_8$ for survey area $\Omega_{\rm s}=25{\rm deg}^2$. The red and blue contours show the constraints obtained by likelihood based on the copula likelihood (Eq. \[gau\_copula\_like\]) and Gaussian likelihood (Eq. \[gau\_like\]), respectively. The cross symbol shows the fiducial cosmological parameters. []{data-label="fig:om_sigz"}](f1.eps "fig:"){width="45.00000%"} -
![Same as Fig. \[fig:om\_sigz\], but two-dimensional marginalized constraints on $\Omega_{\rm m}$ and $w$ for weak lensing plus WMAP 3-yr results. []{data-label="fig:om_wz20"}](f2.eps "fig:"){width="45.00000%"} -
[*Impact on future weak lensing surveys.*]{} Next let us assume a survey area of 2000 deg$^2$, which roughly corresponds to the planned Subaru Weak Lensing Survey [@2006SPIE.6269E...9M]. Because our ray-tracing simulation is limited to 25 deg$^2$, we have to extend it to estimate the mean power spectrum $\mu_i$ and their covariance with help from analytic treatments. When estimating the mean power spectrum $\mu_i$, we use the linear perturbation theory based on WMAP 3-yr fiducial cosmology for multipoles in the range $8 \le \ell < \ell_f=72$, where $\ell_f$ corresponds to the largest angular mode of our ray-tracing simulation (5$\times$ 5 deg$^2$ for an area). For the covariance matrix ${\rm Cov}$, we assume that it has only diagonal elements for multipoles smaller than $\ell_f=72$ because the linear approximation is valid for that multipole range. In the absence of shape noise, the diagonal elements would arise from sample variance and should be equal to the power spectrum squared divided by the number of independent modes in the bin. Also, we assume that the covariances between the scales larger than $\ell_f=72$ and scales smaller than $\ell_f=72$ are zero, which means that the powers at the largest scales are independent from those at smaller scales.
Figure \[fig:om\_wz20\_hsc\] shows the constraints on the ($\Omega_{\rm
m}$, $w$) plane obtained when weak lensing information with the survey area of $2000$ deg$^2$ is combined with WMAP3. In this case, we do not see any significant difference between the results from the two likelihoods. This result is understood as follows. If one considers the larger survey area, $\Upsilon$ becomes larger at a fixed multipole because $\sigma^2$ becomes smaller (see, Eq. \[eq:chi\_square\]). Hence, $\chi^2$ distribution for the convergence power spectrum becomes well approximated by the Gaussian distribution at the fixed multipole. Even though at the lowest multipole bin the distribution may be deviated from Gaussian, the bulk of information comes from the large multipole bins, where the distribution is almost Gaussian. Hence the contours become indistinguishable.
It should be noted, however, that the Gaussianity in the distribution of the power spectrum should be due to our sparse binning of the multipole space. To utilize the information as much as possible, we have to divide the multipole space into a larger number of bins than we have done in this study. In such a case, the distribution of the power spectrum at each multipole bin will deviate from the Gaussian distribution and the copula likelihood will still be useful to model the probability of the power spectrum distribution.
![Same as Fig. \[fig:om\_wz20\], but assumed survey area is 2000 deg$^2$ for weak lensing. This good agreement may not be caused by a physical reason, but rather attributed to the coarse binning.[]{data-label="fig:om_wz20_hsc"}](f3.eps "fig:"){width="45.00000%"} -
. In this Letter we constructed a likelihood function for the convergence power spectrum from weak lensing survey using copula. The copula likelihood function describes more accurately the distribution of the power spectrum derived from ray-tracing simulations and thus utilizes more information than the previous methods in the literature. We then used the copula likelihood model to calculate the allowed range of the cosmological parameters paying particular attention to the dark energy equation of state parameter $w$, and compare it with that derived from the conventional multivariate Gaussian likelihood model. We found that, for the $25$ deg$^2$ weak lensing survey, the results can be different even combined with the CMB data, depending on which likelihood function is used. The difference is as large as a few percent.
For the $2000$ deg$^2$ weak lensing survey, we found the results coincide with each other. This result allows us to use a multivariate Gaussian likelihood model for the future weak lensing survey, which will greatly simplify the parameter estimation analysis. We note, however, that this coincidence might be an artifact from our sparse binning of the multipole space. A full analysis with finer binning is difficult at present because computational cost is high, and we leave it for a future study.
We thank Agnieszka Pollo and Masahiro Takada for comments. We also thank the anonymous referees for careful reading of our manuscript and very useful and constructive suggestions that help to clarify our paper further. M.S. is supported by the JSPS. T.T.T. has been supported by Program for Improvement of Research Environment for Young Researchers from Special Coordination Funds for Promoting Science and Technology. This work is partially supported by the Grant-in-Aid for the Scientific Research Fund No. 20740105 (T.T.T,), No. 21740177, No. 22012004 (K.I.), and Grant-in-Aid for Scientific Research on Priority Areas No. 467 “Probing the Dark Energy through an Extremely Wide and Deep Survey with Subaru Telescope” commissioned by the MEXT of Japan.
[^1]: [email protected]
[^2]: http://pan-starrs.ifa.hawaii.edu/public/
[^3]: http://www.darkenergysurvey.org/
[^4]: http://www.lsst.org/
[^5]: http://jdem.gsfc.nasa.gov/
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'J. Cuypers'
- 'C. Aerts'
- 'P. De Cat'
- 'J. De Ridder'
- 'K. Goossens'
- 'C. Schoenaers'
- 'K. Uytterhoeven'
- 'B. Acke[^1]'
- 'G. Davignon'
- 'J. Debosscher'
- 'L. Decin'
- 'W. De Meester'
- 'P. Deroo[^2]'
- 'R. Drummond'
- 'K. Kolenberg'
- 'K. Lefever'
- 'G. Raskin'
- 'M. Reyniers'
- 'S. Saesen'
- 'B. Vandenbussche'
- 'R. Van Malderen'
- 'T. Verhoelst'
- 'H. Van Winckel'
- 'C. Waelkens'
bibliography:
- 'af.bib'
date: 'March 05, 2009'
title: 'Long-term photometric monitoring with the Mercator telescope. Frequencies and multicolour amplitudes of $\gamma$ Doradus stars [^3] '
---
[ [$\gamma$Dor]{} stars are excellent targets for asteroseismology since the gravity modes present in these stars probe the deep stellar interiors. Mode identification will improve the knowledge of these stars considerably. ]{} [A selected group of [$\gamma$Dor]{} stars and some candidates were observed with the Mercator telescope to find and/or confirm the periodicities in the light variations and to derive reliable amplitude ratios in different pass bands. ]{} [A frequency analysis was performed on all new data obtained in the Geneva photometric system. In order to have more reliable and accurate frequencies, the new data were combined with similar data from the literature and with Hipparcos observations. A set of frequencies that minimized the the residuals in a harmonic fit was searched for while allowing means and amplitudes to vary from one observation set to another.]{} [Frequencies and amplitudes in the photometric passbands of the Geneva system are given for 21 [$\gamma$Dor]{} stars. We report the discovery of as a newly found [$\gamma$Dor]{} star.]{} [Our study provides the first extensive multicolour database for the understanding of gravity modes in F-type stars]{}
Introduction: $\gamma$ Doradus stars ([$\gamma$Dor]{} stars) {#intro}
============================================================
As reviewed in the article by @henry07 the class of [$\gamma$Dor]{} stars has now over 60 members. These stars have spectral types of late A or F, luminosity class IV or V and exhibit periodicity in the light variations with periods in the range of 0.3 to 3 days. Line-profile variability is present as well [@mathias04; @decat06]. There is no doubt that the cause of the variations is pulsation. The modes are high-order gravity modes (g-modes), excited by a flux blocking mechanism at the base of the convective envelope of the stars [@guzik00; @dupret05].
Because the g-modes probe the deep stellar interior, the [$\gamma$Dor]{} stars are excellent targets for asteroseismology, particularly also since they may exhibit solar-like oscillations as well [@michel08]. However, mode identificiation is not simple for these stars since the spectrum of g-modes is very dense and only a few modes seem to have high enough amplitudes to be well observed from the ground. Therefore, observables such as photometric amplitude ratios and line-profile variations are extremely useful to identify the modes. By observing light variations in different passbands the identification of the degree $\ell$ of the pulsation mode becomes possible [@balona79; @watson88; @dupret05].
Since this kind of observable is made available with this paper for 21 [$\gamma$Dor]{} stars, a large number of variable stars of this class will have their modes identified, and parameters for the physics of stellar models, including e.g. a value for the mixing-length parameter, can in principle be derived. Such seismic inferences are the subject of an accompanying paper (Miglio et al, 2008, in preparation).
The instrument and the observations
===================================
The Mercator telescope is a 1.2-meter telescope located on the Roque de los Muchachos observatory on La Palma, Spain. Since the beginning of its scientific operations in spring 2001 it was equipped with the P7 photometer.
The P7 photometer is the refurbished photometer that has been active on the 70-cm Swiss Telescope at La Silla Chile before. It is a two-channel photometer for quasi simultaneous 7-band measurements ($U, B_1, B, B_2, V_1, V, G$) in the Geneva photometric system [@golay66; @rufener88; @rufener89]. The first channel (A) is centred on the star while the second channel (B) is centred on the sky. The filter wheel turns at 4 hertz and a chopper directs both channels alternatively to the photomultiplier. As such, the photomultiplier measures both beams through the seven filters four times per second. In this way quasi-simultaneous measurements of the light in the 7 filters become possible. This makes the instrument very suitable to observe variable stars, since information on the light variations in all filters can be obtained very efficiently.
Furthermore an almost continuous access to the telescope during the year was possible, so the instrument became ideal to monitor stars with periods of the order of days and/or long beat periods of the order of months or years. Here we report on the results of a long-term monitoring of [$\gamma$Dor]{} stars, while @decat07 discussed the results for a sample of variable OB stars.
Some of these stars were observed over two years (time span around 370 days), some over three years (time span around 740 days) or four years (1110 days). Stars observed during only one season were not considered in detail. Up to five observations per star were obtained for about two months per season. This resulted in about 130 measurements per star on average, but there is a large spread as can be seen in Table \[resM\].
Photometric observations in the Geneva system are labeled with two quality indices (from lowest quality 0 to highest quality 4), one for the magnitudes, one for the colours [@rufener89]. In a first analysis we eliminated all observations labeled 0 or 1, but we also carried out a re-analysis with the low-quality observations included. In most cases the difference in the result was negligible. Since some stars were only measured at high airmasses, they almost all automatically received a low index value. Here we used those values as well, since without them the number of observations of those stars would be too low. The errors are slightly larger in such cases, but the results are the same. Observations of standard stars indicate that the errors for the programme stars, which have visual magnitudes between 7 and 9, are in the range $0.005$ to $0.010$ mag.
Selection of the targets
========================
A selection of stars was made out of the lists by @aerts98 and @handler99, of the Hipparcos selection results and from lists of [$\gamma$Dor]{} stars, candidate [$\gamma$Dor]{} or other variables [@koen02]. Since the start of the observations, almost all stars of the list were comfirmed as [$\gamma$Dor]{} stars. For very few of them, multicolour photometry was available and in several cases the multiperiodicity needed confirmation. The list with the 21 stars with more than 25 observations of good quality, spread over more than one season, is given in Table \[resM\]. A number of stars was already (partly) analysed in a previous paper [@deridder04] but for some stars the additional observations proved to be very useful to clarify the multiperiodicity of the light variations.
Fifteen other variable A-F stars were monitored during the same period. Not all showed characteristics of [$\gamma$Dor]{} stars: a few turned out to be $\delta$ Scuti stars, others need some further analysis. The results for these stars will be published in another article.
Analysis methods
================
We used standard methods for the period analysis, but since the light curves of these stars are very sinusoidal, we adopted as a final analysis method a multifrequency least-squares fit with sinusoidal components [@schoenaers04], where we searched for the best frequencies in large intervals around an initial solution. This is an extremely computer intensive procedure since the sum of the residuals is a rapidly oscillating function, especially when up to six frequencies had to be considered. To obtain the initial values we used Lomb-Scargle periodogram methods [@lomb76; @scargle82], and phase dispersion minimization methods [@stellingwerf78], each time followed by a prewhitening for a first analysis. The final search was always done with a multifrequency least-squares fit.
For initial values of the frequencies in the Mercator data we usually looked at the B or B1 data, since we expect the highest amplitudes of the light variations in these filters (e.g. @aerts04). It happened that the amplitude in other filters (typically V1 or G) became too low for the frequencies to be significant there. We re-analysed the Hipparcos data independently, but we looked only for a local minimum in the least-squares fit in the (wide) neighbourhood of the ground-based solution.
If observations were available from other sources, we combined the Geneva colour V, Johnson V and Hipparcos data. We searched in the combined data for a multifrequency least-squares fit with unknown means and amplitudes for each data source to search for the best frequency (frequencies) to describe all data. We did not use statistical weights on individual observations or groups of data: the data from the literature were of similar quality and the measurement errors about the same as for the Geneva photometry. The Hipparcos data often had larger errors, but since these are important for solving the alias problems, we also gave those equal weights.
Although for individual data sets only frequencies peaking four times above the average signal-to-noise level in the periodogram are acceptable, we relaxed this criterion for the frequencies found with the multifrequency fit technique in the combined data. Frequency solutions of the multifrequency fit were accepted, if a corresponding extremum was found in the periodogram of each data set. In this way we accepted frequencies with peaks 1.5 times above the noise in some of the individual data sets, as shown a posteriori.
In several cases the combination of the data solved the alias or ambiguity problems and it gave, in general, more accurate values for the frequencies. In a few cases, it turned out that observations of the same star(s) were carried out, unintentionally, as in multisite campaigns. Here the gain in frequency determination was obvious. Including the Hipparcos data enlarged the time basis to about 5000 days for almost all stars. We did no colour transformation for the Hipparcos data, since our method takes possible amplitude differences into account. Since the Hipparcos data are not always as accurate as the ground-based data and since the observations sometimes have a rather odd time distribution, there was not always the expected gain in frequency accuracy. In most cases however, the frequency errors became of the order of $10^{-6}$[$\mathrm{d}^{-1}$]{}.
Amplitude ratios and other characteristics were computed in the different passbands of the Geneva photometric system with frequencies as close as possible to the results of the combined data set. For this purpose we searched again for the closest local minimum in the residuals of the harmonic fit in the Mercator data. In some cases there were differences between the best solution found in the Mercator data alone and this solution. However, differences were small or caused by choosing the wrong alias in the initial solution, as can be seen by comparing Table \[resM\] and the Tables 5 to 25 (only available electronically).
Results on the individual stars
===============================
= HIP 623
----------
HD 277 is an F2 star intensively observed by @henry01. They found three frequencies and remarked more scatter around light maximum than around light minimum.
The three frequencies found by @henry01 were confirmed, although the $2-f$ alias of their third frequency is more significant in the Mercator data. The high significance of this alias is caused by the limited nightly sampling of our data, combined with the length of the period.
The third frequency as derived from the Hipparcos data (1.3437[$\mathrm{d}^{-1}$]{}) deviates from the third frequency (1.3866[$\mathrm{d}^{-1}$]{}) found by @henry01 and from our third frequency (1.38709[$\mathrm{d}^{-1}$]{}) as well. At the moment we have no explanation for this, but the standard deviation of the residuals remaining in the Hipparcos data after prewhitening with the first three frequencies is still higher (0.021 mag) than can be expected from the errors on the measurements (0.017 mag). This could indicate that more (low amplitude) frequencies are present, but they could not be detected in the individual data sets. In the combined data set (Hipparcos, @henry01 and Mercator data) a fourth frequency at 0.8349[$\mathrm{d}^{-1}$]{}was detected.
We also observed a few times some extra brightening around light maximum in this star. The standard deviation of the residuals in the Mercator data is about 0.015 mag in B1 and 0.011 mag in V. This is an indication of some additional variability.
= HIP 2510
-----------
In previous analyses of the Mercator data [@deridder04; @cuypers06] only two frequencies were found, but now the three frequencies, as given by @henry05, could be confirmed without any doubt. The three-frequency solution is also clearly present in the Hipparcos data of this star. The frequency given by @koen02 for the Hipparcos data (1.78085[$\mathrm{d}^{-1}$]{}) is not present in any of the significant three-frequency solutions.
@fekel03 observed no velocity changes in the 2 spectra they obtained, but @mathias04 did. The latter also mentioned possible line-profile variations in the blue wing of the lines.
= HIP 5674
-----------
Three frequencies are easily recognized in the Mercator data. Two frequencies are in common with the @henry03 results, but their second frequency (2.7640[$\mathrm{d}^{-1}$]{}) is probably an alias of the frequency 1.76031[$\mathrm{d}^{-1}$]{} found both in the Mercator and Hipparcos data.
This star has 136 Mercator measurements over the period 2001 to 2003, but was unfortunately only once observed in 2004. Therefore, the errors on the results are larger than usual.
= HIP 17826
------------
So far only one frequency (2.2565[$\mathrm{d}^{-1}$]{}) has been accepted for this star [@henry03], since the second frequency found by @handler99 in the Hipparcos data could not be confirmed. However, more than one frequency is expected given the large residuals in all data sets. In the Mercator data a secondary frequency at 2.8883[$\mathrm{d}^{-1}$]{} or its alias at 1.8856[$\mathrm{d}^{-1}$]{} is significant. A two-frequency search in the data of @henry03 gave also 2.2565 and 1.8849[$\mathrm{d}^{-1}$]{} as one of the most significant solutions. By combining the Henry and Fekel data, the Mercator data and the Hipparcos data, the two frequencies could be calculated more precisely.
= HIP 32263 = V553 Aur
-----------------------
@handler99 found two frequencies for this star in the Hipparcos data: 0.5244 and 0.8688[$\mathrm{d}^{-1}$]{}. Only one frequency is given by @henry03: 0.9125[$\mathrm{d}^{-1}$]{}. @martin03 suggested the presence of four frequencies in the Hipparcos data. Our analysis confirmd these four frequencies, although for the second frequency an alias is found to be more significant in the Mercator data and the fourth frequency is only marginally significant. The combination of the data sets, with the Hipparcos data included, solves the alias problems and indicates a stable four-frequency solution.
= HIP 37863 = DO Lyn
---------------------
For this star @kaye99 found five frequencies: 1.60146, 1.43678, 1.73671,1.83372, 1.80753[$\mathrm{d}^{-1}$]{} in their multi-site campaign of 1997 and 1998 and they suggested one more near 2.9[$\mathrm{d}^{-1}$]{}. The first three have amplitudes larger than 0.01 mag in V and are well identified. The others have smaller amplitudes. With our multifrequency fit techniques, we found in the data of @kaye99 the frequencies 1.60139, 1.43675, 1.73637,1.83373, 1.80793[$\mathrm{d}^{-1}$]{}, all very close to the given solution.
The most significant frequency in the Hipparcos data is 1.3329[$\mathrm{d}^{-1}$]{}. This frequency is not related to any of the other frequencies, but this could be due to the odd time distribution of the Hipparcos data.
In the Mercator data of 2001 to 2004 the main frequency found in filters B and B1 is 1.5985[$\mathrm{d}^{-1}$]{}. This could be an alias of the first frequency given by @kaye99. In the other filters this frequency cannot unambiguously be identified. A frequency at 1.2032[$\mathrm{d}^{-1}$]{} is present in all filters.
It seems impossible to connect the frequency solutions of the different campaigns for this star. This could mean that the frequencies and/or the amplitudes in this star are not stable over a longer period or, more likely, that a (different) combination of more frequencies is necessary to describe the data adequately.
The star is a double-lined spectroscopic binary [@kaye98] and orbital parameters were given by @kaye99 and @mathias04. Line profile variations were observed in the primary star by @mathias04.
We do not expect a large influence of the secondary component on the pulsational behaviour given the relatively long orbital period of 11.615 days, but currently we have no explanation for the variation of the frequency spectrum from season to season.
= HIP 40791
------------
HD 69715 was confirmed as a [$\gamma$Dor]{} star by @henry05. The two frequencies found by them, 2.4566 and 2.4416[$\mathrm{d}^{-1}$]{} with error 0.0002[$\mathrm{d}^{-1}$]{}, are confirmed in the Mercator data, 2.45642 and 2.44120[$\mathrm{d}^{-1}$]{}with error 0.00007[$\mathrm{d}^{-1}$]{}. In the Hipparcos data a two-frequency solution with 2.0129 and 2.36462[$\mathrm{d}^{-1}$]{} is present, as already indicated by [@martin03], but a well defined local minimum is found for 2.45630 and 2.44139[$\mathrm{d}^{-1}$]{} as well . Therefore, in the combination of all data sets, highly accurate values for these frequencies could be found (errors less than 0.00001[$\mathrm{d}^{-1}$]{}). The time span of the observations obtained by @martin03 for this star was probably too short to separate these two frequencies. As a consequence, it is not surprising that other values for the frequencies were found, although one of their results, 2.425[$\mathrm{d}^{-1}$]{}, is close to the two frequencies given here. @mathias04 observed no line-profile variations in the spectra, pointing to very low velocity amplitudes, probably due to rotational line broadening ($\mathrm{v}\sin{i} = 145~km/s$).
= HIP 43062
------------
This is a star from the list by @koen02 and it is the only star in our sample that is not yet catalogued as variable star. It has all the properties of a [$\gamma$Dor]{} star. The frequency in the Hipparcos data as listed by @koen02 is 1.9057[$\mathrm{d}^{-1}$]{}. This frequency ($1.90570\pm 0.00005$[$\mathrm{d}^{-1}$]{}) is found easily in the Mercator data, where at least one more frequency could be identified ($1.8210 \pm 0.0001$[$\mathrm{d}^{-1}$]{}). Although this frequency could be an alias, we prefer this value since a two-frequency analysis of the Hipparcos data gives $1.9058$ and $1.8212$[$\mathrm{d}^{-1}$]{} as the best two-frequency solution. In the combined data set we found $1.905773$ and $1.820999$[$\mathrm{d}^{-1}$]{}. The next frequency candidate here is $1.845691$[$\mathrm{d}^{-1}$]{}. Phase diagrams constructed with the frequencies found are given in Figure \[74504\].
{width="5.66cm"} {width="5.66cm"} {width="5.66cm"}
= HIP 48895 = HR 3936
----------------------
For this [$\gamma$Dor]{} star, identified as such by @handler99 and @koen02, four frequencies were known from the observations published by @fekel03. In the Hipparcos data we identified three of those four frequencies. In the Mercator data the four frequencies (or some clear aliases) are significant as well, but an additional high amplitude variation with frequency 1.00286[$\mathrm{d}^{-1}$]{} (or an alias near 1.0014[$\mathrm{d}^{-1}$]{}) is present. Its amplitude (0.022 mag in V, but smaller in other filters) is not comparable at all with the little extra power in the periodogram near 1[$\mathrm{d}^{-1}$]{} mentioned by @fekel03 and seems large for a stellar oscillation frequency. It is difficult to explain why a frequency corresponding to one cycle per sidereal day is so significant in this dataset and only marginally or not at all significant in other datasets. It could be an artefact of e.g imperfect extinction corrections, but in that case we would expect this frequency to be significant in other Mercator data sets as well. This is not the case. After prewhitening with this frequency, it was possible to find a well-defined set of four frequencies in the Mercator data and in the combined data as well. In a five-frequency fit, the variations associated with the frequency near 1[$\mathrm{d}^{-1}$]{} still have a significant amplitude. Therefore we calculated the associated amplitudes for this frequency as well.
Indications of the binarity of this star were given by @fekel03 and @henry03. The spectra published by @mathias04 clearly showed the binarity and indications of line-profile variations in the slow rotator. @griffin06 discussed the system extensively and gave revised orbital elements and mass estimates. In principle, the frequency 1.00286[$\mathrm{d}^{-1}$]{} or 1.00139[$\mathrm{d}^{-1}$]{} found in the Mercator data could be a one day alias of a smaller frequency related to the orbital period of this system (33.7days), but this is unlikely in view of the estimated inclination of 45$\degr$ as given by @griffin06.
= HIP 56275
------------
This star is a well known [$\gamma$Dor]{} star. The four frequencies given by @henry03, or daily aliases are present in the Mercator data as well. The frequencies were refined by combining all data including the Hipparcos data, but some of the results can still be one year aliases of the true frequencies. A candidate for a fifth frequency, present also in the combined data set, is 1.54354[$\mathrm{d}^{-1}$]{}. The star is a binary system of which @griffin06 gives high-precision orbital elements.
= HIP 59203
------------
@henry01 found six frequencies in the data for this star. We can easily confirm four of them (or their aliases) in the Mercator data. The six-frequency solution is a local minimum as well and is also present in the combined V data. Therefore, amplitude ratios for all frequencies could be estimated. In the Hipparcos data alone, the frequencies are difficult to find. There 0.7151[$\mathrm{d}^{-1}$]{} appears as a first frequency, although a frequency of 1.3210[$\mathrm{d}^{-1}$]{} is present as well.
= HIP 60571 = DD CVn
---------------------
The frequencies found in the Mercator data are compatible with the results of @breger97 and @henry02. The Hipparcos data are very noisy and could not give independent information. @breger97 indicated possible amplitude variations and this might be a reason for not finding the frequencies in a data set with a small number of observations. However, combining the Hipparcos data and the data of @henry02 with the Mercator data leads to a well-defined three-frequency solution. @mathias04 report that this star is a binary with line-profile variations in the slow rotator.
= HIP 63951
------------
@henry03 found five frequencies (0.8887, 2.0084, 1.7785, 2.8965 and 0.8841[$\mathrm{d}^{-1}$]{}) and they indicated that there could be some alias problems in their final solution. One of their frequencies is doubtful since it differs only by 0.0046[$\mathrm{d}^{-1}$]{} from the main frequency and this is at the limit of reliable extraction from an observational campaign of less than 250 days for this star. In the Mercator data a similar solution could be found (0.8884, 1.0056, 1.7761, 1.8910, 0.8834[$\mathrm{d}^{-1}$]{}). By combining the data of @henry03 with the Mercator data and the Hipparcos observations, some ambiguities regarding day and year aliases could be resolved. A six-frequency solution (0.88838, 1.00420, 1.77757, 1.89378, 0.88681, 1.42221[$\mathrm{d}^{-1}$]{}) was obtained from this analysis. The sixth frequency (or its alias at 2.42[$\mathrm{d}^{-1}$]{}) is also marginally present in Figure 18 of @henry03. The frequency analysis for this star is not decisive. There is one period extremely close to 1 day. The third frequency nearly equals the sum of the first and fifth frequency. As a result not all alias problems nor ambiguities related to possible sum or double frequencies are completely resolved yet. Some indications of closely spaced frequencies near the main frequency are present. The star is very likely to be a binary, as suggested by @henry03 and confirmed by [@mathias04]. It is not clear which component is the [$\gamma$Dor]{} star. Since the components are estimated to have spectral type A9 and F1, both stars could be [$\gamma$Dor]{} stars.
= HIP 89601 = HR 6844 = V2502 Oph
----------------------------------
There is no doubt that this binary star (orbital period: 4.48518 days [@fekel03]) is a multiperiodic variable. Two frequencies (near 0.765 and near 0.697[$\mathrm{d}^{-1}$]{}) are significant in all data sets. The third frequency in the Mercator data is near the second frequency cited in @aerts98. We identified five frequencies in the combined data, including the third frequency given by @henry03. One frequency is close to the $2-f$ alias of the second frequency and this may hamper the frequency analysis of individual ground-based data sets. There is a lot of scatter in the light curves and some excess scatter at maximum brightness.
= HIP 92837
------------
One frequency is clearly present in all the data sets of this star (Hipparcos, @henry05 and Mercator data). The Hipparcos data have a few outliers that strongly influence the period search. No secondary frequency could be identified unambiguously.
= HIP 100859 = HR 7828 = V2121 Cyg
-----------------------------------
This star was sometimes observed on the same dates by @henry05 and by the Mercator telescope team. The three-frequency solution was easily confirmed in the combined data. A candidate for a fourth frequency occurs near 0.28 or its alias at 1.28[$\mathrm{d}^{-1}$]{}, between $f_1$ and $f_2$. The first frequency found by @jankov06 in spectroscopic data ($1.61$[$\mathrm{d}^{-1}$]{}) is not significantly present in any of the photometric data sets.
= HIP 106897 = HR 8276 = NZ Peg
--------------------------------
For this star only three frequencies of the set of five found by @henry01 were significant in the Mercator data. However, a local minimum in the residuals exists for the five-frequency solution both in the Mercator and in the combined data sets. One extra frequency was found in the combined data as well. Also for the six-frequency solution there was convergence to a local minimum in the residuals for the Mercator data. Therefore, amplitudes could be calculated for all six frequencies in the Geneva filters. As expected the amplitudes of frequencies $f_4$, $f_5$ en $f_6$ are small.
= HIP 107558 = HR8330 =V372 Peg
--------------------------------
In the Mercator data the one year alias of the frequency given by @aerts01 and @guinan01 is found. The Hipparcos data on its own are not very useful to resolve the ambiguity but in the combined data there is a slight preference for the frequency $0.38538$[$\mathrm{d}^{-1}$]{}.
= HIP 110163 = PR Peg
----------------------
Only three of the four frequencies (or of their aliases) found by @henry07 are significant in the Mercator data. Their second frequency near 1.12[$\mathrm{d}^{-1}$]{} is not found, but is apparent in the combined data. Furthermore, as already remarked by @henry07, their fourth frequency is equal to the difference between $f_1$ and $f_2$, or, $f_2 = f_1 + f_4$. Adding a second harmonic to the fourth frequency significantly improves the fit. This adds to the puzzling state of this star as already reported by @mathias04, since line-profile variations were observed, but also changes of the equivalent widths of the line. Stellar activity is suggested as an explanation. This could mean that one of the frequencies might be related to rotation.
= HIP 114189 = HR 8799 = V342 Peg
----------------------------------
There is little doubt that this star has at least two frequencies. The first two frequencies (1.9791 and 1.7368[$\mathrm{d}^{-1}$]{}) found in the multisite campaign presented in @zerbi99 are present in the Mercator data as well ($1.9806$ and $1.7326$[$\mathrm{d}^{-1}$]{}). Since their multisite campaign was relatively short, there is agreement within the errors. Combining the data with the Hipparcos measurements reveals another frequency,$0.76762$[$\mathrm{d}^{-1}$]{}, which could be an alias of the frequency $f_4=0.2479$[$\mathrm{d}^{-1}$]{} of @zerbi99.
= HIP 116434
-------------
It is known that for some time series of observations of multiperiodic variations, the technique of consecutive prewhitening will not yield the correct solution (see e.g. @cuypers98). The Hipparcos measurements of this star are another example of this phenomenon. In the Hipparcos series a first frequency appears near 0.90[$\mathrm{d}^{-1}$]{}, but only multifrequency fits lead to a solution compatible with the other sets of observations. In the final three-frequency solution, the frequency near 0.90[$\mathrm{d}^{-1}$]{} is not present anymore. When the Mercator and Hipparcos data were combined (the @henry02 data were not used), very precise values for the frequencies could be found (errors less than 0.00001[$\mathrm{d}^{-1}$]{}). The frequencies are not independent: in the Mercator data the third frequency is the mean of $f_1$ and $f_2$, in the data of @henry02 and in the Hipparcos data the second frequency is the sum of the two others (within the errors). Adding a harmonic to the third frequency of the Mercator data improves the fit. The two main frequencies are also in agreement with values published by @kaye04. Their spectroscopic observations indicate that HD 221866 is a spectroscopic binary with an orbital period of 135 days and a mass ratio of 1.11$\pm$0.03.
Summary of the results
======================
The results of the frequency analysis of the Mercator data are summarized in Table \[resM\]. For the combined data the frequency solution can be found in Table \[res\]. The errors on the frequencies are of the order of the last decimal given, typically $10^{-5}$ - $10^{-6}$[$\mathrm{d}^{-1}$]{}.
For all stars with a well-defined frequency solution in the combined data, the amplitudes in the different filters of the Geneva photometric systems were calculated. For each individual star, we give a summary of the results based on the Mercator data in Tables 5 to 25. The errors on the frequencies (and the periods) were calculated with the expressions given by @montgomery99 or @breger99, but only the largest value over the filters is given. In general the difference is less than a factor of 2. The errors on the amplitudes were calculated with the standard deviations of the final residuals. As a consequence, the errors are the same for each frequency in a multifrequency solution. Since these values are larger than the formal values obtained from a harmonic least-squares fit, they can be considered as a reliable upper limit. The relative errors will increase and this is reflected in the increase of errors on the amplitude ratios from the dominant to the lower amplitude modes (see Figure \[rat277\]).
The mean values of the magnitudes in each filter as obtained from the fit (with errors) are given in each Table as well. The values were used for the calibration of effective temperature, gravity and a metallicity indicator according to @kunzli97, of which a summary is given in Table \[calib\].
Conclusions
===========
Almost all frequencies of known [$\gamma$Dor]{} stars presented in this paper are confirmed and one candidate (HD 74504) can now be considered as an additional member of this class. By combining the Mercator data with data available from the literature (including Hipparcos data), sets of frequencies were enlarged, confirmed and/or more precisely determined. Many [$\gamma$Dor]{} stars have frequencies that are very stable over a long period (1989-2004). Only in a few cases did the frequencies found in the Hipparcos data (1989-1993) deviate from the values found in more recent data. Differences are only a few hundredths of a cycle per day in some cases and yearly aliasing could be the cause, but in a few other stars there was no agreement at all. This may be caused by the coarse sampling of the Hipparcos data, but it could also be an indication of some more hidden frequencies, complex beating patterns and/or changing frequencies.
For the large majority of the [$\gamma$Dor]{} stars the behaviour of the amplitude ratios as a function of the wavelength is very similar. We will explore these observational results in future modelling of their oscillations, with the goal to check if we can constrain the input physics of the models for F-type stars. Our study is the only consistent long-term multicolour photometric database of g-mode pulsations in an extended sample of [$\gamma$Dor]{} stars so far, and it constitutes a useful reference for future interpretations of the oscillatory behaviour of such pulsators.
This research was made possible thanks to the financial support from the Fund for Scientific Research - Flanders (FWO), under projects G.0178.02 and G.0332.06 The Leuven authors additionally are supported by the Research Council of K.U.Leuven under grant GOA/2008/04. The research leading to these results has received funding from the European Research Council under the European Community’s Seventh Framework Programme (FP7/2007–2013)/ERC grant agreement n$^\circ$227224 (PROSPERITY). We are indebted to our colleagues of the Geneva Observatory who reduced the [[mercator]{}]{} data. This research has made use of the NASA’s Astrophysics Data System and the SIMBAD astronomical database operated at the CDS in Strasbourg, France.
[rrlrrllllll]{} HD number & HIP number & Spectral & N & T(d) & $f_1$([$\mathrm{d}^{-1}$]{})& $f_2$([$\mathrm{d}^{-1}$]{})& $f_3$([$\mathrm{d}^{-1}$]{}) & $f_4$([$\mathrm{d}^{-1}$]{}) & $f_5$([$\mathrm{d}^{-1}$]{})&$f_6$([$\mathrm{d}^{-1}$]{})\
& & type & & & & & & &\
\
277 & 623 & F0 & 239 & 1148 & 1.11109 & 1.08130 & 1.38709\
2842 & 2510 & F0V & 109 & 1149 & 1.71270 & 1.53729 & 1.67063\
7169 & 5674 & F2V & 96 & 1149 & 1.8225 & 1.9245\
23874 & 17826 & F0 & 127 & 1045 & 2.2567 & 1.8856\
48271 & 32263 & F0 & 88 & 869 & 0.9123 & 0.8586 & 0.9314 & 1.7327\
62454 & 37863 & F0 & 158 & 847 & 1.5985:\
69715 & 40791 & A5 & 64 & 805 & 2.45642 & 2.44120\
74504 & 43062 & F0 & 122 & 794 & 1.90574 & 1.82101 & 1.84570\
86358 & 48895 & F3V & 100 & 805 & 1.1216 & 1.2930 & 1.1869 & 1.1423 & 1.00286\
100215 & 56275 & Am & 169 & 839 & 1.32191 & 1.42201 & 1.27911 & 1.6159:\
105458 & 59203 & F0III & 340 & 985 & 1.32084 & 1.25034 & 1.40903 & 0.94631 & 1.55279 & 1.09007\
108100 & 60571 & F2V & 196 & 980 & 1.40132 & 1.32726 & 1.34071:\
113867 & 63951 & F0 & 163 & 946 & 0.88841 & 1.00553 & 1.89101 & 1.77613 & 0.88340 & 1.42185\
167858 & 89601 & F2V & 62 & 407 & 0.7650 & 0.6984& 1.0734\
175337 & 92837 & F5 & 40 & 752 & 1.27109\
195068 & 100859 & F5 & 189 & 766 & 1.25054 & 1.29843 & 0.96553 & 0.28517\
206043 & 106897 & F2V & 158 & 1163 & 2.4324 & 2.3596 & 2.5243\
207223 & 107558 & F3V & 115 & 1163 & 0.38287\
211699 & 110163 & F0 & 97 & 1149 & 0.9327 & 1.1963 & 1.1635\
218396 & 114189 & A5V & 95 & 1151 & 1.9806 & 1.7326 & 0.7676\
221866 & 116434 & A3m & 135 & 1149 & 0.8384 & 0.8772 & 0.8572\
\[resM\]
[rrlrrlllllll]{} HD & HIP & Spectral & Source & N & T(d) & $f_1$([$\mathrm{d}^{-1}$]{})& $f_2$([$\mathrm{d}^{-1}$]{})& $f_3$([$\mathrm{d}^{-1}$]{}) & $f_4$([$\mathrm{d}^{-1}$]{}) & $f_5$([$\mathrm{d}^{-1}$]{})&$f_6$([$\mathrm{d}^{-1}$]{})\
number & number & type & & & & & & &\
\
277 & 623 & F0 & HLM & 864 & 5371 & 1.11101 & 1.08099 & 1.38704 & [**0.83469**]{}\
2842 & 2510 & F0V & HLM & 435 & 5377 & 1.71268 & 1.53725 & 1.66799\
7169 & 5674 & F2V & HLM & 548 & 5384 & 1.82259 & 1.92502 & [***1.76031***]{}\
23874 & 17826 & F0 & HLM & 493 & 5385 & 2.25651 & [**1.88562**]{}\
48271 & 32263 & F0 & HLM & 580 & 5109 & [***0.91236***]{} & [***0.87189***]{} & [***0.92314***]{} & [***1.75472***]{}\
62454 & 37863 & F0 & HLM & 264 & 5084 & 1.60156 & 1.43631 & 1.73694 & 1.83413 & 1.81079\
69715 & 40791 & A5 & HLM & 580 & 4816 & 2.456426 & 2.441492\
74504 & 43062 & F0 & H M & 213 & 5113 & 1.905773 & [**1.820999**]{} & [**1.845691**]{}\
86358 & 48895 & F3V & HLM & 598 & 5169 & 1.289696 & 1.185229 & 1.121666 & 1.142454 & [**0.997646**]{}:\
100215 & 56275 & Am & HLM & 946 & 5213 & 1.32188 & 1.42207 & 1.27905 & 1.61480 & [**1.54354**]{}\
105458 & 59203 & F0III & HLM &1043 & 5215 & 1.32083 & 1.25036 & 0.94632 & 1.40903 & 1.09021 & 1.55261\
108100 & 60571 & F2V & HLM & 747 & 5192 & 1.32572 & 1.40136 & 1.36542\
113867 & 63951 & F0 & HLM & 722 & 5180 & 0.88838 & [***1.00420***]{} & [***1.89379***]{} & 1.77757 & 1.42221\
167858 & 89601 & F2V & HLM & 622 & 4532 & 0.76508 & 0.69845 & 1.30185 & [**1.60558**]{} & [**1.07551**]{}\
175337 & 92837 & F5 & HLM & 372 & 4869 & 1.271134\
195068 & 100859 & F5 & HLM & 610 & 4994 & 1.250511 & 1.298422 & 0.965416 & [**0.284863**]{}\
206043 & 106897 & F2V & HLM & 600 & 5366 & 2.35944 & 2.43242 & 2.52427 & 2.26551 & 2.59895 & [**2.46092**]{}\
207223 & 107558 & F3V & H M & 204 & 5351 & [***0.385383***]{}\
211699 & 110163 & F0 & HLM & 617 & 5472 & 0.93280 & 1.12646 & 1.16322 & 0.19377\
218396 & 114189 & A5V & H M & 169 & 5394 & 1.980518 & 1.732528 & [***0.767562***]{}\
221866 & 116434 & A3M & H M & 304 & 5394 & 0.877197 & 0.838478 & 1.715664\
\[res\]
[rrrrrrrrr]{} HD number & U & B1 & B & B2 & V1 & V & G &\
\
277 & 9.2075 & 8.7915 & 7.8194 & 9.2134 & 9.0873 & 8.3666 & 9.4687\
2842 & 8.7691 & 8.3682 & 7.4092 & 8.8170 & 8.7050 & 7.9858 & 9.0944\
7169 & 8.1116 & 7.7034 & 6.7373 & 8.1385 & 8.0028 & 7.2810 & 8.3826\
23874 & 9.0657 & 8.6702 & 7.6947 & 9.0866 & 8.9212 & 8.1980 & 9.2914\
48271 & 8.2670 & 7.8668 & 6.9094 & 8.3160 & 8.1995 & 7.4807 & 8.5860\
62454 & 7.9513 & 7.5519 & 6.5841 & 7.9808 & 7.8535 & 7.1336 & 8.2365\
69715 & 7.9669 & 7.5578 & 6.6032 & 8.0114 & 7.8897 & 7.1708 & 8.2751\
74504 & 9.6144 & 9.2039 & 8.2523 & 9.6614 & 9.5673 & 8.8519 & 9.9629\
86358 & 7.2567 & 6.8881 & 5.9217 & 7.3235 & 7.1841 & 6.4625 & 7.5631\
100215 & 8.7886 & 8.3636 & 7.4040 & 8.8074 & 8.6996 & 7.9816 & 9.0866\
105458 & 8.5229 & 8.1213 & 7.1688 & 8.5793 & 8.4753 & 7.7584 & 8.8683\
108100 & 7.9628 & 7.5625 & 6.5931 & 7.9882 & 7.8371 & 7.1152 & 8.2101\
113867 & 7.6159 & 7.1823 & 6.2260 & 7.6367 & 7.5425 & 6.8253 & 7.9382\
167858 & 7.4093 & 6.9926 & 6.0277 & 7.4350 & 7.3337 & 6.6143 & 7.7259\
175337 & 8.2102 & 7.8103 & 6.8335 & 8.2302 & 8.1012 & 7.3784 & 8.4828\
195068 & 6.5117 & 6.0955 & 5.1233 & 6.5300 & 6.4220 & 5.6994 & 6.8119\
206043 & 6.5605 & 6.1317 & 5.1674 & 6.5809 & 6.4790 & 5.7572 & 6.8713\
207223 & 6.9961 & 6.6026 & 5.6300 & 7.0305 & 6.8966 & 6.1730 & 7.2762\
211699 & 9.9568 & 9.5352 & 8.5691 & 9.9679 & 9.8425 & 9.1226 & 10.2255\
218396 & 6.6524 & 6.2587 & 5.3186 & 6.7531 & 6.6774 & 5.9601 & 7.0853\
221866 & 8.2668 & 7.7669 & 6.8081 & 8.2177 & 8.1670 & 7.4528 & 8.5775\
HD number $T_{\mathrm{eff}}(\mathrm{K})$ $\Delta$ $\log~\mathrm{g}$ $\Delta$ $[M/H]$ $\Delta$
----------- -------------------------------- ---------- ------------------- ---------- --------- ----------
277 6995 60 4.38 0.11 0.11 0.08
2842 7091 61 4.49 0.08 -0.12 0.10
7169 6905 53 4.29 0.13 -0.09 0.09
23874 6720 47 4.04 0.15 -0.06 0.08
48271 7050 60 4.47 0.08 -0.15 0.10
62454 6997 57 4.45 0.11 0.03 0.09
69715 6987 58 4.39 0.10 -0.24 0.10
74504 7221 62 4.50 0.08 -0.14 0.10
86358 6910 54 4.47 0.13 -0.11 0.10
100215 7109 59 4.40 0.07 -0.04 0.09
105458 7144 62 4.50 0.07 -0.20 0.11
108100 6809 50 4.17 0.15 -0.08 0.09
113867 7195 62 4.40 0.08 -0.10 0.10
167858 7177 61 4.46 0.07 0.01 0.09
175337 6993 59 4.44 0.12 0.12 0.08
195068 7133 61 4.46 0.08 0.07 0.09
206043 7144 60 4.41 0.07 -0.07 0.09
207223 6952 54 4.43 0.12 0.01 0.09
211699 6984 56 4.34 0.10 0.00 0.09
218396 7355 67 4.57 0.07 -0.71 0.18
221866 7480 66 4.36 0.07 0.14 0.09
[^1]: Fellow of the Fund for Scientific Research, Flanders (FWO)
[^2]: NASA Post-doctoral fellow
[^3]: Based on observations collected with the Flemish 1.2-meter [Mercator]{} Telescope at Roque de los Muchachos, La Palma
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The neutrino-heating mechanism remains a viable possibility for the cause of the explosion in a wide mass range of supernova progenitors. This is demonstrated by recent two-dimensional hydrodynamic simulations with detailed, energy-dependent neutrino transport. Neutrino-driven explosions were not only found for stars in the 8–10$\,M_\odot$ range with ONeMg cores and in case of the iron core collapse of an 11$\,M_\odot$ progenitor, but also for a “typical” 15$\,M_\odot$ progenitor model. For such more massive stars, however, the explosion occurs significantly later than so far thought, and is crucially supported by large-amplitude bipolar oscillations due to the nonradial standing accretion shock instability (SASI), whose low (dipole and quadrupole) modes can develop large growth rates in conditions where convective instability is damped or even suppressed. The dominance of low-mode deformation at the time of shock revival has been recognized as a possible explanation of large pulsar kicks and of large-scale mixing phenomena observed in supernovae like SN 1987A.'
author:
- 'H.-Thomas Janka'
- Andreas Marek
- 'Francisco-Shu Kitaura'
bibliography:
- 'sample.bib'
title: 'Neutrino-driven explosions twenty years after SN1987A'
---
[ address=[Max-Planck-Institute for Astrophysics, Karl-Schwarzschild-Str. 1, D-85741 Garching, Germany]{} ]{}
[ address=[Max-Planck-Institute for Astrophysics, Karl-Schwarzschild-Str. 1, D-85741 Garching, Germany]{} ]{}
[ address=[Max-Planck-Institute for Astrophysics, Karl-Schwarzschild-Str. 1, D-85741 Garching, Germany]{} ]{}
Introduction
============
Besides the many other surprises Supernova 1987A brought for astronomers, it had a major impact on the theory of stellar core-collapse and explosion mainly by two discoveries. On the one hand the historical detection of two dozen neutrino events in three underground laboratories has confirmed the concept of gravitational instability and neutron star formation, in which the production of electron capture neutrinos and the emission of neutrinos and antineutrinos of all flavors by thermal processes had been predicted for a long time [@Arnett.etal:1989].
On the other hand, the lightcurve and spectra of SN 1987A brought unambiguous evidence that nucleosynthesis products were distributed strongly anisotropically and that large-scale mixing took place during the explosion, for which reason X-rays and $\gamma$-rays from the decay of radioactive cobalt were measured much earlier than expected. Heavy elements were observed to expand with velocities significantly larger than expected from spherically symmetric explosion models. This was interpreted as a clear sign that the onion-shell structure of the progenitor star was destroyed during the explosion [@Arnett.etal:1989]. Meanwhile, twenty years later, the remnant of SN 1987A at the center of the ring system reveals a clear prolate deformation and suggests a global asymmetry of the mass ejection.
Multi-dimensional supernova models showed that sufficiently strong radial mixing of radioactive nuclei requires that hydrodynamic instabilities have developed in layers near the stellar core and already during the earliest stages of the explosion. In fact, simulations of the onset of the explosion demonstrated that strong convective overturn can occur in the Ledoux-unstable region of neutrino energy deposition behind the stalled supernova shock [@Herant.etal:1994; @Burrows.etal:1995; @Janka.Mueller:1996].
Meanwhile it is clear that convection is not the only source of asymmetry during the shock stagnation phase. The standing accretion shock has been recognized to be generically unstable to nonradial deformation, even in situations where convection is damped or suppressed. This so-called “standing accretion shock instability” (SASI; [@Blondin.etal:2003]; for more literature, see [@Scheck.etal:2007]) shows a preferential growth of low shock-deformation modes (dipole, $l=1$, and quadrupole, $l=2$, modes in terms of an expansion in spherical harmonics). The presence of such a low-mode instability has turned out to have important implications for large-scale explosion asymmetries, pulsar kicks, and — as suggested by very recent simulations — for the development of neutrino-driven explosions. Corresponding results will be reported below and implications for SN 1987A will be discussed.
Explosion models with energy-dependent neutrino transport
=========================================================
Numerical method
----------------
The core-collapse and post-bounce calculations presented here were performed in spherical symmetry with the neutrino-hydrodynamics code <span style="font-variant:small-caps;">Vertex</span> (for details, see [@Rampp.Janka:2002; @Buras.etal:2006]). The code module that integrates the nonrelativistic hydrodynamics equations is a conservative, Eulerian implementation of a Godunov-type scheme with higher-order spatial and temporal accuracy. The self-gravity of the stellar gas is treated with an approximation to general relativity as described in [@Marek.etal:2006]. The code was tested against fully relativistic simulations in [@Liebendoerfer.etal:2005; @Marek.etal:2006]. The time-implicit transport routine solves the moment equations for neutrino number, energy, and momentum. It employs a variable Eddington closure factor that is obtained from iterating to convergence a simplified Boltzmann equation coupled to the set of its moment equations. The interactions of neutrinos ($\nu$) and antineutrinos ($\bar\nu$) of all flavors include a state-of-the-art treatment of charged-current and neutral-current interactions with electrons, nucleons, and nuclei (making use of the improved electron capture rates on a very large NSE-ensemble of nuclei as considered by [@Langanke.Martinez-Pinedo.ea:2003]). The most important neutrino-pair processes in SNe as well as reactions between neutrinos of different flavors are taken into account [@Buras.etal:2006; @Marek.etal:2005]).
![[*Left:*]{} Density profile of the ONeMg core and the surrounding He-shell of an 8.8$\,M_\odot$ star, which is considered to be representative of the 8–10$\,M_\odot$ range, compared to progenitor stars with 10.2, 11.2, and 15$\,M_\odot$. Note that due to the lack of data from stellar evolution models, the He-shell outside the oxygen-helium transition at about 1000$\,$km was constructed from hydrostatic equilibrium, using a temperature profile as given by the 10.2$\,M_\odot$ progenitor (A. Heger, private communication). The actual density gradient is even steeper (K. Nomoto, private communication). [*Right:*]{} The mass accretion rate of the collapsing ONeMg core at a function of time after bounce, measured just outside of the supernova shock []{data-label="fig:ONeMgprofile"}](ONeMg-densityprof.eps "fig:"){width=".48\textwidth"} ![[*Left:*]{} Density profile of the ONeMg core and the surrounding He-shell of an 8.8$\,M_\odot$ star, which is considered to be representative of the 8–10$\,M_\odot$ range, compared to progenitor stars with 10.2, 11.2, and 15$\,M_\odot$. Note that due to the lack of data from stellar evolution models, the He-shell outside the oxygen-helium transition at about 1000$\,$km was constructed from hydrostatic equilibrium, using a temperature profile as given by the 10.2$\,M_\odot$ progenitor (A. Heger, private communication). The actual density gradient is even steeper (K. Nomoto, private communication). [*Right:*]{} The mass accretion rate of the collapsing ONeMg core at a function of time after bounce, measured just outside of the supernova shock []{data-label="fig:ONeMgprofile"}](ONeMg-accretionrate.eps "fig:"){width=".48\textwidth"}
![[*Left:*]{} Four snapshots of the explosion of an 8–10$\,M_\odot$ star in a two-dimensional (2D) simulation, which was performed in a $\pm 45^\circ$ wedge around the equatorial plane, using periodic boundary conditions. Time is normalized to bounce. The color coding represents the entropy per nucleon with black corresponding to values of ${\,\hbox{\hbox{$ < $}\kern -0.8em \lower
1.0ex\hbox{$\sim$}}\,}$7$\,k_\mathrm{B}$, red to 10–15$\,k_\mathrm{B}$, orange to 15–20$\,k_\mathrm{B}$, and white to about 25$\,k_\mathrm{B}$. The supernova shock is visible as sharp red/black discontinuity at about 210$\,$km in the upper left panel, while it is already far outside the displayed region at all other times (the corresponding shock radii are roughly 900$\,$km, 5600$\,$km, and 15000$\,$km). [*Right:*]{} Explosion energy as a function of time for the 2D simulation of the left figure compared to two runs in spherical symmetry (1D) with a soft (“L&S”) and a stiff (“W&H”) nuclear equation of state. The steep increase of the explosion energy in the 1D models after about 150$\,$ms is caused by the onset of the expansion of neutrino-heated matter away from the gain radius. Convective overturn leads to more efficient neutrino heating of a larger mass and to an earlier rise of the explosion energy in the 2D simulation []{data-label="fig:ONeMgexpl"}](ONeMg-snapshots.eps "fig:"){width=".50\textwidth"} ![[*Left:*]{} Four snapshots of the explosion of an 8–10$\,M_\odot$ star in a two-dimensional (2D) simulation, which was performed in a $\pm 45^\circ$ wedge around the equatorial plane, using periodic boundary conditions. Time is normalized to bounce. The color coding represents the entropy per nucleon with black corresponding to values of ${\,\hbox{\hbox{$ < $}\kern -0.8em \lower
1.0ex\hbox{$\sim$}}\,}$7$\,k_\mathrm{B}$, red to 10–15$\,k_\mathrm{B}$, orange to 15–20$\,k_\mathrm{B}$, and white to about 25$\,k_\mathrm{B}$. The supernova shock is visible as sharp red/black discontinuity at about 210$\,$km in the upper left panel, while it is already far outside the displayed region at all other times (the corresponding shock radii are roughly 900$\,$km, 5600$\,$km, and 15000$\,$km). [*Right:*]{} Explosion energy as a function of time for the 2D simulation of the left figure compared to two runs in spherical symmetry (1D) with a soft (“L&S”) and a stiff (“W&H”) nuclear equation of state. The steep increase of the explosion energy in the 1D models after about 150$\,$ms is caused by the onset of the expansion of neutrino-heated matter away from the gain radius. Convective overturn leads to more efficient neutrino heating of a larger mass and to an earlier rise of the explosion energy in the 2D simulation []{data-label="fig:ONeMgexpl"}](ONeMg-expenergies.eps "fig:"){width=".48\textwidth"}
Neutrino-driven explosions for progenitors below 10$\,M_\odot$
--------------------------------------------------------------
Recently Kitaura et al. [@Kitaura.etal:2006] reinvestigated the stellar collapse of a $\sim$1.3$\,M_\odot$ core of oxygen, neon, and magnesium, surrounded by a thin ($\sim$0.08$M_\odot$) carbon layer and a very dilute helium shell. The progenitor had 8.8$\,M_\odot$ on the main sequence mass and an initial He-core with 2.2$\,M_\odot$ [@Nomoto:1984]. It can be considered as representative of the lowest-mass progenitors of core-collapse supernovae in the 8–10$\,M_\odot$ range.
Kitaura et al. obained an explosion that set in about 100$\,$ms after core bounce and whose energy was provided by a neutrino-driven wind. The spherically symmetric (1D) simulations confirm qualitatively older calculations by Mayle and Wilson [@Mayle.Wilson:1988], although the recent explosion models are significantly less powerful and important differences with respect to the nucleosynthesis conditions in the ejecta are seen.
Because of the presence of O, Ne, Mg, and C, nuclear burning still proceeds in the outer regions of the stellar core while efficient electron capture (mostly on $^{20}$Ne, $^{24}$Na, and $^{24}$Mg) reduces the electron degeneracy pressure and drives the core to gravitational instability. It is, however, not the presence of the energy release by burning in some shells that makes the explosion of stars with such cores much easier than that of more massive progenitors with iron in the center (the compressed matter in any case is heated to nuclear statistical equilibrium, and the energy released by the burning is efficiently removed by escaping neutrinos). The main reason for the readiness of such low-mass stars to explode by the neutrino-driven mechanism is the decreasing density in the C-layer and the extremely steep density gradient at the transition from the C-shell to the He-mantle (see the left plot in Fig. \[fig:ONeMgprofile\]). This leads to a continuous, fast drop of the mass accretion rate after about 50$\,$ms of post-bounce evolution (Fig. \[fig:ONeMgprofile\], right plot). As a consequence, the stalled prompt shock starts reexpanding and accelerates the very dilute matter in its downstream region. At about 150$\,$ms after bounce material expands outward from regions near the gain radius, where it was exposed to intense neutrino heating. This phase is associated with a steep rise of the explosion energy in Fig. \[fig:ONeMgexpl\] (right panel). Between 200 and 250$\,$ms after bounce a powerful neutrino-driven wind begins to shed off more gas from the surface of the nascent neutron star. From this time on the explosion energy in Fig. \[fig:ONeMgexpl\] shows a more gradual but continuous further increase.
Multi-dimensional effects are obviously not crucial for obtaining neutrino-driven explosions of progenitors with the structure of the considered $\sim\,$9$\,M_\odot$ model. Nevertheless, a simulation performed in two dimensions (2D; i.e., assuming axial symmetry) shows that convective overturn in the neutrino-heated layer between the gain radius (at 90$\,$km) and the shock becomes strong about 80$\,$ms after bounce and has fully developed 20$\,$ms later (see Fig. \[fig:ONeMgexpl\], upper panels of left plot). It carries cooler matter in narrow downdrafts from larger distances to locations closer to the gain radius, where the gas is exposed to more efficient neutrino heating. Therefore a larger gas mass absorbs energy from neutrinos before it accelerates outward in rising high-entropy plumes. This leads to a considerably higher energy of the explosion than in the corresponding 1D simulations (Fig. \[fig:ONeMgexpl\], right plot), but has essentially no effect on the propagation of the supernova shock during this phase, because the shock is already far outside of the convective region. After about 150$\,$ms of post-bounce evolution the radial propagation of the neutrino-heated layers has become so fast that the mixing motions freeze out and the corresponding fluid pattern with characteristic Rayleigh-Taylor mushrooms expands self-similarly with high velocity (Fig. \[fig:ONeMgexpl\], lower panels in the left plot).
The 2D simulation also shows that convection inside the nascent neutron star does not lead to any significant increase of the neutrino luminosities and thus of the neutrino heating behind the shock. The enhanced explosion energy is merely a consequence of the convective activity behind the supernova shock. This is clearly different from the simulations by Mayle & Wilson [@Mayle.Wilson:1988], who obtained models with larger explosion energy by assuming that the neutrino luminosities were boosted by neutron-finger convection below the neutrinosphere.
The rapid outward acceleration also has the consequence that the convective pattern never develops dominant power on the largest scales. The expansion of the gain layer happens so quickly that the convective plumes have no time to merge to structures with lateral wavelengths of more than about 45$^\circ$. Since the shock radius grows continuously with time, also the SASI has no possibility to grow (for more details, see below). Such a situation disfavors the development of a large global asymmetry of the small amount of material that is accelerated during the early stages of the explosion. Therefore the pulsar kick velocities must be expected to remain rather small (roughly ${\,\hbox{\hbox{$ < $}\kern -0.8em \lower
1.0ex\hbox{$\sim$}}\,}$100$\,$km/s) in case of the O-Ne-Mg core collapse events.
![[*Left:*]{} Angular average of the shock radius (volume weighted) for the 2D simulation of the 15$\,M_\odot$ explosion compared to the shock position of a corresponding spherically symmetric simulation. [*Right:*]{} Evolution of the ratio of advection timescale of accreted matter through the gain layer to the neutrino-heating timescale for the exploding 15$\,M_\odot$ model. There is a continuous increase until the critical value of unity is exceeded after about 500$\,$ms of post-bounce evolution. At $t>580\,$ms the beginning strong overall expansion of the postshock layer prevents a reasonable determination of the advection timescale []{data-label="fig:15Msunexpl-rst"}](15Msun-rot1D-rshock.eps "fig:"){width=".48\textwidth"} ![[*Left:*]{} Angular average of the shock radius (volume weighted) for the 2D simulation of the 15$\,M_\odot$ explosion compared to the shock position of a corresponding spherically symmetric simulation. [*Right:*]{} Evolution of the ratio of advection timescale of accreted matter through the gain layer to the neutrino-heating timescale for the exploding 15$\,M_\odot$ model. There is a continuous increase until the critical value of unity is exceeded after about 500$\,$ms of post-bounce evolution. At $t>580\,$ms the beginning strong overall expansion of the postshock layer prevents a reasonable determination of the advection timescale []{data-label="fig:15Msunexpl-rst"}](15Msun-rot-tratio.eps "fig:"){width=".48\textwidth"}
![Radial positions of the shock near the north and south pole as functions of post-bounce time (white lines) in the 2D simulation of the explosion of a 15$\,M_\odot$ star. The color coding represents the entropy per nucleon of the stellar gas. The quasi-periodic shock expansion and contraction due to the SASI can be clearly seen []{data-label="fig:15Msunexpl-rse"}](15Msun-rot-rshentropy.eps){width=".99\textwidth"}
=0.5mm
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Four snapshots from the post-bounce evolution of the exploding 15$\,M_\odot$ star in a 2D simulation. The upper left plot shows the entropy distribution at $t= 119\,$ms after bounce, about 40$\,$ms after the postshock convection has reached the nonlinear regime. The upper right and lower left plots ($t = 454\,$ms and $t = 524\,$ms after bounce) demonstrate the presence of a very strong bipolar oscillation due to the SASI, and the lower right plot ($t = 610\,$ms p.b.) displays the acceleration phase of the strongly aspherical explosion with a large $l=1$ component. Note that the contracting nascent neutron star exhibits a growing prolate deformation because of the rotation assumed in this simulation []{data-label="fig:15Msunexpl-snaps"}](15Msun-rot-119ms.eps "fig:"){width=".48\textwidth"} ![Four snapshots from the post-bounce evolution of the exploding 15$\,M_\odot$ star in a 2D simulation. The upper left plot shows the entropy distribution at $t= 119\,$ms after bounce, about 40$\,$ms after the postshock convection has reached the nonlinear regime. The upper right and lower left plots ($t = 454\,$ms and $t = 524\,$ms after bounce) demonstrate the presence of a very strong bipolar oscillation due to the SASI, and the lower right plot ($t = 610\,$ms p.b.) displays the acceleration phase of the strongly aspherical explosion with a large $l=1$ component. Note that the contracting nascent neutron star exhibits a growing prolate deformation because of the rotation assumed in this simulation []{data-label="fig:15Msunexpl-snaps"}](15Msun-rot-454ms.eps "fig:"){width=".48\textwidth"}
![Four snapshots from the post-bounce evolution of the exploding 15$\,M_\odot$ star in a 2D simulation. The upper left plot shows the entropy distribution at $t= 119\,$ms after bounce, about 40$\,$ms after the postshock convection has reached the nonlinear regime. The upper right and lower left plots ($t = 454\,$ms and $t = 524\,$ms after bounce) demonstrate the presence of a very strong bipolar oscillation due to the SASI, and the lower right plot ($t = 610\,$ms p.b.) displays the acceleration phase of the strongly aspherical explosion with a large $l=1$ component. Note that the contracting nascent neutron star exhibits a growing prolate deformation because of the rotation assumed in this simulation []{data-label="fig:15Msunexpl-snaps"}](15Msun-rot-524ms.eps "fig:"){width=".48\textwidth"} ![Four snapshots from the post-bounce evolution of the exploding 15$\,M_\odot$ star in a 2D simulation. The upper left plot shows the entropy distribution at $t= 119\,$ms after bounce, about 40$\,$ms after the postshock convection has reached the nonlinear regime. The upper right and lower left plots ($t = 454\,$ms and $t = 524\,$ms after bounce) demonstrate the presence of a very strong bipolar oscillation due to the SASI, and the lower right plot ($t = 610\,$ms p.b.) displays the acceleration phase of the strongly aspherical explosion with a large $l=1$ component. Note that the contracting nascent neutron star exhibits a growing prolate deformation because of the rotation assumed in this simulation []{data-label="fig:15Msunexpl-snaps"}](15Msun-rot-610ms.eps "fig:"){width=".48\textwidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
SASI-supported neutrino-driven explosions of stars above 10$\,M_\odot$
----------------------------------------------------------------------
The core structure of stars more massive than about 10$\,M_\odot$ is considerably different from that of lower mass supernova progenitors (see Fig. \[fig:ONeMgprofile\]). Spherically symmetric calculations, carried out over many hundreds of milliseconds after core bounce, have therefore not found explosions happening. Instead, the supernova shock stalls and mass is continuously accreting onto the forming neutron star (see the 1D result in the left plot of Fig. \[fig:15Msunexpl-rst\]).
Hydrodynamic instabilities in the supernova core, however, can change the situation. In 2D simulations Buras et al. [@Buras.etal:2006b] obtained an explosion of an 11.2$\,M_\odot$ progenitor. Numerical tests with different angular wedges and lateral boundary conditions of the polar grid showed that the crucial difference here was the growth of low ($l = 1,\,2$) SASI modes. The associated development of large-amplitude bipolar oscillations pushed the shock to larger radii and thus increased the timescale of accreted matter to fall from the shock (at $R_\mathrm{s}$) to the gain radius $R_\mathrm{g}$. The corresponding advection timescale $$\tau_\mathrm{adv}\, \equiv\, {R_\mathrm{s} - R_\mathrm{g}
\over |\left\langle v_r \right\rangle|}$$ can be considered as a measure of the duration gas is exposed to neutrino heating in the gain layer. When the stalled shock reaches a larger radius $R_\mathrm{s}$, the preshock velocity and average postshock velocity $\left\langle v_r \right\rangle$ are significantly smaller, which leads to a considerably longer advection timescale (roughly $\tau_\mathrm{adv}
\propto R_\mathrm{s}^{3/2}$; Eq. (15) in [@Janka.etal:2001]). Our numerical experiments showed that the presence of postshock convection alone (if the low SASI modes were suppressed by grid constraints) was unable to provide enough support for a neutrino-driven explosion. When SASI oscillations helped increasing the shock radius, however, the crucial ratio of advection timescale to neutrino heating timescale grows and finally exceeds the critical value of unity. The neutrino heating timescale, $$\tau_\mathrm{heat}\,\equiv\,
{E_\mathrm{bind}[R_\mathrm{gain},\,R_\mathrm{shock}] \over
Q_\mathrm{heat}}$$ measures the time it takes neutrinos to deposit (with an integrated rate $Q_\mathrm{heat}$) and energy equal to the binding energy $E_\mathrm{bind}[R_\mathrm{gain},\,R_\mathrm{shock}]$ of the matter in the gain layer.
Very recent simulations show that such a positive feedback between low-mode SASI oscillations and neutrino heating also occurs in a 15$\,M_\odot$ progenitor (model s15s7b2 of [@Woosley.Weaver:1995]). Also there it finally leads to a neutrino-driven explosion, however at much a later time after core bounce (Fig. \[fig:15Msunexpl-rst\]). The particular model considered here includes a modest amount of rotation (the pre-collapse iron core had a rotation period of about 12 seconds as in Sect. 3.4 of [@Buras.etal:2006b]), which explains a growing oblateness of the nascent neutron star (see Fig. \[fig:15Msunexpl-snaps\]). Comparison with non-rotating models, however, reveals that angular momentum dependent effects may cause some quantitative differences (and may to some extent foster the evolution towards an explosion) but do not seem to be the essential ingredient that determines the overall behavior of the collapsing stellar core in the long run[^1].
Figure \[fig:15Msunexpl-rst\] (left) reveals a growth of the average shock radius, which starts at about 350$\,$ms after bounce and is accompanied by a continuous rise of the timescale ratio $\tau_\mathrm{adv}/\tau_\mathrm{heat}$ (Fig. \[fig:15Msunexpl-rst\], right). This rise is caused by an increase of the average advection timescale $\tau_\mathrm{adv}$, while $\tau_\mathrm{heat}$ remains nearly constant. The kinetic energy (also for the lateral component of the velocity) in the gain layer triples during this period of the evolution (while the rotational energy changes only by a modest amount), suggesting that nonradial fluid motions become more and more violent during this phase. Indeed, the bipolar SASI oscillations, which are visible from alternating shock expansion and contraction phases in the northern and southern hemispheres with a period of 10–15$\,$ms, exhibit a growing amplitude for $t_\mathrm{pb} > 350\,$ms (Fig. \[fig:15Msunexpl-rse\]). With a larger average shock radius also more mass is accumulated in the gain layer. At $t {\,\hbox{\hbox{$ > $}\kern -0.8em \lower
1.0ex\hbox{$\sim$}}\,}530\,$ms the critical timescale ratio exceeds unity and a runaway situation is reached. The accelerating overall expansion indicates the onset of a strongly aspherical, neutrino-powered explosion (Fig. \[fig:15Msunexpl-snaps\]).
![[*Left:*]{} G-mode oscillations of the nascent neutron star in the exploding 15$\,M_\odot$ simulation during 610$\,$ms of post-bounce evolution. The plot shows the amplitudes of the $l=1$ to $l=5$ modes of the pressure fluctuations at a radius of $r = 10\,$km expanded in spherical harmonics. Note that the quadrupole mode ($l=2$) has a large and growing amplitude because of the oblateness of the rotating neutron star. [*Middle:*]{} Test simulations with artificially instigated dipole ($l=1$) oscillation of the neutron star. Two different amplitudes of the initially imposed velocity field were used, $5\times 10^7\,$cm/s and $2\times 10^8\,$cm/s, corresponding to a factor of 16 different kinetic energies (as indicated in the plot). The clear presence of many cycles of the dipole oscillation demonstrates the ability of our numerical code to follow such gravity waves, if they are excited. [*Right:*]{} The amplitude of the $l=1$ mode in the lower panel of the middle plot as function of time and radius. Interior of about 10$\,$km the core oscillates with twice the frequency as the mantle outside of $r\approx 25\,$km. In the intermediate, convective layer the gravity waves are damped []{data-label="fig:15Msun-NSosc"}](15Msun-rot-dP-10km.eps "fig:"){width=".32\textwidth"} ![[*Left:*]{} G-mode oscillations of the nascent neutron star in the exploding 15$\,M_\odot$ simulation during 610$\,$ms of post-bounce evolution. The plot shows the amplitudes of the $l=1$ to $l=5$ modes of the pressure fluctuations at a radius of $r = 10\,$km expanded in spherical harmonics. Note that the quadrupole mode ($l=2$) has a large and growing amplitude because of the oblateness of the rotating neutron star. [*Middle:*]{} Test simulations with artificially instigated dipole ($l=1$) oscillation of the neutron star. Two different amplitudes of the initially imposed velocity field were used, $5\times 10^7\,$cm/s and $2\times 10^8\,$cm/s, corresponding to a factor of 16 different kinetic energies (as indicated in the plot). The clear presence of many cycles of the dipole oscillation demonstrates the ability of our numerical code to follow such gravity waves, if they are excited. [*Right:*]{} The amplitude of the $l=1$ mode in the lower panel of the middle plot as function of time and radius. Interior of about 10$\,$km the core oscillates with twice the frequency as the mantle outside of $r\approx 25\,$km. In the intermediate, convective layer the gravity waves are damped []{data-label="fig:15Msun-NSosc"}](15Msun-HW-dP-10km-5e7+2e8.eps "fig:"){width=".32\textwidth"} ![[*Left:*]{} G-mode oscillations of the nascent neutron star in the exploding 15$\,M_\odot$ simulation during 610$\,$ms of post-bounce evolution. The plot shows the amplitudes of the $l=1$ to $l=5$ modes of the pressure fluctuations at a radius of $r = 10\,$km expanded in spherical harmonics. Note that the quadrupole mode ($l=2$) has a large and growing amplitude because of the oblateness of the rotating neutron star. [*Middle:*]{} Test simulations with artificially instigated dipole ($l=1$) oscillation of the neutron star. Two different amplitudes of the initially imposed velocity field were used, $5\times 10^7\,$cm/s and $2\times 10^8\,$cm/s, corresponding to a factor of 16 different kinetic energies (as indicated in the plot). The clear presence of many cycles of the dipole oscillation demonstrates the ability of our numerical code to follow such gravity waves, if they are excited. [*Right:*]{} The amplitude of the $l=1$ mode in the lower panel of the middle plot as function of time and radius. Interior of about 10$\,$km the core oscillates with twice the frequency as the mantle outside of $r\approx 25\,$km. In the intermediate, convective layer the gravity waves are damped []{data-label="fig:15Msun-NSosc"}](15Msun-HW-dP-NS2D.eps "fig:"){width=".32\textwidth"}
Some comments on core g-modes and the acoustic mechanism
--------------------------------------------------------
In view of the recent numerical finding of acoustically-driven explosions, which are initiated by the acoustic power generated by large-amplitude core g-mode oscillations of the accreting neutron star [@Burrows.etal:2006; @Burrows.etal:2007], we have evaluated our long-time 15$\,M_\odot$ simulation for the gravity-wave activity of the forming compact remnant. Figure \[fig:15Msun-NSosc\] (left) displays the g-mode amplitudes of the first terms ($l=1,...,5$) of a spherical harmonics expansion of the pressure fluctuations at a radius of 10$\,$km inside the neutron star. The analysis follows the description in Ref. [@Burrows.etal:2006], see Fig. 7 there. The amplitudes of core g-modes in our model are roughly two orders of magnitude smaller than those seen in the run-up to an explosion in that figure. The acoustic energy flux radiated by the oscillating neutron star in our model is therefore completely negligible compared to neutrino heating behind the shock, which typically deposits energy at a net rate of 3–4$\times 10^{51}\,$erg/s at $t > 200\,$ms after bounce. The acoustic mechanism does not play a role for the evolution of our model and, according to the simulations in[@Burrows.etal:2007], it might become relevant only much later than our model explodes by neutrino-energy deposition.
But is our code able to follow core g-mode oscillations, in particular of $l=1$ type, because in this case the gas in the stellar center participates in the motion? The answer is “yes” (in contrast to statements that can be found in the literature[^2], see [@Burrows.etal:2006; @Burrows.etal:2007]). The middle and right panels of Figure \[fig:15Msun-NSosc\] show results of test simulations in which at some moment of the post-bounce evolution we artificially instigated a large dipole g-mode by imposing an $l=1,\,n=1$ (i.e., we assumed one radial node) perturbation of the $z$-component of the velocity field with varied amplitude and conserved linear momentum. The plots demonstrate that essentially a pure $l=1$ oscillation develops (after some initial relaxation, because our chosen perturbation did not correspond to an eigenfunction), which the code is able to follow through many cycles. We are therefore confident that we should see large core g-mode oscillations, if the anisotropic accretion flow around the neutron star were causing their excitation.
Conclusions
===========
The results of 2D supernova simulations presented in this paper demonstrate the ability of neutrino heating to initiate delayed explosions for progenitors in a wider range of masses. The explosion occurs significantly later than observed in older calculations with approximative neutrino transport. We identified large-amplitude SASI modes to play a crucial, supportive role for the development of the explosion because they enforce shock expansion and thus reduce the average infall velocity in the postshock region, which enables the accreted matter to stay in the neutrino-heating layer for a significantly longer time.
Our simulations, however, were stopped too early (for CPU time reasons) to allow for a final determination of the explosion energy. Accretion of matter by the shock is still going on, in particular in the 11 and 15$\,M_\odot$ stars, and gas is channelled towards the gain radius, where neutrino heating is strongest. A large fraction of this infalling material will start reexpanding, and energy this gas has absorbed from neutrinos and is released by nucleon recombination to alpha particles and iron-group nuclei will contribute to the explosion energy. In order to obtain reliable numbers for the explosion properties, the simulations will have to follow this accretion phase, which might last even for hundreds of milliseconds. Ultimately, however, 3D simulations will be needed. The explosion, its onset and strength, may depend on the additional degrees of freedom that are accessible to the fluid flow in three dimensions. Convective downdrafts and buoyant plumes, vorticity, and spiral modes are different in 3D or even do not exist when the flow is constrained to axisymmetry with all structures being tori around the polar grid axis.
The kind of asphericities seen in case of our 11.2 and 15$\,M_\odot$ explosion models, with a large contribution from an $l=1$ component, were shown to lead to such a big anisotropy of the supernova mass ejection that the neutron star receives a recoil sufficiently strong to explain the high velocities observed for many young pulsars, even those in excess of 1000$\,$km/s [@Scheck.etal:2004; @Scheck.etal:2006]. Moreover, the initial deformation of the supernova shock and the asymmetric ejecta distribution are the seed of subsequent hydrodynamic instabilities at the composition interfaces of the disrupted star after the passage of the supernova shock. These instabilities prevent the strong deceleration of the heavy elements and lead to a highly anisotropic distribution not only of Fe-group nuclei but also of silicon and oxygen. Large-scale mixing takes place, in course of which hydrogen and helium are carried deep into the star and pockets and clumps of heavy elements remain expanding with high velocities as observed in SN 1987A [@Kifonidis.etal:2006].
Even 20 years after the spectacular stellar death it is not clear what caused the explosion of SN 1987A. The ring system was interpreted as a sign for rapid rotation being present in the $\sim$18$\,M_\odot$ progenitor star. In particular the existence of a common axis of the ring system and of the elongated ejecta is a strong indication that rotation has played a role in the dying star, possibly as the consequence of a binary merger event some ten thousand years before the stellar collapse (see P. Podsiadlowski’s talk at this meeting). It is, however, not clear how such a merger has affected the angular momentum evolution of the stellar core. Only if the initial spin period of the core was small (${\,\hbox{\hbox{$ < $}\kern -0.8em \lower 1.0ex\hbox{$\sim$}}\,} 2\,$s according to Ref. [@Burrows.etal:2007b]), the free energy of rotation in the nascent neutron star was sufficiently large to power a supernova explosion by magnetohydrodynamic effects. But if the collapsing core was rotating so rapidly, why then is there no sign now of the energy input from a bright, Crab-like pulsar? A delayed collapse of a transiently existing neutron star to a black hole is disfavored as the solution of this puzzle, because the compact remnant formed in a typical SN 1987A progenitor is not expected to be so heavy that it cannot be stabilized by nuclear equation-of-states that are consistent with measured neutron star masses. Moreover, the pronounced prolate deformation of the now visible supernova ejecta at the center of the ring system may not be an unambiguous signature of very rapid core rotation but could result from a bipolar SASI asymmetry. SN 1987A may not only have been a unique event, it may also have been an uncommon one. We will probably never find out with final certainty. The next galactic supernova, however, will give us a new chance to learn more about the processes that trigger the explosion of a massive star: Tens of thousands of neutrino events will be captured by various underground experiments, and highly sensitive instruments promise to register the gravitational-wave signal produced by a nonspherical bounce and by hydrodynamic instabilities in the supernova core.
We are very grateful to R. Buras, W. Hillebrandt, K. Kifonidis, B. Müller, E. Müller, and M. Rampp for their input to various aspects of the reported project, and A. Heger, K. Nomoto, and S. Woosley for data of their progenitor models. This work was supported by the Deutsche Forschungsgemeinschaft through the Transregional Collaborative Research Center SFB/TR 27 “Neutrinos and Beyond”, the Collaborative Research Center SFB-375 “Astro-Particle Physics”, and the Cluster of Excellence “Origin and Structure of the Universe” (<http://www.universe-cluster.de>). Supercomputer time grants at the John von Neumann Institute for Computing (NIC) in Jülich, at the High Performance Computing Center Stuttgart (HLRS) of the University of Stuttgart, and at the Computer Center in Garching (RZG) are acknowledged.
[9]{}
W.D. Arnett, J.N. Bahcall, R.P. Kirshner, and S.E. Woosley, *Annu. Rev. Astron. Astrophys.* **27**, 629–700 (1989).
M. Herant, W. Benz, W.R. Hix, C.L. Fryer, and S.A. Colgate, *Astrophys. J.* **435**, 339–361 (1994).
A. Burrows, J. Hayes, and B.A. Fryxell, *Astrophys. J.* **450**, 830–850 (1995).
H.-T. Janka, and E. Müller, *Astron. Astrophys.* **306**, 167–198 (1996).
J.M. Blondin, A. Mezzacappa, and C. DeMarino, *Astrophys. J.* **584**, 971–980 (2003).
L. Scheck, H.-T. Janka, T. Foglizzo, and K. Kifonidis, *Astron. Astrophys.*, submitted; arXiv:0704.3001.
M. Rampp, and H.-T. Janka, *Astron. Astrophys.* **396**, 361–392 (2002).
R. Buras, M. Rampp, H.-T. Janka, and K. Kifonidis, *Astron. Astrophys.* **447**, 1049–1092 (2006).
A. Marek, H. Dimmelmeier, H.-T. Janka, E. Müller, and R. Buras, *Astron. Astrophys.* **445**, 273–289 (2006).
M. Liebendörfer, M. Rampp, H.-T. Janka, and A. Mezzacappa, *Astrophys. J.* **620**, 840–860 (2005).
K. Langanke, G. Martínez-Pinedo, J.M. Sampaio, D.J. Dean, W.R. Hix, O.E.B. Messer, A. Mezzacappa, M. Liebendörfer, H.-T. Janka, and M. Rampp, *Phys. Rev. Lett.* **90**, 241102 (2003).
A. Marek, H.-Th. Janka, R. Buras, M. Liebendörfer, and M. Rampp, *Astron. Astrophys.* **443**, 201–210 (2005).
F.S. Kitaura, H.-T. Janka, and W. Hillebrandt, *Astron. Astrophys.* **450**, 345–350 (2006).
K. Nomoto, *Astrophys. J.* **277**, 791–805 (1984).
R. Mayle, and J.R. Wilson, *Astrophys. J.* **334**, 909–926 (1988).
R. Buras, H.-T. Janka, M. Rampp, K. Kifonidis, *Astron. Astrophys.* **457**, 281–308 (2006).
H.-T. Janka, K. Kifonidis, and M. Rampp, “Supernova Explosions and Neutron Star Formation”, in *Physics of Neutron Star Interiors*, edited by D. Blaschke, N.K. Glendenning, and A. Sedrakian, Springer, Berlin, 2001, pp. 333–363; astro-ph/0103015.
S.E. Woosley, and T.A. Weaver, *Astrophys. J. Suppl.* **101**, 181–235 (1995).
A. Burrows, E. Livne, L. Dessart, C.D. Ott, and J. Murphy, *Astrophys. J.* **640**, 878–890 (2006).
A. Burrows, E. Livne, L. Dessart, C.D. Ott, and J. Murphy, *Astrophys. J.* **655**, 416–433 (2007).
L. Scheck, T. Plewa, H.-T. Janka, K. Kifonidis, and E. Müller, *Phys. Rev. Lett.* **92**, 011103 (2004).
L. Scheck, K. Kifonidis, H.-T. Janka, and E. Müller, *Astron. Astrophys.* **457**, 963–986 (2006).
K. Kifonidis, T. Plewa, L. Scheck, H.-T. Janka, and E. Müller, *Astron. Astrophys.* **453**, 661–678 (2006).
A. Burrows, L. Dessart, E. Livne, C.D. Ott, and J. Murphy, *Astrophys. J.*, submitted; astro-ph/0702539.
[^1]: Because of the considerable CPU-time requirements of 2D simulations with our sophisticated, energy-dependent neutrino transport, we could not yet carry the comparative runs of non-rotating models to the very late post-bounce time reached in the case presented here.
[^2]: It is true that in our simulations a few radial zones in the central $\sim$1.5$\,$km of the star are treated in spherical symmetry to get around the most severe CFL constraint for the hydrodynamic timestep. This small central region within a protoneutron star of radius 15–50$\,$km, however, resembles a pinhead in the middle of a cup filled with sloshing tea.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Wesley D. Sacher$^{1*}$'
- 'William M. J. Green$^{2*}$'
- 'Solomon Assefa$^{2}$'
- 'Tymon Barwicz$^{2}$'
- 'Huapu Pan$^{2}$'
- 'Steven M. Shank$^{3}$'
- 'Yurii A. Vlasov$^{2}$'
- 'Joyce K. S. Poon$^{1}$'
bibliography:
- 'references.bib'
title: Breaking the cavity linewidth limit of resonant optical modulators
---
[^1] [^2]
**Microring optical modulators are being explored extensively for energy-efficient photonic communication networks in future high-performance computing systems and microprocessors [@CoteusIBM2011; @CunninghamJSTQE2011], because they can significantly reduce the power consumption of optical transmitters via the resonant circulation of light [@XuNAT2005; @ZortmanCLEO2010; @DongOL2009]. However, resonant modulators have traditionally suffered from a trade-off between their power consumption and maximum operation bit rate, which were thought to depend oppositely upon the cavity linewidth [@GheormaJLT2002; @RabieiJLT2002; @MillerIEEEProc2009]. Here, we break this linewidth limitation using a silicon microring. By controlling the rate at which light enters and exits the microring, we demonstrate modulation free of the parasitic cavity linewidth limitations at up to 40 GHz, more than $6\times$ the cavity linewidth. The device operated at 28 Gb/s using single-ended drive signals less than 1.5 V. The results show that high-$Q$ resonant modulators can be designed to be simultaneously low-power and high-speed, features which are mutually incompatible in typical resonant modulators studied to date.**
![Schematics, image, and tunable transmission spectra of the microring modulator. (a) Schematic of an intracavity modulated microring. (b) Schematic of a coupling modulated microring that uses a $2 \times 2$ MZI-coupler as marked by the box. (c) Optical microscope image of the fabricated SOI microring with the $2 \times 2$ MZI-coupler marked by the box. Measured transmission spectra for (d) tuning the coupling coefficient at a fixed resonance and (e) tuning the resonance wavelength with a fixed coupling coefficient. Independent tuning of the coupling and resonance wavelength using the thermal tuners was achieved.[]{data-label="fig:microring"}](fig1_v4.eps){width="89"}
A microring optical modulator, in its simplest form, consists of a closed waveguide loop coupled to an input/output waveguide. A constant light beam is the input, and a physical parameter of the microring is modulated to produce a time-varying optical output. Two distinct operation modes are intracavity and coupling modulation. The vast majority of microring modulators to date uses intracavity modulation (Fig. \[fig:microring\]a), where the circulating optical field is modulated by the intracavity round-trip phase, $\phi(t)$, and/or loss, $a(t)$, while the coupler parameters are fixed [@RabieiJLT2002; @XuNAT2005; @ZortmanCLEO2010; @DongOL2009]. Because the intracavity optical field amplitude rises and falls at a time-scale set by the photon cavity lifetime, the maximum intracavity modulation bandwidth diminishes with increasing $Q$ [@SacherOE2008]. Additionally, complete on/off modulation (0-100% transmission) requires the stored intracavity optical energy be completely charged and depleted in each switching cycle. Thus, whether in the small- or large- output signal regime, the intracavity modulation bandwidth is inherently limited by the cavity linewidth.
Coupling modulation circumvents this linewidth limitation. In coupling modulation, the intracavity parameters, $\phi$ and $a$, are kept constant, while the through- and cross-coupling coefficients, $\sigma(t)$ and $\kappa(t)$ respectively, are modulated (Fig. \[fig:microring\]b) [@SacherOE2008; @SacherJLT2009]. We term the regime where the modulation rate is greater than the cavity linewidth as “non-adiabatic coupling modulation” [@SacherJLT2009]. “Adiabatic coupling modulation” refers to modulation rates less than the linewidth, when the intracavity optical field in its entirety follows the modulation signal [@YarivPTL2002; @GreenOE2005; @ZhouOE2007; @GillJSTQE2010]. Distinct from the adiabatic regime, intracavity modulation, and $Q$-switching [@DongOL2009b], non-adiabatic coupling modulation does not completely dump the stored intracavity optical energy to generate near 0-100% transmission swings. Instead, it extracts, in the transient, minor fractions of the large intracavity optical field in a high finesse cavity to produce output optical pulses with peak powers that can equal the input optical power. The coupler gates the intracavity optical field as it exits the microring to enable the bandwidth of non-adiabatic coupling modulation to exceed the cavity linewidth [@SacherJLT2009]. Non-adiabatic coupling modulation is a resonantly enhanced phenomenon, since the required changes to the coupling coefficients, and hence the device power consumption, reduce as the stored intracavity optical energy increases [@SacherJLT2009].
In this work, we demonstrate and investigate non-adiabatic coupling modulation using a silicon microring incorporating a $2 \times 2$ Mach-Zehnder interferometer (MZI) as a coupler [@YarivPTL2002; @GreenOE2005; @ZhouOE2007; @GillJSTQE2010], as illustrated in Fig. \[fig:microring\](b). The MZI-coupler provides independent control of the coupling coefficient and resonance wavelength. The differential phase-shift between the MZI-coupler arms changes $\kappa(t)$ and $\sigma(t)$ [@YarivPTL2002; @GreenOE2005], while the common-mode phase-shift changes the round-trip optical path length. Figure \[fig:microring\](c) shows the microring fabricated on a silicon-on-insulator (SOI) substrate using the IBM Silicon CMOS Integrated Nanophotonics process [@AssefaOFC2011]. The MZI had 3 dB directional couplers, 50 $\mathrm{\mu m}$ long thermal tuners, and matched 200 $\mathrm{\mu m}$ long PN diode phase-shifters for push-pull modulation. An identical PN diode phase-shifter was included inside the microring to facilitate direct comparisons between intracavity and coupling modulation with the same device. A MZI with identical PN diode phase-shifters was fabricated as a reference to extract the electro-optic characteristics of the MZI-coupler independent of the microring (see Supplementary Information). The static transmission spectra of the microring in Fig. \[fig:microring\](d)-(e) demonstrate the precise and independent tuning of the coupling coefficient and resonance wavelength achieved by the thermal tuners. An extinction ratio near 30 dB was reached at critical coupling. With no voltage applied to the PN diode phase-shifters, the cavity linewidth at critical coupling was $\Delta \nu \approx 7$ GHz, corresponding to a loaded $Q$ of about 28000 and a finesse of $14$.
![Optical small-signal modulation responses of coupling and intracavity modulation. Each curve is obtained by normalizing the electro-optic $S_{21}$ of the microring to the $S_{21}$ of the reference MZI and referencing to the value at 100 MHz. The microring was biased near critical coupling, with a cavity linewidth $\Delta \nu \approx 6$ GHz. The intracavity modulation response for a $\sim 1.3$ GHz detuning from resonance (blue) has a 3 dB bandwidth of 4.4 GHz, similar to the linewidth. A $\sim 5$ GHz detuning produces a resonant sideband peak near the value of the detuning (red), and the 3 dB bandwidth is extended to $\sim 13$ GHz. The coupling modulation response (black) does not roll-off to 40 GHz (more than $6\times$ the linewidth). []{data-label="fig:smallsig"}](S21_new_v2b.eps){width="89"}
The small-signal optical modulation characteristics in Fig. \[fig:smallsig\] demonstrate that the coupling modulation bandwidth significantly exceeds the traditional cavity linewidth limit. In the figure, the resonant optical modulation response was isolated from the electrical characteristics of the measurement setup and PN diode phase-shifters by normalizing the electro-optic $S_{21}$ parameter of the microring modulator to the $S_{21}$ of the reference MZI (details in the Supplementary Information). The phase-shifters were operated in depletion with a DC bias of -1 V. Under bias, the device was near critical coupling and had a cavity linewidth of $\Delta \nu \approx 6$ GHz. The input laser wavelength was on resonance for coupling modulation and was detuned slightly off-resonance for intracavity modulation to obtain an appreciable modulation depth. The black curve in Fig. \[fig:smallsig\] shows that in contrast to intracavity modulation, the coupling modulation response did not roll off to 40 GHz, more than $6\times$ the cavity linewidth. The maximum frequency measured here was limited by the instrumentation. The flat response indicates that the frequency roll-off characteristics of the non-adiabatic coupling modulation resembled those of the non-resonant reference MZI. The slight decrease in the modulation depth near the frequency corresponding to the cavity linewidth was due to a slight under-coupling of the microring [@SacherOE2008].
In comparison, the intracavity modulation response with input light that is 1.3 GHz detuned from resonance (blue curve) had a 3 dB bandwidth of about 4.4 GHz, confirming that the intrinsic modulation bandwidth was limited by the cavity linewidth. For larger detunings, a peak occurs when a modulation sideband is on resonance and becomes comparable to or greater than the carrier in amplitude within the microring [@SacherOE2008; @GillJSTQE2010]. The peak is exaggerated for large detunings, because the circulating amplitude of an off-resonant carrier is small. This effect is shown by the $\sim 5$ GHz detuning measurement (red curve). Although the modulation sideband peak extended the 3 dB bandwidth to about 13 GHz, intracavity modulation of a highly detuned carrier is not practical, because the absolute modulation depth and the linearity of the modulator are compromised.
Large-signal data modulation and optical eye diagram measurements provide further evidence that the coupling modulation bandwidth is not similarly limited by the cavity linewidth as intracavity modulation. To reduce the required drive voltage, the PN diode phase-shifters were in forward bias, with a DC offset of 0.28 V, and driven by pre-emphasized non-return-to-zero (NRZ) $2^{31}-1$ pseudo-random bit sequence (PRBS) signals. The electrical pre-emphasis extended the modulation bandwidth of the PN diode phase-shifters far beyond their minority carrier lifetime limit of $\sim 1$ GHz [@XuOE2007; @GreenOE2007; @ManipatruniLEOS2007]. The single-ended voltage swing of the pre-emphasized bits was 1.5 $\mathrm{V}_{pp}$, and the non-emphasized bits were between 0.24-0.3 $\mathrm{V}_{pp}$. Fig. \[fig:eye\](a) (left) shows the eye diagram of the pre-emphasized 28 Gb/s drive signal. The optical output of the reference MZI driven with this signal in push-pull mode (Fig. \[fig:eye\](a), right) shows that the MZI-coupler optical output did not contain remnants of the pre-emphasis which could potentially extend the microring modulation bandwidth beyond the limits of the resonant optical dynamics.
![Eye diagrams of large-signal coupling and intracavity modulation. (a) Eye diagrams of the pre-emphasized electrical drive signals at 28 Gb/s (left) and the resultant optical output of the reference MZI (right). No remnants of the pre-emphasis are present in the optical output. (b) Coupling (left) and intracavity (right) modulation eye diagrams at 6-28 Gb/s for bias points near critical coupling ($\Delta \nu \approx 6-7$ GHz). The coupling modulation eye is open at 28 Gb/s, but the intracavity modulation eye closes at bit rates greater than roughly $2\times$ the linewidth. (c) Intracavity modulation eye diagrams of an over-coupled microring ($\Delta \nu \approx 9$ GHz). The eye opening is larger than in (b), confirming that the intracavity modulation bandwidth depends on the cavity linewidth.[]{data-label="fig:eye"}](fig3.eps){width="87"}
Figure \[fig:eye\](b) summarizes the coupling and intracavity modulation eye diagrams of the microring at bit rates between 6 and 28 Gb/s. The cavity linewidths at the operating bias points were $\Delta \nu \approx 6-7$ GHz. At each bit rate, identical drive signals were applied to the coupler or intracavity phase-shifters, except the MZI-coupler was driven in push-pull while the intracavity phase-shifter was driven single-ended. For coupling modulation, the microring was modulated between critical coupling and under-coupling with the input light on resonance. For intracavity modulation, the microring was biased at critical coupling and the input wavelength was slightly detuned from resonance. The pre-emphasis ratio and detuning were optimized to maximize the eye opening for each case. At 6 and 12.5 Gb/s, both the coupling and intracavity modulation eye diagrams had extinction ratios of 10-13 dB and a maximum optical transmission $> 40\%$. As the bit rate increased, the coupling modulation eye remained wide open up to 28 Gb/s, whereas the intracavity modulation eye was closed at 22 Gb/s. However, because of the modest finesse and PN diode phase-shifter efficiency, the extinction ratio of coupling modulation at 22 and 28 Gb/s decreased to 10 dB, and the maximum optical transmission was only about 10 to 20% of the off-resonance transmission.
To check that the intracavity modulation eye closure was due to the cavity linewidth and not to an electrical artifact, we over-coupled the microring to increase its linewidth to 9 GHz at the expense of modulator efficiency and extinction ratio. Fig. \[fig:eye\](c) shows that the eye opening increased at 22 Gb/s, but remained closed at 28 Gb/s. The eye diagrams show that intracavity modulation suffered from severe inter-symbol interference (ISI) at bit rates greater than roughly $2\times$ the cavity linewidth (i.e. $> 12.5$ Gb/s), while coupling modulation at up to $4\times$ the linewidth was not similarly affected. Coupling modulation should function well beyond 28 Gb/s; however, the measurements were limited by the instrumentation. The results in Fig. \[fig:smallsig\] and \[fig:eye\] demonstrate that the long-held cavity linewidth limit to the intracavity modulation bandwidth can be broken with coupling modulation.
A potential drawback to coupling modulation is the ISI originating from the *low* frequency content of the drive signal, which can significantly deplete the stored optical energy in the cavity. Fortuitously, this ISI can be mitigated. One suggestion is to modulate two output couplers to maintain a constant intracavity optical power [@PopovicIPR2010] at the expense of device complexity, cavity finesse, and power efficiency for large-signal modulation. Here, we propose a more direct approach of encoding the electrical data to produce a DC-balanced drive signal. An example is the 8b/10b code, a typical line code for the Ethernet and InfiniBand communication standards. To illustrate the effect of the ISI and encoding, Fig. \[fig:scaling\](a) compares a series of computed eye diagrams. The simulations use a NRZ PRBS $2^{17}-1$ pattern, a 5 GHz cavity linewidth, and a 250 $\mathrm{\mu m}$ round-trip length. The top row shows that uncoded intracavity modulation requires a linewidth that is at least half the bit rate, in agreement with our experimental results. The middle row shows that for uncoded coupling modulation in the non-adiabatic regime, the transmission swing in the eye opening can be about 25%. Larger drive signals would increase the ISI. As in our experiment, the resonator is driven between under-coupling and critical coupling. The bottom row shows that with an 8b/10b encoded drive signal, non-adiabatic coupling modulation with an eye opening of about 90% and low ISI are possible at 100 Gb/s, characteristics that cannot be attained by intracavity modulation. Significantly, the ISI in coupling modulation diminishes as the bit rate increases, since the low frequency content of the modulation signal is reduced.
![ISI and efficiency scaling of coupling and intracavity modulation. The calculations assume a group index of 4.3, a NRZ PRBS $2^{17}-1$ data signal, a resonant input for coupling modulation, and critical coupling for intracavity modulation. (a) Computed eye diagrams at several bit rates assuming $\Delta\nu = 5$ GHz and a round-trip length of 250 $\mathrm{\mu m}$ for (top) intracavity modulation and (center) coupling modulation driven by an uncoded NRZ signal, and (bottom) coupling modulation driven by a NRZ 8b/10b encoded signal. With DC-balanced encoding, coupling modulation can achieve a 0-90% swing at 100 Gb/s. (b) The coupling modulation efficiency, $\eta_c$, versus microring waveguide loss and cavity linewidth for several round-trip lengths, $L$. The calculations assume a 8b/10b encoded drive signal and a 0-90% output swing. The intracavity efficiency, $\eta_i$, of a 5 $\mathrm{\mu m}$ radius microring with the same output swing at 40 Gb/s and 100 Gb/s, using linewidths of 20 GHz and 50 GHz respectively, are marked for comparison. Coupling modulation becomes increasingly efficient over intracavity modulation as the $Q$ factor and bit rate increase. []{data-label="fig:scaling"}](fig4.eps){width="89"}
Although our results demonstrate that the coupling modulation bandwidth can be substantially larger than the intracavity modulation bandwidth, an essential question remains of whether coupling modulation of a narrow linewidth resonator can be more power efficient than intracavity modulation of a small resonator with a broad linewidth. Here, we find the general efficiency scaling relationships to show that coupling modulation can indeed be more efficient, particularly for high bit rate and large-signal modulation. Thus, coupling modulation can be suitable for energy-efficient large capacity optical links.
We define an efficiency metric, $\eta = \Delta \phi_{MZI}/\Delta \phi_{ring}$, where $\Delta \phi_{ring}$ and $\Delta \phi_{MZI}$ are respectively the phase-shifts of a microring and a MZI biased at quadrature required to produce the same output transmission swing, assuming identical phase-shifters. $\Delta \phi$ is single-ended for intracavity modulation and is applied push-pull as $\pm \Delta \phi/2$ for coupling modulation. Referencing to $\Delta \phi_{MZI}$ allows for a general comparison between coupling and intracavity modulation independent of material platforms. The phase-shifts are related to the power consumption of the device by the physical mechanism of electro-optic modulation and any drive circuitry.
For intracavity modulation, from the microring transmission function [@YarivPTL2002], we find the efficiency, $\eta_i$, is roughly proportional to the *intracavity power* or finesse, $F$: $$\label{eqn:etai}
\eta_i \approx k_i F,$$ where $k_i \lesssim 0.42$ depends on the high and low transmission levels as well as the ratio of the round-trip loss to the coupling. As examples, at critical coupling and $F \gtrsim 5$, $k_i = 0.24$ for a 0-90% output swing and $k_i = 0.41$ for a 20-30% swing. $k_i$ is lower for large-signal modulation than in the small-signal regime because the microring transmission spectrum flattens at wavelengths detuned from the resonance. In contrast, the efficiency of coupling modulation, $\eta_c$, scales with the *intracavity field*. For a MZI-coupler and a resonant optical input, $$\label{eqn:etac}
\eta_c \approx k_c\sqrt{F},$$ where $k_c \lesssim 1$ depends on the transmission levels and $F$ is the finesse at critical coupling. For the same transmission levels, $k_c$ is lower in the non-adiabatic than adiabatic regime due to a reduced intracavity optical field amplitude. In the adiabatic regime and for $F \gtrsim 5$, we find $k_c = 0.85$ for a 0-90% output swing and $k_c = 0.73$ for a 20-30% swing. In the non-adiabatic regime and $F \gtrsim 20$, $k_c = 0.41$ for a 0-90% output swing. It follows from Eq. \[eqn:etai\] and \[eqn:etac\] that adiabatic coupling modulation is more efficient than intracavity modulation in cavities with the same $F$ and phase-shifters when $F \lesssim (k_c/k_i)^2$. While intracavity modulation is more efficient in the small-signal regime if the resonator has at least a moderate finesse (e.g. 20-30% swing, $F \gtrsim 5$), adiabatic coupling modulation is more efficient for large-signal modulation as $k_i$ diminishes. For example, $\eta_c > \eta_i$ for a 0-90% output swing when $F \lesssim 10$ and for a 0-99% swing when $F \lesssim 87$.
The efficiency scaling becomes especially favourable to non-adiabatic coupling modulation over intracavity modulation at high $Q$ factors and high bit rates. Because the minimum cavity linewidth is dictated by the desired modulation rate, improvements in $\eta_i$ of intracavity modulation via the finesse must come from the cavity size reduction. However, a small cavity size prohibits the inclusion of tuning structures needed to attain large extinction ratios and large-signal swings in practice. Thus, the key advantage of coupling modulated microrings is that they can be kept larger (to accommodate tunable couplers), while $\eta_c$ can, in principle, be arbitrarily boosted by increasing $Q$ to raise the finesse. Figure \[fig:scaling\](b) shows $\eta_c$ versus the cavity linewidth for several round-trip lengths assuming a 0-90% output swing. The values of $\eta_i$ for SOI microrings with a 5 $\mathrm{\mu m}$ radius and the same output swings at 40 Gb/s and 100 Gb/s are marked. As the round-trip length of a coupling-modulated microring increases from 50 $\mathrm{\mu m}$ to 500 $\mathrm{\mu m}$, a narrower linewidth is required for $\eta_c > \eta_i$.
By combining the benefits of resonant enhancement with the large bandwidths of non-resonant modulators, coupling modulation, for the first time, opens the avenue toward ultra-low power yet high-speed modulation of ultra-high-$Q$ resonators. Such resonators on silicon chips can possess finesses exceeding 10000 [@LeeNP2012].
Methods
=======
The optical modulators were fabricated using the IBM Silicon CMOS Integrated Nanophotonics process [@AssefaOFC2011] on a 200 mm-diameter SOI wafer with a 2 $\mathrm{\mu m}$-thick buried-oxide layer and a 220 nm-thick top silicon layer. Fully-etched silicon access waveguides and partially-etched PN diode rib waveguides were defined and planarized with silicon dioxide through a shallow trench isolation module. Typical CMOS ion implantation conditions were used to form a lateral PN diode junction at the center of each phase-shifter. The diode junction was designed with a nominal carrier concentration of $5\times 10^{17}$ cm$^{-3}$ in the P- and N-type regions. After a rapid thermal activation anneal, silicide ohmic contacts to the phase-shifters were formed. The silicide also formed the resistive thermal tuners for the MZI bias control. Finally, tungsten vias and copper metal interconnects were formed to electrically contact the phase-shifters and thermal heaters. Dies were prepared with cleaved facets for on-/off-chip optical coupling using tapered optical fibers. The measurements were taken with a swept wavelength tunable laser (Agilent 81682A). A 67 GHz vector network analyzer (VNA, Agilent E8361A) was used for the measurements in Fig. \[fig:smallsig\]. For intracavity modulation, a single-ended signal was applied to the intracavity phase-shifter. For coupling modulation, the MZI phase-shifters were driven in push-pull. To generate a differential drive signal, the VNA output was fed into a fanout circuit (Hittite HMC842LC4B), which had a bandwidth of about 32 GHz. Custom 40 GHz RF probes (GGB Industries) contacted the devices. The output was detected with a 40 GHz InGaAs photoreceiver (Archcom Technology).
To generate the drive signals for the data modulation experiments of Fig. \[fig:eye\], the output of a PRBS generator (Centellax PPG12500, TG1P4A) was fed into a pre-emphasis converter (Anritsu MP1825B-002), which operated up to 28 Gb/s. The pre-emphasized signals were directly fed to the RF probes. The optical output of the modulator was amplified using an erbium doped fiber amplifier, bandpass filtered (full width at half maximum bandwidth of 0.8 nm), and captured on a digital communications analyzer with a 28 Gb/s optical module (Agilent 86100C, 86106B). All eye patterns were obtained using $2^{31}-1$ PRBS and single-tap pre-emphasis. Analytic forms for $k_i$ and $k_c$ in Eq. \[eqn:etai\] and \[eqn:etac\] in the adiabatic regime were derived from the static transmission [@YarivPTL2002]. An analytic form for $k_c$ in the non-adiabatic regime was derived by assuming a periodic square-wave drive signal and solving for the average intracavity field with a rate equation. This approximation neglects the exact shape and low-frequency content of the drive signal. Figure \[fig:scaling\] was calculated using the microring modulator model in [@SacherOE2008; @SacherJLT2009].
Acknowledgments
===============
W.D.S. and J.K.S.P. thank the University of Toronto Emerging Communications Technology Institute for access to the VNA and RF signal generators. W.M.J.G. and W.D.S. thank D. M. Gill at IBM Research for helpful discussions. W.D.S. and J.K.S.P. are grateful for the financial support of the Natural Sciences and Engineering Research Council of Canada.
Supplementary information
=========================
This section provides details about the reference Mach-Zehnder interferometer (MZI) and the small-signal modulation response ($S_{21}$) measurements. The $S_{21}$ normalization procedure used in Fig. 2 is also discussed.
Reference MZI
-------------
To isolate the modulation responses of intracavity and coupling modulation from the electrical characteristics of the PN diode phase-shifters and experimental setup, a reference MZI device was fabricated and measured. Specifically, measurements of the reference MZI were used to normalize the $S_{21}$ measurements in Fig. 2 and to ensure the pre-emphasis in the pseudo-random bit sequence (PRBS) drive signals only equalized the roll-off due to the modulation bandwidth limitations of the PN diode phase-shifters and measurement setup in Fig. 3(b). The reference MZI and microring modulator measured in this work are shown in Fig. \[fig:microscopePics\]. The reference MZI is nominally identical to the output coupler of the microring (i.e., designed to have identical waveguides, PN diode phase-shifters, thermal tuners, and wiring). The reference MZI and microring modulator were on the same die and separated by about 620 $\mathrm{\mu}$m.
![Optical microscope images of (a) the reference MZI and (b) the microring modulator.[]{data-label="fig:microscopePics"}](Fig1_v2.eps){width="87"}
$S_{21}$ measurements and normalization
---------------------------------------
Push-pull reference MZI and coupling modulation electro-optical $S_{21}$ measurements were performed using the experimental setup illustrated in Fig. \[fig:setup\](a). The $S_{21}$ parameter was collected from the voltage of the photodetector referenced to the vector network analyzer (VNA) output. A high-speed fanout circuit split the single-ended RF output of the VNA into push-pull drive signals for the microring coupler or reference MZI. Single-arm drive reference MZI and intracavity modulation $S_{21}$ measurements were performed using the experimental setup shown in Fig. \[fig:setup\](b). Push-pull drive signals were not required for single-arm drive of the reference MZI and intracavity modulation, so the fanout was not included.
![Schematics of the $S_{21}$ measurement setups for (a) coupling modulation and push-pull modulation of the reference MZI and (b) intracavity modulation and single-arm modulation of the reference MZI.[]{data-label="fig:setup"}](Fig2.eps){width="87"}
The reference MZI $S_{21}$ measurements were performed at the quadrature bias point, which was established by applying DC voltages to the thermal tuners in the MZI arms. Experimentally, it was found that the shape of the reference MZI $S_{21}$ curves was insensitive to changes in the MZI bias point. The quadrature bias point was chosen to maximize the modulation efficiency of the MZI, and thus, reduce the relative noise in the $S_{21}$ measurements of the reference MZI. The intracavity modulation $S_{21}$ measurements were taken at critical coupling with roughly 1.3 GHz and 5 GHz bias detunings. The coupling modulation $S_{21}$ measurements were taken with the input wavelength on resonance and a slightly under-coupled bias. In all cases, the PN diode phase-shifters were biased at $-1$ V.
The responses of the RF cables, RF adapters, and bias tees were de-embedded from the $S_{21}$ data; however, the responses of the fanout, RF probes, and on-chip wiring remained embedded in the $S_{21}$ data. This raw $S_{21}$ data is shown in Fig. \[fig:S21\] (i.e., the first two curves in (a) and the first three curves in (b)); each raw $S_{21}$ curve is referenced to its value at a 100 MHz modulation frequency. $S_{21,MZI,push-pull}$ and $S_{21,MZI,single-arm}$ are defined as the raw push-pull and single-arm drive reference MZI responses, respectively. $S_{21,cplng}$ and $S_{21,intracav}$ are defined as the raw responses of the coupling and intracavity modulation, respectively. To isolate the modulation responses of the optical cavity dynamics from those of the PN diode phase-shifters, on-chip wiring, RF probes, and fanout, the following normalization procedure was applied:
\[eqn:norm\] $$S_{21,cplng,norm} = \frac{S_{21,cplng}}{S_{21,MZI,push-pull}},$$ $$S_{21,intracav,norm} = \frac{S_{21,intracav}}{S_{21,MZI,single-arm}}.$$
$S_{21,cplng,norm}$ and $S_{21,intracav,norm}$ are the normalized $S_{21}$ data for coupling and intracavity modulation, respectively, and are presented in Fig. 2 and repeated in Fig. \[fig:S21\]. $S_{21,cplng,norm}$ and $S_{21,intracav,norm}$ represent the modulation responses of the microring due solely to optical cavity dynamics from coupling and intracavity modulation, respectively.
![$S_{21}$ measurements for (a) coupling modulation and push-pull modulation of the reference MZI and (b) intracavity modulation and single-arm modulation of the reference MZI. $S_{21}$ measurements with only the RF cables, RF adapters, and bias tees de-embedded are shown for the microring and reference MZI (i.e., the raw $S_{21,cplng}$, $S_{21,MZI,push-pull}$, $S_{21,intracav}$, and $S_{21,MZI,single-arm}$). Microring $S_{21}$ measurements normalized by the reference MZI response using Eq. \[eqn:norm\] are also shown (i.e., $S_{21,cplng,norm}$ and $S_{21,intracav,norm}$).[]{data-label="fig:S21"}](Fig3_v3.eps){width="87"}
[^1]: **Corresponding authors:** J. K. S. Poon, W. M. J. Green, and W. D. Sacher
[^2]: **Contributions:** W.D.S. performed the measurements. W.D.S., W.M.J.G., and J.K.S.P. conceived and designed experiments and analyzed the data. W.D.S. and J.K.S.P. performed the analytical modeling of the modulators. W.M.J.G., S.A., T.B., S.M.S., and Y.A.V. developed the modulator integration process and qualified the devices with the help of H.P.. W.D.S., W.M.J.G., S.A., Y.A.V. and J.K.S.P. wrote the paper.\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The problem of deciding if a Traveling Salesman Problem (TSP) tour is minimal was proved to be [[coNP]{}]{}–complete by Papadimitriou and Steiglitz. We give an alternative proof based on a polynomial time reduction from [[3SAT]{}]{}. Like the original proof, our reduction also shows that given a graph $G$ and an Hamiltonian path of $G$, it is [[NP]{}–complete]{} to check if $G$ contains an Hamiltonian cycle (Restricted Hamiltonian Cycle problem).'
author:
- |
Marzio De Biasi\
``
bibliography:
- 'tspmindecision.bib'
title: 'Minimal TSP Tour is coNP–Complete'
---
Introduction
============
The *Traveling Salesman Problem* ([TSP]{}) is a well–known problem from graph theory [@PapadimitriouComplexity],[@GJ]: we are given $n$ cities and a nonnegative integer distance $d_{ij}$ between any two cities $i$ and $j$ (assume that the distances are symmetric, i.e. for all $i,j, d_{ij} = d_{ji}$). We are asked to find the *shortest tour* of the cities, that is a permutation $\pi$ of $[1..n]$ such that $\sum_{i=1}^n d_{\pi(i),\pi(i+1)}$ (where $\pi(n+1) = \pi(n)$) is as small as possible. Its decision version is the following:
[[TSPDecision]{}]{}: If a nonnegative integer bound $B$ (the traveling salesman’s “budget”) is given along with the distances, does there exist a tour of all the cities having total length no more than $B$?
[[TSPDecision]{}]{} is [[NP]{}–complete]{} (we assume that the reader is familiar with the theory of [[NP]{}]{}–completeness, for a good introduction see [@GJ] or [@Sipser]). In [@PapadimitriouComplexity] two other problems are introduced:
[[TSPExact]{}]{}: Given the distances $d_{ij}$ among the $n$ cities and an nonnegative integer $B$, is the length of the shortest tour *equal* to $B$; and\
\
[[TSPCost]{}]{}: Given the distances $d_{ij}$ among the $n$ cities calculate the *length* of the shortest tour.
[[TSPExact]{}]{} is [[DP]{}]{}–complete (a language $L$ is in the class [[DP]{}]{} if and only if there are two languages $L_1 \in {{\sf NP}}$ and $L_2 \in {{\sf coNP}}$ and $L = L_1 \cap L_2$); [[TSPCost]{}]{} and [TSP]{} are both [$\sf{FP^{NP}}$]{}–complete ([$\sf{FP^{NP}}$]{} is the class of all functions from strings to strings that can be computed by a polynomial–time Turing machine with a [[SAT]{}]{} oracle) [@PapadimitriouComplexity].
Recently a post by Jean Francois Puget: “[No, The TSP Isn’t NP Complete](https://www.ibm.com/developerworks/community/blogs/jfp/entry/no_the_tsp_isn_t_np_complete)” and the subsequent reply by Lance Fortnow: “[Is Traveling Salesman NP-Complete?](http://blog.computationalcomplexity.org/2014/01/is-traveling-salesman-np-complete.html)” [@blog:tspmindecision] (re–)raised the question of the correct interpretation of the statement “TSP is [[NP]{}–complete]{}”; indeed, if we are given a tour, checking that it is the shortest tour seems not to be in [[NP]{}]{}. A question about the complexity of the following problem:
[[TSPMinDecision]{}]{}: Given a set of $n$ cities, the distance between all city pairs and a tour $T$, is T visiting each city exactly once and is T of minimal length?
was posted on , a question and answer site for professional researchers in theoretical computer science and related fields [@cstheory:tspmindecision].
We gave an answer with a first sketch of the proof that [[TSPMinDecision]{}]{} is [[coNP]{}]{}–complete, but after formalising and publishing it on arXiv, [we discovered that the result is not new](http://cstheory.stackexchange.com/a/21644/3247) and it originally appeared in [@papasteitsp] (see also Section 19.9 in [@combopt]). The proof given by Papadimitriou and Steiglitz is different: they prove that the Restricted Hamiltonian Cycle (RHC) problem is [[NP]{}–complete]{} starting from an instance of the Hamiltonian cycle problem $G$ and modifying $G$ into a new graph $G'$ that contains an Hamiltonian path, and has an Hamiltonian cycle if and only if the original $G$ has an Hamiltonian cycle. Our alternative proof is a chain of reductions from [[3SAT]{}]{} to the problem of finding a tour shorter than a given one, and it may be interesting in and of itself, so we decided not to withdraw the paper.
Minimal TSP tour is coNP–complete {#sec:proof}
=================================
Proving that [[TSPMinDecision]{}]{} is [[coNP]{}]{}–complete is equivalent to proving the [[NP]{}]{}–completeness of the following:
\
[ **Instance**: A complete graph $G = (V, E)$ with positive integer distances $d_{ij}$ between the nodes, and a simple cycle $C$ that visits all the nodes of $G$. ]{}
[**Question**: Is there a simple cycle $D$ that visits all the nodes of $G$ such the total length of the tour $D$ in $G$ is strictly less than the total of the tour $C$ in $G$? ]{}
\[thm:tspanothertour\] [[TSPAnotherTour]{}]{} is [[NP]{}–complete]{}.
It is easy to see that a valid solution to the problem can be verified in polynomial time: just check if the tour $D$ visits all the cities and if its length is strictly less than the length of the given tour $C$, so the problem is in [[NP]{}]{}. To prove its hardness we give a polynomial time reduction from [[3SAT]{}]{}; given a [[3CNF]{}]{} formula $\varphi$ with $n$ variables $x_1,\ldots,x_n$ and $m$ clauses $C_1,...,C_m$; we introduce a new dummy variable $z$ and add it to every clause: $(x_{i_1} \lor x_{i_2} \lor x_{i_3} \lor z)$. We obtain a [[4CNF]{}]{} formula $\varphi^z$ that has at least one satisfying assignment (just set $u=true$). Note that every satisfying assignment of $\varphi^z$ in which $z =false$ is also a satisfying assignment of $\varphi$.
From $\varphi^z$ we generate an undirected graph $G = \{V,E\}$ following the same standard transformation used to prove that the Hamiltonian cycle problem is [[NP]{}–complete]{}: for every clause we add a node $c_j$, for every variable $x_i$ we add a *diamond–like* component, and we add a directed edge from one of the nodes of the diamond to the node $c_j$ if $x_i$ appears in $C_j$ as a positive literal; a directed edge from $c_j$ to one of the nodes of the diamond if $x_i$ appears in $C_j$ as a negative literal. Starting from the top we can choose to traverse the diamonds corresponding to variables $x_1,x_2,...,x_n,u$ from left to right (i.e. set $x_i$ to $true$) or from right to left (i.e. set $x_i$ to $false$). The resulting directed graph $G$ has an Hamiltonian cycle if and only if the original formula is satisfiable. For the details of the construction see [@Sipser] or [@AroraBarak].
We focus on the diamond corresponding to the dummy variable $z$; let $e_z$ be the edge that must be traversed if we assign to $u$ the value of $true$ (see Figure \[fig:reduction\]).
![Reduction from [[3SAT]{}]{} to directed Hamiltonian cycle.[]{data-label="fig:reduction"}](figreduction.pdf){width="7cm"}
We can transform $G$ to an undirected graph $G' = \{V',E'\}$ replacing each node $u \in V$ with three linked nodes $u_1, u_2, u_3 \in V'$ and modify the edges according to the standard reduction used to prove the [[NP]{}]{}-completeness of UNDIRECTED HAMILTONIAN CYCLE from DIRECTED HAMILTONIAN CYCLE [@Sipser]: we use $u_1$ for the incoming edges of $u$, and $u_3$ for the outgoing edges, i.e. we replace every directed edge $(u \to v) \in E$ with $(u_3 \to v_1) \in E'$. We have $G'$ has an Hamiltonian cycle if and only if $G$ has an Hamiltonian cycle if and only if $\varphi^z$ is satisfiable.
Finally we transform $G'$ into an instance of [[TSPAnotherTour]{}]{} assigning length $1$ to all edges except edge $e_z$ which has length $2$; and we complete the graph adding the missing edges and setting their length to $3$.
The dummy variable $z$ guarantees that we can easily find a tour $T$: just travel the diamonds from left to right without worrying of the clause nodes; when we reach the diamond corresponding to $z$, traverse it from left to right (i.e. assign to $z$ the value of $true$), and include all the $c_j$s. By construction the total length of the tour $T$ is exactly $|V'|+1$: all edges have length 1 except $e_u$ that has length $2$.
Another tour $D$ can have a length strictly less than $|V'|+1$ only if it doesn’t use the edge $e_u$; so if it exists we can derive a valid satisfying assignment for the original formula $\varphi$, indeed by construction $\varphi$ is satisfiable if and only if there exists a satisfying assignment for $\varphi^z$ in which $z=false$. In the opposite direction if there exists a valid satisfying assignment for $\varphi$ we can easily find a tour $D$ of length $|V'|$: just traverse the diamonds according to the truth values of the variables $x_i$ and traverse the diamond corresponding to $z$ from right to left.
So there is another tour $D$ of total length strictly less than $T$ if and only if the original [[3SAT]{}]{} formula $\varphi$ is satisfiable.
Hence we have:
[[TSPMinDecision]{}]{} is [[coNP]{}]{}–complete.
The reduction used to prove Theorem \[thm:tspanothertour\] “embeds” the $\sf{NP}$–completeness proof of the *Restricted Hamiltonian Cycle problem* (RHC) [@combopt]:
\[cor:ham\] Given a graph $G$ and an Hamiltonian path in it, it is [[NP]{}–complete]{} to decide if $G$ contains an Hamiltonian cycle as well.
In the reduction above, after the creation of the undirected graph $G'$, if we remove the edge $e_z$, we are sure that an Hamiltonian path exists from one endpoint of $e_z$ to the other (just delete $e_z$ from the Hamiltonian cycle that can be constructed setting $z = true$). An Hamiltonian cycle in $E \setminus \{e_z\}$ *must* use the edge corresponding to $z = false$, so it exists if and only if the original [[3SAT]{}]{} formula $\varphi$ is satisfiable.
Conclusion
==========
We are optimist: if someone – out there – shouts: “TSP is NP–complete” we are confident that he really means: “The decision version of TSP is NP–complete”; and we hope that, soon or later, someone – out there – will shout “We already know that there is \[not\] a polynomial time algorithm that solves TSP because $\sf{P}$ is \[not\] equal to $\sf{NP}$” :-)
Acknowledgements {#acknowledgements .unnumbered}
================
Thanks to Pálvölgyi Dömötör for the nice hint about Theorem \[cor:ham\], and to Marcus Ritt for pointing out the original Papadimitriou and Steiglitz’s paper.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We implement a decision procedure for answering questions about a class of infinite words that might be called (for lack of a better name) “Fibonacci-automatic”. This class includes, for example, the famous Fibonacci word ${\bf f} = 01001010\cdots$, the fixed point of the morphism $0 \rightarrow 01$ and $1 \rightarrow 0$. We then recover many results about the Fibonacci word from the literature (and improve some of them), such as assertions about the occurrences in $\bf f$ of squares, cubes, palindromes, and so forth. As an application of our method we prove a new result: there exists an aperiodic infinite binary word avoiding the pattern $x x x^R$. This is the first avoidability result concerning a nonuniform morphism proven purely mechanically.'
author:
- 'Chen Fei Du[^1], Hamoon Mousavi$^1$, Luke Schaeffer[^2], and Jeffrey Shallit$^1$'
title: 'Decision Algorithms for Fibonacci-Automatic Words, with Applications to Pattern Avoidance'
---
\[theorem\][Corollary]{} \[theorem\][(Almost) Theorem]{} \[theorem\][Lemma]{} \[theorem\][Proposition]{}
\[theorem\][Definition]{} \[theorem\][Example]{} \[theorem\][Conjecture]{} \[theorem\][Open Problem]{} \[theorem\][Procedure]{}
\[theorem\][Remark]{}
Decidability {#decide}
============
As is well-known, the logical theory $\operatorname{Th}(\Enn,+)$, sometimes called Presburger arithmetic, is decidable [@Presburger:1929; @Presburger:1991]. Büchi [@Buchi:1960] showed that if we add the function $V_k(n) = k^e$, for some fixed integer $k \geq 2$, where $e = \max \{ i \ : \ k^i \, | \, n \}$, then the resulting theory is still decidable. This theory is powerful enough to define finite automata; for a survey, see .
As a consequence, we have the following theorem (see, e.g., [@Shallit:2013]):
There is an algorithm that, given a proposition phrased using only the universal and existential quantifiers, indexing into one or more $k$-automatic sequences, addition, subtraction, logical operations, and comparisons, will decide the truth of that proposition. \[one\]
Here, by a $k$-automatic sequence, we mean a sequence $\bf a$ computed by deterministic finite automaton with output (DFAO) $M = (Q, \Sigma_k, \Delta, \delta, q_0, \kappa) $. Here $\Sigma_k := \lbrace 0,1,\ldots, k-1 \rbrace$ is the input alphabet, $\Delta$ is the output alphabet, and outputs are associated with the states given by the map $\kappa:Q \rightarrow \Delta$ in the following manner: if $(n)_k$ denotes the canonical expansion of $n$ in base $k$, then ${\bf a}[n] = \kappa(\delta(q_0, (n)_k))$. The prototypical example of an automatic sequence is the Thue-Morse sequence ${\bf t} = t_0 t_1 t_2 \cdots$, the fixed point (starting with $0$) of the morphism $0 \rightarrow 01$, $1 \rightarrow 10$.
It turns out that many results in the literature about properties of automatic sequences, for which some had only long and involved proofs, can be proved purely mechanically using a decision procedure. It suffices to express the property as an appropriate logical predicate, convert the predicate into an automaton accepting representations of integers for which the predicate is true, and examine the automaton. See, for example, the recent papers . Furthermore, in many cases we can explicitly enumerate various aspects of such sequences, such as subword complexity .
Beyond base $k$, more exotic numeration systems are known, and one can define automata taking representations in these systems as input. It turns out that in the so-called Pisot numeration systems, addition is computable , and hence a theorem analogous to Theorem \[one\] holds for these systems. See, for example, . It is our contention that the power of this approach has not been widely appreciated, and that many results, previously proved using long and involved ad hoc techniques, can be proved with much less effort by phrasing them as logical predicates and employing a decision procedure. Furthermore, many enumeration questions can be solved with a similar approach.
We have implemented a decision algorithm for one such system; namely, Fibonacci representation. In this paper we report on our results obtained using this implementation. We have reproved many results in the literature purely mechanically, as well as obtained new results, using this implementation.
The paper is organized as follows. In Section \[fibrep\], we briefly recall the details of Fibonacci representation. In Section \[proofsf\] we report on our mechanical proofs of properties of the infinite Fibonacci word; we reprove many old results and we prove some new ones. In Section \[finitefib\] we apply our ideas to prove results about the finite Fibonacci words. In Section \[rotefib\] we study a special infinite word, the Rote-Fibonacci word, and prove many properties of it, including a new avoidability result. In Section \[other\] we look briefly at another sequence, the Fibonacci analogue of the Thue-Morse sequence. In Section \[additive\] we apply our methods to another avoidability problem involving additive squares. In Section \[enumer\] we report on mechanical proofs of some enumeration results. Some details about our implementation are given in the last section.
Fibonacci representation {#fibrep}
========================
Let the Fibonacci numbers be defined, as usual, by $F_0 = 0$, $F_1 = 1$, and $F_n = F_{n-1} + F_{n-2}$ for $n \geq 2$. (We caution the reader that some authors use a different indexing for these numbers.)
It is well-known, and goes back to Ostrowski [@Ostrowski:1922], Lekkerkerker [@Lekkerkerker:1952], and Zeckendorf [@Zeckendorf:1972], that every non-negative integer can be represented, in an essentially unique way, as a sum of Fibonacci numbers $(F_i)_{i\geq 2}$, subject to the constraint that no two consecutive Fibonacci numbers are used. For example, $43 = F_9 + F_6 + F_2$. Also see [@Carlitz:1968; @Fraenkel:1985].
Such a representation can be written as a binary string $a_1 a_2 \cdots a_n$ representing the integer $\sum_{1 \leq i \leq n} a_i F_{n+2-i}$. For example, the binary string $10010001$ is the Fibonacci representation of $43$.
For $w = a_1 a_2 \cdots a_n \in \Sigma_2^*$, we define $[a_1 a_2 \cdots a_n]_F := \sum_{1 \leq i \leq n} a_i F_{n+2-i}$, even if $a_1 a_2 \cdots a_n$ has leading zeroes or consecutive $1$’s. By $(n)_F$ we mean the [*canonical*]{} Fibonacci representation for the integer $n$, having no leading zeroes or consecutive $1$’s. Note that $(0)_F = \epsilon$, the empty string. The language of all canonical representations of elements of $\Enn$ is $\epsilon + 1(0+01)^*$.
Just as Fibonacci representation is the analogue of base-$k$ representation, we can define the notion of [*Fibonacci-automatic sequence*]{} as the analogue of the more familiar notation of $k$-automatic sequence . We say that an infinite word ${\bf a} = (a_n)_{n \geq 0}$ is Fibonacci-automatic if there exists an automaton with output $M = (Q, \Sigma_2, q_0, \delta, \kappa, \Delta)$ that $a_n = \kappa(\delta(q_0, (n)_F))$ for all $n \geq 0$. An example of a Fibonacci-automatic sequence is the infinite Fibonacci word, $${\bf f} = f_0 f_1 f_2 \cdots = 01001010\cdots$$ which is generated by the following 2-state automaton:
(q\_0) [$q_0/{\tt 0}$]{}; (q\_1) \[right=of q\_0\] [$q_1/{\tt 1}$]{};
(q\_0) edge \[loop above\] node [0]{} () (q\_0.10) edge node [1]{} (q\_1.170) (q\_1.190) edge node [0]{} (q\_0.350);
To compute $f_i$, we express $i$ in canonical Fibonacci representation, and feed it into the automaton. Then $f_i$ is the output associated with the last state reached (denoted by the symbol after the slash). Another characterization of Fibonacci-automatic sequences can be found in [@Shallit:1988a].
A basic fact about Fibonacci representation is that addition can be performed by a finite automaton. To make this precise, we need to generalize our notion of Fibonacci representation to $r$-tuples of integers for $r \geq 1$. A representation for $(x_1, x_2,\ldots, x_r)$ consists of a string of symbols $z$ over the alphabet $\Sigma_2^r$, such that the projection $\pi_i(z)$ over the $i$’th coordinate gives a Fibonacci representation of $x_i$. Notice that since the canonical Fibonacci representations of the individual $x_i$ may have different lengths, padding with leading zeroes will often be necessary. A representation for $(x_1, x_2, \ldots, x_r)$ is called canonical if it has no leading $[0,0,\ldots 0]$ symbols and the projections into individual coordinates have no occurrences of $11$. We write the canonical representation as $(x_1, x_2, \ldots, x_r)_F$. Thus, for example, the canonical representation for $(9,16)$ is $[0,1][1,0][0,0][0,1][0,0][1,0]$.
Thus, our claim about addition in Fibonacci representation is that there exists a deterministic finite automaton (DFA) $M_{\rm add}$ that takes input words of the form $[0,0,0]^* (x,y,z)_F$, and accepts if and only if $x +y =z$. Thus, for example, $M_{\rm add}$ accepts $[0,0,1][1,0,0][0,1,0][1,0,1]$, since the three strings obtained by projection are $0101, 0010, 1001$, which represent, respectively, $4$, $2$, and $6$ in Fibonacci representation. This result is apparently originally due to Berstel [@Berstel:1982]; also see .
Since this automaton does not appear to have been given explicitly in the literature and it is essential to our implementation, we give it here. The states of $M_{\rm add}$ are $Q = \lbrace 0,1,2,\ldots, 16 \rbrace$, the input alphabet is $\Sigma_2 \times \Sigma_2 \times \Sigma_2$, the final states are $F = \lbrace 1,7,11 \rbrace$, the initial state is $q_0 = 1$, and the transition function $\delta$ is given below. The automaton is incomplete, with any unspecified transitions going to a non-accepting dead state that transitions to itself on all inputs. This automaton actually works even for non-canonical expansions having consecutive $1$’s; an automaton working only for canonical expansions can easily be obtained by intersection with the appropriate regular languages. The state $0$ is a “dead state” that can safely be ignored.
\[0,0,0\] \[0,0,1\] \[0,1,0\] \[0,1,1\] \[1,0,0\] \[1,0,1\] \[1,1,0\] \[1,1,1\]
---- ----------- ----------- ----------- ----------- ----------- ----------- ----------- -----------
0 0 0 0 0 0 0 0 0
1 1 2 3 1 3 1 0 3
2 4 5 6 4 6 4 7 6
3 0 8 0 0 0 0 0 0
4 5 0 4 5 4 5 6 4
5 0 0 0 0 0 0 9 0
6 2 10 1 2 1 2 3 1
7 8 11 0 8 0 8 0 0
8 3 1 0 3 0 3 0 0
9 0 0 5 0 5 0 4 5
10 0 0 9 0 9 0 12 9
11 6 4 7 6 7 6 13 7
12 10 14 2 10 2 10 1 2
13 0 15 0 0 0 0 0 0
14 0 0 0 0 0 0 16 0
15 0 3 0 0 0 0 0 0
16 0 0 0 0 0 0 5 0
: Transition table for $M_{\rm add}$ for Fibonacci addition
We briefly sketch a proof of the correctness of this automaton. States can be identified with certain sequences, as follows: if $x,y,z$ are the identical-length strings arising from projection of a word that takes $M_{\rm add}$ from the initial state $1$ to the state $t$, then $t$ is identified with the integer sequence $([x0^n]_F + [y0^n]_F - [z0^n]_F)_{n \geq 0}$. With this correspondence, we can verify the following table by a tedious induction. In the table $L_n$ denotes the familiar Lucas numbers, defined by $L_n = F_{n-1} + F_{n+1}$ for $n \geq 0$ (assuming $F_{-1} = 1$). If a sequence $(a_n)_{n \geq 0}$ is the sequence identified with a state $t$, then $t$ is accepting iff $a_0 = 0$.
state sequence
------- ---------------------------
1 [**0**]{}
2 $(-F_{n+2})_{n \geq 0}$
3 $(F_{n+2})_{n \geq 0}$
4 $(-F_{n+3})_{n \geq 0}$
5 $(-F_{n+4})_{n \geq 0}$
6 $(-F_{n+1})_{n \geq 0}$
7 $(F_n)_{n \geq 0}$
8 $(F_{n+1})_{n \geq 0}$
9 $(-L_{n+2})_{n \geq 0}$
10 $(-2F_{n+2})_{n \geq 0}$
11 $(-F_n)_{n \geq 0}$
12 $(-2F_{n+1})_{n \geq 0}$
13 $(L_{n+1})_{n \geq 0}$
14 $(-3F_{n+2})_{n \geq 0}$
15 $(2F_{n+1})_{n \geq 0}$
16 $(-2F_n-3L_n)_{n \geq 0}$
: Identification of states with sequences
Note that the state $0$ actually represents a set of sequences, not just a single sequence. The set corresponds to those representations that are so far “out of synch” that they can never “catch up” to have $x +y = z$, no matter how many digits are appended.
We note that, in the spirit of the paper, this adder itself can, in principle, be checked mechanically (in $\operatorname{Th}(\Enn, 0)$, of course!), as follows:
First we show the adder $\cal A$ is specifying a function of $x$ and $y$. To do so, it suffices to check that $$\forall x \ \forall y \ \exists z \ {\cal A}(x,y,z)$$ and $$\forall x \ \forall y \ \forall z \ \forall z' \ {\cal A}(x,y,z) \wedge {\cal A}(x,y,z') \implies z = z' .$$ The first predicate says that there is at least one sum of $x$ and $y$ and the second says that there is at most one.
If both of these are verified, we know that $\cal A$ computes a function $A= A(x,y)$.
Next, we verify associativity, which amounts to checking that $$\forall x \ \forall y \ \forall z \ A(A(x,y),z) = A(x,A(y,z)) .$$ We can do this by checking that $$\forall x\ \forall y \ \forall z \ \forall w
\ \forall r \ \forall s \ \forall t \ ({\cal A}(x,y,r) \ \wedge \
{\cal A}(r,z,t) \ \wedge \ {\cal A}(y,z,s) ) \ \implies \ {\cal A}(x,s,t) .$$
Finally, we ensure that $\cal A$ is an adder by induction. First, we check that $\forall x \ A(x,0) = x$, which amounts to $$\forall x\ \forall y \ {\cal A}(x,0,y) \iff x = y .$$
Second, we check that if $A(x,1) = y$ then $x < y$ and there does not exist $z$ such that $x < z < y$. This amounts to $$\forall x, y, {\cal A}(x,1,y) \implies ((x < y) \ \wedge \
\neg \exists z \ (x<z) \ \wedge \ (z<y) ) .$$
This last condition shows that $A(x,1) = x+1$. By associativity $A(x,y+1) = A(x,A(y,1)) = A(A(x,y),1) = A(x,y) + 1$. By induction, $A(x,y) = A(x,0)+y = x+y$, so we are done.
Another basic fact about Fibonacci representation is that, for canonical representations containing no two consecutive $1$’s or leading zeroes, the radix order on representations is the same as the ordinary ordering on $\Enn$. It follows that a very simple automaton can, on input $(x,y)_F$, decide whether $x < y$.
Putting this all together, we get the analogue of Theorem \[one\]:
\[proc:Fib-auto-decide\] \
[**Input:**]{} $m,n \in \Enn$, $m$ DFAOs witnessing Fibonacci-automatic words ${\bf w}_1,{\bf w}_2,\dots,{\bf w}_m$, a first-order proposition with $n$ free variables $\varphi(v_1,v_2,\dots,v_n)$ using constants and relations definable in $\operatorname{Th}(\Enn,0,1,+)$ and indexing into ${\bf w}_1,{\bf w}_2,\dots,{\bf w}_m$.\
[**Output:**]{} DFA with input alphabet $\Sigma_2^n$ accepting $\{ (k_1,k_2,\dots,k_n)_F {\;:\;}\varphi(k_1,k_2,\dots,k_n) \text{ holds} \}$.
We remark that there was substantial skepticism that any implementation of a decision procedure for Fibonacci-automatic words would be practical, for two reasons:
- first, because the running time is bounded above by an expression of the form $$2^{2^{\Ddots^{ 2^{p(N)}}}}$$ where $p$ is a polynomial, $N$ is the number of states in the original automaton specifying the word in question, and the number of exponents in the tower is one less than the number of quantifiers in the logical formula characterizing the property being checked.
- second, because of the complexity of checking addition (15 states) compared to the analogous automaton for base-$k$ representation (2 states).
Nevertheless, we were able to carry out nearly all the computations described in this paper in a matter of a few seconds on an ordinary laptop.
Mechanical proofs of properties of the infinite Fibonacci word {#proofsf}
==============================================================
Recall that a word $x$, whether finite or infinite, is said to have period $p$ if $x[i] = x[i+p]$ for all $i$ for which this equality is meaningful. Thus, for example, the English word ${\tt alfalfa}$ has period $3$. The [*exponent*]{} of a finite word $x$, written $\exp(x)$, is $|x|/P$, where $P$ is the smallest period of $x$. Thus $\exp({\tt alfalfa}) = 7/3$.
If $\bf x$ is an infinite word with a finite period, we say it is [*ultimately periodic*]{}. An infinite word $\bf x$ is ultimately periodic if and only if there are finite words $u, v$ such that $x = uv^\omega$, where $v^\omega= vvv \cdots$.
A nonempty word of the form $xx$ is called a [*square*]{}, and a nonempty word of the form $xxx$ is called a [*cube*]{}. More generally, a nonempty word of the form $x^n$ is called an $n$’th power. By the [*order*]{} of a square $xx$, cube $xxx$, or $n$’th power $x^n$, we mean the length $|x|$.
The infinite Fibonacci word ${\bf f} = 01001010 \cdots = f_0 f_1 f_2 \cdots$ can be described in many different ways. In addition to our definition in terms of automata, it is also the fixed point of the morphism $\varphi(0) = 01$ and $\varphi(1) = 0$. This word has been studied extensively in the literature; see, for example, [@Berstel:1980b; @Berstel:1986b].
In the next subsection, we use our implementation to prove a variety of results about repetitions in $\bf f$.
Repetitions {#repe-subsec}
-----------
The word $\bf f$ is not ultimately periodic.
We construct a predicate asserting that the integer $p \geq 1$ is a period of some suffix of $\bf f$: $$(p \geq 1) \ \wedge \ \exists n \ \forall i \geq n\ {\bf f}[i] =
{\bf f}[i+p] .$$ (Note: unless otherwise indicated, whenever we refer to a variable in a predicate, the range of the variable is assumed to be $\Enn = \lbrace 0, 1, 2, \ldots \rbrace$.) From this predicate, using our program, we constructed an automaton accepting the language $$L = 0^*\ \lbrace (p)_F \ : \ (p \geq 1) \ \wedge \ \exists n
\ \forall i \geq n \ {\bf f}[i] = {\bf f}[i+p] \rbrace .$$ This automaton accepts the empty language, and so it follows that ${\bf f}$ is not ultimately periodic.
Here is the log of our program:
p >= 1 with 4 states, in 60ms
i >= n with 7 states, in 5ms
F[i] = F[i + p] with 12 states, in 34ms
i >= n => F[i] = F[i + p] with 51 states, in 15ms
Ai i >= n => F[i] = F[i + p] with 3 states, in 30ms
p >= 1 & Ai i >= n => F[i] = F[i + p] with 2 states, in 0ms
En p >= 1 & Ai i >= n => F[i] = F[i + p] with 2 states, in 0ms
overall time: 144ms
The largest intermediate automaton during the computation had 63 states.
A few words of explanation are in order: here “[F]{}” refers to the sequence $\bf f$, and “[E]{}” is our abbreviation for $\exists$ and “[A]{}” is our abbreviation for $\forall$. The symbol “[=>]{}” is logical implication, and “[&]{}” is logical and.
From now on, whenever we discuss the language accepted by an automaton, we will omit the $0^*$ at the beginning.
We recall an old result of Karhumäki [@Karhumaki:1983 Thm. 2]:
$\bf f$ contains no fourth powers.
We create a predicate for the orders of all fourth powers occurring in $\bf f$: $$(n > 0) \ \wedge \ \exists i \ \forall t<3n \ {\bf f}[i+t] = {\bf f}[i+n+t] .$$
The resulting automaton accepts nothing, so there are no fourth powers.
n > 0 with 4 states, in 46ms
t < 3 * n with 30 states, in 178ms
F[i + t] = F[i + t + n] with 62 states, in 493ms
t < 3 * n => F[i + t] = F[i + t + n] with 352 states, in 39ms
At t < 3 * n => F[i + t] = F[i + t + n] with 3 states, in 132ms
Ei At t < 3 * n => F[i + t] = F[i + t + n] with 2 states, in 0ms
n > 0 & Ei At t < 3 * n => F[i + t] = F[i + t + n] with 2 states, in 0ms
overall time: 888ms
The largest intermediate automaton in the computation had 952 states.
Next, we move on to a description of the orders of squares occurring in $\bf f$. An old result of Séébold [@Seebold:1985b] (also see ) states
All squares in $\bf f$ are of order $F_n$ for some $n \geq 2$. Furthermore, for all $n \geq 2$, there exists a square of order $F_n$ in $\bf f$. \[squares\]
We create a predicate for the lengths of squares: $$(n > 0) \ \wedge \ \exists i \ \forall t<n \ {\bf f}[i+t] = {\bf f}[i+n+t] .$$
When we run this predicate, we obtain an automaton that accepts exactly the language $10^*$. Here is the log file:
n > 0 with 4 states, in 38ms
t < n with 7 states, in 5ms
F[i + t] = F[i + t + n] with 62 states, in 582ms
t < n => F[i + t] = F[i + t + n] with 92 states, in 12ms
At t < n => F[i + t] = F[i + t + n] with 7 states, in 49ms
Ei At t < n => F[i + t] = F[i + t + n] with 3 states, in 1ms
n > 0 & Ei At t < n => F[i + t] = F[i + t + n] with 3 states, in 0ms
overall time: 687ms
The largest intermediate automaton had 236 states.
We can easily get much, much more information about the square occurrences in $\bf f$. The positions of all squares in $\bf f$ were computed by Iliopoulos, Moore, and Smyth , but their description is rather complicated and takes 5 pages to prove. Using our approach, we created an automaton accepting the language $$\{ (n,i)_F \ : \ (n > 0) \ \wedge \ \forall t<n \ {\bf f}[i+t] = {\bf f}[i+n+t]
\} .$$
This automaton has only 6 states and efficiently encodes the orders and starting positions of each square in $\bf f$. During the computation, the largest intermediate automaton had 236 states. Thus we have proved
The language $$\{ (n,i)_F \ : \ \text{there is a square of order $n$ beginning at
position $i$ in {\bf f}} \}$$ is accepted by the automaton in Figure \[squareorders\].
![Automaton accepting orders and positions of all squares in $\bf f$[]{data-label="squareorders"}](fibsquares.pdf){width="6.5in"}
Next, we examine the cubes in $\bf f$. Evidently Theorem \[squares\] implies that any cube in $\bf f$ must be of order $F_n$ for some $n$. However, not every order occurs.
The cubes in $\bf f$ are of order $F_n$ for $n \geq 4$, and a cube of each such order occurs.
We use the predicate $$(n > 0) \ \wedge \ \exists i \ \forall t<2n \ {\bf f}[i+t] = {\bf f}[i+n+t] .$$
When we run our program, we obtain an automaton accepting exactly the language $(100)0^*$, which corresponds to $F_n$ for $n \geq 4$.
n > 0 with 4 states, in 34ms
t < 2 * n with 16 states, in 82ms
F[i + t] = F[i + t + n] with 62 states, in 397ms
t < 2 * n => F[i + t] = F[i + t + n] with 198 states, in 17ms
At t < 2 * n => F[i + t] = F[i + t + n] with 7 states, in 87ms
Ei At t < 2 * n => F[i + t] = F[i + t + n] with 5 states, in 1ms
n > 0 & Ei At t < 2 * n => F[i + t] = F[i + t + n] with 5 states, in 0ms
overall time: 618ms
The largest intermediate automaton had 674 states.
Next, we encode the orders and positions of all cubes. We build a DFA accepting the language $$\{ (n,i)_F \ : \ (n > 0) \ \wedge \ \forall t<2n \ {\bf f}[i+t] = {\bf f}[i+n+t]
\} .$$
The language $$\{ (n,i)_F \ : \ \text{there is a cube of order $n$ beginning at
position $i$ in {\bf f}} \}$$ is accepted by the automaton in Figure \[cubeorders\].
![Automaton accepting orders and positions of all cubes in $\bf f$[]{data-label="cubeorders"}](fibcubes.pdf){width="6.5in"}
Finally, we consider all the maximal repetitions in $\bf f$. Let $p(x)$ denote the length of the least period of $x$. If ${\bf x} = a_0 a_1 \cdots$, by ${\bf x}[i..j]$ we mean $a_i a_{i+1} \cdots a_j$. Following Kolpakov and Kucherov , we say that ${\bf f}[i..i+n-1]$ is a [*maximal repetition*]{} if
- $p({\bf f}[i..i+n-1]) \leq n/2$;
- $p({\bf f}[i..i+n-1]) < p({\bf f}[i..i+n]) $;
- If $i > 0$ then $p({\bf f}[i..i+n-1]) < p({\bf f}[i-1..i+n-1])$.
The factor ${\bf f}[i..i+n-1]$ is a maximal repetition of $\bf f$ iff $(n,i)_F$ is accepted by the automaton depicted in Figure \[maxreps2\].
![Automaton accepting occurrences of maximal repetitions in $\bf f$[]{data-label="maxreps2"}](output_maxreps2.pdf){width="3.5in"}
An [*antisquare*]{} is a nonempty word of the form $x \overline{x}$, where $\overline{x}$ denotes the complement of $x$ ($1$’s changed to $0$’s and vice versa). Its order is $|x|$. For a new (but small) result we prove
The Fibonacci word $\bf f$ contains exactly four antisquare factors: $01, 10, 1001, $ and $10100101$.
The predicate for having an antisquare of length $n$ is $$\exists i \ \forall k < n \ {\bf f}[i+k] \not= {\bf f}[i+k+n] .$$ When we run this we get the automaton depicted in Figure \[antisquare\], specifying that the only possible orders are $1$, $2$, and $4$, which correspond to words of length $2$, $4$, and $8$.
![Automaton accepting orders of antisquares in $\bf f$[]{data-label="antisquare"}](fib-antisquares.pdf){width="5.5in"}
Inspection of the factors of these lengths proves the result.
Palindromes and antipalindromes
-------------------------------
We now turn to a characterization of the palindromes in $\bf f$. Using the predicate $$\exists i \ \forall j<n \ {\bf f}[i+j] = {\bf f}[i+n-1-j],$$ we specify those lengths $n$ for which there is a palindrome of length $n$. Our program then recovers the following result of Chuan [@Chuan:1993b]:
There exist palindromes of every length $\geq 0$ in $\bf f$.
We could also characterize the positions of all nonempty palindromes. The resulting 21-state automaton is not particularly enlightening, but is included here to show the kind of complexity that can arise.
![Automaton accepting orders and positions of all nonempty palindromes in $\bf f$[]{data-label="palindrome-orders"}](output_palindrome-positions.pdf){width="4.8in"}
Although the automaton in Figure \[palindrome-orders\] encodes all palindromes, more specific information is a little hard to deduce from it. For example, let’s prove a result of Droubay [@Droubay:1995]:
The Fibonacci word $\bf f$ has exactly one palindromic factor of length $n$ if $n$ is even, and exactly two palindromes of length $n$ if $n$ odd.
First, we obtain an expression for the lengths $n$ for which there is exactly one palindromic factor of length $n$. $$\begin{gathered}
\exists i \ (\forall t<n \ {\bf f}[i+t] = {\bf f}[i+n-1-t])
\ \wedge \ \\
\forall j \ (\forall s<n\ {\bf f}[j+s] = {\bf f}[j+n-1-s]) \implies
( \forall u<n\ {\bf f}[i+u] = {\bf f}[j+u])\end{gathered}$$ The first part of the predicate asserts that ${\bf f}[i..i+n-1]$ is a palindrome, and the second part asserts that any palindrome ${\bf f}[j..j+n-1]$ of the same length must in fact be equal to ${\bf f}[i..i+n-1]$.
When we run this predicate through our program we get the automaton depicted below in Figure \[onepal\].
![Automaton accepting lengths with exactly one palindrome[]{data-label="onepal"}](1-pal-lengths.pdf){width="6.5in"}
It may not be obvious, but this automaton accepts exactly the Fibonacci representations of the even numbers. The easiest way to check this is to use our program on the predicate $\exists i \ n = 2i$ and verify that the resulting automaton is isomorphic to that in Figure \[onepal\].
Next, we write down a predicate for the existence of exactly two distinct palindromes of length $n$. The predicate asserts the existence of two palindromes ${\bf x}[i..i+n-1]$ and ${\bf x}[j..j+n-1]$ that are distinct and for which any palindrome of the same length must be equal to one of them.
$$\begin{gathered}
\exists i\ \exists j\ (\forall t<n\ {\bf f}[i+t] = {\bf f}[i+n-1-t])
\ \wedge \
(\forall s<n\ {\bf f}[j+s] = {\bf f}[j+n-1-s])
\ \wedge \ \\
(\exists m<n\ {\bf f}[i+m] \not= {\bf f}[j+m])
\ \wedge \ \\
( \forall u (\forall k < n\ {\bf f}[u+k] = {\bf f}[u+n-1-k]) \implies
(( \forall l<n\ {\bf f}[u+l] = {\bf f}[i+l]) \ \vee \ (\forall p<n \ {\bf f}[u+p] = {\bf f}[j+p]))) \end{gathered}$$
Again, running this through our program gives us an automaton accepting the Fibonacci representations of the odd numbers. We omit the automaton.
The prefixes are factors of particular interest. Let us determine which prefixes are palindromes:
The prefix ${\bf f}[0..n-1]$ of length $n$ is a palindrome if and only if $n = F_i - 2$ for some $i \geq 3$.
We use the predicate $$\forall i<n\ {\bf f}[i] = {\bf f}[n-1-i]$$ obtaining an automaton accepting $\epsilon + 1 + 10(10)^*(0+01)$, which are precisely the representations of $F_i - 2$.
Next, we turn to the property of “mirror invariance”. We say an infinite word $\bf w$ is mirror-invariant if whenever $x$ is a factor of $\bf w$, then so is $x^R$. We can check this for $\bf f$ by creating a predicate for the assertion that for each factor $x$ of length $n$, the factor $x^R$ appears somewhere else: $$\forall i \geq 0 \ \exists j \text{ such that }
{\bf f}[i..i+n-1] = {\bf f}[j..j+n-1]^R .$$ When we run this through our program we discover that it accepts the representations of all $n \geq 0$. Here is the log:
t < n with 7 states, in 99ms
F[i + t] = F[j + n - 1 - t] with 264 states, in 7944ms
t < n => F[i + t] = F[j + n - 1 - t] with 185 states, in 89ms
At t < n => F[i + t] = F[j + n - 1 - t] with 35 states, in 182ms
Ej At t < n => F[i + t] = F[j + n - 1 - t] with 5 states, in 2ms
Ai Ej At t < n => F[i + t] = F[j + n - 1 - t] with 3 states, in 6ms
overall time: 8322ms
Thus we have proved:
The word ${\bf f}$ is mirror invariant.
An [*antipalindrome*]{} is a word $x$ satisfying $x = \overline{x^R}$. For a new (but small) result, we determine all possible antipalindromes in $\bf f$:
The only nonempty antipalindromes in $\bf f$ are $01$, $10$, $(01)^2$, and $(10)^2$.
Let us write a predicate specifying that ${\bf f}[i..i+n-1]$ is a nonempty antipalindrome, and further that it is a first occurrence of such a factor: $$(n > 0) \ \wedge\ (\forall j<n \ {\bf f}[i+j] \not= {\bf f}[i+n-1-j]) \ \wedge \
(\forall i' < i \ \exists j<n\ {\bf f}[i'+j] \not= {\bf f}[i+j]) .$$
When we run this through our program, the language of $(n,i)_F$ satisfying this predicate is accepted by the following automaton:
![Automaton accepting orders and positions of first occurrences of nonempty antipalindromes in $\bf f$[]{data-label="antipal"}](antipal.pdf){width="5in"}
It follows that the only $(n,i)$ pairs accepted are $(2,0), (2,1), (4,3), (4,4)$, corresponding, respectively, to the strings $01$, $10$, $(01)^2$, and $(10)^2$.
Special factors
---------------
Next we turn to special factors. It is well-known (and we will prove it in Theorem \[sturmcomp\] below), that ${\bf f}$ has exactly $n+1$ distinct factors of length $n$ for each $n \geq 0$. This implies that there is exactly one factor $x$ of each length $n$ with the property that both $x0$ and $x1$ are factors. Such a factor is called [*right-special*]{} or sometimes just [*special*]{}. We can write a predicate that expresses the assertion that the factor ${\bf f}[i..i+n-1]$ is the unique special factor of length $n$, and furthermore, that it is the first occurrence of that factor, as follows: $$\begin{gathered}
(\forall i' < i \ \exists s < n \ {\bf f}[i'+s] \not= {\bf f}[i+s])
\ \wedge \
\exists j \ \exists k \ ((\forall t < n\ {\bf f}[j+t] = {\bf f}[i+t]) \\
\wedge \
(\forall u < n\ {\bf f}[k+u] = {\bf f}[i+u]) \ \wedge \
({\bf f}[j+n] \not= {\bf f}[k+n])) .\end{gathered}$$
The automaton depicted below in Figure \[special\] accepts the language $$\{ (i,n)_F \ : \ \text{the factor } {\bf f}[i..i+n-1]
\text{ is the first occurrence of the unique special factor of length $n$} \} .$$
![Automaton accepting first positions and lengths of special factors in $\bf f$[]{data-label="special"}](output_special-factors.pdf){width="3.5in"}
Furthermore it is known (e.g., [@Pirillo:1997 Lemma 5]) that
The unique special factor of length $n$ is ${\bf f}[0..n-1]^R$.
We create a predicate that says that if a factor is special then it matches ${\bf f}[0..n-1]^R$. When we run this we discover that all lengths are accepted.
Least periods
-------------
We now turn to least periods of factors of ${\bf f}$; see [@Saari:2007] and and .
Let $P$ denote the assertion that $n$ is a period of the factor ${\bf f}[i..j]$, as follows: $$\begin{aligned}
P(n,i,j) &=& {\bf f}[i..j-n] = {\bf f}[i+n..j] \\
&=& \forall \ t\ \text{ with $i \leq t \leq j-n$ we have }
{\bf f}[t] = {\bf f}[t+n] .\end{aligned}$$ Using this, we can express the predicate $LP$ that $n$ is the least period of ${\bf f}[i..j]$: $$LP(n,i,j) = P(n,i,j) \text{ and } \forall n'
\text{ with } 1 \leq n' < n \ \neg P(n',i,j).$$ Finally, we can express the predicate that $n$ is a least period as follows $$L(n) = \exists i, j \geq 0 \text{ with $0 \leq i+n \leq j-1$ }
LP(n, i, j) .$$
Using an implementation of this, we can reprove the following theorem of Saari [@Saari:2007 Thm. 2]:
If a word $w$ is a nonempty factor of the Fibonacci word, then the least period of $w$ is a Fibonacci number $F_n$ for $n \geq 2$. Furthermore, each such period occurs.
We ran our program on the appropriate predicate and found the resulting automaton accepts $10^+$, corresponding to $F_n$ for $n \geq 2$.
Furthermore, we can actually encode information about all least periods. The automaton depicted in Figure \[leastp\] accepts triples $(n,p,i)$ such that $p$ is a least period of ${\bf f}[i..i+n-1]$.
![Automaton encoding least periods of all factors in $\bf f$[]{data-label="leastp"}](output_all-least-periods.pdf){width="5.5in"}
We also have the following result, which seems to be new.
Let $n \geq 1$, and define $\ell(n)$ to be the smallest integer that is the least period of some length-$n$ factor of $\bf f$. Then $\ell(n) = F_j$ for $j \geq 1$ if $L_j-1 \leq n \leq L_{j+1}-2$, where $L_j$ is the $j$’th Lucas number defined in Section \[fibrep\]. \[allpers\]
We create an automaton accepting $(n,p)_F$ such that (a) there exists at least one length-$n$ factor of period $p$ and (b) for all length-$n$ factors $x$, if $q$ is a period of $x$, then $q \geq p$. This automaton is depicted in Figure \[least-period-over\] below.
![Automaton encoding smallest period over all length-$n$ factors in $\bf f$[]{data-label="least-period-over"}](leastper.pdf){width="6.5in"}
The result now follows by inspection and the fact that $(L_j-1)_F = 10 (01)^{(j-2)/2}$ if $j \geq 2$ is even, and $100 (10)^{(j-3)/2}$ if $j \geq 3$ is odd.
Quasiperiods
------------
We now turn to quasiperiods. An infinite word $\bf a$ is said to be [*quasiperiodic*]{} if there is some finite nonempty word $x$ such that ${\bf a}$ can be completely “covered” with translates of $x$. Here we study the stronger version of quasiperiodicity where the first copy of $x$ used must be aligned with the left edge of $\bf w$ and is not allowed to “hang over”; these are called [*aligned covers*]{} in . More precisely, for us ${\bf a} = a_0 a_1 a_2 \cdots$ is quasiperiodic if there exists $x$ such that for all $i \geq 0$ there exists $j\geq 0$ with $i-n < j \leq i$ such that $a_j a_{j+1} \cdots a_{j+n-1} = x$, where $n = |x|$. Such an $x$ is called a [*quasiperiod*]{}. Note that the condition $j \geq 0$ implies that, in this interpretation, any quasiperiod must actually be a prefix of $\bf a$.
The quasiperiodicity of the Fibonacci word $\bf f$ was studied by Christou, Crochemore, and Iliopoulos , where we can (more or less) find the following theorem:
A nonempty length-$n$ prefix of $\bf f$ is a quasiperiod of $\bf f$ if and only if $n$ is not of the form $F_n - 1$ for $n \geq 3$.
In particular, the following prefix lengths are not quasiperiods: $1$, $2$, $4$, $7$, $12$, and so forth.
We write a predicate for the assertion that the length-$n$ prefix is a quasiperiod: $$\forall i \geq 0 \ \exists j \text{ with } i-n < j \leq i
\text{ such that } \forall t<n \ {\bf f}[t] = {\bf f}[j+t] .$$ When we do this, we get the automaton in Figure \[quasi\] below. Inspection shows that this DFA accepts all canonical representations, except those of the form $1(01)^*(\epsilon + 0)$, which are precisely the representations of $F_n - 1$.
![Automaton accepting lengths of prefixes of $\bf f$ that are quasiperiods[]{data-label="quasi"}](output_quasiperiods.pdf){width="4in"}
Unbordered factors
------------------
Next we look at unbordered factors. A word $y$ is said to be a [*border*]{} of $x$ if $y$ is both a nonempty proper prefix and suffix of $x$. A word $x$ is [*bordered*]{} if it has at least one border. It is easy to see that if a word $y$ is bordered iff it has a border of length $\ell$ with $0 < \ell \leq |y|/2$.
The only unbordered nonempty factors of $\bf f$ are of length $F_n$ for $n \geq 2$, and there are two for each such length. For $n \geq 3$ these two unbordered factors have the property that one is a reverse of the other.
We can express the property of having an unbordered factor of length $n$ as follows $$\exists i\ \forall j, 1 \leq j \leq n/2, \
\exists t<j\ {\bf f}[i+t] \not= {\bf f}[i+n-j+t] .$$
Here is the log:
j >= 1 with 4 states, in 155ms
2 * j <= n with 16 states, in 91ms
j >= 1 & 2 * j <= n with 21 states, in 74ms
t < j with 7 states, in 17ms
F[i + t] != F[i + n - j + t] with 321 states, in 10590ms
t < j & F[i + t] != F[i + n - j + t] with 411 states, in 116ms
Et t < j & F[i + t] != F[i + n - j + t] with 85 states, in 232ms
j >= 1 & 2 * j <= n => Et t < j & F[i + t] != F[i + n - j + t] with 137 states, in 19ms
Aj j >= 1 & 2 * j <= n => Et t < j & F[i + t] != F[i + n - j + t] with 7 states, in 27ms
Ei Aj j >= 1 & 2 * j <= n => Et t < j & F[i + t] != F[i + n - j + t] with 3 states, in 0ms
overall time: 11321ms
The automaton produced accepts the Fibonacci representation of $0$ and $F_n$ for $n \geq 2$.
Next, we make the assertion that there are exactly two such factors for each appropriate length. We can do this by saying there is an unbordered factor of length $n$ beginning at position $i$, another one beginning at position $k$, and these factors are distinct, and for every unbordered factor of length $n$, it is equal to one of these two. When we do this we discover that the representations of all $F_n$ for $n \geq 2$ are accepted.
Finally, we make the assertion that for any two unbordered factors of length $n$, either they are equal or one is the reverse of the other. When we do this we discover all lengths except length $1$ are accepted. (That is, for all lengths other than $F_n$, $n \geq 2$, the assertion is trivially true since there are no unbordered factors; for $F_2 = 1$ it is false since $0$ and $1$ are the unbordered factors and one is not the reverse of the other; and for all larger $F_i$ the property holds.)
Recurrence, uniform recurrence, and linear recurrence
-----------------------------------------------------
We now turn to various questions about recurrence. A factor $x$ of an infinite word $\bf w$ is said to be [*recurrent*]{} if it occurs infinitely often. The word $\bf w$ is recurrent if every factor that occurs at least once is recurrent. A factor $x$ is [*uniformly recurrent*]{} if there exists a constant $c = c(x)$ such that any factor ${\bf w}[i..i+c]$ is guaranteed to contain an occurrence of $x$. If all factors are uniformly recurrent then $\bf w$ is said to be uniformly recurrent. Finally, ${\bf w}$ is [*linearly recurrent*]{} if the constant $c(x)$ is $O(|x|)$.
The word [**f**]{} is recurrent, uniformly recurrent, and linearly recurrent.
A predicate for all length-$n$ factors being recurrent: $$\forall i \geq 0\ \forall j \geq 0\ \exists k > j\ \forall t<n \
{\bf f}[i+t] = {\bf f}[k+t] .$$ This predicate says that for every factor $z = {\bf f}[i..i+n-1]$ and every position $j$ we can find another occurrence of $z$ beginning at a position $k > j$. When we run this we discover that the representations of all $n \geq 0$ are accepted. So $\bf f$ is recurrent.
A predicate for uniform recurrence: $$\forall i\ \exists \ell\ \forall j \ \exists s, \ j \leq s \leq j+l-n \
\forall p<n \ {\bf f}[s+p] = {\bf f}[i+p] .$$ Once again, when we run this we discover that the representations of all $n \geq 0$ are accepted. So $\bf f$ is uniformly recurrent.
A predicate for linear recurrence with constant $C$: $$\forall i\ \forall j \ \exists s, \ j \leq s \leq j+Cn \
\forall p<n \ {\bf f}[s+p] = {\bf f}[i+p] .$$ When we run this with $C = 4$, we discover that the representations of all $n \geq 0$ are accepted (but, incidentally, not for $C = 3$). So $\bf f$ is linearly recurrent.
We can decide the property of linear recurrence for Fibonacci-automatic words even without knowing an explicit value for the constant $C$. The idea is to accept those pairs $(n,t)$ such that there exists a factor of length $n$ with two consecutive occurrences separated by distance $t$. Letting $S$ denote the set of such pairs, then a sequence is linearly recurrent iff $\limsup_{(n,t)\in S} t/n < \infty$, which can be decided using an argument like that in . However, we do not know how to compute, in general, the exact value of the $\limsup$ for Fibonacci representation (which we do indeed know for base-$k$ representation), although we can approximate it arbitrarily closely.
Lyndon words
------------
Next, we turn to some results about Lyndon words. Recall that a nonempty word $x$ is a [*Lyndon word*]{} if it is lexicographically less than all of its nonempty proper prefixes.[^3] We reprove some recent results of Currie and Saari and Saari [@Saari:2014].
Every Lyndon factor of $\bf f$ is of length $F_n$ for some $n \geq 2$, and each of these lengths has a Lyndon factor.
Here is the predicate specifying that there is a factor of length $n$ that is Lyndon: $$\exists i\ \forall j, 1 \leq j < n, \
\exists t < n-j \ (\forall u<t \ {\bf f}[i+u]={\bf f}[i+j+u]) \ \wedge \
{\bf f}[i+t] < {\bf f}[i+j+t] .$$ When we run this we get the representations $10^*$, which proves the result.
For $n \geq 2$, every length-$n$ Lyndon factor of $\bf f$ is a conjugate of ${\bf f}[0..n-1]$.
Using the predicate from the previous theorem as a base, we can create a predicate specifying that every length-$n$ Lyndon factor is a conjugate of ${\bf f}[0..n-1]$. When we do this we discover that all lengths except $1$ are accepted. (The only lengths having a Lyndon factor are $F_n$ for $n \geq 2$, so all but $F_2$ have the desired property.)
Critical exponents
------------------
Recall from Section \[proofsf\] that $\exp(w) = |w|/P$, where $P$ is the smallest period of $w$. The [*critical exponent*]{} of an infinite word $\bf x$ is the supremum, over all factors $w$ of $\bf x$, of $\exp(w)$.
A classic result of is
The critical exponent of $\bf f$ is $2+ \alpha$, where $\alpha = (1+\sqrt{5})/2$.
Although it is known that the critical exponent is computable for $k$-automatic sequences , we do not yet know this for Fibonacci-automatic sequences (and more generally Pisot-automatic sequences). However, with a little inspired guessing about the maximal repetitions, we can complete the proof.
For each length $n$, the smallest possible period $p$ of a factor is given by Theorem \[allpers\]. Hence the critical exponent is given by $\lim_{j \rightarrow \infty}
(L_{j+1}-2)/F_j$, which is $2+\alpha$.
We can also ask the same sort of questions about the [*initial critical exponent*]{} of a word $\bf w$, which is the supremum over the exponents of all prefixes of $\bf w$.
The initial critical exponent of $\bf f$ is $1+\alpha$.
We create an automaton $M_{\rm ice}$ accepting the language $$L = \{ (n,p)_F \ : \ {\bf f}[0..n-1] \text{ has least period } p \} .$$ It is depicted in Figure \[ice\] below. From the automaton, it is easy to see that the least period of the prefix of length $n \geq 1$ is $F_j$ for $j \geq 2$ and $F_{j+1}-1 \leq n \leq F_{j+2} - 2$. Hence the initial critical exponent is given by $\limsup_{j \rightarrow \infty} (F_{j+2} - 2)/F_j$, which is $1+\alpha$.
![Automaton accepting least periods of prefixes of length $n$[]{data-label="ice"}](leastperprefixes.pdf){width="6.5in"}
The shift orbit closure
-----------------------
The [*shift orbit closure*]{} of a sequence $\bf x$ is the set of all sequences $\bf t$ with the property that each prefix of $\bf t$ appears as a factor of $\bf x$. Note that this set can be much larger than the set of all suffixes of $\bf x$.
The following theorem is well known :
The lexicographically least sequence in the shift orbit closure of $\bf f$ is $0{\bf f}$, and the lexicographically greatest is $1 {\bf f}$.
We handle only the lexicographically least, leaving the lexicographically greatest to the reader.
The idea is to create a predicate $P(n)$ for the lexicographically least sequence ${\bf b} = b_0 b_1 b_2 \cdots$ which is true iff $b_n = 1$. The following predicate encodes, first, that $b_n = 1$, and second, that if one chooses any length-($n+1$) factor $t$ of $\bf f$, then $b_0 \cdots b_n$ is equal or lexicographically smaller than $t$.
$$\begin{gathered}
\exists j \ {\bf f}[j+n]=1 \ \wedge \
\forall k \ (( \forall s \leq n \ {\bf f}[j+s] = {\bf f}[k+s] ) \ \vee \ \\
(\exists i\leq n\ {\text s. t. }\ {\bf f}[j+i] < {\bf f}[k+i]
\ \wedge \ ( \forall t<i \ {\bf f}[j+t]={\bf f}[k+t] )))\end{gathered}$$
When we do this we get the following automaton, which is easily seen to generate the sequence $0 {\bf f}$.
![Automaton accepting lexicographically least sequence in shift orbit closure of ${\bf f}$[]{data-label="lexleastorbit"}](lexleastorbit.pdf){width="6.5in"}
Minimal forbidden words
-----------------------
Let ${\bf x}$ be an infinite word. A finite word $z = a_0 \cdots a_n$ is said to be [*minimal forbidden*]{} if $z$ is not a factor of $\bf x$, but both $a_1 \cdots a_n$ and $a_0 \cdots a_{n-1}$ are .
We can characterize all minimal forbidden words as follows: we create an automaton accepting the language $$\begin{gathered}
\{ (i,n)_F \ : \ {\bf f}[i..i+n-1] \, \overline{{\bf f}[n]}
\text{ is not a factor of $\bf f$ and } \\
{\bf f}[i+1..i+n-1] \, \overline{{\bf f}[n]} \text{ is a factor }
\text{and } i \text{ is as small as possible } \}.\end{gathered}$$
When we do so we find the words accepted are $$[1,1] ([0,0][1,1])^* (\epsilon + [0,0]) .$$ This corresponds to the words $${\bf f}[F_n - 1..2F_n -3] \, \overline{{\bf f}[2F_n -2]}$$ for $n \geq 3$. The first few are $$11, 000, 10101, 00100100, 1010010100101, \ldots .$$
Grouped factors
---------------
Cassaigne [@Cassaigne:1998] introduced the notion of *grouped factors*. A sequence ${\bf a} = (a_i)_{i \geq 0}$ has grouped factors if, for all $n \geq 1$, there exists some position $m = m(n)$ such that ${\bf a}[m..m+\rho(n)+n-2]$ contains all the $\rho(n)$ length-$n$ blocks of $\bf a$, each block occurring exactly once. One consequence of his result is that the Fibonacci word has grouped factors.
We can write a predicate for the property of having grouped factors, as follows: $$\begin{gathered}
\forall n \geq 1 \quad \exists m, s \geq 0 \quad
\forall i \geq 0 \\
\exists j \text{ s.t. } m \leq j \leq m+s \text{ and } {\bf a}[i..i+n-1] = {\bf a}[j..j+n-1] \text{ and } \\
\forall j', \ m \leq j' \leq m+s, \quad j \not= j'
\text{ we have } {\bf a}[i..i+n-1] \not= {\bf a}[j'..j'+n-1] .\end{gathered}$$
The first part of the predicate says that every length-$n$ block appears somewhere in the desired window, and the second says that it appears exactly once.
(This five-quantifier definition can be viewed as a response to the question of Homer and Selman , “...in what sense would a problem that required at least three alternating quantifiers to describe be natural?")
Using this predicate and our decision method, we verified that the Fibonacci word does indeed have grouped factors.
Mechanical proofs of properties of the finite Fibonacci words {#finitefib}
=============================================================
Although our program is designed to answer questions about the properties of the infinite Fibonacci word $\bf f$, it can also be used to solve problems concerning the finite Fibonacci words $(X_n)$, defined as follows: $$X_n = \begin{cases}
\epsilon, & \text{if $n = 0$}; \\
1, & \text{if $n = 1$}; \\
0, & \text{if $n = 2$}; \\
X_{n-1} X_{n-2}, & \text{if $n > 2$}.
\end{cases}$$ Note that $|X_n| = F_n$ for $n \geq 1$. (We caution the reader that there exist many variations on this definition in the literature, particularly with regard to indexing and initial values.) Furthermore, we have $\varphi(X_n) = X_{n+1}$ for $n \geq 1$.
Our strategy for the the finite Fibonacci words has two parts:
- Instead of phrasing statements in terms of factors, we phrase them in terms of occurrences of factors (and hence in terms of the indices defining a factor).
- Instead of phrasing statements about finite Fibonacci words, we phrase them instead about [*all*]{} length-$n$ prefixes of $\bf f$. Then, since $X_i = {\bf f}[0..F_i - 1]$, we can deduce results about the finite Fibonacci words by considering the case where $n$ is a Fibonacci number $F_i$.
To illustrate this idea, consider one of the most famous properties of the Fibonacci words, the [*almost-commutative*]{} property: letting $\eta(a_1 a_2 \cdots a_n) =
a_1 a_2 \cdots a_{n-2} a_n a_{n-1}$ be the map that interchanges the last two letters of a string of length at least $2$, we have
$X_{n-1} X_n = \eta(X_n X_{n-1})$ for $n \geq 2$.
We can verify this, and prove even more, using our method.
Let $x = {\bf f}[0..i-1]$ and $y = {\bf f}[0..j-1]$ for $i > j > 1$. Then $xy = \eta(yx)$ if and only if $i = F_n$, $j = F_{n-1}$ for $n \geq 3$.
The idea is to check, for each $i > j > 1$, whether $${\bf f}[0..i-1] {\bf f}[0..j-1] = \eta({\bf f}[0..j-1] {\bf f}[0..i-1]).$$ We can do this with the following predicate: $$\begin{gathered}
(i>j) \ \wedge \ (j\geq 2) \ \wedge \ (\forall t,\ j\leq t<i,\ {\bf f}[t]=
{\bf f}[t-j])
\ \wedge \\
(\forall s \leq j-3\ {\bf f}[s]={\bf f}[s+i-j]) \ \wedge \ ({\bf f}[j-2]={\bf f}[i-1])
\ \wedge \ ({\bf f}[j-1]={\bf f}[i-2]) .\end{gathered}$$ The log of our program is as follows:
i > j with 7 states, in 49ms
j >= 2 with 5 states, in 87ms
i > j & j >= 2 with 12 states, in 3ms
j <= t with 7 states, in 3ms
t < i with 7 states, in 17ms
j <= t & t < i with 19 states, in 6ms
F[t] = F[t - j] with 16 states, in 31ms
j <= t & t < i => F[t] = F[t - j] with 62 states, in 31ms
At j <= t & t < i => F[t] = F[t - j] with 14 states, in 43ms
i > j & j >= 2 & At j <= t & t < i => F[t] = F[t - j] with 12 states, in 9ms
s <= j - 3 with 14 states, in 72ms
F[s] = F[s + i - j] with 60 states, in 448ms
s <= j - 3 => F[s] = F[s + i - j] with 119 states, in 14ms
As s <= j - 3 => F[s] = F[s + i - j] with 17 states, in 58ms
i > j & j >= 2 & At j <= t & t < i => F[t] = F[t - j] & As s <= j - 3 => F[s] = F[s + i - j] with 6 states, in 4ms
F[j - 2] = F[i - 1] with 20 states, in 34ms
i > j & j >= 2 & At j <= t & t < i => F[t] = F[t - j] & As s <= j - 3 => F[s] = F[s + i - j] & F[j - 2] = F[i - 1] with 5 states, in 1ms
F[j - 1] = F[i - 2] with 20 states, in 29ms
i > j & j >= 2 & At j <= t & t < i => F[t] = F[t - j] & As s <= j - 3 => F[s] = F[s + i - j] & F[j - 2] = F[i - 1] & F[j - 1] = F[i - 2] with 5 states, in 1ms
overall time: 940ms
The resulting automaton accepts $[1,0][0,1][0,0]^+$, which corresponds to $i = F_n$, $j = F_{n-1}$ for $n \geq 4$.
An old result of Séébold [@Seebold:1985b] is
If $uu$ is a square occurring in $\bf f$, then $u$ is conjugate to some finite Fibonacci word.
Assertion $\operatorname{conj}(i,j,k,\ell)$ means ${\bf f}[i..j]$ is a conjugate of ${\bf f}[k..\ell]$ (assuming $j-i = \ell-k$) $$\operatorname{conj}(i,j,k,\ell) :=
\exists m \ {\bf f}[i..i+\ell-m] = {\bf f}[m..\ell] \text{ and }
{\bf f}[i+\ell-m+1..j] = {\bf f}[k..m-1].$$
Predicate: $$({\bf f}[i..i+n-1] = {\bf f}[i+n..i+2n-1])
\implies \operatorname{conj}(i,i+n-1,0,n-1)$$
This asserts that any square $uu$ of order $n$ appearing in $\bf f$ is conjugate to ${\bf f}[0..n-1]$. When we implement this, we discover that all lengths are accepted. This makes sense since the only lengths corresponding to squares are $F_n$, and for all other lengths the base of the implication is false.
We now reprove an old result of de Luca [@deLuca:1981]. Recall that a primitive word is a non-power; that is, a word that cannot be written in the form $x^n$ where $n$ is an integer $\geq 2$.
All finite Fibonacci words are primitive.
The factor ${\bf f}[i..j]$ is a power if and only if there exists $d$, $0 < d < j-i+1 $, such that ${\bf f}[i..j-d] = {\bf f}[i+d..j]$ and ${\bf f}[j-d+1..j] = {\bf f}[i..i+d-1]$. Letting $\operatorname{pow}(i,j)$ denote this predicate, the predicate $$\neg \operatorname{pow}(0,n-1)$$ expresses the claim that the length-$n$ prefix ${\bf f}[0..n-1]$ is primitive. When we implement this, we discover that the prefix of every length is primitive, except those prefixes of length $2 F_n$ for $n \geq 4$.
A theorem of Chuan [@Chuan:1993b Thm. 3] states that the finite Fibonacci word $X_n$, for $n \geq 5$, is the product of two palindromes in exactly one way: where the first factor of length $F_{n-1} -2$ and the second of length $F_{n-2} + 2$. (Actually, Chuan claimed this was true for all Fibonacci words, but, for example, for $010$ there are evidently two different factorizations of the form $(\epsilon)(010)$ and $(010)\epsilon$.) We can prove something more general using our method, by generalizing:
If the length-$n$ prefix ${\bf f}[0..n-1]$ of $\bf f$ is the product of two (possibly empty) palindromes, then $(n)_F$ is accepted by the automaton in Figure \[pal2\] below.
![Automaton accepting lengths of prefixes that are the product of two palindromes[]{data-label="pal2"}](product2pal.pdf){width="6.5in"}
Furthermore, if the length-$n$ prefix ${\bf f}[0..n-1]$ of $\bf f$ is the product of two (possibly empty) palindromes in exactly one way, then $(n)_F$ is accepted by the automaton in Figure \[pal2u\] below.
![Automaton accepting lengths of prefixes that are the product of two palindromes in exactly one way[]{data-label="pal2u"}](product2pals1way.pdf){width="6.5in"}
Evidently, this includes all $n$ of the form $F_j$ for $j \geq 5$.
For the first, we use the predicate $$\exists p\leq n\ \left( (\forall t<p\ {\bf f}[t] = {\bf f}[p-1-t]) \
\wedge\ (\forall u< n-p\ {\bf f}[p+u] = {\bf f}[n-1-u]) \right) .$$
For the second, we use the predicate $$\begin{gathered}
\exists p\leq n\ ( (\forall t<p\ {\bf f}[t] = {\bf f}[p-1-t]) \
\wedge\
(\forall u< n-p\ {\bf f}[p+u] = {\bf f}[n-1-u]) )) \ \wedge \
\\
(\forall q\leq n \ ( (\forall m<q\ {\bf f}[m] = {\bf f}[q-1-m]) \ \wedge\
(\forall v < n-q\ {\bf f}[q+v] = {\bf f}[n-1-v]) ) \implies p=q ) . \end{gathered}$$
A result of Cummings, Moore, and Karhumäki states that the borders of the finite Fibonacci word ${\bf f}[0..F_n - 1]$ are precisely the words ${\bf f}[0..F_{n-2k} - 1]$ for $2k < n$. We can prove this, and more:
Consider the pairs $(n,m)$ such that $1 \leq m < n$ and ${\bf f}[0..m-1]$ is a border of ${\bf f}[0..n-1]$. Their Fibonacci representations are accepted by the automaton below in Figure \[borders\].
![Automaton encoding borders of prefixes of $\bf f$[]{data-label="borders"}](output_borders.pdf){width="3in"}
We use the predicate $$(n > m) \ \wedge \ (m \geq 1) \ \wedge \ \forall i<m \ {\bf f}[i] = {\bf f}[n-m+i] .$$ By following the paths with first coordinate of the form $10^+$ we recover the result of Cummings, Moore, and Karhumäki as a special case.
Avoiding the pattern $x x x^R$ and the Rote-Fibonacci word {#rotefib}
==========================================================
In this section we show how to apply our decision method to an interesting and novel avoidance property: avoiding the pattern $x x x^R$ . An example matching this pattern in English is a factor of the word [bepepper]{}, with $x = {\tt ep}$. Here, however, we are concerned only with the binary alphabet $\Sigma_2 = \lbrace 0, 1 \rbrace$.
Although avoiding patterns with reversal has been considered before (e.g., ), it seems our particular problem has not been studied.
If our goal is just to produce some infinite word avoiding $x x x^R$, then a solution seems easy: namely, the infinite word $(01)^\omega$ clearly avoids $x x x^R$, since if $|x| = n$ is odd, then the second factor of length $n$ cannot equal the first (since the first symbol differs), while if $|x| = n$ is even, the first symbol of the third factor of length $n$ cannot be the last symbol of $x$. In a moment we will see that even this question seems more subtle than it first appears, but for the moment, we’ll change our question to
*Are there infinite aperiodic binary words avoiding $x x x^R$?*
To answer this question, we’ll study a special infinite word, which we call the [*Rote-Fibonacci word*]{}. (The name comes from the fact that it is a special case of a class of words discussed in 1994 by Rote [@Rote:1994].) Consider the following transducer $T$:
(q\_0) [$q_0$]{}; (q\_1) \[right=of q\_0\] [$q_1$]{};
(q\_0.10) edge node [[0]{}/[00]{}, [1]{}/[0]{}]{} (q\_1.170) (q\_1.190) edge node [[0]{}/[11]{}, [1]{}/[1]{}]{} (q\_0.350);
This transducer acts on words by following the transitions and outputting the concatenation of the outputs associated with each transition. Thus, for example, the input $01001$ gets transduced to the output $00100110$.
The Rote-Fibonacci word $${\bf r} = 001001101101100100110110110010010011011001001001101100100100 \cdots = r_0 r_1 r_2 \cdots$$ has the following equivalent descriptions:
0\. As the output of the transducer $T$, starting in state $0$, on input $\bf f$.
1\. As $\tau(h^\omega(a))$ where $h$ and $\tau$ are defined by $$\begin{aligned}
h(a) &= a b_1 &\quad \tau(a) = 0 \\
h(b) &= a & \quad \tau(b) = 1 \\
h(a_0) &= a_2 b & \quad \tau(a_0) = 0 \\
h(a_1) &= a_0 b_0 & \quad \tau(a_1) = 1 \\
h(a_2) &= a_1 b_2 & \quad \tau(a_2) = 1 \\
h(b_0) &= a_0 & \quad \tau(b_0) = 0 \\
h(b_1) &= a_1 & \quad \tau(b_1) = 0 \\
h(b_2) &= a_2 & \quad \tau(b_2) = 1 \\\end{aligned}$$
2\. As the binary sequence generated by the following DFAO, with outputs given in the states, and inputs in the Fibonacci representation of $n$.
\(a) [$a/{\tt 0}$]{}; (b\_1) \[right=of a\] [$b_1/{\tt 0}$]{}; (a\_1) \[right=of b\_1\] [$a_1/{\tt 1}$]{}; (b\_0) \[right=of a\_1\] [$b_0/{\tt 0}$]{}; (b) \[right=of b\_0\] [$b/{\tt 1}$]{}; (a\_0) \[right=of b\] [$a_0/{\tt 0}$]{}; (a\_2) \[right=of a\_0\] [$a_2/{\tt 1}$]{}; (b\_2) \[right=of a\_2\] [$b_2/{\tt 1}$]{};
\(a) edge \[loop above\] node [0]{} () edge node [1]{} (b\_1) (b\_1) edge node [0]{} (a\_1) (a\_1) edge node [1]{} (b\_0) (a\_1.315) edge \[bend right=20\]node \[pos=0.1,swap\] [0]{} (a\_0.225) (b\_0.45) edge \[bend left=45\] node \[pos=0.5,swap\] [0]{} (a\_0.135) (b.210) edge \[bend left=24\] node \[pos=0.1,swap\] [0]{} (a\_1.330) (a\_0) edge node \[swap\] [1]{} (b) (a\_0) edge node \[swap\] [0]{} (a\_2) (a\_2.135) edge \[bend right=24\]node \[pos=0.1,swap\] [0]{} (a\_1.45) (a\_2.10) edge node [1]{} (b\_2.170) (b\_2.190) edge node [0]{} (a\_2.350);
3\. As the limit, as $n \rightarrow \infty$, of the sequence of finite Rote-Fibonacci words $(R_n)_n$ defined as follows: $R_0 = 0$, $R_1 = 00$, and for $n \geq 3$ $$R_n =
\begin{cases}
R_{n-1} R_{n-2}, & \text{ if $n \equiv 0$ (mod 3);} \\
R_{n-1} \overline{R_{n-2}}, & \text{ if $n \equiv 1, 2$ (mod 3).}
\end{cases}$$
4\. As the sequence obtained from the Fibonacci sequence ${\bf f} = f_0 f_1 f_2 \cdots = 0100101001001 \cdots$ as follows: first, change every $0$ to $1$ and every one to $0$ in ${\bf f}$, obtaining $\overline{\bf f} = 1011010110110 \cdots$. Next, in $\overline{\bf f}$ change every second $1$ that appears to $-1$ (which we write as $\mone$ for clarity): $1 0 \mone 1 0 \mone 0 1 \mone 0 1 \mone 0 \cdots$. Now take the running sum of this sequence, obtaining $1101100100100 \cdots$, and finally, complement it to get $\bf r$.
5\. As $\rho(g^\omega (a))$, where $g$ and $\rho$ are defined as follows $$\begin{aligned}
g(a) &= abcab \quad & \rho(a) = 0 \\
g(b) &= cda \quad & \rho(b) = 0 \\
g(c) &= cdacd \quad & \rho(c) = 1 \\
g(d) &= abc \quad & \rho(d) = 1\end{aligned}$$
$(0) \iff (3)$: Let $T_0 (x)$ (resp., $T_1 (x)$) denote the output of the transducer $T$ starting in state $q_0$ (resp., $q_1$) on input $x$. Then a simple induction on $n$ shows that $T_0 (X_{n+1}) = R_n$ and $T_1(X_{n+1}) = \overline{R_n}$. We give only the induction step for the first claim: $$\begin{aligned}
T_0 (X_{n+1}) &= T_0 (X_n X_{n-1}) \\
&= \begin{cases}
T_0 (X_n) T_0 (X_{n-1}), & \text{if $|X_n|$ is even}; \\
T_0 (X_n) T_1 (X_{n-1}), & \text{if $|X_n|$ is odd};
\end{cases} \\
&= \begin{cases}
R_{n-1} R_{n-2}, & \text{if $n \equiv 0$ (mod 3)}; \\
R_{n-1} \overline{R_{n-2}}, & \text{if $n \not\equiv 0$ (mod 3)};
\end{cases} \\
&= R_n .\end{aligned}$$ Here we have used the easily-verified fact that $|X_n|= F_n$ is even iff $n \equiv 0$ (mod $3$).
$(1) \iff (3)$: we verify by a tedious induction on $n$ that for $n \geq 0$ we have $$\begin{aligned}
\tau(h^n(a)) &= \tau(h^{n+1} (a)) = R_n \\
\tau(h^n(a_i)) &= \tau(h^{n+1} (b_i)) = \begin{cases}
R_i, & \text{if $n \equiv i$ (mod 3)}; \\
\overline{R_i}, & \text{if $n \not\equiv i$ (mod 3)}.
\end{cases}\end{aligned}$$
$(2) \iff (4)$: Follows from the well-known transformation from automata to morphisms and vice versa (see, e.g., ).
$(3) \iff (4)$: We define some transformations on sequences, as follows:
- $C(x)$ denotes $\overline{x}$, the complement of $x$;
- $s(x)$ denotes the sequence arising from a binary sequence $x$ by changing every second $1$ to $-1$;
- $a(x)$ denotes the running sum of the sequence $x$; that is, if $x = a_1 a_2 a_3 \cdots $ then $a(x)$ is $a_1 (a_1 + a_2) (a_1 + a_2 +a_3) \cdots$.
Note that $$a (s (xy)) =
\begin{cases}
a(s(x)) \ a(s(y)), & \text{if $|x|_1$ even}; \\
a(s(x)) \ C(a(s(y))), & \text{if $|x|_1$ odd}.
\end{cases}$$ Then we claim that $ C(R_n) = a(s(C(X_{n+2})))$. This can be verified by induction on $n$. We give only the induction step: $$\begin{aligned}
a(s(C(X_{n+2}))) &= a(s( C(X_{n+1}) C(X_{n}) )) \\
&= \begin{cases}
a(s(C(X_{n+1}))) \ a(s(C(X_{n}))), & \text{ if $C(X_{n+1})_1$ even}; \\
a(s(C(X_{n+1}))) \ C(a(s(C(X_{n})))), & \text{ if $C(X_{n+1})_1$ odd};
\end{cases} \\
&= \begin{cases}
C(R_{n-1}) \ C(R_{n-2}), & \text{ if $n \equiv 0$ (mod 3)}; \\
C(R_{n-1}) \ R_{n-2}, & \text{ if $n \not\equiv 0$ (mod 3)};
\end{cases} \\
&= R_{n}.\end{aligned}$$
$(3) \iff (5)$: Define $\gamma$ by $$\begin{aligned}
\gamma(a) &= \gamma(a_0) = a \\
\gamma(b_0) &= \gamma(b_1) = b \\
\gamma(a_1) &= \gamma(a_2) = c \\
\gamma(b) &= \gamma(b_2) = d .\end{aligned}$$ We verify by a tedious induction on $n$ that for $n \geq 0$ we have $$\begin{aligned}
g^n (a) &= \gamma(h^{3n} (a)) = \gamma(h^{3n} (a_0)) \\
g^n (b) &= \gamma(h^{3n} (b_0)) = \gamma(h^{3n} (b_1)) \\
g^n (c) &= \gamma(h^{3n} (a_1)) = \gamma(h^{3n} (a_2)) \\
g^n (d) &= \gamma(h^{3n} (b)) = \gamma(h^{3n} (b_2)) .\end{aligned}$$
The first differences $\Delta {\bf r}$ of the Rote-Fibonacci word $\bf r$, taken modulo $2$, give the complement of the Fibonacci word $\overline{f}$, with its first symbol omitted. \[rotecor\]
Note that if ${\bf x} = a_0 a_1 a_2 \cdots$ is a binary sequence, then $\Delta(C({\bf x})) = -\Delta({\bf x})$. Furthermore $\Delta(a(x)) = a_1 a_2 \cdots$. Now from the description in part 4, above, we know that ${\bf r} = C(a(s(C({\bf f}))))$. Hence $\Delta({\bf r}) = \Delta ( C(a(s(C({\bf f}))))) =
-\Delta( a(s(C({\bf f})))) = \operatorname{dr}(-s(C({\bf f})))$, where $\operatorname{dr}$ drops the first symbol of its argument. Taking the last result modulo $2$ gives the result.
We are now ready to prove our avoidability result.
The Rote-Fibonacci word $\bf r$ avoids the pattern $x x x^R$. \[rote-avoid-thm\]
We use our decision procedure to prove this. A predicate is as follows: $$\exists i\ \forall t<n\ ({\bf r}[i+t]={\bf r}[i+t+n]) \ \wedge \ ({\bf r}[i+t]={\bf r}[i+3n-1-t]) .$$ When we run this on our program, we get the following log:
t < n with 7 states, in 36ms
R[i + t] = R[i + t + n] with 245 states, in 1744ms
R[i + t] = R[i + 3 * n - 1 - t] with 1751 states, in 14461ms
R[i + t] = R[i + t + n] & R[i + t] = R[i + 3 * n - 1 - t] with 3305 states, in 565ms
t < n => R[i + t] = R[i + t + n] & R[i + t] = R[i + 3 * n - 1 - t] with 2015 states, in 843ms
At t < n => R[i + t] = R[i + t + n] & R[i + t] = R[i + 3 * n - 1 - t] with 3 states, in 747ms
Ei At t < n => R[i + t] = R[i + t + n] & R[i + t] = R[i + 3 * n - 1 - t] with 2 states, in 0ms
overall time: 18396ms
Then the only length $n$ accepted is $n = 0$, so the Rote-Fibonacci word $\bf r$ contains no occurrences of the pattern $x x x^R$.
We now prove some interesting properties of $\bf r$.
The minimum $q(n)$ over all periods of all length-$n$ factors of the Rote-Fibonacci word is as follows: $$q(n) =
\begin{cases}
1, & \text{if $1 \leq n \leq 2$;} \\
2, & \text{if $n = 3$;} \\
F_{3j+1}, & \text{if $j \geq 1$ and $L_{3j} \leq n < L_{3j+2}$;} \\
L_{3j+1}, & \text{if $j \geq 1$ and $L_{3j+2} \leq n < L_{3j+2}+F_{3j-2}$;} \\
F_{3j+2}+L_{3j}, & \text{if $j \geq 2$ and $L_{3j+2} + F_{3j-2} \leq n
< L_{3j+2}+ F_{3j-1}$;} \\
2F_{3j+2}, & \text{if $L_{3j+2}+F_{3j-1} \leq n < L_{3j+3}$} .
\end{cases}$$ \[rfperiods-thm\]
To prove this, we mimic the proof of Theorem \[allpers\]. The resulting automaton is displayed below in Figure \[leastper-rote\].
![Automaton accepting least periods of prefixes of length $n$[]{data-label="leastper-rote"}](output_least-period-over-all-factors-rote.pdf){width="4in"}
The critical exponent of the Rote-Fibonacci word is $2+\alpha$. \[critical-rote\]
An examination of the cases in Theorem \[rfperiods-thm\] show that the words of maximum exponent are those corresponding to $n = L_{3j+2}-1$, $p = F_{3j+1}$. As $j \rightarrow \infty$, the quantity $n/p$ approaches $2 + \alpha$ from below.
All squares in the Rote-Fibonacci word are of order $F_{3n+1}$ for $n \geq 0$, and each such order occurs. \[rote3n\]
We use the predicate $$(n \geq 1) \ \wedge \ \exists i \ \forall j<n\ ({\bf r}[i+j] =
{\bf r}[i+j+n]) .$$ The resulting automaton is depicted in Figure \[rotesquares\]. The accepted words correspond to $F_{3n+1}$ for $n \geq 0$.
![Automaton accepting orders of squares in the Rote-Fibonacci word[]{data-label="rotesquares"}](rotesquares.pdf){width="6.5in"}
We now turn to problems considering prefixes of the Rote-Fibonacci word $\bf r$.
A length-$n$ prefix of the Rote-Fibonacci word $\bf r$ is an antipalindrome iff $n = F_{3i+1} - 3$ for some $i \geq 1$.
We use our decision method on the predicate $$\forall j<n\ {\bf r}[j] \not= {\bf r}[n-1-j] .$$ The result is depicted in Figure \[rote-antipal\]. The only accepted expansions are given by the regular expression $\epsilon + 1(010101)^* 0 (010+101000)$, which corresponds to $F_{3j+1} - 3$. We use the predicate $$(n \geq 1) \ \wedge \ \exists i \ \forall j<n \ {\bf r}[i+j] =
{\bf r}[i+j+n]) .$$ The resulting automaton is depicted in Figure \[rote-antipal\]. The accepted words correspond to $F_{3n+1}$ for $n \geq 0$.
![Automaton accepting lengths of antipalindrome prefixes in the Rote-Fibonacci word[]{data-label="rote-antipal"}](rote-antipal-prefixes.pdf){width="6.5in"}
A length-$n$ prefix of the Rote-Fibonacci word is an antisquare if and only if $n = 2F_{3k+2}$ for some $k \geq 1$.
The predicate for having an antisquare prefix of length $n$ is $$\forall k < n \ {\bf r}[i+k] \not= {\bf r}[i+k+n] .$$ When we run this we get the automaton depicted in Figure \[rote-antisquare-prefix\].
![Automaton accepting orders of antisquares that are prefixes of $\bf f$[]{data-label="rote-antisquare-prefix"}](rote-antisquare-prefix.pdf){width="5.5in"}
The Rote-Fibonacci word has subword complexity $2n$.
Follows from Corollary \[rotecor\] together with [@Rote:1994 Thm. 3].
The Rote-Fibonacci word is mirror invariant. That is, if $z$ is a factor of $\bf r$ then so is $z^R$. \[rotemi\]
We use the predicate $$\forall i \ \exists j \ \forall t < n \
{\bf r}[i+t] = {\bf r}[j+n-1-t] .$$ The resulting automaton accepts all $n$, so the conclusion follows. The largest intermediate automaton has 2300 states and the calculation took about 6 seconds on a laptop.
The Rote-Fibonacci word avoids the pattern $x x^R x^R$.
Suppose $x x^R x^R$ occurs in $\bf r$. Then by Theorem \[rotemi\] we know that $(x x^R x^R)^R = x x x^R$ occurs in $\bf f$. But this is impossible, by Theorem \[rote-avoid-thm\].
As it turns out, the Rote-Fibonacci word has (essentially) appeared before in several places. For example, in a 2009 preprint of Monnerot-Dumaine [@Monnerot-Dumaine:2009], the author studies a plane fractal called the “Fibonacci word fractal”, specified by certain drawing instructions, which can be coded over the alphabet $S, R, L$ by taking the fixed point $g^\omega (a)$ and applying the coding $\gamma(a) = S$, $\gamma(b) = R$, $\gamma(c) = S$, and $\gamma(d) = L$. Here $S$ means “move straight one unit”, “$R$” means “right turn one unit” and “$L$” means “left turn one unit”.
More recently, Blondin Massé, Brlek, Labbé, and Mendès France studied a remarkable sequence of words closely related to $\bf r$ . For example, in their paper “Fibonacci snowflakes” they defined a certain sequence $q_i$ which has the following relationship to $g$: let $\xi(a) = \xi(b) = L$, $\xi(c) = \xi(d) = R$. Then $$R \xi(g^n(a)) = q_{3n+2} L .$$
Conjectures and open problems about the Rote-Fibonacci word
-----------------------------------------------------------
In this section we collect some conjectures we have not yet been able to prove. We have made some progress and hope to completely resolve them in the future.
Every infinite binary word avoiding the pattern $x x x^R$ has critical exponent $\geq 2+\alpha$.
Let $z$ be a finite nonempty primitive binary word. If $z^\omega$ avoids $x x x^R$, then $|z| = 2 F_{3n+2}$ for some integer $n \geq 0$. Furthermore, $z$ is a conjugate of the prefix ${\bf r}[0..2F_{3n+2} - 1]$, for some $n \geq 0$. Furthermore, for $n \geq 1$ we have that $z$ is a conjugate of $y \overline{y}$, where $y = \tau(h^{3n} (a))$.
We can make some partial progress on this conjecture, as follows:
Let $k \geq 1$ and define $n = 2F_{3k+2}$. Let $z = {\bf r}[0..n-1]$. Then $z^\omega$ contains no occurrence of the pattern $x x x^R$.
We have already seen this for $k = 0$, so assume $k \geq 1$.
Suppose that $z^\omega$ does indeed contain an occurrence of $x x x^R$ for some $|x| = \ell > 0$. We consider each possibility for $\ell$ and eliminate them in turn.
Case I: $\ell \geq n$.
There are two subcases:
Case Ia: $n \nodiv \ell$: In this case, by considering the first $n$ symbols of each of the two occurrences of $x$ in $x x x^R$ in $z^\omega$, we see that there are two different cyclic shifts of $z$ that are identical. This can only occur if ${\bf r}[0..n-1]$ is a power, and we know from Theorem \[rote3n\] and Corollary \[critical-rote\] that this implies that $n = 2F_{3k+1}$ or $n = 3F_{3k+1}$ for some $k \geq 0$. But $2F_{3k+1} \not= 2F_{3k'+2}$ and $3F_{3k+1} \not= 2F_{3k'+2}$ provided $k, k' > 0$, so this case cannot occur.
Case Ib: $n \divides \ell$: Then $x$ is a conjugate of $z^e$, where $e = \ell/n$. By a well-known result, a conjugate of a power is a power of a conjugate; hence there exists a conjugate $y$ of $z$ such that $x = y^e$. Then $x^R = y^e$, so $x$ and hence $y$ is a palindrome. We can now create a predicate that says that some conjugate of ${\bf r}[0..n-1]$ is a palindrome: $$\exists i<n \ \wedge\ (\forall j<n \ \operatorname{cmp}(i+j,n+i-1-j))$$ where $$\begin{gathered}
\operatorname{cmp}(k,k') :=
(((k<n) \ \wedge\ (k'<n)) \implies ({\bf r}[k] = {\bf r}[k'])) \ \wedge \ \\
(((k<n)\ \wedge \ (k' \geq n)) \implies ({\bf r}[k] = {\bf r}[k'-n])) \ \wedge \ \\
(((k \geq n)\ \wedge \ (k'<n)) \implies ({\bf r}[k-n] = {\bf r}[k'])) \ \wedge \ \\
(((k \geq n)\ \wedge \ (k' \geq n)) => ({\bf r}[k-n] = {\bf r}[k'-n])) .\end{gathered}$$
When we do this we discover the only $n$ with Fibonacci representation of the form $10010^i$ accepted are those with $i \equiv 0, 2$ (mod $3$), which means that $2F_{3k+2}$ is not among them. So this case cannot occur.
Case II: $\ell < n$.
There are now four subcases to consider, depending on the number of copies of $z$ needed to “cover” our occurrence of $x x x^R$. In Case II.$j$, for $1 \leq j \leq 4$, we consider $j$ copies of $z$ and the possible positions of $x x x^R$ inside that copy.
Because of the complicated nature of comparing one copy of $x$ to itself in the case that one or both overlaps a boundary between different copies of $z$, it would be very helpful to be able to encode statements like ${\bf r}[k \bmod n] = {\bf r}[\ell \bmod n]$ in our logical language. Unfortunately, we cannot do this if $n$ is arbitrary. So instead, we use a trick: assuming that the indices $k,
k'$ satisfy $0 \leq k, k' < 2n$, we can use the $\operatorname{cmp}(k,k')$ predicate introduced above to simulate the assertion ${\bf r}[k \bmod n] = {\bf r}[k'
\bmod n]$. Of course, for this to work we must ensure that $0 \leq k,
k' < 2n$ holds.
The cases are described in Figure \[rotecon\]. We assume that that $|x| = \ell$ and $x x x^R$ begins at position $i$ of $z^\omega$. We have the inequalities $i < n$ and $\ell < n$ which apply to each case. Our predicates are designed to compare the first copy of $x$ to the second copy of $x$, and the first copy of $x$ to the $x^R$.
![Cases of the argument[]{data-label="rotecon"}](rotecases2.pdf){width="5in"}
Case 1: If $xxx^R$ lies entirely within one copy of $z$, it also lies in $\bf r$, which we have already seen cannot happen, in Theorem \[rote-avoid-thm\]. This case therefore cannot occur.
Case 2: We use the predicate $$\exists i \ \exists \ell \
(i+3\ell \geq n) \ \wedge \ (i+3\ell < 2n) \ \wedge \
(\forall j < \ell\ \operatorname{cmp}(i+j, i+\ell+j ) ) \ \wedge \
(\forall k < \ell\ \operatorname{cmp}(i+k,i+3\ell-1-k) )$$ to assert that there is a repetition of the form $x x x^R$.
Case 3: We use the predicate $$\exists i \ \exists \ell \
(i + 3\ell \geq 2n) \ \wedge \ (i+3\ell < 3n) \ \wedge \
(\forall j < \ell\ \operatorname{cmp}(i+j, i+\ell+j-n) ) \ \wedge \
(\forall k < \ell\ \operatorname{cmp}(i+k,i+3\ell-1-k-n) ) ) .$$
Case 4: We use the predicate $$\exists i \ \exists \ell \
(i+3 \ell \geq 3n) \ \wedge \ (i+3\ell < 4n) \ \wedge \
(\forall j < \ell\ \operatorname{cmp}(i+j, i+\ell+j-n) ) \ \wedge \
(\forall k < \ell\ \operatorname{cmp}(i+k, i+3\ell-1-k-2n ) ) .$$
When we checked each of the cases 2 through 4 with our program, we discovered that $n = 2F_{3k+2}$ is never accepted. Actually, for cases (2)–(4) we had to employ one additional trick, because the computation for the predicates as stated required more space than was available on our machine. Here is the additional trick: instead of attempting to run the predicate for all $n$, we ran it only for $n$ whose Fibonacci representation was of the form $10010^*$. This significantly restricted the size of the automata we created and allowed the computation to terminate. In fact, we propagated this condition throughout the predicate.
We therefore eliminated all possibilities for the occurrence of $x x x^R$ in $z^\omega$ and so it follows that no $x x x^R$ occurs in $z^\omega$.
How many binary words of length $n$ avoid the pattern $x x x^R$? Is it polynomial in $n$ or exponential? How about the number of binary words of length $n$ avoiding $x x x^R$ and simultaneously avoiding $(2+\alpha)$-powers?
Consider finite words of the form $x x x^R$ having no proper factor of the form $w w w^R$.
For $n = F_{3k+1}$ there are $4$ such words of length $n$. For $n = F_{3k+1} \pm F_{3k-2}$ there are $2$ such words. Otherwise there are none.
For $k \geq 3$ the $4$ words of length $n = F_{3k+1}$ are given by ${\bf r}[p_i..p_i+n-1]$, $i = 1,2,3,4$, where $$\begin{aligned}
(p_1)_F &= 1000 (010)^{k-3} 001 \\
(p_2)_F &= 10 (010)^{k-2} 001 \\
(p_3)_F &= 1001000 (010)^{k-3} 001 \\
(p_4)_F &= 1010 (010)^{k-2} 001 \end{aligned}$$
For $k \geq 3$ the $2$ words of length $n = F_{3k+1}-F_{3k-2}$ are given by ${\bf r}[q_i..q_i+n-1]$, $i = 1,2$, where $$\begin{aligned}
(q_1)_F &= 10 (010)^{k-3} 001 \\
(q_2)_F &= 10000 (010)^{k-3} 001 \end{aligned}$$
For $k \geq 3$ the $2$ words of length $n = F_{3k+1}+F_{3k-2}$ are given by ${\bf r}[s_i..s_i+n-1]$, $i = 1,2$, where $$\begin{aligned}
(s_1)_F &= 10 (010)^{k-3} 001 \\
(s_2)_F &= 1000 (01)^{k-2} 001\end{aligned}$$
Other sequences {#other}
===============
In this section we briefly apply our method to some other Fibonacci-automatic sequences, obtaining several new results.
Consider a Fibonacci analogue of the Thue-Morse sequence $${\bf v} = (v_n)_{n \geq 0} = 0111010010001100010111000101 \cdots$$ where $v_n$ is the sum of the bits, taken modulo $2$, of the Fibonacci representation of $n$. This sequence was introduced in [@Shallit:1988a Example 2, pp. 12–13].
We recall that an [*overlap*]{} is a word of the form $axaxa$ where $x$ may be empty; its order is defined to be $|ax|$. Similarly, a [*super-overlap*]{} is a word of the form $abxabxab$; an example of a super-overlap in English is the word [tingalingaling]{} with the first letter removed.
The only squares in $\bf v$ are of order $4$ and $F_n$ for $n \geq 2$, and a square of each such order occurs. The only cubes in $\bf v$ are the strings $000$ and $111$. The only overlaps in $\bf v$ are of order $F_{2n}$ for $n \geq 1$, and an overlap of each such order occurs. There are no super-overlaps in $\bf v$.
As before. We omit the details.
We might also like to show that $\bf v$ is recurrent. The obvious predicate for this property holding for all words of length $n$ is $$\forall i\ \exists j\ ((j>i) \wedge ( \forall t \ ((t<n) \implies
({\bf v}[i+t]={\bf v}[j+t])))) .$$ Unfortunately, when we attempt to run this with our prover, we get an intermediate NFA of 1159 states that we cannot determinize within the available space.
Instead, we rewrite the predicate, setting $k := j-i$ and $u := i+t$. This gives $$\forall i\ \exists j \ (j>i) \wedge
\forall k \ \forall u \
((k \geq 1) \wedge (i=j+k) \wedge (u \geq i) \wedge (u < n+i))
\implies {\bf v}[u]={\bf v}[u+k] .$$ When we run this we discover that $\bf v$ is indeed recurrent. Here the computation takes a nontrivial 814007 ms, and the largest intermediate automaton has 625176 states. This proves
The word $\bf v$ is recurrent.
Another quantity of interest for the Thue-Morse-Fibonacci word $\bf v$ is its subword complexity $\rho_{\bf v}(n)$. It is not hard to see that it is linear. To obtain a deeper understanding of it, let us compute the first difference sequence $d(n) = \rho_{\bf v}(n+1) - \rho_{\bf v}(n)$. It is easy to see that $d(n)$ is the number of words $w$ of length $n$ with the property that both $w0$ and $w1$ appear in $\bf v$. The natural way to count this is to count those $i$ such that $t:= {\bf v}[i..i+n-1]$ is the first appearance of that factor in $\bf v$, and there exists a factor ${\bf v}[k..k+n]$ of length $n+1$ whose length-$n$-prefix equals $t$ and whose last letter ${\bf v}[k+n]$ differs from ${\bf v}[i+n]$. $$(\forall j<i \ \exists t<n \ {\bf v}[i+t] \not= {\bf v}[j+t]) \ \wedge \
(\exists k\ (\forall u <n\ {\bf v}[i+u]={\bf v}[k+u]) \wedge
{\bf v}[i+n] \not= {\bf v}[k+n]).$$ Unfortunately the same blowup appears as in the recurrence predicate, so once agin we need to substitute, resulting in the predicate
$$\begin{gathered}
(\forall j<i \ \exists k\geq 1\ \exists v\
(i=j+k) \wedge (v \geq j) \wedge (v<n+j) \wedge {\bf v}[u] \not= {\bf v}[u+k] )
\wedge \\
(\exists l>i \ {\bf v}[i+n] \not= {\bf v}[l+n] ) \wedge \\
(\forall k' \ \forall u' \
(k'\geq 1) \wedge (l = i+k') \wedge (u' \geq i) \wedge (v' < n+i)
\implies {\bf v}[k'+u']={\bf v}[u'] ) .\end{gathered}$$
From this we obtain a linear representation of rank $46$. We can now consider all vectors of the form $u \{ M_0, M_1 \}^*$. There are only finitely many and we can construct an automaton out of them computing $d(n)$.
The first difference sequence $(d(n))_{n \geq 0}$ of the subword complexity of $\bf v$ is Fibonacci-automatic, and is accepted by the following machine.
![Automaton computing $d(n)$[]{data-label="tmf-specialf"}](tmf-specialf.pdf){width="6.5in"}
Combining two representations and avoidability {#additive}
==============================================
In this section we show how our decidability method can be used to handle an avoidability question where two different representations arise.
Let $x$ be a finite word over the alphabet $\Enn^* = \lbrace 1, 2, 3 \ldots \rbrace$. We say that $x$ is an [*additive square*]{} if $x = x_1 x_2$ with $|x_1| = |x_n|$ and $\sum x_1 = \sum x_2$. For example, with the usual association of ${\tt a} = 1$, ${\tt b} = 2$, and so forth, up to ${\tt z} = 26$, we have that the English word [baseball]{} is an additive square, as [base]{} and [ball]{} both sum to $27$.
An infinite word ${\bf x}$ over $\Enn^*$ is said to [*avoid additive squares*]{} if no factor is an additive square. It is currently unknown, and a relatively famous open problem, whether there exists an infinite word over a [*finite*]{} subset of $\Enn^*$ that avoids additive squares .., although it is known that additive cubes can be avoided over an alphabet of size $4$ . (Recently this was improved to alphabet size $3$; see [@Rao:2013].)
However, it is easy to avoid additive squares over an [*infinite*]{} subset of $\Enn^*$; for example, any sequence that grows sufficiently quickly will have the desired property. Hence it is reasonable to ask about the [*lexicographically least*]{} sequence over $\Enn^*$ that avoids additive squares. Such a sequence begins $$1 2 1 3 1 2 1 4 2 1 2 5 2 1 3 1 2 1 3 4 1 2 1 7 2 \cdots ,$$ but we do not even know if this sequence is unbounded.
Here we consider the following variation on this problem. Instead of considering arbitrary sequences, we start with a sequence ${\bf b} = b_0 b_1 b_2 \cdots$ over $\Enn^+$ and from it construct the sequence $S({\bf b}) = a_1 a_2 a_3 \cdots$ defined by $${\bf a}[i] = {\bf b}[\nu_2 (i)]$$ for $i \geq 1$, where $\nu_2(i)$ is the exponent of the largest power of $2$ dividing $i$. (Note that ${\bf a}$ and ${\bf b}$ are indexed differently.) For example, if ${\bf b} = 123\cdots$, then ${\bf a} = 1213121412131215 \cdots$, the so-called “ruler sequence”. It is known that this sequence is squarefree and is, in fact, the lexicographically least sequence over $\Enn^*$ avoiding squares .
We then ask: what is the lexicographically least sequence avoiding additive squares that is of the form $S({\bf b})$? The following theorem gives the answer.
\[thm:lex-least-add-sq\] The lexicographically least sequence over $\Enn {\setminus}\{0\}$ of the form $S({\bf b})$ that avoids additive squares is defined by ${\bf
b}[i] {\mathrel{\mathop:}=}F_{i+2}$. \[additive-thm\]
First, we show that ${\bf a} {\mathrel{\mathop:}=}S({\bf b}) = \prod_{k=1}^\infty {\bf b}[\nu_2(k)] = \prod_{k=1}^\infty F_{\nu_2(k)+2}$ avoids additive squares.
For $m,n,j \in \Enn$, let $A(m,n,j)$ denote the number of occurrences of $j$ in $\nu_2(m+1), \dots, \nu_2(m+n)$.
(a): Consider two consecutive blocks of the same size say $a_{i+1} \cdots a_{i+n}$ and $a_{i+n+1} \cdots a_{i+2n}$. Our goal is to compare the sums $\sum_{i < j \leq i+n} a_j$ and $\sum_{i+n < j \leq i+2n} a_j$.
First we prove
Let $m,j \geq 0$ and $n \geq 1$ be integers. Let $A(m, n,j)$ denote the number of occurrences of $j$ in $\nu_2 (m+1), \ldots, \nu_2 (m+n)$. Then for all $m, m' \geq 0$ we have $|A(m', n, j) - A(m, n, j)| \leq 1$. \[flemm\]
We start by observing that the number of positive integers $\leq n$ that are divisible by $t$ is exactly $\lfloor n/t \rfloor$. It follows that the number $B(n,j)$ of positive integers $\leq n$ that are divisible by $2^j$ but not by $2^{j+1}$ is $$B(n,j) = \lfloor {n \over {2^j}} \rfloor - \lfloor {n \over {2^{j+1}}} \rfloor .
\label{fl}$$ Now from the well-known identity $$\lfloor x \rfloor + \lfloor x + {1 \over 2} \rfloor = \lfloor 2x \rfloor,$$ valid for all real numbers $x$, substitute $x = n/2^{j+1}$ to get $$\lfloor {n \over {2^{j+1}}} \rfloor +
\lfloor {n \over {2^{j+1}}} + {1 \over 2} \rfloor = \lfloor {n \over {2^j}} \rfloor ,$$ which, combined with , shows that $$B(n,j) = \lfloor {n \over {2^{j+1}}} + {1 \over 2} \rfloor .$$ Hence $${n \over {2^{j+1}}} - {1 \over 2} \leq B(n,j) <
{n \over {2^{j+1}}} + {1 \over 2} .
\label{flooreq}$$
Now the number of occurrences of $j$ in $\nu_2(m+1), \ldots, \nu_2(m+n)$ is $A(m,n,j) = B(m+n,j)-B(m,j)$. From we get $${n \over {2^{j+1}}} - 1 < A(m,n,j) < {n \over {2^{j+1}}} + 1
\label{flreq2}$$ for all $m \geq 0$. Since $A(m,n,j)$ is an integer, the inequality implies that $|A(m',n,j)-A(m,n,j)| \leq 1$ for all $m, m'$.
Note that for all $i,n \in \Enn$, we have $\sum_{k=i}^{i+n-1} {\bf a}[k] = \sum_{j=0}^{{\left\lfloor\log_2(i+n)\right\rfloor}} A(i,n,j) F_{j+2}$, so for adjacent blocks of length $n$, $\sum_{k=i+n}^{i+2n-1} {\bf a}[k] - \sum_{k=i}^{i+n-1} {\bf a}[k] = \sum_{j=0}^{{\left\lfloor\log_2(i+2n)\right\rfloor}} (A(i+n,n,j)-A(i,n,j)) F_{j+2}$. Hence, ${\bf a}[i {\ldotp\ldotp}i+2n-1]$ is an additive square iff $\sum_{j=0}^{{\left\lfloor\log_2(i+2n)\right\rfloor}} (A(i+n,n,j)-A(i,n,j)) F_{j+2} = 0$, and by above, each $A(i+n,n,j)-A(i,n,j) \in \{-1,0,1\}$.
The above suggests that we can take advantage of “unnormalized” Fibonacci representation in our computations. For $\Sigma \subseteq \Zee$ and $w \in \Sigma^*$, we let the unnormalized Fibonacci representation ${\left\langlew\right\rangle}_{uF}$ be defined in the same way as ${\left\langlew\right\rangle}_F$, except over the alphabet $\Sigma$.
In order to use Procedure \[proc:Fib-auto-decide\], we need two auxiliary DFAs: one that, given $i,n \in \Enn$ (in any representation; we found that base 2 works), computes ${\left\langleA(i+n,n,\_)-A(i,n,\_)\right\rangle}_{uF}$, and another that, given $w \in \{{\tt -1},{\tt 0},{\tt 1}\}^*$, decides whether ${\left\langlew\right\rangle}_{uF} = 0$. The first task can be done by a 6-state (incomplete) DFA $M_\text{add22F}$ that accepts the language $\{ z \in (\Sigma_2^2 \times \{{\tt -1},{\tt 0},{\tt 1}\})^* {\;:\;}\forall j (\pi_3(z)[j] = A({\left\langle\pi_1(z)\right\rangle}_2+{\left\langle\pi_2(z)\right\rangle}_2,{\left\langle\pi_2(z)\right\rangle}_2,j) - A({\left\langle\pi_1(z)\right\rangle}_2,{\left\langle\pi_2(z)\right\rangle}_2,j))\}$. The second task can be done by a 5-state (incomplete) DFA $M_\text{1uFisZero}$ that accepts the language $\{ w \in \{{\tt -1},{\tt 0},{\tt 1}\}^* {\;:\;}{\left\langlew\right\rangle}_{uF} = 0 \}$.
We applied a modified Procedure \[proc:Fib-auto-decide\] to the predicate $n \geq 1 {\wedge}\exists w ({\tt add22F}(i,n,w) {\wedge}{\tt 1uFisZero}(w))$ and obtained as output a DFA that accepts nothing, so ${\bf a}$ avoids additive squares.
Next, we show that ${\bf a}$ is the lexicographically least sequence over $\Enn {\setminus}\{0\}$ of the form $S({\bf b})$ that avoids additive squares.
Note that for all ${\bf x},{\bf y} \in \Enn {\setminus}\{0\}$, $S({\bf x}) < S({\bf y})$ iff ${\bf x} < {\bf y}$ in the lexicographic ordering. Thus, we show that if any entry ${\bf b}[s]$ with ${\bf b}[s] > 1$ is changed to some $t \in [1,{\bf b}[s]-1]$, then ${\bf a} = S({\bf b})$ contains an additive square using only the first occurrence of the change at ${\bf a}[2^s-1]$. More precisely, we show that for all $s,t \in \Enn$ with $t \in [1,F_{s+2}-1]$, there exist $i,n \in \Enn$ with $n \geq 1$ and $i+2n < 2^{s+1}$ such that either ($2^s-1 \in [i,i+n-1]$ and $\sum_{k=i+n}^{i+2n-1} {\bf a}[k] - \sum_{k=i}^{i+n-1} {\bf a}[k] + t = 0$) or ($2^s-1 \in [i+n,i+2n-1]$ and $\sum_{k=i+n}^{i+2n-1} {\bf a}[k] - \sum_{k=i}^{i+n-1} {\bf a}[k] - t = 0$).
Setting up for a modified Procedure \[proc:Fib-auto-decide\], we use the following predicate, which says “$r$ is a power of $2$ and changing ${\bf a}[r-1]$ to any smaller number results in an additive square in the first $2r$ positions", and six auxiliary DFAs. Note that all arithmetic and comparisons are in base 2. $$\begin{aligned}
&{\tt powOf2}(r) {\wedge}\forall t ((t \geq 1 {\wedge}t<r {\wedge}{\tt canonFib}(t)) {\rightarrow}\exists i \exists n (n \geq 1 {\wedge}i+2n < 2r {\wedge}{} \\
&\quad ((i<r {\wedge}r\leq i+n {\wedge}\forall w ({\tt add22F}(i,n,w) {\rightarrow}\forall x ({\tt bitAdd}(t,w,x) {\rightarrow}{\tt 2uFisZero}(x)))) {\vee}{} \\
&\quad \hphantom{(}(i+n<r {\wedge}r \leq i+2n {\wedge}\forall w ({\tt add22F}(i,n,w) {\rightarrow}\forall x ({\tt bitSub}(t,w,x) {\rightarrow}{\tt 2uFisZero}(x))))))).\end{aligned}$$ $$\begin{aligned}
L(M_\text{powOf2}) &= \{w \in \Sigma_2^* {\;:\;}\exists n (w=(2^n)_2)\}. \\
L(M_\text{canonFib}) &= \{w \in \Sigma_2^* {\;:\;}\exists n (w=(n)_F)\}. \\
L(M_\text{bit(Add/Sub)}) &= \{z \in (\Sigma_2 \times \{{\tt -1},{\tt 0},{\tt 1}\} \times \{{\tt -1},{\tt 0},{\tt 1},{\tt 2}\})^* {\;:\;}\forall i (\pi_1(z)[i] \pm \pi_2(z)[i] = \pi_3(z)[i]) \}. \\
L(M_\text{2uFisZero}) &= \{w \in \{{\tt -1},{\tt 0},{\tt 1},{\tt 2}\}^* {\;:\;}{\left\langlew\right\rangle}_{uF} = 0\}.\end{aligned}$$ We applied a modified Procedure \[proc:Fib-auto-decide\] to the above predicate and auxiliary DFAs and obtained as output $M_\text{powOf2}$, so ${\bf a}$ is the lexicographically least sequence over $\Enn {\setminus}\{0\}$ of the form $S({\bf b})$ that avoids additive squares.
Enumeration {#enumer}
===========
Mimicking the base-$k$ ideas in , we can also mechanically enumerate many aspects of Fibonacci-automatic sequences. We do this by encoding the factors having the property in terms of paths of an automaton. This gives the concept of [*Fibonacci-regular sequence*]{} as previously studied in . Roughly speaking, a sequence $(a(n))_{n \geq 0}$ taking values in $\Enn$ is Fibonacci-regular if the set of sequences $$\{ (a([xw]_F)_{w \in \Sigma_2^*} \ : \ x \in \Sigma_2^* \}$$ is finitely generated. Here we assume that $a([xw]_F)$ evaluates to $0$ if $xw$ contains the string $11$. Every Fibonacci-regular sequence $(a(n))_{n \geq 0}$ has a [*linear representation*]{} of the form $(u, \mu, v)$ where $u$ and $v$ are row and column vectors, respectively, and $\mu:\Sigma_2 \rightarrow \Enn^{d \times d}$ is a matrix-valued morphism, where $\mu(0) = M_0$ and $\mu(1) = M_1$ are $d \times d$ matrices for some $d \geq 1$, such that $$a(n) = u \cdot \mu(x) \cdot v$$ whenever $[x]_F = n$. The [*rank*]{} of the representation is the integer $d$. As an example, we exhibit a rank-$6$ linear representation for the sequence $a(n) = n+1$: $$\begin{aligned}
u &= [1 \ 2 \ 2 \ 3 \ 3 \ 2] \\
M_0 &= \left[ \begin{array}{cccccc}
1 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 1 & 1 & 0 \\
0 & 0 & 1 & 1 & 1 & 1 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0
\end{array} \right ] \\
M_1 &= \left[ \begin{array}{cccccc}
0 & 0 & 0 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0
\end{array} \right ] \\
v &= [1 \ 0 \ 0 \ 0 \ 0 \ 0 ]^T .\end{aligned}$$ This can be proved by a simple induction on the claim that $$u \cdot \mu(x) = [ x_F + 1 \ (1x)_F + 1 \ (10x)_F - x_F
\ (100x)_F - x_F \ (101x)_F - (1x)_F \ (1001x)_F - (101x)_F ]$$ for strings $x$.
Recall that if $\bf x$ is an infinite word, then the subword complexity function $\rho_{\bf x} (n)$ counts the number of distinct factors of length $n$. Then, in analogy with , we have
If $\bf x$ is Fibonacci-automatic, then the subword complexity function of $\bf x$ is Fibonacci-regular.
Using our implementation, we can obtain a linear representation of the subword complexity function for $\bf f$. To do so, we use the predicate $$\{ (n,i)_F \ : \ \forall i' < i \ {\bf f}[i..i+n-1] \not=
{\bf f}[i'..i'+n-1] \} ,$$ which expresses the assertion that the factor of length $n$ beginning at position $i$ has never appeared before. Then, for each $n$, the number of corresponding $i$ gives $\rho_{\bf f}(n)$. When we do this for $\bf f$, we get the following linear representation $(u', \mu', v')$ of rank $10$: $$\begin{aligned}
u' &= [0 \ 0 \ 0 \ 1 \ 0 \ 0 \ 0 \ 0 \ 0\ 0] \\
M'_0 &= \left[ \begin{array}{ccccccccccc}
0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0\\
1 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 1 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 1 & 0 & 0\\
1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0
\end{array}
\right] \\
M'_1 &= \left[ \begin{array}{ccccccccccc}
0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0
\end{array}
\right] \\
v' &= [1\ 0\ 1\ 1\ 1\ 1\ 1\ 1\ 1\ 1]^T \end{aligned}$$
To show that this computes the function $n+1$, it suffices to compare the values of the linear representations $(u, \mu, v)$ and $(u', \mu', v')$ for all strings of length $\leq 10 + 6 = 16$ (using ). After checking this, we have reproved the following classic theorem of Morse and Hedlund :
The subword complexity function of $\bf f$ is $n+1$. \[sturmcomp\]
We now turn to a result of Fraenkel and Simpson . They computed the exact number of squares appearing in the finite Fibonacci words $X_n$; this was previously estimated by [@Crochemore:1981].
There are two variations: we could count the number of distinct squares in $X_n$, or what Fraenkel and Simpson called the number of “repeated squares” in $X_n$ (i.e., the total number of [*occurrences*]{} of squares in $X_n$).
To solve this using our approach, we generalize the problem to consider any length-$n$ prefix of $X_n$, and not simply the prefixes of length $F_n$.
We can easily write down predicates for these. The first represents the number of distinct squares in ${\bf f}[0..n-1]$: $$\begin{gathered}
L_{\rm ds} :=
\{ (n,i,j)_F \ : \ (j \geq 1) \text{ and } (i+2j \leq n) \text{ and }
{\bf f}[i..i+j-1] = {\bf f}[i+j..i+2j-1] \\
\text{ and } \forall i' < i \
{\bf f}[i'..i'+2j-1] \not= {\bf f}[i..i+2j-1] \} .\end{gathered}$$ This predicate asserts that ${\bf f}[i..i+2j-1]$ is a square occurring in ${\bf f}[0..n-1]$ and that furthermore it is the first occurrence of this particular string in ${\bf f}[0..n-1]$.
The second represents the total number of occurrences of squares in ${\bf f}[0..n-1]$: $$L_{\rm dos} := \{ (n,i,j)_F \ : \ (j \geq 1) \text{ and }
(i+2j \leq n) \text{ and }
{\bf f}[i..i+j-1] = {\bf f}[i+j..i+2j-1] \} .$$ This predicate asserts that ${\bf f}[i..i+2j-1]$ is a square occurring in ${\bf f}[0..n-1]$.
We apply our method to the second example, leaving the first to the reader. Let $b(n)$ denote the number of occurrences of squares in ${\bf f}[0..n-1]$. First, we use our method to find a DFA $M$ accepting $L_{\rm dos}$. This (incomplete) DFA has 27 states.
Next, we compute matrices $M_0$ and $M_1$, indexed by states of $M$, such that $(M_a)_{k,l}$ counts the number of edges (corresponding to the variables $i$ and $j$) from state $k$ to state $l$ on the digit $a$ of $n$. We also compute a vector $u$ corresponding to the initial state of $M$ and a vector $v$ corresponding to the final states of $M$. This gives us the following linear representation of the sequence $b(n)$: if $x = a_1 a_2 \cdots a_t$ is the Fibonacci representation of $n$, then $$b(n) = u M_{a_1} \cdots M_{a_t} v ,
\label{linrep}$$ which, incidentally, gives a fast algorithm for computing $b(n)$ for any $n$.
Now let $B(n)$ denote the number of square occurrences in the finite Fibonacci word $X_n$. This corresponds to considering the Fibonacci representation of the form $10^{n-2}$; that is, $B(n+1) = b([10^{n-1}]_F)$. The matrix $M_0$ is the following $27 \times 27$ array $$\left[
\begin{array}{ccccccccccccccccccccccccccc}
1&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0 \\
1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0 \\
1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1 \\
1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0 \\
0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0 \\
1&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0 \\
1&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&1&0&0&1&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&1&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0 \\
1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&1&0&0&1&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1 \\
0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0 \\
0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&1&0&0&0&0&0&0
\end{array}
\right]$$ and has minimal polynomial $$X^4 (X-1)^2(X+1)^2(X^2-X-1)^2.$$ It now follows from the theory of linear recurrences that there are constants $c_1, c_2, \ldots, c_8$ such that $$B(n+1) = (c_1n + c_2) \alpha^n + (c_3n+c_4) \beta^n + c_5n+c_6 +
(c_7n+c_8)(-1)^n$$ for $n \geq 3$, where $\alpha = (1+\sqrt{5})/2$, $\beta = (1-\sqrt{5})/2$ are the roots of $X^2 - X - 1$. We can find these constants by computing $B(4), B(5), \ldots, B(11)$ (using Eq. ) and then solving for the values of the constants $c_1, \ldots, c_8$.
When we do so, we find $$\begin{aligned}
c_1 &= {2 \over 5} \quad\quad & c_2 &= {-{2\over{25}}}\sqrt{5} - 2 \\
c_3 &= {2 \over 5} \quad\quad & c_4 &= {{2\over{25}}}\sqrt{5} - 2 \\
c_5 &= 1 \quad\quad & c_6 &= 1 \\
c_7 &= 0 \quad\quad & c_8 &= 0\end{aligned}$$
A little simplification, using the fact that $F_n = (\alpha^n - \beta^n)/(\alpha - \beta)$, leads to
Let $B(n)$ denote the number of square occurrences in $X_n$. Then $$B(n+1) = {4 \over 5} n F_{n+1} - {2 \over 5} (n+6) F_{n} - 4F_{n-1} + n + 1$$ for $n \geq 3$.
This statement corrects a small error in Theorem 2 in (the coefficient of $F_{n-1}$ was wrong; note that their $F$ and their Fibonacci words are indexed differently from ours), which was first pointed out to us by Kalle Saari.
In a similar way, we can count the cube occurrences in $X_n$. Using analysis exactly like the square case, we easily find
Let $C(n)$ denote the number of cube occurrences in the Fibonacci word $X_n$. Then for $n \geq 3$ we have $$C(n) = (d_1 n+ d_2) \alpha^n + (d_3 n+d_4) \beta^n + d_5n + d_6$$ where $$\begin{aligned}
d_1 &= {{3-\sqrt{5}}\over {10}} \quad\quad & d_2 &= {{17}\over {50}}\sqrt{5} - {3 \over 2} \\
d_3 &= {{3+\sqrt{5}}\over {10}} \quad\quad & d_4 &= -{{17}\over {50}}\sqrt{5} - {3 \over 2} \\
d_5 &= 1 \quad\quad & d_6 &= -1 .\end{aligned}$$
We now turn to a question of Chuan and Droubay. Let us consider the prefixes of $\bf f$. For each prefix of length $n$, form all of its $n$ shifts, and let us count the number of these shifts that are palindromes; call this number $d(n)$. (Note that in the case where a prefix is a power, two different shifts could be identical; we count these with multiplicity.)
Chuan [@Chuan:1993b Thm. 7, p. 254] proved
For $i > 2$ we have $d(F_i) = 0$ iff $i \equiv \modd{0} {3}$. \[chuan93-thm\]
Along the way we actually prove a lot more, characterizing $d(n)$ for all $n$, not just those $n$ equal to a Fibonacci number.
We start by showing that $d(n)$ takes only three values: $0$, $1$, and $2$. To do this, we construct an automaton accepting the language $$\{ (n,i)_F \ : \ (0 \leq i < n) \ \wedge\ {\bf f}[i..n-1]{\bf f}[0..i-1]
\text{ is a palindrome } \} .$$ From this we construct the linear representation $(u, M_0, M_1, v)$ of $d(n)$ as discussed above; it has rank $27$.
The range of $c$ is finite if the monoid ${\cal M} = \langle M_0, M_1 \rangle$ is finite. This can be checked with a simple queue-based algorithm, and $\cal M$ turns out to have cardinality $151$. From these a simple computation proves $$\lbrace uMv \ : \ M \in {\cal M} \rbrace
= \lbrace 0, 1, 2 \rbrace,$$ and so our claim about the range of $c$ follows.
Now that we know the range of $c$ we can create predicates $P_0(n), P_1(n), P_2(n)$ asserting that (a) there are no length-$n$ shifts that are palindromes (b) there is exactly one shift that is a palindrome and (c) more than one shift is a palindrome, as follows: $$P_0 : \neg \exists i, (0 \leq i < n), {\bf f}[i..n-1]{\bf f}[0..i-1]
\text{ is a palindrome }$$ $$P_1 : \exists i, (0 \leq i < n), {\bf f}[i..n-1]{\bf f}[0..i-1]
\text{ is a palindrome and } \neg\exists j \not= i
(0 \leq j < n), {\bf f}[j..n-1]{\bf f}[0..j-1]$$ $$P_2 : \exists i, j, 0 \leq i < j < n
{\bf f}[i..n-1]{\bf f}[0..i-1] \text{ and }
{\bf f}[j..n-1]{\bf f}[0..j-1] \text{ are both palindromes }$$ For each one, we can compute a finite automaton characterizing the Fibonacci representations of those $n$ for which $d(n)$ equals, respectively, $0$, $1$, and $2$.
For example, we computed the automaton corresponding to $P_0$, and it is displayed in Figure \[noshifts\] below.
![Automaton accepting lengths of prefixes for which no shifts are palindromes[]{data-label="noshifts"}](output_no-shifts-are-pals.pdf){width="4in"}
By tracing the path labeled $10^*$ starting at the initial state labeled $18$, we see that the “finality” of the states encountered is ultimately periodic with period $3$, proving Theorem \[chuan93-thm\].
To finish this section, we reprove a result of Kolpakov and Kucherov . Recalling the definition of maximal repetition from Section \[repe-subsec\], they counted the number $\operatorname{mr}(F_n)$ of occurrences of maximal repetitions in the prefix of $\bf f$ of length $F_n$:
For $n \geq 5$ we have $\operatorname{mr}(F_n) = 2F_{n-2} - 3$.
We create an automaton for the language $$\lbrace (n,i,j)_F \ : \ 0 \leq i \leq j < n \text{ and } {\bf f}[i..j]
\text{ is a maximal repetition of } {\bf f}[0..n-1] \rbrace ,$$ using the predicate $$\begin{gathered}
(i \leq j) \ \wedge\ (j<n)\ \wedge \
\exists p \text{ with } 1 \leq p \leq (j+1-i)/2 \text{ such that } \\
( (\forall k\leq j-i-p \ {\bf f}[i+k]= {\bf f}[i+k+p]) \ \wedge \ \\
(i \geq 1) \implies (\forall q \text{ with } 1 \leq q \leq p \
\exists \ell \leq j-i-q+1 \ {\bf f}[i-i+\ell] \not= {\bf f}[i-1+\ell+q])
\ \wedge\ \\
(j+1\leq n-1) \implies (\forall r \text{ with } 1 \leq r \leq p\
\exists m \leq j+1-r-i \ {\bf f}[i+m] \not= {\bf f}[i+m+r] ) ) .\end{gathered}$$ Here the second line of the predicate specifies that there is a period $p$ of ${\bf f}[i..j]$ corresponding to a repetition of exponent at least $2$. The third line specifies that no period $q$ of ${\bf f}[i-1..j]$ (when this makes sense) can be $\leq p$, and the fourth line specifies that no period $r$ of ${\bf f}[i..j+1]$ (when $j+1 \leq n-1$) can be $\leq p$.
From the automaton we deduce a linear representation $(u, \mu, v)$ of rank 59. Since $(F_n)_F = 10^{n-2}$, it suffices to compute the minimal polynomial of $M_0 = \mu(0)$. When we do this, we discover it is $X^4(X^2 - X - 1)(X-1)^2(X+1)^2$. It follows from the theory of linear recurrences that $$\operatorname{mr}(F_n) = e_1 \alpha^n + e_2 \beta^n + e_3 n + e_4 + (e_5n + e_6)(-1)^n$$ for constants $e_1, e_2, e_3, e_4, e_5, e_6$ and $n \geq 6$. When we solve for $e_1, \ldots, e_6$ by using the first few values of $\operatorname{mr}(F_n)$ (computed from the linear representation or directly) we discover that $e_1 = (3\sqrt{5} - 5)/5$, $e_2 = (-3\sqrt{5} -5)/5$, $e_3 = e_5 = e_6 = 0$, and $e_4 = -3$. From this the result immediately follows.
In fact, we can prove even more.
For $n \geq 0$ the difference $\operatorname{mr}(n+1) - \operatorname{mr}(n)$ is either $0$ or $1$. Furthermore there is a finite automaton with 10 states that accepts $(n)_F$ precisely when $\operatorname{mr}(n+1) - \operatorname{mr}(n) = 1$.
Every maximal repetition ${\bf f}[i..j]$ of ${\bf f}[0..n-1]$ is either a maximal repetition of ${\bf f}[0..n]$ with $j \leq n-1$, or is a maximal repetition with $j = n-1$ that, when considered in ${\bf f}[0..n]$, can be extended one character to the right to become one with $j = n$. So the only maximal repetitions of ${\bf f}[0..n]$ not (essentially) counted by $\operatorname{mr}(n)$ are those such that $$\begin{gathered}
{\bf f}[i..n] \text{ is a maximal repetition of } {\bf f}[0..n]
\text{ and } \\
{\bf f}[i..n-1] \text{ is {\it not\/} a maximal repetition of }
{\bf f}[0..n-1].
\label{condit}\end{gathered}$$
We can easily create a predicate asserting this latter condition, and from this obtain the linear representation of $\operatorname{mr}(n+1) - \operatorname{mr}(n)$: $$\begin{aligned}
u &= [0\ 0\ 0\ 0\ 0\ 1\ 0\ 0\ 0\ 0\ 0\ 0\ ] \\
\mu(0) &= \left[ \begin{array}{cccccccccccc}
0&0&0&0&0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&0&0&0&1&0\\
0&0&0&1&0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&1&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&1&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&1&0\\
0&0&0&0&1&0&0&0&0&0&0&0\\
0&0&0&0&1&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&1\\
0&0&0&0&0&0&1&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&1&0&0
\end{array} \right] \\
\mu(1) &= \left[ \begin{array}{cccccccccccc}
0&0&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&1&0&0&0\\
0&1&1&0&0&0&0&0&0&0&0&0\\
1&1&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0&0&0&0\\
0&1&1&0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&1&0&0&0
\end{array} \right] \\
v &= [0\ 0\ 0\ 0\ 1\ 0\ 0\ 0\ 1\ 0\ 0\ 1 ] \\\end{aligned}$$
We now use the trick we previously used for the proof of Theorem \[chuan93-thm\]; the monoid generated by $\mu(0)$ and $\mu(1)$ has size $61$ and for each matrix $M$ in this monoid we have $u M v \in \lbrace 0, 1 \rbrace$. It follows that $\operatorname{mr}(n+1) - \operatorname{mr}(n) \in \lbrace 0, 1 \rbrace$ for all $n \geq 0$.
Knowing this, we can now build an automaton accepting those $n$ for which there exists an $i$ for which holds. When we do so we get the automaton depicted below in Figure \[maxrepp\].
![Automaton accepting $(n)_F$ such that $\operatorname{mr}(n+1) - \operatorname{mr}(n) = 1$[]{data-label="maxrepp"}](kk-prefixes.pdf){width="6in"}
Abelian properties
==================
Our decision procedure does not apply, in complete generality, to abelian properties of infinite words. This is because there is no obvious way to express assertions like $\psi(x) = \psi(x')$ for two factors $x, x'$ of an infinite word. (Here $\psi:\Sigma^* \rightarrow \Enn^{|\Sigma|}$ is the Parikh map that sends a word to the number of occurrences of each letter.) Indeed, in the $2$-automatic case it is provable that there is at least one abelian property that is inexpressible [@Schaeffer:2013 §5.2].
However, the special nature of the Fibonacci word $\bf f$ allows us to mechanically prove some assertions involving abelian properties. In this section we describe how we did this.
By an [*abelian square of order $n$*]{} we mean a factor of the form $x x'$ where $\psi(x) = \psi(x')$, where $n = |x|$. In a similar way we can define abelian cubes and higher powers.
We start with the elementary observation that $\bf f$ is defined over the alphabet $\lbrace 0, 1 \rbrace$. Hence, to understand the abelian properties of a factor $x$ it suffices to know $|x|$ and $|x|_0$. Next, we observe that the map that sends $n$ to $a_n := |{\bf f}[0..n-1]|_0$ (that is, the number of $0$’s in the length-$n$ prefix of $\bf f$), is actually [*synchronized*]{} (see ). That is, there is a DFA accepting the Fibonacci representation of the pairs $(n,a_n)$. In fact we have the following
Suppose the Fibonacci representation of $n$ is $e_1 e_2 \cdots e_i$. Then $a_n = [e_1 e_2 \cdots e_{i-1}]_F + e_i$. \[fibr\]
First, we observe that an easy induction on $m$ proves that $|X_m|_0 = F_{m-1}$ for $m \geq 2$. We will use this in a moment.
The theorem’s claim is easily checked for $n = 0,1$. We prove it for $F_{m+1} \leq n < F_{m+2}$ by induction on $m$. The base case is $m = 1$, which corresponds to $n = 1$.
Now assume the theorem’s claim is true for $m-1$; we prove it for $m$. Write $(n)_F = e_1 e_2 \cdots e_m$. Then, using the fact that ${\bf f}[0..F_{m+2}-1] = X_{m+2} = X_{m+1} X_m$, we get $$\begin{aligned}
|{\bf f}[0..n-1]|_0 &= |{\bf f}[0..F_{m+1}-1]|_0 + |{\bf f}[F_{m+1}..n-1]|_0 \\
&= |X_{m+1}|_0 + |{\bf f}[0..n-1-F_{m+1}]|_0 \\
&= F_m + |{\bf f}[0..n-1-F_{m+1}|_0 \\
&= F_m + [e_2\cdots e_{m-1}]_F + e_m \\
&= [e_1 \cdots e_{m-1}]_F + e_m ,\end{aligned}$$ as desired.
In fact, the synchronized automaton for $(n,a_n)_F$ is given in the following diagram:
![Automaton accepting $(n,a_n)_F$[]{data-label="synchro"}](synchrofib.pdf){width="6in"}
Here the missing state numbered $2$ is a “dead” state that is the target of all undrawn transitions.
The correctness of this automaton can be checked using our prover. Letting $\operatorname{ZC}(x,y)$ denote $1$ if $(x,y)_F$ is accepted, it suffices to check that
1. $ \forall x\ \exists y\ \operatorname{ZC}(x,y)= 1$ (that is, for each $x$ there is at least one corresponding $y$ accepted);
2. $\forall x\ \forall y\ \forall z\ (\operatorname{ZC}(x,y) = \operatorname{ZC}(x,z)) \implies
y = z$ (that is, for each $x$ at most one corresponding $y$ is accepted);
3. $\forall x \ \forall y \ ((\operatorname{ZC}(x,y)=1) \ \wedge \ ({\bf f}[x] = 1))
\implies (\operatorname{ZC}(x+1,y+1) = 1)$;
4. $\forall x \ \forall y \ ((\operatorname{ZC}(x,y)=1) \ \wedge \ ({\bf f}[x] = 0))
\implies (\operatorname{ZC}(x+1,y) = 1)$;
Another useful automaton computes, on input $n, i, j$ the function $$\operatorname{FAB}(n,i,j) :=
|{\bf f}[i..i+n-1]|_0 - |{\bf f}[j..j+n-1]|_0 =
a_{i+n}-a_i - a_{j+n}+a_j.$$ From the known fact that the factors of $\bf f$ are “balanced” we know that $\operatorname{FAB}$ takes only the values $-1, 0, 1$. This automaton can be deduced from the one above. However, we calculated it by “guessing” the right automaton and then verifying the correctness with our prover.
The automaton for $\operatorname{FAB}(n,i,j)$ has 30 states, numbered from $1$ to $30$. Inputs are in $\Sigma_2^3$. The transitions, as well as the outputs, are given in the table below.
$q$ $[0,0,0]$ $[0,0,1]$ $[0,1,0]$ $[0,1,1]$ $[1,0,0]$ $[1,0,1]$ $[1,1,0]$ $[1,1,1]$ $\tau(q)$
----- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- -----------
1 1 2 3 4 4 5 6 7 0
2 8 1 9 3 3 4 10 6 0
3 11 12 1 2 2 13 4 5 0
4 14 11 8 1 1 2 3 4 0
5 15 11 16 1 1 2 3 4 1
6 17 18 8 1 1 2 3 4 $-1$
7 19 18 16 1 1 2 3 4 0
8 1 2 3 4 4 20 6 21 0
9 11 12 1 2 2 22 4 20 0
10 18 23 1 2 2 13 4 5 $-1$
11 1 2 3 4 4 5 24 25 0
12 8 1 9 3 3 4 26 24 0
13 16 1 27 3 3 4 10 6 1
14 1 2 3 4 4 20 24 28 0
15 2 13 4 5 5 20 25 28 $-1$
16 2 13 4 5 5 20 7 21 $-1$
17 3 4 10 6 6 21 24 28 1
18 3 4 10 6 6 7 24 25 1
19 4 5 6 7 7 21 25 28 0
20 15 14 16 8 8 1 9 3 1
21 19 17 16 8 8 1 9 3 0
22 16 8 27 9 9 3 29 10 1
23 9 3 29 10 10 6 26 24 1
24 17 18 14 11 11 12 1 2 $-1$
25 19 18 15 11 11 12 1 2 0
26 18 23 11 12 12 30 2 13 $-1$
27 12 30 2 13 13 22 5 20 $-1$
28 19 17 15 14 14 11 8 1 0
29 18 23 1 2 2 22 4 20 $-1$
30 16 1 27 3 3 4 26 24 1
: Automaton to compute $\operatorname{FAB}$
Once we have guessed the automaton, we can verify it as follows:
1. $\forall i \ \forall j\ \operatorname{FAB}[0][i][j]=0$. This is the basis for an induction.
2. Induction steps:
- $\forall i\ \forall j\ \forall n \ ({\bf f}[i+n]={\bf f}[j+n])
\implies (\operatorname{FAB}[n][i][j]=\operatorname{FAB}[n+1][i][j])$.
- $\forall i\ \forall j\ \forall n\ (({\bf f}[i+n]=0) \wedge
({\bf f}[j+n]=1)) \implies
(((\operatorname{FAB}[n][i][j]=-1) \wedge (\operatorname{FAB}[n+1][i][j]=0)) \vee
((\operatorname{FAB}[n][i][j]=0) \wedge (\operatorname{FAB}[n+1][i][j]=1))) $
- $\forall i\ \forall j\ \forall n\ (({\bf f}[i+n]=0) \wedge
({\bf f}[j+n]=1)) \implies
(((\operatorname{FAB}[n][i][j]=1) \wedge (\operatorname{FAB}[n+1][i][j]=0)) \vee
((\operatorname{FAB}[n][i][j]=0) \wedge (\operatorname{FAB}[n+1][i][j]=-1))) $.
As the first application, we prove
The Fibonacci word $\bf f$ has abelian squares of all orders.
We use the predicate $$\exists i \ (\operatorname{FAB}[n][i][i+n] = 0) .$$ The resulting automaton accepts all $n \geq 0$. The total computing time was 141 ms.
Cummings and Smyth counted the total number of all occurrences of (nonempty) abelian squares in the Fibonacci words $X_i$. We can do this by using the predicate $$(k>0) \wedge (i+2k \leq n) \wedge (\operatorname{FAB}[k][i][i+k]=0),$$ using the techniques in Section \[enumer\] and considering the case where $n = F_i$.
When we do, we get a linear representation of rank 127 that counts the total number $w(n)$ of occurrences of abelian squares in the prefix of length $n$ of the Fibonacci word.
To recover the Cummings-Smyth result we compute the minimal polynomial of the matrix $M_0$ corresponding to the predicate above. It is $$x^4 (x-1)(x+1)(x^2+x+1)(x^2-3x+1)(x^2-x+1)(x^2+x-1)(x^2-x-1).$$
This means that $w(F_n)$, that is, $w$ evaluated at $10^{n-2}$ in Fibonacci representation, is a linear combination of the roots of this polynomial to the $n$’th power (more precisely, the $(n-2)$th, but this detail is unimportant). The roots of the polynomial are $$-1, 1, (-1 \pm i \sqrt{3})/2, (3 \pm \sqrt{5})/2, (1 \pm i \sqrt{3})/2,
(-1 \pm \sqrt{5})/2, (1 \pm \sqrt{5})/2.$$ Solving for the coefficients as we did in Section \[enumer\] we get
For all $n \geq 0$ we have $$\begin{gathered}
w(F_n) =
c_1 \left({{3+\sqrt{5}}\over 2}\right)^n + c_1 \left({{3-\sqrt{5}}\over 2}\right)^n
+
c_2 \left( {{1+\sqrt{5}}\over 2} \right)^n + c_2 \left( {{1-\sqrt{5}}\over 2} \right)^n
+ \\
c_3 \left({{1+i\sqrt{3}}\over 2} \right)^n + \overline{c_3} \left({{1-i\sqrt{3}}\over 2} \right)^n
+
c_4 \left({{-1+i\sqrt{3}}\over 2} \right)^n + \overline{c_4} \left({{-1-i\sqrt{3}}\over 2} \right)^n
+
c_5 (-1)^n,\end{gathered}$$ where $$\begin{aligned}
c_1 &= 1/40 \\
c_2 &= -\sqrt{5}/20 \\
c_3 &= (1 - i\sqrt{3})/24 \\
c_4 &= i\sqrt{3}/24 \\
c_5 &= -2/15,\end{aligned}$$ and here $\overline{x}$ denotes complex conjugate. Here the parts corresponding to the constants $c_3, c_4, c_5$ form a periodic sequence of period 6.
Next, we turn to what is apparently a new result. Let $h(n)$ denote the total number of distinct factors (not occurrences of factors) that are abelian squares in the Fibonacci word $X_n$.
In this case we need the predicate $$(k \geq 1) \wedge (i+2k \leq n) \wedge
(\operatorname{FAB}[k][i][i+k]=0) \wedge
(\forall j<i \
(\exists t<2k\ ({\bf f}[j+t]\not= {\bf f}[i+t]))).$$
We get the minimal polynomial $$x^4(x+1)(x^2+x+1)(x^2-3x+1)(x^2-x+1)(x^2+x-1)(x^2-x-1)(x-1)^2.$$ Using the same technique as above we get
For $n \geq 2$ we have $h(n) = a_1c_1^n + \cdots + a_{10}c_{10}^n $ where $$\begin{aligned}
a_1 &= (-2+\sqrt{5})/20 \\
a_2 &= (-2-\sqrt{5})/20 \\
a_3 &= (5-\sqrt{5})/20 \\
a_4 &= (5+\sqrt{5})/20 \\
a_5 &= 1/30 \\
a_6 &= -5/6 \\
a_7 &= (1/12)-i \sqrt{3}/12 \\
a_8 &= (1/12)+i \sqrt{3}/12 \\
a_9 &= (1/6) + i \sqrt{3}/12 \\
a_{10}&= (1/6) - i \sqrt{3}/12 \end{aligned}$$ and $$\begin{aligned}
c_1 &= (3+\sqrt{5})/2 \\
c_2 &= (3-\sqrt{5})/2 \\
c_3 &= (1+\sqrt{5})/2 \\
c_4 &= (1-\sqrt{5})/2 \\
c_5 &= -1 \\
c_6 &= 1 \\
c_7 &= (1/2)+i \sqrt{3}/2 \\
c_8 &= (1/2)-i \sqrt{3}/2 \\
c_9 &= (-1/2)+i \sqrt{3}/2 \\
c_{10} &= (-1/2)-i \sqrt{3}/2 .\end{aligned}$$
For another new result, consider counting the total number $a(n)$ of distinct factors of length $2n$ of the infinite word $\bf f$ that are abelian squares.
This function is rather erratic. The following table gives the first few values:
$n$ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
-------- --- --- --- --- --- --- --- ---- --- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- -- -- -- -- -- -- -- -- -- --
$a(n)$ 1 3 5 1 9 5 5 15 3 13 13 5 25 9 15 25 1 27 19 11
We use the predicate $$(n \geq 1) \wedge (\operatorname{FAB}[n][i][i+n]=0) \wedge
(\forall j < i\ ( \exists t<2n \ ({\bf f}[j+t]\not= {\bf f}[i+t]))).$$ to create the matrices and vectors.
$a(n) = 1$ infinitely often and $a(n) = 2n-1$ infinitely often. More precisely $a(n) = 1$ iff $(n)_F = 1$ or $(n)_F = (100)^i 101$ for $i \geq 0$, and $a(n) = 2n-1$ iff $(n)_F = 10^i$ for $i \geq 0$.
For the first statement, we create a DFA accepting those $(n)_F$ for which $a(n) = 1$, via the predicate $$\forall i\ \forall j \ ((\operatorname{FAB}[n][i][i+n]=0) \wedge (\operatorname{FAB}[n][j][j+n]=0)) \implies
(\forall t<2n \ ({\bf f}[j+t] = {\bf f}[i+t])).$$ The resulting $6$-state automaton accepts the set specified.
For the second result, we first compute the minimal polynomial of the matrix $M_0$ of the linear representation. It is $x^5 (x-1)(x+1)(x^2-x-1)$. This means that, for $n \geq 5$, we have $a(F_n) = c_1 + c_2 (-1)^n + c_3 \alpha^n + c_4 \beta^n$ where, as usual, $\alpha = (1+\sqrt{5})/2$ and $\beta=(1-\sqrt{5})/2$. Solving for the constants, we determine that $a(F_n) = 2F_n - 1$ for $n \geq 2$, as desired.
To show that these are the only cases for which $a(n) = 2n-1$, we use a predicate that says that there are not at least three different factors of length $2n$ that are not abelian squares. Running this through our program results in only the cases previously discussed.
Finally, we turn to abelian cubes. Unlike the case of squares, some orders do not appear in $\bf f$.
The Fibonacci word $\bf f$ contains, as a factor, an abelian cube of order $n$ iff $(n)_F$ is accepted by the automaton below.
![Automaton accepting orders of abelian cubes in $\bf f$[]{data-label="fibabelcube"}](fibabelcube.pdf){width="6.5in"}
Theorem \[fibr\] has the following interesting corollary.
Let $h:\lbrace 0, 1 \rbrace^* \rightarrow \Delta^*$ be an arbitrary morphism such that $h(01) \not= \epsilon$. Then $h({\bf f})$ is an infinite Fibonacci-automatic word.
From Theorem \[fibr\] we see that there is a predicate $\operatorname{ZC}(n,n')$ which is true if $n' = |{\bf f}[0..n-1]|_0$ and false otherwise, and this predicate can be implemented as a finite automaton taking the inputs $n$ and $n'$ in Fibonacci representation.
Suppose $h(0) = w$ and $h(1) = x$. Now, to show that h([**f**]{}) is Fibonacci-automatic, it suffices to show that, for each letter $a \in \Delta$, the language of “fibers” $$L_a = \{ (n)_F : (h({\bf f}))[n] = a \}$$ is regular.
To see this, we write a predicate for the $n$ in the definition of $L_a$, namely $$\begin{gathered}
\exists q\ \exists r_0 \ \exists r_1 \ \exists m \
(q \leq n < q+ |h({\bf f}[m])|) \ \wedge \ \operatorname{ZC}(m,r_0) \ \wedge \
(r_0+r_1=m) \wedge \\
(r_0 |w| + r_1 |x| = q) \ \wedge \
(( {\bf f}[m]=0 \ \wedge \ w[n-q] = a) \ \vee \ ({\bf f}[m] = 1 \ \wedge\ x[n-q] = a) ) .\end{gathered}$$
Notice that the predicate looks like it uses multiplication, but this multiplication can be replaced by repeated addition since $|w|$ and $|x|$ are constants here.
Unpacking this predicate we see that it asserts the existence of $m$, $q$, $r_0$, and $r_1$ having the meaning that
- the $n$’th symbol of h([**f**]{}) lies inside the block $h({\bf f}[m])$ and is in fact the $(n-q)$’th symbol in the block (with the first symbol being symbol 0)
- ${\bf f}[0..m-1]$ has $r_0$ 0’s in it
- $ {\bf f}[0..m-1]$ has $r_1$ 1’s in it
- the length of $h({\bf f}[0..m-1])$ is $q$
Since everything in this predicate is in the logical theory $(\Enn, +, <, F)$ where $F$ is the predicate for the Fibonacci word, the language $L_a$ is regular.
Notice that everything in this proof goes through for other numeration systems, provided the original word has the property that the Parikh vector of the prefix of length $n$ is synchronized.
Details about our implementation
================================
Our program is written in JAVA, and was developed using the [Eclipse]{} development environment.[^4] We used the [dk.brics.automaton]{} package, developed by Anders M[ø]{}ller at Aarhus University, for automaton minimization.[^5] [Maple 15]{} was used to compute characteristic polynomials.[^6] The [GraphViz]{} package was used to display automata.[^7]
Our program consists of about 2000 lines of code. We used Hopcroft’s algorithm for DFA minimization.
A user interface is provided to enter queries in a language very similar to the language of first-order logic. The intermediate and final result of a query are all automata. At every intermediate step, we chose to do minimization and determinization, if necessary. Each automaton accepts tuples of integers in the numeration system of choice. The built-in numeration systems are ordinary base-$k$ representations and Fibonacci base. However, the program can be used with any numeration system for which an automaton for addition and ordering can be provided. These numeration system-specific automata can be declared in text files following a simple syntax. For the automaton resulting from a query it is always guaranteed that if a tuple $t$ of integers is accepted, all tuples obtained from $t$ by addition or truncation of leading zeros are also accepted. In Fibonacci representation, we make sure that the accepting integers do not contain consecutive $1$’s.
The program was tested against hundreds of different test cases varying in simplicity from the most basic test cases testing only one feature at a time, to more comprehensive ones with many alternating quantifiers. We also used known facts about automatic sequences and Fibonacci word in the literature to test our program, and in all those cases we were able to get the same result as in the literature. In a few cases, we were even able to find small errors in those earlier results.
The source code and manual will soon be available for free download.
Acknowledgments
===============
We thank Kalle Saari for bringing our attention to the small error in . We thank Narad Rampersad and Michel Rigo for useful suggestions.
Eric Rowland thought about the proof of Theorem \[additive-thm\] with us in 2010, and was able to prove at that time that the word $1213121512131218\cdots$ avoids additive squares. We acknowledge his prior work on this problem and thank him for allowing us to quote it here.
[10]{}
C. Ahlbach, J. Usatine, C. Frougny, and N. Pippenger. Efficient algorithms for [Zeckendorf]{} arithmetic. (2013), 249–256.
J.-P. Allouche, N. Rampersad, and J. Shallit. Periodicity, repetitions, and orbits of an automatic sequence. (2009), 2795–2803.
J.-P. Allouche, K. Scheicher, and R. F. Tichy. Regular maps in generalized number systems. (2000), 41–58.
J.-P. Allouche and J. Shallit. . Cambridge University Press, 2003.
J. Berstel. Mots de [Fibonacci]{}. (1980–81), 57–78.
J. Berstel. Fonctions rationnelles et addition. In M. Blab, editor, [*Théorie des Langages, École de printemps d’informatique théorique*]{}, pp. 177–183. LITP, 1982.
J. Berstel. Fibonacci words—a survey. In G. Rozenberg and A. Salomaa, editors, [*The Book of L*]{}, pp. 13–27. Springer-Verlag, 1986.
J. Berstel and C. Reutenauer. , Vol. 137 of [*Encylopedia of Mathematics and Its Applications*]{}. Cambridge University Press, 2011.
B. Bischoff, J. D. Currie, and D. Nowotka. Unary patterns with involution. (2012), 1641–1652.
B. Bischoff and D. Nowotka. Pattern avoidability with involution. In [*WORDS 2011*]{}, pp. 65–70, 2011. Available at <http://rvg.web.cse.unsw.edu.au/eptcs/content.cgi?WORDS2011>.
A. Blondin [Massé]{}, S. Brlek, A. Garon, and S. [Labbé]{}. Two infinite families of polyominoes that tile the plane by translation in two distinct ways. (2011), 4778–4786.
A. Blondin [Massé]{}, S. Brlek, S. [Labbé]{}, and M. [Mendès]{} France. Fibonacci snowflakes. (2011), 141–152.
A. Blondin [Massé]{}, S. Brlek, S. [Labbé]{}, and M. [Mendès]{} France. Complexity of the [Fibonacci]{} snowflake. (2012), 257–260.
J.-P. Borel and F. Laubie. Quelques mots sur la droite projective [réelle]{}. (1993), 23–51.
T. C. Brown and A. R. Freedman. Arithmetic progressions in lacunary sets. (1987), 587–596.
V. [Bruyère]{} and G. Hansel. Bertrand numeration systems and recognizability. (1997), 17–43.
V. [Bruyère]{}, G. Hansel, C. Michaux, and R. Villemaire. Logic and $p$-recognizable sets of integers. (1994), 191–238. Corrigendum, [*Bull. Belg. Math. Soc.*]{} [**1**]{} (1994), 577.
J. R. [Büchi]{}. Weak secord-order arithmetic and finite automata. (1960), 66–92. Reprinted in S. Mac Lane and D. Siefkes, eds., [*The Collected Works of J. Richard [Büchi]{}*]{}, Springer-Verlag, 1990, pp. 398–424.
L. Carlitz. Fibonacci representations. (1968), 193–220.
A. Carpi and V. D’Alonzo. On the repetitivity index of infinite words. (2009), 145–158.
A. Carpi and V. D’Alonzo. On factors of synchronized sequences. (2010), 3932–3937.
A. Carpi and C. Maggi. On synchronized sequences and their separators. (2001), 513–524.
J. Cassaigne. Sequences with grouped factors. In [*Developments in Language Theory [III]{}*]{}, pp. 211–222. Aristotle University of Thessaloniki, 1998.
J. Cassaigne, J. Currie, L. Schaeffer, and J. Shallit. Avoiding three consecutive blocks of the same size and same sum. Preprint, 2013.
E. Charlier, N. Rampersad, and J. Shallit. Enumeration and decidable properties of automatic sequences. (2012), 1035–1066.
M. Christou, M. Crochemore, and C. S. Iliopoulos. Quasiperiodicities in [Fibonacci]{} strings. To appear in [*Ars Combinatoria*]{}. Preprint available at [ http://arxiv.org/abs/1201.6162]{}, 2012.
W.-F. Chuan. Symmetric [Fibonacci]{} words. (1993), 251–255.
A. Cobham. Uniform tag sequences. (1972), 164–192.
M. Crochemore. An optimal algorithm for computing the repetitions in a word. (1981), 244–250.
L. J. Cummings, D. Moore, and J. [Karhumäki]{}. Borders of [Fibonacci]{} strings. (1996), 81–87.
L. J. Cummings and W. F. Smyth. Weak repetitions in strings. (1997), 33–48.
J. D. Currie. Pattern avoidance with involution. Available at <http://arxiv.org/abs/1105.2849>, 2011.
J. D. Currie, N. Rampersad, and K. Saari. Suffix conjugates for a class of morphic subshifts. In J. [Karhumäki]{}, A. [Lepistö]{}, and L. Zamboni, editors, [ *WORDS 2013*]{}, Vol. 8079 of [*Lecture Notes in Computer Science*]{}, pp. 95–106. Springer-Verlag, 2013.
J. D. Currie and K. Saari. Least periods of factors of infinite words. (2009), 165–178.
A. de Luca. A combinatorial property of the [Fibonacci]{} words. (1981), 193–195.
X. Droubay. Palindromes in the [Fibonacci]{} word. (1995), 217–221.
D. D. A. Epple and J. Siefken. Collapse: a [Fibonacci]{} and [Sturmian]{} game. Available at <http://www.siefkenj.com/tmp/Fibonacci-4.pdf>, 2014.
A. S. Fraenkel. Systems of numeration. (1985), 105–114.
A. S. Fraenkel and J. Simpson. The exact number of squares in [Fibonacci]{} words. (1999), 95–106.
C. Frougny. Linear numeration systems of order two. (1988), 233–259.
C. Frougny. Fibonacci representations and finite automata. (1991), 393–399.
C. Frougny. Representations of numbers and finite automata. (1992), 37–60.
C. Frougny and B. Solomyak. On representation of integers in linear numeration systems. In M. Pollicott and K. Schmidt, editors, [*Ergodic Theory of [$\Zee^d$]{} Actions (Warwick, 1993–1994)*]{}, Vol. 228 of [*London Mathematical Society Lecture Note Series*]{}, pp. 345–368. Cambridge University Press, 1996.
D. Goc, D. Henshall, and J. Shallit. Automatic theorem-proving in combinatorics on words. In N. Moreira and R. Reis, editors, [*CIAA 2012*]{}, Vol. 7381 of [*Lecture Notes in Computer Science*]{}, pp. 180–191. Springer-Verlag, 2012.
D. Goc, H. Mousavi, and J. Shallit. On the number of unbordered factors. In A.-H. Dediu, C. Martin-Vide, and B. Truthe, editors, [*LATA 2013*]{}, Vol. 7810 of [*Lecture Notes in Computer Science*]{}, pp. 299–310. Springer-Verlag, 2013.
D. Goc, K. Saari, and J. Shallit. Primitive words and [Lyndon]{} words in automatic and linearly recurrent sequences. In A.-H. Dediu, C. Martin-Vide, and B. Truthe, editors, [*LATA 2013*]{}, Vol. 7810 of [*Lecture Notes in Computer Science*]{}, pp. 311–322. Springer-Verlag, 2013.
D. Goc, L. Schaeffer, and J. Shallit. The subword complexity of $k$-automatic sequences is $k$-synchronized. In M.-P. Béal and O. Carton, editors, [*DLT 2013*]{}, Vol. 7907 of [*Lecture Notes in Computer Science*]{}, pp. 252–263. Springer-Verlag, 2013.
M. Guay-Paquet and J. Shallit. Avoiding squares and overlaps over the natural numbers. (2009), 6245–6254.
L. Halbeisen and N. [Hungerbühler]{}. An application of [Van der Waerden’s]{} theorem in additive number theory. (2000), \#A7. .
C. Holton and L. Q. Zamboni. Directed graphs and substitutions. (2001), 545–564.
S. Homer and A. L. Selman. . Springer-Verlag, 2nd edition, 2011.
C. S. Iliopoulos, D. Moore, and W. F. Smyth. A characterization of the squares in a [Fibonacci]{} string. (1997), 281–291.
J. [Karhumäki]{}. On cube-free $\omega$-words generated by binary morphisms. (1983), 279–297.
R. Kolpakov and G. Kucherov. On maximal repetitions in words. In G. Ciobanu and G. [Păun]{}, editors, [*Fundamentals of Computation Theory: FCT ’99*]{}, Vol. 1684 of [*Lecture Notes in Computer Science*]{}, pp. 374–385. Springer-Verlag, 1999.
C. G. Lekkerkerker. Voorstelling van natuurlijke getallen door een som van getallen van [Fibonacci]{}. (1952), 190–195.
F. Mignosi and G. Pirillo. Repetitions in the [Fibonacci]{} infinite word. (1992), 199–204.
A. Monnerot-Dumaine. The [Fibonacci]{} word fractal. Published electronically at <http://hal.archives-ouvertes.fr/hal-00367972/fr/>, 2009.
M. Morse and G. A. Hedlund. Symbolic dynamics [II]{}. [Sturmian]{} trajectories. (1940), 1–42.
A. Ostrowski. Bemerkungen zur [Theorie]{} der [Diophantischen]{} [Approximationen]{}. (1922), 77–98,250–251. Reprinted in [*Collected Mathematical Papers*]{}, Vol. 3, pp. 57–80.
G. Pirillo. Fibonacci numbers and words. (1997), 197–207.
G. Pirillo and S. Varricchio. On uniformly repetitive semigroups. (1994), 125–129.
M. Presburger. die [Volständigkeit]{} eines gewissen [Systems]{} der [Arithmetik]{} ganzer [Zahlen]{}, in welchem die [Addition]{} als einzige [Operation]{} hervortritt. In [*Sparawozdanie z I Kongresu matematyków krajów slowianskich*]{}, pp. 92–101, 395. Warsaw, 1929.
M. Presburger. On the completeness of a certain system of arithmetic of whole numbers in which addition occurs as the only operation. (1991), 225–233.
N. Rampersad and J. Shallit. Words avoiding reversed subwords. (2005), 157–164.
M. Rao. On some generalizations of abelian power avoidability. Preprint, 2013.
G. Rote. Sequences with subword complexity $2n$. (1994), 196–213.
K. Saari. Periods of factors of the [Fibonacci]{} word. In [*WORDS 07*]{}, 2007.
K. Saari. Lyndon words and [Fibonacci]{} numbers. (2014), 34–44.
L. Schaeffer. Deciding properties of automatic sequences. Master’s thesis, University of Waterloo, 2013.
L. Schaeffer and J. Shallit. The critical exponent is computable for automatic sequences. (2012), 1611–1626.
P. [Séébold]{}. . PhD thesis, [Université]{} P. et M. Curie, Institut de Programmation, Paris, 1985.
J. O. Shallit. A generalization of automatic sequences. (1988), 1–16.
J. Shallit. Decidability and enumeration for automatic sequences: a survey. In A. A. Bulatov and A. M. Shur, editors, [*CSR 2013*]{}, Vol. 7913 of [*Lecture Notes in Computer Science*]{}, pp. 49–63. Springer-Verlag, 2013.
E. Zeckendorf. des nombres naturels par une somme de nombres de [Fibonacci]{} ou de nombres [Lucas]{}. (1972), 179–182.
[^1]: School of Computer Science, University of Waterloo, Waterloo, ON N2L 3G1, Canada; , [[email protected]]{}, [[email protected]]{} .
[^2]: Computer Science and Artificial Intelligence Laboratory, The Stata Center, MIT Building 32, 32 Vassar Street, Cambridge, MA 02139 USA; [[email protected]]{} .
[^3]: There is also a version where “prefixes” is replaced by “suffixes”.
[^4]: Available from [http://www.eclipse.org/ide/]{} .
[^5]: Available from [http://www.brics.dk/automaton/]{} .
[^6]: Available from [http://www.maplesoft.com]{} .
[^7]: Available from [http://www.graphviz.org]{} .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The emergence of multicellularity and developmental programs are among the major problems of evolutionary biology. Traditionally, research in this area has been based on the combination of data analysis and experimental work on one hand and theoretical approximations on the other. A third possibility is provided by computer simulation models, which allow to both simulate reality and explore alternative possibilities. These in silico models offer a powerful window to the possible and the actual by means of modeling how virtual cells and groups of cells can evolve complex interactions beyond a set of isolated entities. Here we present several examples of such models, each one illustrating the potential for artificial modeling of the transition to multicellularity.'
author:
- 'Ricard V. Solé [^1]'
- 'Salva Duran-Nebreda'
title: In silico transitions to multicellularity
---
The physics of multicellularity
===============================
The transition to multicellularity is tied to the emergence of interactions among previously isolated cells. As a problem in complexity (Anderson 1972) we could say that a multicellular system defines a level of organization whose global properties cannot be reduced to the properties of the individual units. This statement is relevant for many reasons. First, because the presence of an evolutionary innovation necessarily requires the cooperation between previously unrelated units (Schuster 1996). Once such interactions are in place, a network of connected elements needs to be considered in order to understand what is now at work. In early phases predating the transition to complex multicellular life forms, the network involved cell-cell as well as cell-substrate interactions.
In dynamical terms, the transition required the emergence of cooperation among elements, which share a common space where they relate to each other and respond to environmental changes in a concerted manner. Most theoretical and computational approaches to this problem do not take into account the fact that these systems are formed by physical objects and it might seem not so relevant when dealing with the generic mechanisms associated to cooperative dynamics. As usual, the level of detail that is used in a model scales with the type of question we wish to answer. If we search for general principles defining the appearance of cooperative aggregates, general models considering population dynamics and gene interactions are enough. But multicellularity also connects us with the first steps towards developmental programs and a previous step to other major innovations. Such innovations are always associated with novelties in the ways cells and tissues interact. The spatial arrangement of cells and the diversity potential provided by space and thus a explicit introduction of spatial degrees of freedom is essential.
Meaningful models of evolutionary dynamics of multicellular systems need to consider the role of generic physical mechanisms of morphogenesis that are not the result of complex regulatory processes. In this context, physical forces including gravity, adhesion or diffusion, and their generative potential, are considered (Newman and Comper 1990; Goodwin 1994). The interplay between these mechanisms allows for spontaneous pattern formation through segregation of cell types, differential cell growth and mortality. Some of these generic, pattern-forming mechanisms likely predate the early history of both pre-cellular and multicellular life forms(Forgacs and Newman 2005; Solé et al 2007; Solé 2009), along with others controlled by genetic circuits .
Using some of these mechanisms, an evolutionary model of form can be constructed. Moreover, since some of these mechanisms seem to strongly constrain the repertoire of potential structures that can be generated, they also offer a powerful framework to understand the origins of convergent designs (Alberch, 1980). In this context, as pointed out by John Tyler Bonner, simple explanations based on mathematical and computational models can help to grasp the principles of multicellular organization (Bonner 2001; Forgacs and Newman 2005). As noticed in (Newman and Bhat 2008) the interplay between physical constraints and genetic regulatory mechanisms has been traditionally overlooked in most studies, with few exceptions (see for example Eggenberger 1997; Coen et al 2004; Cummings 2006; Doursat 2008; Kaandorp et al. 2008).
Although physics and embodiment are usually discussed in the context (or at the level of) organisms or tissues, there is another level of embodiment that requires attention: the external world, whose fluctuations and properties influence the repertoire of adaptations that can be available. Thus we could add other factors playing a role in the early stages of multicellularity, including the ecological context and the physics of the environment should also be taken into account.
Here we present five examples of [*in silico*]{} models illustrating different aspects of the emergence of multicellularity. They involve different levels of complexity and address different questions, although all illustrate the generative potential of the role played by the basic physics that pervades the interactions among virtual cells.
The dynamics of cell sorting can be easily modelled by considering a simplified lattice model where the movement of cells is constrained by their local preferences. Specifically, let us consider a $L \times L$ lattice $\Omega$, i. e. a discrete set of sites: $$\Omega = \{ (i,j) \in Z^2 \vert 1 \le i, j, \le L 1\}$$ Each site $(i,i) \in \Omega$ is characterized by a “state”, indicated as $S(i,j)$. This state can be $0$ if the site is empty and either $1$ or $2$ if occupied by cells. The two possible “cell” states indicate different cell types with different adhesion properties. Cells can be more or less prone to remain together and also might tend to either avoid or attach to the external medium. Since cell-cell (and cell-medium) interactions are necessarily local, (figure 1) a given cell can only interact with a maximum of eight nearest neighbors. If $J(S(i,j),S(k,l))$ indicates the energy associated to the interactions between the sites $(i,j)$ and $(k,l)$, a matrix of interaction coefficients will be defined, namely $${\bf J} =
\left( \begin{array}{ccc}
J(0,0) & J(0,1) & J(0,2) \\
J(1,0) & J(1,1) & J(1,2) \\
J(2,0) & J(2,1) & J(2,2) \end{array} \right)$$
which is obviously symmetric, i. e. $$J(a,b)=J(b,a)$$ and $$J(0,0)=0.$$
The model allows cells to move to a neighboring position by switching the two local states provided that the final state is more likely to happen, i. e. consistent with the optimization of cell preferences. This can be obtained using an energy function, defined as $${\cal H} = \sum_{i,j} \sum_{k,l} \left (1 - \delta_{s(i,j)s(k,l)} \right ) J(s(i,j), s(k,l))$$ where the sum is performed only over nearest pairs. Here $\delta_{mn}=1$ when $m=n$ and zero otherwise, and thus the term $1-\delta_{mn}$ just discards pairs of sites with identical states.
Each step, we choose a random site, another neighboring site and compute the new energy ${\cal H}'$ and compare it to the original one ${\cal H}$. If the difference $$\Delta {\cal H} = {\cal H}' - {\cal H}$$ is positive, an increase in energy would occur and thus is discarded. Instead, when $$\Delta {\cal H}<0,$$the largest the difference the more likely is the change to happen. This is a simplified model and more sophisticated approaches were developed by Glazier and co-workers, involving “cells” composed of connected sites defining a specific cell, whose shapes can change in realistic ways (Graner and Glazier 1992; Glazier and Graner 1993).
Evolving differential adhesion
==============================
Our first example is a discrete model of evolution where digital “creatures” composed by many simplified cells interact through discrete adhesion forces. Hogeweg’s model considers a population of of model organisms that is evolved using a [*genetic algorithm*]{} (Mitchell 1998; Forrest 1993) that allows to search over shape space under different selection pressures (Hogeweg 2000a, 2000b). Hogeweg’s approach considers the growth and development of a simulated embryo. The model description includes an internal boolean gene network (see Kauffman 1993), the evolution of which leads to different adhesion among cells, cell division and death caused by stretching and compression, cell migration and differentiation.
Adhesion is introduced using very simplified but effective physical models (Graner and Glazier 1992; Glazier and Graner 1993; Sawill and Hogeweg 1997) and is one of the main players influencing the evolutionary dynamics of these virtual metazoans and their potential for diversification, consistent with the role played by development in the context of morphological radiations. Cell adhesion can easily promote the spatial organization of an initially disorganized, mixed group of cells. This is a very robust, repeatable and predictable mechanism of organization that amplifies initial disorder experienced by a mixed set of cells that move in space and aggregate with other cells under differential adhesion. The preferential adhesion mechanism is responsible for the sorting of cells in space, in a way that corresponds to the behavior of immiscible fluids (Forgacs and Newman 2005).
As pointed out by Hogeweg (2000b) differential cell adhesion is regulated by the gene network affecting cell behavior and the communication between cells through cell-cell interactions. The model considers different types of fitness functions but the only strong pressure is directed to maximizing the diversity of cell types and thus there is no explicit search for special spatial arrangements or predefined developmental programs. Hogeweg’s work revealed the existence of a neutral landscape of possible phenotypes that pervades the punctuated nature of transitions (Hogeweg 2000b; see also Fontana and Schuster 1998). Long periods of stasis are characterized by slow increases in fitness as small variations in phenotype are achieved. Selection for diverse gene expression patterns is used (see also Solé et al. 2003), a choice that can be justified by an observed, well-known trend. The number of cell types is a good measure of complexity, which is known to increase through metazoan evolution (Carroll 2005). Increases in cell type number provide a high potential for further evolution of anatomical and functional complexity, essentially through division of labor and the formation of specialized tissues (Maynard-Smith and Szathmáry 1995).
Since the imposed selection pressure is rather generic, no special constraints are posed on the way genes interact and influence cell arrangements; no particular, predefined architectures and developmental plans are favored. The model is able to evolve complex forms, and in the process of evolving them, different remarkable changes take place. Complex shapes and some familiar ways of obtaining them (such as tissue engulfing, budding, etc) appeared and complex interactions between apoptosis or migration emerged. As pointed out in Hogeweg (2000a), morphogenesis itself emerges as a byproduct of optimization for cell diversity. It is worth noting that other works involving cell type richness as a fitness function favor the explosion of pattern forming motifs as soon as a threshold of genetic complexity is reached (Solé et al. 2003).
Multicellularity for free
=========================
One of the most relevant and striking features of the transition to multicellularity is the fact that it took place multiple times in different lineages in the history of evolution (Abedin 2010; King 2004; Bonner 2001; Knoll 2011; Niklas and Newman 2013). For several authors this suggests that there is a certain component of inevitability in this process (Buss 1984; Grossberg and Strahmann 2007), that cooperation and specialization are such powerful innovations that convergent evolution into multicellularity ensued. In this vein, if Hogeweg’s model proves that selecting for cell diversification can lead to co-select unexpected, emergent properties and behavior, this second model by Kaneko and colleagues (Kaneko and Yomo 1998) shows that even when there is no selection at work, cells can easily drift into multicellular phenotypes including differentiation and spatial patterning.
As in the last example, the model starts with a single cell embryo that can develop into simple yet hallmark structures (see figure 4), just by allowing cell-cell and cell-environment interactions. The former are simulated by non-preferential adhesion between close cells and the latter by transport and processing of chemical species present in the virtual environment, which cells use to grow and divide. The internal workings of the cells are modeled by sets of coupled ordinary differential equations which describe autoregulatory transcription factor networks (see box 2). Remarkably, Kaneko’s work shows that even if the internal networks, the initial state of the cell and the environment composition are randomized, a significant fraction of all possible cases develops consistently into virtual organisms with life-like features.
Cellular states are defined by the set of concentration of each species, namely ${C_{k}^1(t), ... , C_{k}^l(t)}$ for the concentration of the [*l*]{} species in the [*k*th]{} cell at time [*t*]{}. A reaction matrix with notation $W(R, P, Q)$ is constructed, each position taking the value $1$ if there exists a reaction from chemical $R$ to chemical $P$ catalyzed by $Q$ and $0$ otherwise. The reaction matrix is generated randomly before each simulation and remains fixed during that simulation. Concentration change in this simple set of 3 species, namely Reactant $R$, Product $P$ and Catalyzer $Q$. The equations read:
$${\partial C_{k}^P \over \partial t} = W(R, P, Q)C_{k}^R (C_{k}^Q)^2$$ $${\partial C_{k}^R \over \partial t} = -W(R, P, Q)C_{k}^R (C_{k}^Q)^2$$
The environment is simulated by a lattice of equal sized patches, each one of them characterized by ${C^1(x,y,t), ... , C^l(x,y,t)}$. Two different modes of material transfer are considered, passive diffusion and active transport. Diffusion takes place among neighbor finite elements and between occupied finite elements and the cells in them. It is assumed proportional to the difference of concentrations. Active transport, on the other hand, is considered to be only proportional to the concentration in the local finite element and modulated by the internal state of the cell. Thus the term accounting for diffusion and transport of the [*i*th]{} chemical in the [*k*th]{} cell is: $${\partial C_{k}^i \over \partial t} = D_{e} \bigtriangledown^2 C^i(x,y) + D_{c}\sigma_{i}(C^i(x,y)-C_{k}^i)+T_{c}C^i(x,y) \sum_{j=1}^l C_{k}^j$$
Where $D_{e}$ and $D_{c}$ stand for environment and cell diffusion coefficients respectively. The term $\sigma_{i}$ represents the membrane permeability to the [*i*th]{} chemical and, as a simplification, can only take two values (0 or 1), fixed at random before the simulation. $T_{c}$ is the transport constant equal for all species and cells. The boundary conditions of the system are maintained to a high concentration of a source metabolite, which can be interpreted as feeding the tank where the simulated cells are placed with raw materials.
Division is considered to be the direct consequence of the accumulation of key, predefined chemicals inside the cell. A threshold function is defined, namely $$\int_{t_{0}}^{t_{0}+\tau}C_{k}^ldt \geq D$$ as the criterion for cell division and splitting of resources between daughter cells, which are generated close to the position of the mother cell. Likewise, if the threshold condition $$\sum_{i} C_{k}^i < S$$ is met, cells are removed from the simulation.
Kaneko and colleagues observed that from the subset of configurations displaying cell growth and intermediate connectivity (case study in Kaneko and Yomo 1999 contains 9 paths for a system of 32 chemical species), most led to periodic or quasi-periodic changes in metabolite concentrations inside the cells, analogous to natural cell cycles observed in both unicellular and multicellular organisms. Beyond a certain population threshold, synchronization and phase stabilization appeared among the cells in the same cluster, the first steps towards cooperation and collective action in cell societies. Further increase in the numbers within the ensemble enables the divergence of the mean chemical concentrations and the cycle itself (i.e. the differentiation of the cells). These new trajectories in the phase diagram were stabilized by mutually reinforcing metabolite exchanges between neighboring cells.
This process of generation of variants from previously equal individuals, dubbed isologous diversification, provides a solid testbed to study both the community effect (Carnac and Gurdon 1997) and positional information (Wolpert 1969) theories. Moreover, after undergoing a first round of differentiation, cells could subsequently change into previously unavailable cell types, thus creating a tree-like hierarchical structure observed in natural developmental lineages. Some of this virtual organisms displayed other key features of living systems like robustness to noise during development (Kaneko and Yomo 1999), resistance to “injury” through regeneration of spatial patterning (Furusawa and Kaneko 1998) and even life cycles exhibiting senile/proliferative stages (Furusawa and Kaneko 1998).
In conclusion, Kaneko’s model demonstrates that even in the absence of evolution or selection, valid unicellular genotypes have the potential to create complex, emergent phenomena given that size thresholds can be surpassed. Whether this phenotypical changes suppose an increase in fitness is not of relevance here, but the feasibility to become multicellular and display potential task allocation as a side effect of cell-cell interactions. This results articulate a clear connection to Kauffman’s work (Kauffman 1993) and the realization that some of the high order features observed in natural systems can arise not as a result of natural selection but the unavoidable properties of systems with high epistatic connectivity.
Evolving multicellular aggregates
=================================
Multicellularity has been a recurrent novelty in the story of life and some clues to its origins can be found (at least at the functional level) in living unicellular systems, such as bacteria or yeast. Many unicellular species display multicellular traits (Shapiro 1998; Bonner 2001) associated to the presence of signals that provide the source of coherent population responses. As a consequence of these responses (mainly to stress signals) multicellular aggregates can form and display some degree of specialization and/or differentiation. Following these observations, a very promising approach to study the feasibility of the transition to multicellularity is the use of artificial selection in natural systems.
This strategy was put forward in a recent set of experiments (Ratcliff et al. 2012), in which the authors sequentially subcultured *S. cerevisiae* cells with the fastest sedimentation in order to force the selection of cooperating aggregates. Yeast is a specially interesting candidate to explore the potential first steps due to the fact that it already presents some pre-adaptations thought to be relevant in this major transition in evolution (Maynard-Smith and Szathmáry 1995; Szathmáry 1994) and its biology and multicellular states are enough well described so that new emergent phenomena can be easily distinguished from them. Remarkably, after just 60 selection events -in a timescale much shorter than previously thought- the so called snowflake phenotype appeared consistently in all cultures. These are roughly spherical clusters of cells formed not by aggregation but by defective separation of cells after division.
The clonal formation of the clusters ensures limited conflict of interest among the elements of the new multicellular individual, as discussed elsewhere (see Grossberg and Strahmann 2007; Bonner 2001; Michod 2000). Later on, the authors studied the role played by cellular fate in cluster reproduction. It was found that clusters did not reproduce through events associated to single cells but instead involved a group-level set of events that lead to the generation of a propagule. This was achieved through a division of labor in the form of the active control of apoptosis, which caused the asymmetrical splitting of the cluster once it reached a threshold size.
In order to test alternative explanations to some of the phenomena observed -particularly the presence of a group-level reproduction and its relation to division of labor- and test other potential scenarios for the rise of multicellular ensembles, a simple embodied computational model was created (Duran-Nebreda and Solé 2014). In it yeast cells are modeled as point elements in a bidimensional lattice which can fail to separate correctly after division, thus remaining attached by a spring. Cell’s movement is modeled by a biased random-walk. No explicit genetic network is implemented, instead cells inherit the mother’s parameters with small deviations. Two causes of cell death are tested, apoptosis as well as a simpler alternative based on the depletion of resources.
[*Cell death caused by nutrient depletion*]{} To take into account this process, cells are placed in a bidimensional lattice that holds nutrient. Cells have an energy value $M_{ni}$, a division threshold $M_{ni}^c$, a counter on the number of divisions $\Delta_{ni}$ and an attachment probability to daughter cells once they divide $p_{ni}$. Nutrient concentration change in the finite element $R_{ij}$ is given by the following equation: $${\partial R_{ij} \over \partial t} = D \bigtriangledown^2 R_{ij} - \rho \Theta_{ij} R_{ij}$$ The Heaviside function $\Theta_{ij} $ is used to indicate the presence or absence of cells in that particular patch of the lattice (so $\Theta_{ij} =1$ if a cell is present and zero otherwise). The diffusion operator $\nabla^2 R_{ij}$ is numerically computed following a standard discretization form: $$\nabla^2 R_{ij} = {\partial^2 R_{ij} \over \partial x^2} + {\partial^2 R_{ij} \over \partial y^2} \approx \left [ R_{ij} - {1\over4} \sum_{kl} R_{kl} \right]$$ where $D$ accounts for the coefficient diffusion and $\rho$ the intake of nutrients from the culture medium. The energy change for $C_{ij}$ is: $${\partial M_{ij} \over \partial t} = \rho R_{ij} - \beta_c M_{ij} (1+\kappa \Delta_{ni})$$ Where $\beta_c$ represents the maintenance costs. If the energy value of a particular cell reaches its division threshold, a new cell is created and the original energy value is split asymmetrically between the cells. Conversely, cells die if: $M_{ni} \leq \delta_c$, Where $\delta_c$ is the energy limit cells can withstand.
The conclusions extracted from the simulations draw a slightly different picture on the possible first steps towards multicellularity. In it, physical constraints previously linked to a decrease in fitness for the aggregates -namely the added difficulty to attain enough nutrients to survive caused by limited diffusion in the core- are shown to work in favor of reproductive fitness, debilitating the cluster structure and promoting splitting after a certain size threshold is achieved.
Physical forces and ecological scales
=====================================
An important component of a consistent theory of multicellularity, particularly in relation to the emergence of not just cooperation, but also developmental programs, requires considering community ecology in embodied models. Such models introduce physics and spatial interactions, and they naturally incorporate some selection forces, since the explicit physics introduce strong constraints to the potential forms and multicellular aggregates that can be obtained. If cell aggregates explicitly move, adhere to substrates, develop cooperation through cell-cell shared nutrients and resist external fluctuations, the evolution of adhesion and other variable features will occur under well-defined selection pressures. If these changes take place in a physical environment where available resources spread and are consumed, ecosystem-level processes might take place, some particular structures, such as budding, coherent multicellular shapes, differentiated aggregates or even life cycles could emerge. Moreover, potential changes in grazing efficiencies and the rise of predators can trigger arm races tied to changes in developmental programs, as it is likely the case for the transition between the Ediacaran and Cambrian biotas (Marshall 2006; Fedonkin 2007; Erwin and Valentine 2013).
![Basic scheme of the components of the CHIMERA model (Solé and Valverde 2013). The system (a) is confined within a rigid cube. Nutrient particles fall from the top layer experiencing physical forces. Cells also experience the same forces, as described by Newton’s laws (see Box 4). Additionally, both cells and particles get degraded. Cells can interact with the boundaries of the system as well as between them. Cells increase in mass every time they collide with a nutrient particle if they have the right internal metabolism. We also display the mechanical forces acting between cells (b) and the interactions between cells and the boundaries (c-d). In both cases, adhesion forces stabilize interactions within some range, but interpenetration is forbidden.[]{data-label="chimera"}](chimeracube)
The CHIMERA model was build as an advanced framework introducing artificial cells as particles in a physical world where Newtonian forces, along with selection pressures, genes and metabolism are taken into account.
[*Cells and particles*]{} Our starting point is a population of single-cell organisms, where each cell in the initial population is identical. Cells and particles are simulated with rigid bodies moving within a fluid-like environment. A cell (particle) has spherical geometry with radius $R_i$, mass $M_i$, spatial position ${\bf
r}_i$ and velocity $${\bf v_i} = d{\bf r}_i/dt .$$ The motion of a cell is described by the standard second law: $$M_i {d {\bf v}_i \over dt} = {\bf F}_i
\label{eq_cell_force}$$ Cell movement is obtained by numerical integration of the Newton equations. Cell velocity at time $t + \Delta t $ is thus: $${\bf v}_i(t + \Delta t) = {\bf v}_i (t ) + {{\bf F}_i/{M_i}} {\Delta t}$$ where $\Delta t$ is the size of the integration step, and the total force acting on $M_i$ will be: $${\bf F}_i = {\bf F}_e + {\bf F}_i^c + {\bf F}_i^n + {\bf F}_i^l -
k_d {\bf v}_i + M_i {\bf g}$$
applied to any cell is the sum of environmental forces ${\bf F}_e$, the gravitational field ${\bf g}$, the collision force ${\bf F}_i^c$, the cell-wall adhesion ${\bf F}_i^n$ and the cell-cell adhesion ${\bf F}_i^l$ term, along with the drag force associated to the medium, i. e. $-k_d {\bf v}_i $.
[*Cell-substrate adhesion*]{} Attachment of cells to surfaces may provide a favorable micro-environment for cell aggregates to develop. If $D({\bf r}_i)$ indicates the cell-to-wall distance, when a cell with adhesion strength to the substrate $J_i^f > 0$ is closer than the adhesion range $\delta_w > 0$, i.e., $$D({\bf x}_i) < \delta_w ,$$ we attach a spring connecting the cell ${\bf x}_i$ with its projection on the wall ${\bf x}_i^w$ (Figure 6). Now, the wall spring exerts the following attraction force:
$${\bf F}_i^n = - k_s \left ( ||{\bf r}_i- {\bf r}_i^w || - d_s \right ) {
{{\bf r}_i- {\bf r}_i^w} \over { ||{{\bf r}_i- {\bf r}_i^w}||} }
\label{cellwall_force}$$
where $d_s$ is the spring equilibrium distance, $k_s$ is the spring constant and ${\bf F}_n^i = 0$ when the cell is not attached to any spring. Existing cell-wall springs can be removed with certain probability $1- q(i)$ or when the spring length is above the maximal length, i.e., $$||{\bf r}_i- {\bf r}_i^w|| > d_s^m .$$ As we will see, cells can evolve cell-wall adhesion $J_i^f$ in order to maximize the intake of nutrient particles.
[*Cell-cell adhesion*]{} Cells can form aggregates by attaching to other cells. Each cell has an intrinsic probability $J_i^c$ to create a new adhesion link. Given two close cells located at ${\bf r}_i$ and ${\bf r}_j$, we will set an adhesion string connecting them with probability $(J_i^c + J_j^c )/2$. The adhesion force to any cell is the sum of forces contributed by all the active cell-cell adhesion springs: $${\bf F}_i^l = - \sum_j k_l \left (||{{\bf r}_i- {\bf r}_j}|| - d_l
\right ) {{\bf r}_{ij} \over ||{\bf r}_{ij}|| }
\label{cellcell_force}$$ where ${\bf r}_{ij} = {\bf r}_i- {\bf r}_j$, $d_l$ is the spring equilibrium distance, and $k_l$ is the spring constant. Adhesion springs break spontaneously with rate $\delta \approx 0.001$ or when the spring is very large, i.e., $$||{\bf r}_i- {\bf r}_j|| > d_l.$$
In order to incorporate all these components, we need to build a complex simulation framework able to capture the essential physics, the presence of a population of interacting artificial agents and mechanisms of evolving the parameters that weight different metabolic and adhesion properties. Such type of model belongs to the tradition of so called artificial life approaches (Ray 1991; Langton 1991, 1995; Sipper 1995; Adami 1998) which involve the study of artificial life-like systems in artificial environments (along with a wet version associated with the construction of living systems using genetic engineering techniques). Computational models, which can reproduce realistic scenarios or completely ignore them, provide an ideal framework to explore the generative potential of an evolving set of rules allowing structures to emerge through time. Following this view, an embodied model of evolution, the so called CHIMERA model (Solé and Valverde 2013) was introduced as a way of including newtonian physics, fluctuations, evolution and ecology in a unified fashion.
![(a) in the basic, newtonian CHIMERA model, the evolution of the system under enough nutrient levels drives the population from a few layers placed at the bottom to an inverted system where cells occupy and adhere to the upper wall. This occurs thanks to a broad distribution of efficiencies (i. e. evolved generalists) exploiting all resources, together with an increased adhesion between cells and specially between cells and surfaces. The main plot shows a fast growth in cell-floor adhesion towards its maximum value. Two snapshots of the system are also shown, before and after the transition. In (b) an example of an evolved multicellular aggregate is shown. This was obtained by using a more detailed implementation of the cell-cell interactions that allow aggregates to emerge Solé and Valverde 2013).[]{data-label="chimera"}](chimeraresults)
The model was intended as an approach to the pre-Mendelian universe, which can be approached by studying the interplay between physical forces such as gravity, diffusion and adhesion and generic pattern-forming mechanisms. In its simplest version (Solé and Valverde 2013) Chimera considers a cubic world involving a fluid-like medium with gravity and turbulence (figure 7) where an initial set of identical “cells” exploit one of a number of potential energy particles falling from the upper side of the cube. A set of rules is then used to evolve the system:
1. Movement: both particles and cells experience both a gravitational and a fluctuating velocity field (the later associated to turbulence). Particles are removed from the system with some probability.
2. Each cell carries a given set of internal parameters and variables: they have a given size and mass and they have a list of possible particle types that they can take and the efficiency of the grazing for each particle type.
3. Cells can attach to the surface of the walls with some probability. When they attach, a spring is used to properly define the physical interaction. Another adhesion probability is used for cell-cell adhesion. At the beginning all are set to zero.
4. If a cell interacts (collides) with a given particle, it ingests it if the efficiency for metabolizing that type of particle is non-zero. If taken, the mass of the particle gets transformed into mass of the cell.
5. Once a maximum cell size is reached, the cell splits into two daughter cells. Moreover, if the cell goes below a minimal value, it dies and it disintegrate. Detritus particles are also allowed to be part of the nutrient intake of cells.
6. Each time a cell divides, mutations can occur in the daughter cell. Metabolic efficiencies and adhesion rates can change.
The model is able to display complex forms of pre-multicellularity in terms of loose aggregates of cooperating cells. These aggregates evolve adhesion rates, both between cells but especially in relation to the physical substrate. Two major trends take place here. From the point of view of grazing efficiency, individual cells tend to become generalists: their metabolism changes in order to exploit all types of nutrient particles, although at the price of being less efficient. Secondly, in order to gather more particles, there is an advantage in providing a larger cell surface, which is possible provided that cells increase their attachment probability (adhesion force) and occupy available space on the lateral walls. This tendency, once started, is rapidly amplified and we can see (figure 8a) that eventually the cells “discover” the source of energy and occupy it. This is done by an increase of cell-floor adhesion but also increasing cell-cell adhesion and (when fluctuations are large) cooperation.
Eventually, the flow of particles is almost shut down except for cell mortality and the sedimentation of cells. Cell death creates detritus particles, which are also exploited with low efficiency. But one a critical amount of detritus gets accumulated at the bottom, a new community of specialized detritivores emerges. This new ecosystem with two layers is stable over time and represents a dramatic example of how ecosystem engineering emerges: the feedbacks between organisms and environment trigger a control of the later by the first, with a deep reorganization of energy flows (Jones et al 1994, 1997; Hastings 2006; Erwin 2008; Erwin and Tweedt 2012).
The outcome of the simulation reveals how a microscopic process of evolution associated with cell adhesion, combined with a community adaptation to the environment leads to a major innovation. Using simpler physics (Palsson and Othmer 2000; Palsson 2001; Ericson 2005; Sandersius and Newman 2008) where cells have volumes but their behavior is closer to point particles does not allow the formation of complex aggregates (Box 4). However, a more sophisticated and realistic definition of forces and spatial interactions allows the formation of multicellular aggregates (figure 8b) thus suggesting that much more can be obtained even under these simplified scenarios (Solé and Valverde 2013).
Combinatorial explosions and the Cambrian conundrum
===================================================
As a final example, we consider the problem of how complex and diverse spatial patterns of gene expression (and thus cell types) can emerge as a consequence of gene networks in development. Specifically, we consider an abstract model of pattern formation where a one-dimensional digital organism is composed of $C$ cells each carrying the same gene network.
A simple, discrete model of pattern formation can be implemented by defining a set of $N=G+H$ genes interacting through a one-dimensional domain involving $C$ cells (Solé et al., 2003). Here gene states will be Boolean: genes are either active ($1$) or inactive ($0$). $G$ genes interacting within the cell, whose state at a given time $t$ will be indicated as $g_i^j(t)$, where $i=1, ..., G$ is gene number and $j=1, ..., C$ is cell number (ordered from anterior to posterior). The second group are generically labeled microhormones and actually involve (implicitly) some local mediators communicating neighboring cells.
The state of these $H$ hormones will be indicated as $h_i^j(t)$. Hormones can receive inputs from any of the first $G$ units, but they can only make output to genes in other cells. Two matrices will be required in order to define the whole spectrum of links between different elements. These two matrices will be indicated by ${\bf W}=(W_{ik})$ and ${\bf C}=(C_{ik})$, defining interactions among the $G$ genes and between genes and hormones, respectively.
The basic set of equations of our gene network model read: $$g_i^j(t+1) = \Phi \left [ \sum_{l=1}^G W_{il}g_l^j(t) +
\sum_{k=1}^H C_{ik} \delta \left( h_k^{j+1}(t) , h_k^{j-1}(t) \right) \right]$$ where $\delta(x,y)=1$ if both $x=y=1$ and zero otherwise (i.e. an “OR” function). Similarly, hormones receive inputs only from inside the cell, $$h_i^j(t+1)= \Phi \left [
\sum_{l=1}^G W_{il}g_l^j(t)
\right ]$$ with additional, specific equations for the boundaries. The function $\Phi(x)$ is a threshold function, i. e. $\Phi(x)=1$ if $x>0$ and zero (inactive) otherwise. Specific equations are also defined for the two ends of the system.
Finally, we need to define an initial condition. The simplest choice that can be defined involves the activation of a single gene at the anterior ($j=1$) end of the digital organism. Specifically, we set $g_{i}^j(0)=h_{i}^j(0)=0$ for all $j=1, ..., N_c$, except $g_1^1(0)=1$. Thise choices corresponds to maternal signals confined to the chosen end. Such initial change will propagate to the rest of the tissue provided that the network allows it.
The gene network includes both gene-gene interactions within the cell and between cells. In other words, we take into account regulatory interactions taking place within each cell together with cell communication through given signals (to be called hormones). These models have been extensively used since the early days of theoretical biology (Kauffman 1993) and and provide a simple way of approaching the problem of defining cell types and thus multicellular assemblies.
In its simplest form, we can define a gene network in terms of a set of $n$ genes whose states $g_i$ are confined to two possibilities, namely $g_i \in \{ 0, 1 \}$. This binary approximation implies that genes are ON-OFF elements, which of course is a simplified picture. If the effect of gene $g_i$ on gene $g_j$ is given by a weight $W_{ij}$. If the gene interaction describes an activation or inhibition, $W_{ij}$ will be positive or negative, respectively. No interaction is given by $W_{ij}=0$. For simplicity, a discrete space of weights is used, namely $W_{ij} \in \{-1,0,+1\}$. In this way, a full exploration of the potential set of states can be performed.
The state of a gene will change as a consequence of its interactions with other genes. This state is updated in discrete time units following: $$g_i^j(t+1) = \Phi \left [ \sum_{l=1}^n W_{il}g_l^j(t) \right]$$ which essentially tells us that the gene will become active or inactive if the global input acting on it is positive or negative, respectively. These networks can generate extremely simple (say, all genes inactive or active) or very complex dynamics (when chaotic changes occur). But for some ranges of connectivities, it leads to a rich diversity of stable states. If a cell type $T$ is identified as a string $\bf S_T$ of active and inactive genes, namely ${\bf S_T}=(g_1, ..., g_n)$ (Kauffman 1993) we have a potential of $2^n$ alternative strings. We can expand this formalism to take into account space, if multiple cells are also taken into account (Box 5). As shown in figure 9a, this is easily implemented and a spatial pattern can be described, for each gene, in terms of its expression level ($0$ or $1$) in different cells. A detailed analysis of this type of model reveals that some types of patterns are easily found (Solé et al 2003; Munteanu and Solé 2008; Tusscher and Hogeweg 2011). This is the case of regular stripe patterns, which is a rather common one.
![Transitions from simple to complex (diverse) patterns in a minimal model of gene network model of biological pattern formation. The model considers a population of digital organisms composed by a linear chain of $C$ cells, each carrying the same genetic network (see Box 5). The complexity of this network, measured in terms of the number of genes $G$ associated to internal switches and the number of cell-cell signals $H$ determines the amount of patterns that can be achieved through evolution. In (a) dark and light cells indicate the high and low expression of one of the genes, respectively. In figures (b-c) the complete set of patterns generated in this way is shown for different sets of genes. For $H=2,G=1$ or $H=1,G=2$ only a few patterns are accessible (b). Once we have $H=G=2$, all possible patterns can be reached (c). []{data-label="creature"}](EVODEVOmodel)
What is the generative potential of a given gene network complexity level? More precisely, if we start from a small network with $G=1$ genes and $H=1$ hormones and expand its number, what are the consequences? These questions are relevant when we consider the early events that shaped the gene networks affecting developmental processes in the history of multicellular life. Once again, thinking in the Cambrian explosion of life, several factors can be considered (Marshall 2006; Erwin and Valentine 2013). These include external, abiotic factors as well as internal ones and they are likely to interact among them. But it would be interesting to know, even under some basic abstract model, if some particular elements can play a key role in promoting sudden changes in the amount of achievable morphologies. In order to analyze this problem, we use the presence of a genotype-phenotype mapping $\Omega$, namely $$\Omega : {\bf W} \longrightarrow \{ \Phi \}$$ between the set of matrices $\bf W$ and the corresponding set of patterns $\{ \Phi \}$. In other words, for a given network $W_a$, the arrangement of ON and OFF genes defining a stable pattern $P_a^*$ can be written as $P_a^* = \Phi(W_a) \in \{ \Phi \}$.
Using our basic model, it is possible to explore the space of potential phenotypes through single mutations in the genotype, as defined by the matrix of gene-gene interactions. Each step in the simulation, we evolve the system by changing single elements in the matrix. Patterns that are not stable are discarded and a change in the matrix is accepted if the diversity of cells within the organism is increased (or at least remains the same). This movement in sequence space is known as an [*adaptive walk*]{} (Kauffman and Levin 1987).
Pattern-forming gene networks display sharp thresholds affecting their combinatorial potential. For small numbers of elements, i. e. when $H+G<4$ and $H, G \le 2$ the range of possible spatial patterns is rather limited (figure 9b) but once the critical number $H=G=2$ is reached, all patterns become accessible (Solé et al. 2003). This is a very interesting finding, since it provides a possible logical explanation for the rapid diversification of developmental paths when genetic complexity thresholds are crossed. Along with other influences, a small increase in regulatory complexity can account for a sudden jump in the achievable diversity of developmental pathways.
Although these results are obtained from a toy model of regulatory interactions and ignore other pattern-forming factors, such as tissue organization, morpho-dynamic processes or cell division, sorting and apoptosis, the basic conclusions are likely to be robust: a relatively small increase in underlying genomic complexity can lead to rich morphogenetic potential (Marshall 2006). In earlier models of evolution on fitness landscapes (Niklas 1994) high diversity is linked to the presence of multiple optima on a morphological landscape. If such optima are easily reached, a diverse range of structures is expected to be obtained. An interesting feature of the space of spatial patterns defined by the gene network model is that it displays neutrality: large, neutral networks percolate sequence space allowing for efficient exploration of the phenotype space. The structure of this pattern forming network space is actually very similar to other found in RNA folding (see Solé et al 2003; Munteanu and Solé 2008). This result tells us something important. As soon as we reach the critical threshold of network complexity, not only multiple patterns become accessible. The intrinsic evolvability of the system is also very high.
Discussion and prospects
========================
In silico models of evolutionary change should be a natural component of our exploration of macroevolutionary patterns and the tempo and mode of evolutionary transitions. Despite their limitations, they offer, along with experimental dynamics using microbial populations (Lenski and Travisano 1994; Elena and Lenski 2003) what no other approach can: an opportunity to recreate potential scenarios and formulate hypothesis about how complexity developed over time. Here we have summarized the outcomes of different models of artificially evolved “organisms” or even communities where different types of transitions in complexity are achieved thanks to the (nonlinear) interplay between genetic and physical -and even ecological- components. Although they are all far from a realistic representation of developmental body plans or true communities, the previous results reveal a great generative potential implicit in the simple rules. In all cases, multicellular complexity undergoes increases or even jumps and some remarkable results can be highlighted. These include, for example, the emergence of some ontogenetic processes resulting from an evolutionary algorithm searching for diverse cell types. Such processes typically incorporate cell-cell interactions that provide the capacity for tissue reorganization and shape changes together with cell differentiation. But even cellular and ecological scales can become related once evolving adhesion provides the exploratory potential for community-level processes to unfold. This connection between such disparate scales provides a novel way of re-considering the problem of hierarchies in evolution (Eldredge 1985; McShea 2001).
![Virtual creatures (a) can be obtained using physical models of interacting subparts defining a class of artificial life form and evolving under given selection constraints (rendering by Zach Winkler using the Stellar Alchemy program). Although they suggest that we can approach early life forms such as the Cambrian creatures like Opabinia (b), they strongly differ in the ways their complexity is generated (rendering by Nobu Tamura). []{data-label="creature"}](3Devolved)
Some more sophisticated models have been created which are capable of evolving complex creatures with multiple connected components (figure 10a). These models involve a more or less detailed physical context, both in terms of the elements used to describe the virtual creature and the physics of the environment. Starting from Karl Sims work, these model involve a set of connected components (often parallelepipeds) linked to each other through strings (Sims 1994). The final outcome of this simulations often reminds us of some type of living creature. However, there is an essential difference: in the evolved artificial creatures there is no developmental program at work and thus there is no genotype-phenotype mapping. This is no minor drawback, since developmental programs are the essential component required to properly understand and model evolutionary paths. Over the last years, novel approaches to this problem incorporating some type of morphogenetic rules are being considered (Doursat 2008; Jin 2011).
In the CHIMERA framework, Our artificial creatures are autogenic engineers (Jones 1994): they change their environment mainly via their own physical structures. The success of our model might be due to the complete set of key components that we allow to interact freely. By using space, diverse ecosystems can be build through spatial segregation of subpopulations. By allowing simple components of pattern formation or aggregate generation it is possible to introduce simple forms of cooperative dynamics. By embedding the virtual creatures within an ecosystem where physics plays a role, selection pressures restrict the repertoire of cellular aggregates that can be formed.
Future work should address the potential for generating complex structures perhaps similar to the Ediacaran fauna and test the role payed by both internal and external innovation triggers. The first includes for example the emergence of predators and the resulting arm races, which are known to be a major player in expanding morphological complexity. The second deals with extinctions caused by geological or astronomical shocks which deeply altered communities and whole ecosystems. The aftermath of the extinction provides a unique microscope to see different evolutionary processes in action. Such recovery patterns have been studied both from field data and modeling (Benton and Twitchett 2003; Erwin 1998, 2001; Solé et al 2002; Chen and Benton 2012; Yedid et al 2012) and offer an additional test for studying how organismal and ecological complexity react to stress.
As a final point, it is worth mentioning that another avenue to address the dawn of multicellular systems is provided by synthetic biology (Benner and Sismour 2005; Solé et al. 2007; Cheng and Lu 2012) which is considered by some researchers as the [*wet*]{} version of artificial life. By engineering unicellular systems, it is possible to build novel forms of cell-cell communication and thus create (and perhaps re-create) novel forms of multicellular assemblies, able to perform novel functions and even complex computations (Regot et al. 2011; Macia et al 2012; Chuang 2012). Given the potential offered by genetic engineering techniques to alter the logic of cell-cell exchanges, we have a unique opportunity of exploring the landscape of transitions from uni- to multicellular forms of organization.
[**Acknowledgements**]{}
We thank the members of the Complex Systems Lab for useful discussions. This work has been supported by grants of the James McDonnell Foundation, the Botin Foundation and by the Santa Fe Institute.
**The present document is a chapter of the book titled *Evolutionary transitions to multicellular life* (Springer) Edited by: A. Nedelcu and I. Ruiz-Trillo. The final publication is available at http://link.springer.com/**
[**References**]{}
1. Abedin, M. and King, N. 2010. Diverse evolutionary paths to cell adhesion. Trends in Cell Biology 20, 734-742.
2. Adami, C. 1998. Introduction to Artificial Life. Springer Verlag, New York.
3. Alberch, P. 1980. Ontogenesis and morphological diversification. Amer. Zool. 20, 653-667.
4. Anderson, P. 1972. More is different. Science 177, 393-396
5. Benner, S.A. and Sismour, A. M. 2005. Synthetic biology. Nat. Rev. Genet. 6, 533-543.
6. Benton M.J., and Twitchett R.J. 2003. How to kill (almost) all life: the end-Permian extinction event. Trends Ecol. Evol. 18, 358-365.
7. Bonner, J. T. 2001. First signals: the evolution of multicellular development. Princeton University Press. Princeton.
8. Buss, L.W. 1987. The evolution of individuality. Princeton University Press. Princeton.
9. Carnac, G. and Gurdon J.B. 1997. The community effect in Xenopus myogenesis is promoted by dorsalizing factors. Int. J. Dev. Biol. 41, 521-524.
10. Carroll, S.B. 2005. Endless forms most beautiful: The New Science of Evo Devo and the Making of the Animal Kingdom. W. W. Norton and Co.
11. Chen, Z.Q. and Benton, M.J. 2012. The timing and pattern of biotic recovery following the end-Permian mass extinction.
12. Cheng, A. and Lu, T.K. 2012. Synthetic Biology: An Emerging Engineering Discipline Annu. Rev. Biomed. Eng. 14, 155-178.
13. Chuang, J.S. 2012. Engineering multicellular traits in synthetic microbial populations. Curr. Opin. Chem. Biol. 16, 370-378.
14. Coen, E., Rolland-Lagan, A.G., Matthews, M., Bangham, J. A. and Prusinkiewicz, P. 2004. The genetics of geometry. Proc. Natl. Acad. Sci. USA 101, 4728-4735.
15. Cummings, F. 2006. On the origin of pattern and form in early metazoans. Int. J. Dev. Biol. 50, 193-208.
16. Doursat, R. 2008. Organically grown architectures: creating decentralized, autonomous systems by embryomorphic engineering. In: Organic Computing, R. P. Wrtz, ed.: pp. 167-200. Springer-Verlag.
17. Duran-Nebreda, S. and Solé, R.V. 2014. Emergence of multicellularity in a model of cell growth, death and aggregation under size-dependent selection. *Submitted to* Proceedings of the Royal Society B.
18. Eggenberger, P. 1997. Evolving morphologies of simulated 3d organisms based on differential gene expression. In. Fourth Europ. Conf. Artificial Life, P. Husbands and I. Harvey, editors. pp. 205-213. MIT Press.
19. Eldredge, N. 1985. Unfinished Synthesis: Biological Hierarchies and Modern Evolutionary Thought. Oxford University Press. Oxford.
20. Elena, S.F.and Lenski, R.E. 2003. Evolution experiments with microorganisms: the dynamics and genetic bases of adaptation. Nat. Rev. Genet. 4(6):457-69.
21. Ericson, C. 2005. Real-Time Collision Detection. Morgan Kaufmann.
22. Erwin, D. H. 1998. The end and the beginning: recoveries from mass extinctions. Trends Ecol. Evol. 13, 344-349.
23. Erwin, D. H. 2001. Lessons from the past: biotic recoveries from mass extinction. Proc. Natl Acad. Sci. USA 98, 5399-5403.
24. Erwin, D.H. and Tweedt, S. 2012. Ecological drivers of the Ediacaran-Cambrian diversification of Metazoa. Evol. Ecol. 26, 417-433
25. Erwin, D.H. and Valentine, J. 2013. The Cambrian Explosion: The Construction of Animal Biodiversity. Roberts and Company.
26. Fedonkin, M.A. 2007. The Rise of Animals: Evolution and Diversification of the Kingdom Animalia. JHU Press.
27. Forgacs, G. and Newman, S. A. 2005. Biological physics of the developing embryo. Cambridge Univ. Press, Cambridge.
28. Forrest, S. Genetic algorithms: Principles of natural selection applied to computation. Science 1993, 261, 872-878.
29. Furusawa, C. and Kaneko, K. 1998. Emergence of multicellular organisms with Dynamic differentiation and spatial pattern. Artificial life 4, volume 1.
30. Furusawa, C. and Kaneko, K. 2003. Robust development as a consequence of generated positional information. J. Theor. Biol. 224, 413-435.
31. Fontana, W. and Schuster, P. 1998. Continuity in evolution: on the nature of transitions. Science 29, 280(5368):1451-5.
32. Goodwin, B. C. 1994. How the leopard got its spots. Princeton U. Press, Princeton.
33. Glazier, J. A. and Graner, F. 1993. Simulation of the differential adhesion driven rearrangement of biological cells. Phys. Rev. E 47, 2128-2154.
34. Graner, F. and Glazier, J.A. 1992. Simulation of biological cell sorting using a two-dimensional extended Potts model. Phys. Rev. Lett. 69, 2013-2016.
35. Grossberg, R.K. and and Strahmann, R.R. 2007. The Evolution of Multicellularity: A Minor Major Transition? Annu. Rev. Ecol. Evol. Syst. 38, 621-654.
36. Hastings, A. 2006. Ecosystem engineering in space and time. Ecol. Lett. 10, 153-164.
37. Hogeweg, P. 2000. Evolving Mechanisms of Morphogenesis: on the Interplay between Differential Adhesion and Cell Differentiation. J. Theor. Biol. 203, 317-333
38. Hogeweg, P. 2000. Shapes in the shadow: Evolutionary dynamics of morphogenesis. Artif. Life, 6: 85-101.
39. Jin, Y. 2011. Morphogenetic robotics: an emerging new field in developmental robotics. IEEE Trans. Syst. Man Cyber. 41,145-160.
40. Jones, C.G. , Lawton, J.M. and Shachak, M. 1994. Organisms as ecosystem engineers. OIKOS 69, 373-370.
41. Jones, C.G. , Lawton, J.M. and Shachak, M. 1997. Positive and negative effects of organisms as physical ecosystem engineers. Ecology 78, 1946-1957.
42. Kaandorp, J.A., Blom, J.G., Verhoef, J., Filatov, M., Postma, M. and Müller, W.E.G. 2008. Modelling genetic regulation of growth and form in a branching sponge. Proc. R. Soc. B 275, 2569-2575.
43. Kaneko, K. and Yomo, T. 1999. Isologous diversification for robust development of cell society. J. theor. Biol 199, 243-256.
44. Kauffman, S.A. and Levin S. 1987. Towards a general theory of adaptive walks on rugged landscapes. J. Theor. Biol. 128, 11-45.
45. Kaufmann, S.A. 1993. The Origins of Order: Self-Organization and Selection in Evolution. Oxford University Press.
46. King, N. 2004. The unicellular ancestry of animal development. Developmental Cell. 7, 313-325.
47. Knoll, A.H. 2011. The Multiple Origins of Complex Multicellularity. Annu. Rev. Earth Planet. Sci. 39, 217-239.
48. Langton, C.G. 1991. Life at the Edge of Chaos. in [*Artificial Life II*]{}, Addison-Wesley.
49. Langton,G.G. (ed) 1995. Artificial life: an overview. MIT Press, Cambridge MA
50. Lenski, R.E. and Travisano, M. 1994. Dynamics of adaptation and diversification: a 10,000-generation experiment with bacterial populations. Proc Natl Acad Sci U S A. 91, 6808-14.
51. Macía, J., Posas, F. and Solé R.V. 2012. Distributed computation: the new wave of synthetic biology devices. Trends Biotechnology 30(6):342-9.
52. Marshall, C.R. 2006. Explaining the Cambrian “Explosion” of animals. Annu. Rev. Earth Planet. Sci. 34, 355-384.
53. Maynard Smith, J. and Szathmáry, E. 1995. The major transitions in evolution. Oxford University Press. Oxford.
54. Michod, R. E. 2000. Darwinian Dynamics: Evolutionary Transitions in Fitness and Individuality. Princeton University Press. Princeton.
55. Mitchell, M. 1998. An Introduction to Genetic Algorithms. Bradford Books.
56. Munteanu, A. and Solé, R.V. 2008. Neutrality and robustness in evo-devo: emergence of lateral inhibition. PLoS Comput. Biol. 4, e1000226.
57. McShea, D.W. 2001. The hierarchical structure of organisms: a scale and documentation of a trend in the maximum. Paleobiology 27, 405-423.
58. Newman, S.A. and Baht, R. 2008. Dynamical patterning modules: physico-genetic determinants of morphological development and evolution. Phys. Biol. 5, 015008.
59. Newman, S.A. and Comper, W.D. 1990.’Generic’ physical mechanisms of morphogenesis and pattern formation.Development, 110(1):1-18.
60. Niklas, K.J. 1994. Morphological evolution through complex domains of fitness. Proc. Natl. Acad. Sci. USA 91, 6772-79.
61. Niklas, K.J. and Newman, S.A. 2013 The origins of multicellular organisms. Evol Dev 15:41-52
62. Palsson, E. 2001. A Three-dimensional model of cell movement in multicellular systems. Future Gener. Comp. Syst. 17, 835-852.
63. Palsson, E. and Othmer, H.G. 2000. A model for individual and collective cell movement in Distyostlium discoideum. Proc. Natl. Acad. Sci. USA 97, 10448-10453.
64. Ratcliff, W.C., Denison, R.F., Borrello, M. and Travisano, M. 2012. Experimental evolution of multicellularity. Proc. Natl. Acad. Sci USA. 109, 1595-1600.
65. Ray, T.S. 1991. An approach to the synthesis of life. In: Langton, C., Taylor, C. and Farmer, D. (eds) Artificial Life II. pp. 371-408.
66. Regot, S., Macia. J., Conde. N., Furukawa. K., Kjellén. J., Peeters. T., Hohmann. S., de Nadal. E., Posas. F. and Solé. R. 2011. Distributed biological computation with multicellular engineered networks. Nature 469(7329):207-11.
67. Sandersius, S.A. and Newman, T.J. 2008. Modelling cell rheology with the subcellular element model. Physical Biology 5, 015002.
68. Savill, N.J. and Hogeweg, P. 1997. Modeling morphogenesis: from single cells to crawling slugs. J Theor Biol 184, 229-235.
69. Schuster, P. 1996. How does complexity arise in evolution? Complexity 2, 22-30.
70. Shapiro, J.A. 1998. Thinking about bacterial populations as multicellular organisms. Annu. Rev. Microbiol. 52, 81-104.
71. Sims, K. 1994. Evolving 3D morphology and behavior by competition. Artif. Life 1, 353-372.
72. Sipper, M. 1995. Using artificial life using a simple, general cellular model. J. Artif. Life 2, 1-35.
73. Solé, R.V. 2009. Evolution and self-assembly of protocells. Int. J. Biochem. Cell. Biol. 41, 274-284.
74. Solé, R.V., Montoya, J.M. and Erwin, D.H. 2002. Recovery after mass extinction: evolutionary assembly in large-scale biosphere dynamics. Phil. Trans. Roy. Soc. London B 357, 697-707.
75. Solé, R.V. , Fernandez, P. and Kauffman, S.A. 2003. Adaptive walks in a gene network model of morphogenesis: insights into the Cambrian explosion. Int. J. Dev. Biol. 47: 685 - 693.
76. Solé, R.V. , Munteanu, A., Rodriguez-Caso, C. and Macia, J. 2007. Synthetic protocell biology: from reproduction to computation. Phil. Trans. R. Soc. B 362, 1727-1739.
77. Solé, R.V. and Valverde, S. 2013. Before the endless forms: embodied model of transition from single cells to aggregates to ecosystem engineering. PLoS One, 8(4):e59664. doi: 10.1371/journal.pone.0059664.
78. Szathmáry, E. 1994. Toy models for simple forms of multicellularity, soma and germ. J. Theor. Biol. 169, 125-132.
79. Tusscher, K.H. and Hogeweg, P. 2011. Evolution of networks for body plan patterning; interplay of modularity, robustness and evolvability. PLoS Comput Biol. 7, e1002208.
80. Yedid, G., Stredwick, J., Ofria, C.A. and Agapow, P.M. 2012. A comparison of the effects of random and selective mass extinctions on erosion of evolutionary history in communities of digital organisms. PLoS ONE 7, e37233.
81. Wolpert, L. 1969. Positional information and the spatial pattern of cellular differentiation. J. Theor. Biol. 25, 1-47.
[^1]: Corresponding author
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Relativistic heavy-ion collisions can produce extremely strong magnetic field in the collision regions. The spatial variation features of the magnetic fields are analyzed in detail for non-central Pb - Pb collisions at LHC $\sqrt{s_{NN}}$= 900, 2760 and 7000 GeV and Au-Au collisions at RHIC $\sqrt{s_{NN}}$ = 62.4, 130 and 200 GeV. The dependencies of magnetic field on proper time, collision energies and impact parameters are investigated in this paper. It is shown that a enormous with highly inhomogeneous spatial distribution magnetic field can indeed be created in off-central relativistic heavy-ion collisions in RHIC and LHC energy regions. The enormous magnetic field is produced just after the collision, and the magnitude of magnetic field of LHC energy region is larger that of RHIC energy region at the small proper time. It is found that the magnetic field in the LHC energy region decreases more quickly with the increase of the proper time than that of RHIC energy region.\
0.2cm Keywowds: Spatial distribution of chiral magnetic field, Non-central collision, chiral magnetic field
author:
- 'Yang Zhong$^{1,2}$'
- 'Chun-Bin Yang$^{1,3}$'
- 'Xu Cai$^{1,3}$'
- 'Sheng-Qin Feng$^{2,3}$'
title: The spatial distributions of magnetic field in the RHIC and LHC energy regions
---
Introduction {#intro}
============
The Chiral Magnetic Effect (CME) is the phenomenon of electric charge separation along the external magnetic field that is introduced by the chirality imbalance [@lab1; @lab2]. It is proposed by Ref. [@lab3; @lab4; @lab5; @lab6; @lab7] that off-central relativistic heavy-ion collisions can create strong transient magnetic fields due to the fast, oppositely directed motion of two colliding nuclei. The magnetic field perpendicular to the reaction plane is aligned. Extremely strong (electromagnetic) magnetic fields are present in non-central collisions, albeit for a very short time. Thus, relativistic heavy-ion collisions provide a unique terrestrial environment to study QCD in strong magnetic field surroundings [@lab8; @lab9; @lab10; @lab11]. This so-called chiral magnetic effect may serve as a sign of the local P and CP violation of QCD. By using relativistic heavy-ion collisions at the Relativistic Heavy-Ion Collider (RHIC) and the Large Hadron Collider (LHC), one can investigate the behavior of QCD at extremely high-energy densities.
In non-central collisions opposite charge quarks would tend to be emitted in opposite directions relative to the system angular momentum [@lab9; @lab12; @lab13; @lab14]. This asymmetry in the emission of quarks would be reflected in an analogous asymmetry between positive- and negative-pion emission directions. This phenomenon is introduced by the large (electro-) magnetic field produced in non-central heavy-ion collisions. The same phenomenon can also be depicted in terms of induction of electric field by the (quasi) static magnetic field, which happens in the occurrence of these topologically nontrivial vacuum solutions. The induced electric field is parallel to the magnetic field and leads to the charge separation in that direction. Thus, the charge separation can be viewed as a nonzero electric dipole moment of the system.
Experimentally, RHIC [@lab15; @lab16; @lab17; @lab18; @lab19] and LHC [@lab20] have published the measurements of CME by the two-particle or three-particle correlations of charged particles with respect to the reaction plane, which are qualitatively consistent with the CME. A clear signal compatible with a charge dependent separation relative to the reaction plane is observed, which shows little or no collision energy dependence when compared to measurements at RHIC energies. This provides a new insight for understanding the nature of the charge-dependent azimuthal correlations observed at RHIC and LHC energies.
Recent years, lots of attentions [@lab21; @lab22; @lab23; @lab24; @lab25] have been paid to the chiral magnetic effect (CME). It is shown that this effect originates from the existence of nontrivial topological configurations of gauge fields and their interplay with the chiral anomaly which results in an asymmetry between left- and right-handed quarks. The created strong magnetic field coupled to a chiral asymmetry can induce an electric charge current along the direction of a magnetic field. The strong magnetic field will separate particles of opposite charges with respect to the reaction plane. Recently, possible CME and topological charge fluctuations have been recognized by QCD lattice calculations in gauge theory [@lab26; @lab27] and in QCD + QED with dynamical $2 + 1$ quark flavors [@lab28]. Thus, such topological and CME effects in QCD might be recognized in relativistic heavy-ion collisions directly in the presence of very intense external electromagnetic fields.
Lots of analytical and numerical calculations indicate existence of extremely powerful electromagnetic fields in relativistic heavy-ion collisions [@lab1; @lab4; @lab6; @lab29]. They are the strongest electromagnetic fields that exist in nature [@lab1; @lab4; @lab6; @lab29]. Ref. [@lab30; @lab31] has discussed the electromagnetic response of the plasma produced by relativistic heavy-ion collisions. It is found that the effects to have an important impact on the field dynamics. An exact analytical and numerical solution for the space and time dependencies of an electromagnetic field produced in heavy-ion collisions was presented in Ref [@lab32]. It was confirmed that nuclear matter plays a crucial role [@lab33] in its time evolution.
In Ref. [@lab34; @lab35], we used the Wood-Saxon nucleon distribution instead of uniform distribution to improve the calculation of the magnetic field of the central point for non-central collision in the RHIC and LHC energy regions. In this paper, we will use the improved magnetic field model to calculate the spatial distribution feature of the chiral magnetic field in the RHIC and LHC energy regions. The dependencies of the spatial features of magnetic fields on the collision energies, centralities, and collision time will be systematically investigated, respectively.
The paper is organized as follows. The key points of the improved model of magnetic field are described in Sec. II. The calculation results of the magnetic field are present in Sec. III. A summary is given in Sec. IV.
The Improved model of chiral Magnetic field
===========================================
The improved model of magnetic field mainly contains three parts:
\(1) As shown in Fig.1, two similar relativistic heavy nuclei with charge $Z$ and radius $R$ are traveling in the positive and negative $z$ direction with rapidity $Y_0$. At $t=0$ they go through a non-central collision with impact parameter $b$ at the origin point. The center of the two nuclei are taken at $x=\pm b/2$ at time $t=0$ so that the direction of $b$ lies along the $x$ axis. The region in which the two nuclei overlap contains the participants, the regions in which they do not overlap contain the spectators.
As the nuclei are nearly traveling with the speed of light in ultra-relativistic heavy-ion collision experiments, the Lorentz contraction factor $\gamma$ is so large that the two included nuclei can be taken as pancake shape(as the $z = 0$ plane). We use the Wood-Saxon nuclear distribution instead of uniform nuclear distribution [@lab1]. The Wood-Saxon nuclear distribution forms is: $$\begin{aligned}
n_A(r)=\frac{n_0}{1+\exp{(\frac{r-R}{d})}},
\label{eq:eq1} %Eq.1\end{aligned}$$
here $d$ = 0.54 fm, $n_0$ = 0.17 fm$^{-3}$ and the radius $R$=1.12 A$^{1/3}$ fm. Considering the Lorentz contraction, the density of the two-dimensional plane can be given by:
$$\begin{aligned}
\rho_{\pm}(\vec{x}^\prime_\bot)=N\cdot\int_{-\infty}^{\infty}dz'\frac{n_0}{1+\exp(\frac{\sqrt{(x'\mp{b/2})^2+y'^{2}+z'^{2}}-{\rm R}}{d})},
\label{eq:eq2} %Eq.2\end{aligned}$$
where $N$ is the normalization constant. The number densities of the colliding nuclei can be normalized as
$$\begin{aligned}
\int{d}\vec{x}^\prime_\bot\rho_{\pm}(\vec{x}^\prime_\bot)=1.
\label{eq:eq3} %Eq.3\end{aligned}$$
\(2) Secondly, in order to study the strength of the magnetic field caused by the two relativistic traveling nuclei, we can split the contribution of particles to the magnetic field in the time $t>0$. The specific forms of expression for the contribution of particles to the magnetic field in the following way
$$\begin{aligned}
\vec{B}=\vec{B}^+_s+\vec{B}^-_s+\vec{B}^+_p+\vec{B}^-_p
\label{eq:eq4} %Eq.4\end{aligned}$$
where $\vec{B}^\pm_s$ and $\vec{B}^\pm_p$ are the the contributions of the spectators and the participants moving in the positive or negative $z$ direction, respectively. For spectators, we assume that they do not scatter at all and that they keep traveling with the beam rapidity $Y_0$. Combining with Eq.(2), we use the density above and give
$$\begin{aligned}
\lefteqn{e\vec{B}^\pm_s(\tau,\eta,\vec{x}_\bot)=\pm Z\alpha_{EM}\sinh(Y_0\mp\eta)
\int{d}^2\vec{x}^\prime_\bot\rho_{\pm}(\vec{x}^\prime_\bot)}\nonumber\\
&&\times[1-\theta_\mp(\vec{x}^\prime_\bot)]\frac{(\vec{x}^\prime_\bot-\vec{x}_\bot)\times\vec{e}_z}
{[(\vec{x}^\prime_\bot-\vec{x}_\bot)^2+\tau^2\sinh(Y_0\mp\eta)^2]^{3/2}},
\label{eq:eq5} %Eq.5\end{aligned}$$
where $\tau=(t^2-z^2)^{1/2}$ is the proper time, $\eta=\frac{1}{2}\ln[(t+z)/(t-z)]$ is the space-time rapidity, and
$$\begin{aligned}
\theta_\mp(\vec{x}^\prime_\bot)=\theta[R^2-(\vec{x}^\prime_\bot\pm\vec{b}/2)^2].
\label{eq:eq6} %Eq.6\end{aligned}$$
In the other hand, the distribution of participants that remain traveling along the beam axis is given by $$\begin{aligned}
f(Y)=\frac{a}{2\sinh(aY_0)}{\rm e}^{aY}, \hskip1cm -Y_{0}\leq{Y}\leq{Y_{0}}.
\label{eq:eq7} %Eq.7\end{aligned}$$
Experimental data gives $a\approx1/2$, which is consistent with the baryon junction stopping mechanism. The contribution of the participants to the magnetic field can be given by
$$\begin{aligned}
e\vec{B}^\pm_p(\tau,\eta,\vec{x}_\bot)=\pm Z\alpha_{EM}\int{\rm d}^2\vec{x}^\prime_\bot
\int{\rm d}Y f(Y)\sinh(Y\mp\eta)\nonumber\\
\times\rho_{\pm}(\vec{x}^\prime_\bot)\theta_\mp(\vec{x}^\prime_\bot)
\frac{(\vec{x}^\prime_\bot-\vec{x}_\bot)\times\vec{e}_z}
{[(\vec{x}_\bot^\prime-\vec{x}_\bot)^2+\tau^2\sinh(Y\mp\eta)^2]^{\frac{3}{2}}}
\label{eq:eq8} %Eq.8\end{aligned}$$
\(3) In the third part, in order to study the spatial distribution of the magnetic field, we will calculate the $eB_{x}$ and $eB_{y}$ components of the chiral magnetic field from spectator and participant nuclei. The specific forms of the contribution of $eB_{x}$ and $eB_{y}$ components from the spectator and participant nuclei are given as follows: $$\begin{aligned}
\lefteqn{eB^\pm_{sy}(\tau,\eta,\vec{x}_\bot)=\mp Z\alpha_{EM}\sinh(Y_0\mp\eta)
\int{d}^2\vec{x}^\prime_\bot\rho_{\pm}(\vec{x}^\prime_\bot)}\nonumber\\
&&\times[1-\theta_\mp(\vec{x}^\prime_\bot)]
\frac{(x^\prime-x)}
{[(\vec{x}^\prime_\bot-\vec{x}_\bot)^2+\tau^2\sinh(Y_0\mp\eta)^2]^{3/2}},
\label{eq:eq9} %Eq.9\end{aligned}$$
where $eB_{sy}$ is the $y$ component of magnetic field from spectators, and the $x$ component of magnetic field from spectators is given by: $$\begin{aligned}
\lefteqn{eB^\pm_{sx}(\tau,\eta,\vec{x}_\bot)=\pm Z\alpha_{EM}\sinh(Y_0\mp\eta)
\int{d}^2\vec{x}^\prime_\bot\rho_{\pm}(\vec{x}^\prime_\bot)}\nonumber\\
&&\times[1-\theta_\mp(\vec{x}^\prime_\bot)]
\frac{(y^\prime-y)}
{[(\vec{x}^\prime_\bot-\vec{x}_\bot)^2+\tau^2\sinh(Y_0\mp\eta)^2]^{3/2}},
\label{eq:eq10} %Eq.10\end{aligned}$$
In the other hand, the $y$ component of magnetic field from participants is given by: $$\begin{aligned}
eB^\pm_{py}(\tau,\eta,\vec{x}_\bot)=\mp Z\alpha_{EM}\int{\rm d}^2\vec{x}^\prime_\bot
\int{\rm d}Y f(Y)\sinh(Y\mp\eta)\nonumber\\
\times\rho_{\pm}(\vec{x}^\prime_\bot)\theta_\mp(\vec{x}^\prime_\bot)
\frac{(x^\prime-x)}
{[(\vec{x}_\bot^\prime-\vec{x}_\bot)^2+\tau^2\sinh(Y\mp\eta)^2]^{\frac{3}{2}}}
\label{eq:eq11} %Eq.11\end{aligned}$$
and the $x$ component of magnetic field from participants is given by: $$\begin{aligned}
eB^\pm_{px}(\tau,\eta,\vec{x}_\bot)=\pm Z\alpha_{EM}\int{\rm d}^2\vec{x}^\prime_\bot
\int{\rm d}Y f(Y)\sinh(Y\mp\eta)\nonumber\\
\times\rho_{\pm}(\vec{x}^\prime_\bot)\theta_\mp(\vec{x}^\prime_\bot)
\frac{(y^\prime-y)}
{[(\vec{x}_\bot^\prime-\vec{x}_\bot)^2+\tau^2\sinh(Y\mp\eta)^2]^{\frac{3}{2}}}
\label{eq:eq12} %Eq.12\end{aligned}$$
The calculation results
=======================
In order to study the dependencies of magnetic field $eB$ on proper time, we show the dependencies of magnetic field $eB$ (at central point $(x, y) = (0, 0)$) on proper time $\tau$ at $\sqrt{s_{NN}}$ = 200 GeV for Au - Au collisions with b=8fm and $\sqrt{s_{NN}}$ = 2760 GeV for Pb - Pb collisions with b = 8 fm, respectively. From Fig.2(a, b), one can find that at small proper time the magnetic field is mainly from the contribution of spectator nucleons, but as the proper time increases, more and more large contribution of the magnetic field is from participant nucleon. Figure 2(c,d) show the comparisons of the magnetic field and the ratio of $(eB)_{p}/(eB)$ at $\sqrt{s_{NN}}$ = 200 GeV and $\sqrt{s_{NN}}$ = 2760 GeV. One can find that at smaller proper time $\tau$($\tau<8\times 10^{-3}$ fm) the magnetic field at $\sqrt{s_{NN}}$ = 2760 GeV is greater than that of $\sqrt{s_{NN}}$ = 200 GeV, but when $\tau> 8 \times 10^{-3}$ fm, the magnetic field at $\sqrt{s_{NN}}$ = 2760 GeV is less than that of $\sqrt{s_{NN}}$ = 200 GeV. From Fig.2(d) one can find that the contribution of magnetic field from participant nucleons increases with the increase of proper time.
Figure 3 shows the dependencies of magnetic field $eB$ (at central point $(x, y)=(0, 0)$ )on central of mass energy $\sqrt{s_{NN}}$ at different proper time $\tau$. It is argued that at smaller proper time ($\tau$ = 0.001 and 0.0001fm) the magnetic fields increase with the increase of the CMS energy ($\sqrt{s_{NN}}$), but with the increase of proper time ($\tau$), the magnetic field decreases sharply with increasing collision energy of $\sqrt{s_{NN}}$. It is found that when $\tau$ = 3 fm and $\sqrt{s_{NN}} > 200$ GeV, the magnetic field approaches zero.
For consistency with the experimental results, we take Au-Au collision with RHIC energy region and Pb-Pb collision with LHC energy region. When studying the spatial distribution characteristics of magnetic field, we choose the spatial regions of -10.0 fm $\leq x\leq$ 10.0 fm and -10.0 fm $\leq y\leq$ 10.0 fm.
Figure 4 shows the magnetic field spatial distributions of $eB_{y}$ with different collision energies $\sqrt{s_{NN}}$ = 62.4 GeV, 130 GeV and 200 GeV and proper time $\tau = 0.0001$ fm. The collision energies shown in Fig. 4 are in RHIC energy region. The spatial distributions of $eB_{y}$ show obviously axis symmetry characteristics along $x = 0$ and $y = 0$ axes. There is a peak around central point $(x, y) = (0, 0)$, and the magnetic field get smaller and smaller when the location go farther away from the center position. When $\tau = 0.0001$ fm,
On both sides of $y = 0$ line, there are two symmetrical peaks. These two peaks are almost connected when $\sqrt{s_{NN}}$ = 62.4 GeV. As the collision energy increases, the two peaks start to separate and expose the valley between the two peaks when $\sqrt{s_{NN}}$ = 130 GeV and 200 GeV. The maximum of magnetic field $eB_{y}$ in RHIC energy region reaches $2.2\times 10^{5}$ MeV$^{2}$.
Compared with Fig.4, Fig.5 shows the magnetic field spatial distributions of $eB_{y}$ in the LHC energy region. When the collision energy rises up to 900 GeV in LHC energy region, the distribution features of magnetic field have some differences from that of the RHIC energy region. For example the magnetic field distribution peak around $x = 0$ and $y = 0$ becomes flat at $\sqrt{s_{NN}}$ = 900 GeV, and begin to appear the phenomenon of two peaks. The maximum of magnetic field $eB_{y}$ in LHC energy region reaches $2.0\times 10^{6}$ MeV$^{2}$, which is larger than that of RHIC energy region at $\tau$ = 0.0001 fm.
From Fig.2 to Fig.5, we argue that the magnetic field spatial distributions of $eB_{y}$ are highly inhomogeneous. The distribution features in the RHIC energy region is different from that of the LHC energy region. It is argued that at smaller proper time ($\tau$ = 0.001 and 0.0001fm) the magnetic fields increase with the increase of the CMS energy ($\sqrt{s_{NN}}$), but with the increase of proper time ($\tau$), the magnetic field decreases sharply with increasing collision energy of central of mass $\sqrt{s}$.
The above we make a discussion of magnetic field spatial distributions with the collision energy and impact parameter relations, we will make a study of magnetic field with the proper time. The magnitude of magnetic field is presented as:
$$\begin{aligned}
eB=\sqrt{(eB_{x})^{2}+ (eB_{y}^{2})}
\label{eq:eq13} %Eq.13\end{aligned}$$
Sometimes, one often takes the $y$ component $eB_{y}$ to approximately replace $eB$. This is the reason that $eB_{y}$ is usually larger than $eB_{x}$. In order to verify the rationality of the substitution, we need a detailed study the relation between $eB_{y}$ and $eB$. Figure 6 shows the dependencies of the ratio of $eB_{y}/(eB)$ on $x$ and $y$ at $\sqrt{s_{NN}}$= 200 GeV and at different proper time $\tau$ = 0.02, 0.2 and 2.0 fm, respectively. The Fig.6(a, c and e) are for $eB_{y}/(eB)$ with $y$ at different proper time. From Fig.6(a, c and e), one can figure out that the ratio of $eB_{y}/(eB)$ with $y$ change is between 0.9 to 1.0. In this case, one can approximate the $eB_{y}$ instead of $eB$. Compared with the relation of ratio $eB_{y}/(eB)$ with $y$, the relationship of ratio $eB_{y}/(eB)$ with $x$ shown as Fig.6(b, d and f) is obviously different. The main different is the dip located at $x = 0$. The minimum value of the ratio at $x = 0$ can be decreased to 0.5.
In order to study the spatial distribution of magnetic field on proper time, we show the dependencies of magnetic field $eB_{y}$ and $eB_{x}$ (at points $(x,y) = (5,5)$ and $(x, y) = (10, 10)$ ) on proper time $\tau$ at $\sqrt{s}$= 200 GeV for Au-Au collisions with b=8fm and $\sqrt{s}$ = 2760 GeV and 7000 GeV for Pb - Pb collisions with b = 8fm, respectively. From Fig.7, one can find that at small proper time the the magnetic field increases with the increase of the collision energy, but the magnetic field of $\sqrt{s}$ = 7000 GeV decrease more quickly than that of $\sqrt{s}$ = 200 GeV with the increase of proper time. Fig.7(c,d) show that there is a relatively flat region with proper time at point $(x, y) = (10,10)$ than that at point $(x,y) = (5,5)$.
Summary and Conclusion
======================
It is shown that an enormous magnetic field can indeed be created in off-central heavy-ion collisions. The magnetic field distributions of $eB_{x}$ and $eB_{y}$ are highly inhomogeneous, and $eB_{x}$ and $eB_{y}$ distributions are completely different. The enormous magnetic field is produced just after the collision, and the magnitude of magnetic field of LHC energy region is larger that of RHIC energy region at the small proper time($\tau < 8.0 \times 10^{-3}$ fm). We are really surprised to find that the magnetic field in the LHC energy region decreases more quickly with the increase of the proper time than that of RHIC energy region. As the proper time $\tau$ increases to a certain value $8.0 \times 10^{-3}$ fm, the magnitude of magnetic field in the RHIC energy region begin to be larger than that of LHC energy region.
The dependencies of the ratio of $eB_{y}/(eB)$ on $x$ and $y$ at different collision energies at RHIC and LHC and at different proper time are analyzed in this paper. In most cases, the ratio $eB_{y}/(eB)$ approaches $1$, so this is a good approximate by using $eB_{y}$ to approximately replace $eB$. But one should note that the ratio $eB_{y}/(eB)$ is between $0.5 \sim 1.0$ along $x = 0$ line.
We systematically study the spatial distribution features of chiral magnetic field in relativistic heavy-ion collisions at energies reached at LHC and RHIC with the improved model of chiral magnetic field in this paper. The feature of chiral magnetic fields at $\sqrt{s_{NN}}$= 900, 2760 and 7000 GeV in the LHC energy region and $\sqrt{s_{NN}}$ = 62.4, 130 and 200 GeV in the RHIC energy region are systematically studied.
The dependencies of the magnetic field on proper time for at RHIC and LHC energy regions, respectively. Comparing with that of RHIC energy region, one finds that the magnitudes of the magnetic fields with proper time fall more rapidly at LHC energy region. The variation characteristics of magnetic field with impact parameter at RHIC energy region are different from that of LHC energy region. The maximum position is located in the small proper time ($\tau \sim 0.0001$ fm), more off-central collisions and $\sqrt{s_{NN}}\sim 7000$ GeV. The maximum of magnetic field in our calculation is about $eB \simeq 2 \times 10^{7} MeV^{2}$ when $\tau = 0.0001$ $b\simeq 8 fm$ and $\sqrt{s_{NN}}\sim 7000$ GeV.
Acknowledgments
===============
This work was supported by the National Natural Science Foundation of China (Grants Nos. 11375069, 11435054, 11075061, and 11221504), also by the Open innovation fund of the Ministry of Education of China under Grant No. QLPL2014P01.
D.E. Kharzeev, L.D. McLerran, H.J. Warringa, Nucl. Phys. A 803, 227 (2008). arXiv:0711.0950 \[hep-ph\]. V. Skokov, A.Y. Illarionov, V. Toneev, Internat. J. Mod. Phys. A 24, 5925 (2009). arXiv:0907.1396 \[nucl-th\]. K. Fukushima, D. E. Kharzeev, and H. J. Warringa, Phys. Rev. Lett. 104, 212001 (2010). W. T. Deng and X. G. Huang, Phys. Rev. C 85, 044907 (2012). K. Fukushima, D. E. Kharzeev, and H. J. Warringa, Nucl. Phys. A 836, 311 (2010). A. Bzdak and V. Skokov, Phys. Lett. B 710, 171 (2012) D. E. Kharzeev, Ann. Phys. (N.Y.) 325, 205 (2010). D. E. Kharzeev, D. T. Son, Phys. Rev. C 106, 062301 (2011). D. E. Kharzeev and A. Zhitnitsky, Nucl. Phys. A 797, 67 (2007). K. Fukushima, D. E. Kharzeev, and H. J. Warringa, Phys. Rev. D 78, 074033 (2008). Y. Burnier, D. E. Kharzeev, J. Liao and H.U. Yee, Phys. Rev. Lett. 107, 052303 (2011). D. Kharzeev, Phys. Lett. B 633, 260 (2006). S. Voloshin and Y. Zhang, Z. Phys. C 70, 665 (1996). S. A. Voloshin, Phys. Rev. C 70, 057901 (2004). B. I. Abelev, M. M. Aggarwal, Z. Ahammed et al., Phys. Rev. Lett. 103, 251601 (2009). B. I. Abelev, M. M. Aggarwal, Z. Ahammed et al., Phys. Rev. C 81, 054908 (2010). I. V. Selyuzhenkov and STAR Collaboration, Romanian Reports in Physics, 58, 49 (2006). S. A. Voloshin and STAR Collaboration, Nucl. Phys. A 830, 377 (2009). N. N. Ajitanand, R. A. Lacey, A. Taranenko, and J. M.Alexander et al., Phys. Rev. C 83, 011901 (2011). B. Abelev, J. Adam, D. Adamova et al., Phys. Rev. Lett. 110, 012301 (2009). D. Kharzeev, R. D. Pisarski, and M. N. G. Tytgat, Phys. Rev. Lett. 81, 512 (1998). D. Kharzeev, R. D. Pisarski, and M. N. G. Tytgat, Phys. Rev. D 61, 111901 (2000). D. Kharzeev, Phys. Lett. B 633, 260 (2006). K. Fukushima, M. Ruggieri, and R. Gatto, Phys. Rev. D 81, 114031 (2010); D. E. Kharzeev and H. J. Warringa, Phys. Rev. D 80, 034028 (2009). P. V. Buividovich, M. N. Chernodub, E. V. Luschevskaya, and M. I. Polikarpov, Phys. Rev. D 80, 054503 (2009) P. V. Buividovich, M. N. Chernodub, E. V. Luschevskaya, and M. I. Polikarpov, Phys. Rev. D 81, 036007 (2010). M. Abramczyk, T. Blum, G. Petropoulos, and R. Zhou, PoS (LAT2009) 181 (2009). K. Tuchin, Advances in High Energy Physics, 2013, 490495 (2014). B. G. Zakharov, Phys. Lett. B 737, 262 (2014). U. G$\ddot{u}$rsoy, D.E. Kharzeev, and K. Rajagopal, Phys. Rev. C 89, 054905 (2014) K. Tuchin, Phys. Rev. C 88, 024911 (2013). K. Tuchin, Phys. Rev. C 82, 034904 (2010); 83, 039903(2011). Y. J. Mo, S. Q. Feng and Y. F. Shi, Phys. Rev. C 88, 024901 (2013). Y. Zhong, C. B. Yang, X. Cai and S. Q. Feng, Advances in High Energy Physics, 2014, 193039 (2014).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate two classes of transformations of cosine similarity and Pearson and Spearman correlations into metric distances, utilising the simple tool of metric-preserving functions. The first class puts anti-correlated objects maximally far apart. Previously known transforms fall within this class. The second class collates correlated and anti-correlated objects. An example of such a transformation that yields a metric distance is the sine function when applied to centered data.'
address: 'EMBL-EBI, Hinxton, Cambridge, UK'
author:
- 'Stijn van Dongen and Anton J. Enright'
bibliography:
- 'the.bib'
title: Metric distances derived from cosine similarity and Pearson and Spearman correlations
---
Results
=======
We derive metric distances from the sample Pearson coefficient, sample Spearman coefficient, and cosine similarity. Using $A$ to denote any of these, it is already known that $\theta = \arccos(A(x,y))$ yields a metric distance, known as the angular distance. We further obtain the correlation distance $\sin({{\scriptstyle\frac{1}{2}}}\theta)$, or equivalently $\sqrt{{{\scriptstyle\frac{1}{2}}}(1-A(x,y))}$. Both distances place anti-correlated objects maximally far apart. A second class of metric distances is obtained that collate correlated and anti-correlated objects. Examples are the acute angular distance ${{\scriptstyle\frac{1}{2}}}\pi - |{{\scriptstyle\frac{1}{2}}}\pi - \theta|$ and the absolute correlation distance $\sin(\theta)$, or equivalently $\sqrt{1-A(x,y)^2}$.
Background
==========
The Pearson correlation coefficient, Spearman correlation coefficient and the cosine similarity are staples of data analysis. The Pearson and Spearman coefficients measure strength of association between two variables $X$ and $Y$. The Pearson coefficient, commonly denoted by $\rho$, is defined as the covariance of the two variables divided by the product of their respective standard deviations. $$\label{eq::pearson}
\rho_{X,Y} = \frac{{\mathrm cov}(X,Y)}{\sigma_X \sigma_Y}$$ The Spearman coefficient is obtained by applying the Pearson coefficient to rank-transformed data. Both are unaffected by linear transformations of the data. Given vectors $x$ and $y$, respectively sampling $X$ and $Y$ and each of length $n$, the sample Pearson coefficient $r_{x,y}$ is obtained by estimating the population covariance and standard deviations from the samples, as defined in Equation (\[eq::sample:pearson\]). Here $\overline{x}$ and $\overline{y}$ denote the sample means. $$\label{eq::sample:pearson}
r_{x,y} = \frac{\sum (x_i - \overline{x}) (y_i - \overline{y})}{\sqrt{\sum (x_i - \overline{x})^2} \sqrt{\sum (y_i - \overline{y})^2}}$$ The cosine similarity is a standard measure used in information retrieval. It is the cosine of the angle between two Euclidean vectors, and thus unaffected by scalar transformations in the data. It is defined below in Equation (\[eq::cosine\]) for vectors $x$ and $y$. $$\label{eq::cosine}
\frac{\sum x_i y_i}{\sqrt{\sum {x_i}^2}\sqrt{\sum {y_i}^2}}$$ These measures are related; Pearson is identical to the cosine applied to centered data (centered cosine), as evident from equations (\[eq::sample:pearson\]) and (\[eq::cosine\]). For the purpose of this paper the terminology of *vectors* and *samples* is used interchangeably. We are not concerned with statistical properties of the Pearson coefficient under certain models, but solely interested in its properties as a function mapping Euclidean spaces to the interval $[-1,1]$. We will henceforth refer to Pearson, Spearman, and cosine similarity as $P$, $S$, and $C$, and use $A$ to indicate all of them are applicable.
Where dissimilarities are used, it is desirable that they satisfy the triangular inequality and are thus a metric distance. Informally, this means that detours take longer: the distance from $a$ to $c$ should always be shorter than the distance from $a$ to $b$ plus the distance from $b$ to $c$. Metric distances abound in data analysis, formalizing a property that is intuitively expected and that allows stringent reasoning about data points. Several methods require this, such as for building $M$-trees [@Ciaccia97m-tree:an] and accelerated algorithms that use the triangle inequality to skip computations by tracking bounds [@Brin:1995:NNS:645921.673006; @hamerly:making].
Metric distances
================
A metric distance takes as input two objects and outputs a real number. It requires four properties. These are i) all distances are nonnegative, ii) the distance of an object to itself is zero and distinct objects are never at distance zero, iii) the distance between two objects is the same in both directions, and iv) the distance satisfies the property that detours are longer, more commonly stated as the triangle inequality. More formally, given a distance $d$, it states that $d(x,y) \leq d(x,z) + d(z,y)$ for all objects $x$, $y$, and $z$. In this formulation, the distance between $x$ and $y$ is compared to the distance when using $z$ as a detour.
In the analysis of distances derived from correlations and cosine similarity we will use a class of functions called *metric preserving*. A function $f$ is metric preserving if the distance $d_f(x,y) = f(d(x,y))$ is again metric for any metric $d$. More specifically, we shall make use of an important subclass of metric-preserving functions, namely those that are *concave* and increasing. A function $f$ is called concave on an interval $I$ if for all $x$ and $y$ in $I$ and for $t$ in $[0,1]$ the inequality $$\label{eq::chord}
f(tx+(1-t)y) \geq t f(x) + (1-t)f(y)$$ holds. We refer to this as the chord condition. It is the formal way of stating that the chord drawn from $[x,f(x)]$ to $[y,f(y)]$ does not exceed $f$ in the interval $[x,y]$. It essentially means that $f$ is curving inward on $I$, as shown in the figure below.\
The following lemma, relating *concave* functions to metric preserving functions is well-known (see e.g. [@mpf::corrazza]). We include a short detailed proof as it is an important prerequisite to this paper, consisting of several steps gathered here for ease of reference. It shows subadditivity to be the key property making certain concave functions also metric-preserving.
**Lemma 1** For $f$ to be metric preserving it is sufficient if $f(0)=0$, and $f(x)$ is both *increasing* and *concave* for $x>0$.
**Proof** We first prove that functions that are concave for $x>0$ and satisfy $f(0) \geq 0$ are also *subadditive* for $x\geq 0$ (that is, $f(a+b)\leq f(a)+f(b)$ for $a,b \geq 0$). This follows by setting $y=0$ in the chord condition (\[eq::chord\]) and using the postulate $f(0) \geq 0$. We obtain the scalar inequality $t f(x) \leq f(t x)$, for $0
\leq t \leq 1$. We then rewrite $f(a+b)$ as ${{\scriptstyle\frac{a}{a+b}}} f(a+b) + {{\scriptstyle\frac{b}{a+b}}} f(a+b)$, noting that ${{\scriptstyle\frac{a}{a+tb}}}$ and ${{\scriptstyle\frac{b}{a+tb}}}$ both lie in $[0,1]$. Using the scalar inequality just derived we bound the rewritten expression from above by $f({{\scriptstyle\frac{a}{a+b}}}(a+b)) + f({{\scriptstyle\frac{b}{a+b}}}(a+b))$, equaling $f(a)+f(b)$.
The proof of the lemma can now be concluded. We need to prove that $d_f$ is a metric distance, i.e. $d_f(x,y) \leq d_f(x,z) + d_f(z,y)$ for all $x,y,z$. First, we use that $f$ is increasing and $d(x,y) \leq d(x,z) + d(z,y)$ (because $d$ is a metric distance) to obtain $$f(d(x,y)) \leq f(d(x,z) + d(z,y))$$ Finally, given that $f$ is concave and $f(0) = 0$ we know that $f$ is also subadditive and thus $$f(d(x,z) + d(z,y)) \leq f(d(x,z)) + f(d(z,y))$$\
The following lemma yields a quick way to determine whether a function is concave.
**Lemma 2** A function $f$ that is twice differentiable on an interval $I$ is concave on $I$ if $f''(x) \leq 0$ for $x \in I$.
The lemma can heuristically be understood as $f''(x) \leq 0$ implies that the rate of acceleration of $f$ is slowing. Hence $f$ curves inward, implying it is concave. The lemma is part of standard calculus, and for a formal proof we refer to [@hlp]. If $f$ is twice-differentiable, increasing, and satisfies $f''(x)\leq 0$ for $x>0$ with $f(0)=0$ it is thus metric-preserving, and we will use this later.
From correlations to distances
==============================
The first three properties of a metric distance are easily obtained when transforming one of the $A$ measures to a dissimilarity by a natural transformation such as $d(x,y) = 1 - A(x,y)$. However, the dissimilarity thus obtained does not guarantee the triangle inequality. We show below why this is the case using generic principles rather than explicit calculations, and why transformations such as $d(x,y): x,y \rightarrow \sqrt{1-A(x,y)^2}$ and $d(x,y): x,y \rightarrow \sqrt{{{\scriptstyle\frac{1}{2}}}(1 - A(x,y))}$ do result in a metric distance.
Currently two metric distances are known to derive from the triple $(P, S, C)$, namely the angle $\theta$ between vectors, and derived from it, $\sqrt{2-2 \cos(\theta)}$, which may be obtained as $\sqrt{2-2\;A(x,y)}$. For the angle $\theta$ the triangle inequality derives from Proposition `XI.20` of Euclid’s The Elements and the fact that three vectors in a high-dimensional space can be embedded in three-dimensional space. It follows that $\arccos(A(x,y))$ yields a metric distance, where $A$ may be any of $P$, $S$, or $C$. It is known (e.g. [@sun::angular]) that $\sqrt{2-2 \cos(\theta)}$ is equal to the Euclidean distance between the two unit-scaled object vectors $x$ and $y$. This follows from (using $\left\|x\right\| = 1$ and $\left\|y\right\| = 1$)
$$\begin{aligned}
\left\|x - y\right\|^2 &= \sum (x_i - y_i)^2\\
\; &= \left\|x\right\|^2 + \left\|y\right\|^2 - \left\|x\right\| \left\|y\right\| x \cdot y\\
&= 2 - \cos(\theta)\end{aligned}$$
It can additionally be observed using a trigonometric identity for $\sin({{\scriptstyle\frac{1}{2}}}\theta)$ ([@abrste], page 72) that $\sqrt{2-2 \cos(\theta)}$ is equal to $\sqrt{2}\sin({{\scriptstyle\frac{1}{2}}}\theta)$ in the interval $[0,\pi]$ and is seen to be concave on that interval by considering its second derivative. Hence $\sqrt{2}\;\sin({{\scriptstyle\frac{1}{2}}}\theta)$ is a metric-preserving function for the angular distance (but not metric-preserving in general).
We formalise this finding and derive another class of metric distances derived from $P$, $S$, and $C$ whose members collate correlated and anti-correlated objects. The canonical representative of this class is the sine function $\sin$. In the lemma below we do not use generic metric-preserving functions, as stronger results can be obtained by utilising traits of the angular distance. However, the functions used share on certain intervals of interest the general traits of an important class of metric-preserving functions, namely being concave and increasing.
**Lemma 3** i) A function $f$ of the angular distance that satisfies $f(0) = 0$, is defined on $[0, \pi]$, and is either a) increasing and concave on the interval $[0, \pi]$, or b) increasing and concave on the interval $[0, {{\scriptstyle\frac{1}{2}}}\pi]$ and satisfies $f(x) = f(\pi-x)$ ($f$ is symmetric around ${{\scriptstyle\frac{1}{2}}}\pi$), is a metric preserving distance for the angular distance. In case b) this requires disregarding the directionality of vectors and collating a vector and its sign-reversed counterpart into a single object.
Examples of such functions in case a) are\
$f_1: x \rightarrow x$\
$f_2: x \rightarrow \sin({{\scriptstyle\frac{1}{2}}}x)$
Examples of such functions in case b) are\
$f_3: x \rightarrow {{\scriptstyle\frac{1}{2}}}\pi - |{{\scriptstyle\frac{1}{2}}}\pi - x|$\
$f_4: x \rightarrow \sin(x)$
These lead to distances that can be computed, again using $A$ to denote any of $(P,S,C)$, as\
$\;\;\;d_1: x,y \rightarrow f_1(A(x,y)) = \arccos(A(x,y))$\
$\;\;\;d_2: x,y \rightarrow f_2(A(x,y)) = \sqrt{{{\scriptstyle\frac{1}{2}}}(1-A(x,y))}$\
(angular distance and correlation distance, respectively), and\
$\;\;\;d_3: x,y \rightarrow f_3(A(x,y)) = {{\scriptstyle\frac{1}{2}}}\pi - |{{\scriptstyle\frac{1}{2}}}\pi - \arccos(A(x,y))|$\
$\;\;\;d_4: x,y \rightarrow f_4(A(x,y)) = \sqrt{1-A(x,y)^2}$\
(acute angular distance and absolute correlation distance, respectively).
ii\) A function $g$ of the angular distance that satisfies $g(0) = 0$ and is increasing and strictly convex on some interval $[0, \epsilon]$, where $\epsilon$ is positive, yields a dissimilarity that violates the triangular inequality. An example of such a function is $g: x \rightarrow 1-\cos(x)$, or equivalently, $1-A(x,y)$.
**Proof** i) Name the three vectors $a$, $b$, and $c$ with angles $\alpha$, $\beta$, and $\gamma$ between the pairs $(b,c)$, $(a,c)$, and $(a,b)$ respectively. In scenario a) we set out to prove that $f(\gamma) \leq f(\alpha) + f(\beta)$ and may use the inequality $\gamma \leq \alpha + \beta$ because the angular distance is a metric. In scenario a), if $\alpha+\beta \leq \pi$ we use subadditivity to deduce $f(\gamma) \leq f(\alpha+\beta) \leq f(\alpha) + f(\beta)$. In the other case it is easy to see that $f(\alpha) + f(\beta) \geq f(\pi)$, either by considering the concave function obtained by extending $f: x \rightarrow f(\pi)$ for $x > \pi$, or by explicit calculation. As $f(\pi)$ is the maximal value of $f$ in $[0,\pi]$ it follows that $f(\gamma) \leq f(\pi) \leq f(\alpha) + f(\beta)$.
In scenario b) we may assume that $\alpha$ and $\beta$ are both smaller than ${{\scriptstyle\frac{1}{2}}}\pi$ because of the following. By sign-reversing $a$ we obtain vectors $-a, b, c$ and angles $\alpha, \pi-\beta, \pi-\gamma$. This transform leaves the values of $f$ on the transformed angles invariant, and the triangular inequality can now be applied to $\alpha', \beta', \gamma'$ = $\alpha, \pi-\beta, \pi-\gamma$. Thus we may sign-reverse any of the three input vectors while preserving the inequality to be proven. By choosing which of $a$, $b$, or $c$ to flip we can always make sure that both $\alpha'$ and $\beta'$ are smaller than ${{\scriptstyle\frac{1}{2}}}\pi$. The inequality $f(\gamma) \leq f(\alpha) + f(\beta)$ is the same as $f(\gamma') \leq f(\alpha') + f(\beta')$, where $\alpha'$, $\beta'$, $\gamma'$ are the angles corresponding with a triple of vectors $(a', b', c')$, allowing the use of the triangle inequality $\gamma' \leq \alpha' + \beta'$. If one of $\gamma'$ or $\pi-\gamma'$ is smaller than either of $\alpha'$ or $\beta'$ there is nothing to prove because $f$ is increasing on $[0, {{\scriptstyle\frac{1}{2}}}\pi]$ and symmetric around ${{\scriptstyle\frac{1}{2}}}\pi$. If $\gamma'$ is bigger than ${{\scriptstyle\frac{1}{2}}}\pi$, we observe that $\alpha' + \beta' \geq \gamma' \geq \pi-\gamma'$ and we can choose to work with $\gamma'' = \pi-\gamma'$ rather than $\gamma'$. If $\gamma'$ is smaller than ${{\scriptstyle\frac{1}{2}}}\pi$, we simply set $\gamma'' = \gamma'$. This leaves us to prove $f(\gamma'') \leq f(\alpha') + f(\beta')$ where $\gamma''$, $\alpha'$, and $\beta'$ are all smaller than ${{\scriptstyle\frac{1}{2}}}\pi$, where $\gamma''$ is larger than both $\alpha'$ and $\beta'$, and where $\alpha'+\beta' \geq \gamma''$. The same reasoning as under a) now applies, restricted to the interval $[0,{{\scriptstyle\frac{1}{2}}}\pi]$.
ii\) Pick vectors $a$, $b$ and $c$ lying in the cartesian plane, such that the angles satisfy $\gamma = \alpha + \beta$, $\gamma < \epsilon$. Then $g(\gamma) = g(\alpha+\beta) > g(\alpha) + g(\beta)$ (by super-additivity of strictly convex functions with $f(0) \leq 0$).
Notes
=====
For a distance $d$ and a metric-preserving function $f$ the distance $d_f$ is ordinally equivalent with $d$, that is, rankings of distances are preserved. The correlation distance $d_2$ is ordinally equivalent to the angular distance $d_1$ and the acute angular distance $d_3$ is equivalent to the absolute correlation distance $d_4$.
Further distances can be obtained by composition of concave functions; for example $f_5: x \rightarrow \sin(x)^p$, where $0<p\leq1$, also yields a distance. Such distances are again ordinally equivalent to the absolute correlation distance and preserve rankings of distances.
Acknowledgments
===============
The authors are grateful to Leopold Parts and Roberto Álvarez for critical reading and insightful comments.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The abundance of iron is measured from emission line complexes at 6.65 keV (Fe line) and 8 keV (Fe/Ni line) in [*RHESSI*]{} X-ray spectra during solar flares. Spectra during long-duration flares with steady declines were selected, with an isothermal assumption and improved data analysis methods over previous work. Two spectral fitting models give comparable results, viz. an iron abundance that is lower than previous coronal values but higher than photospheric values. In the preferred method, the estimated Fe abundance is $A({\rm Fe}) = 7.91 \pm 0.10$ (on a logarithmic scale, with $A({\rm H}) = 12$), or $2.6 \pm 0.6$ times the photospheric Fe abundance. Our estimate is based on a detailed analysis of 1,898 spectra taken during 20 flares. No variation from flare to flare is indicated. This argues for a fractionation mechanism similar to quiet-Sun plasma. The new value of $A({\rm Fe})$ has important implications for radiation loss curves, which are estimated.'
author:
- 'K. J. H. Phillips'
- 'B. R. Dennis'
title: The Solar Flare Iron Abundance
---
INTRODUCTION
============
The solar abundance of iron remains an important parameter and topic in solar physics. Iron is the most abundant of all elements with $Z > 14$, and is a large contributor to the radiation loss at coronal temperatures. Recent determinations of the photospheric abundance give $A({\rm Fe}) = 7.50\pm 0.04$ [@asp09] and $7.52 \pm 0.06$ [@caf11] (on a logarithmic scale where $A({\rm H}) = 12$), in near-agreement with the meteoritic abundance, $A({\rm Fe}) = 7.45\pm 0.01$ [@lod09]. The iron abundance in the corona has been determined from X-ray or extreme ultraviolet Fe emission lines, formed by collisional excitation. As the excitation rates are a function of electron temperature $T_e$, the thermal structure of the emitting coronal plasma must be modeled for correct interpretation of line fluxes, and ionization fractions and excitation rate coefficients must be known. Examples of the coronal Fe abundance include [@par77] ($A({\rm Fe}) = 7.65$), [@flu99] (7.65), [@whi00] (8.19), and [@den08] (7.86), i.e. enhancement factors over the photospheric value of between 1.4 and 4.9. This large range of abundance determinations may indicate time variations in the coronal abundance [@syl84] or measurement uncertainties of $\sim 0.5$ in $A({\rm Fe})$. The Fe abundance from solar energetic particles (SEPs) in the interplanetary medium varies by large factors; a baseline list of abundances for gradual events [@rea95] gives $A({\rm Fe}) = 7.93$, or 2.65 times photospheric. Evidence for systematic differences between photospheric and coronal abundances has been cited by [@mey85] and [@fel92]. According to [@fel00] elements with first ionization potential FIP $\lesssim 10$ eV like Fe have coronal abundances enhanced by factors of 4, apart from low-altitude flares and energetic spray-like events for which the coronal and photospheric abundances are equal. Various models have been put forward to explain the fractionation and its dependence on FIP; they generally involve a mechanism that separates ions and neutral atoms in the chromosphere where low-FIP elements are partly ionized but high-FIP elements are neutral. The mechanisms include magnetic fields carrying ions rising into the corona as active regions develop [@hen98], and the presence of a ponderomotive force in the acceleration of Alfvén waves [@lam09].
Flare spectra in the photon energy range $\sim 3$ keV to 17 MeV have been obtained from the [*Reuven Ramaty High Energy Solar Spectroscopic Imager*]{} ([*RHESSI*]{}) since its launch on 2002 February 5, allowing analysis of thermal spectra (range $\sim 6$–30 keV) with $\sim 1$ keV spectral resolution (FWHM). This range includes thermal continuum emission (free–free and free-bound) and two line complexes [@phi04]. The Fe line complex at $\sim 6.65$ keV is made up of lines and dielectronic satellites of and lower stages, with minor contributions from the Lyman-$\alpha$ lines [@fel80; @tan82; @lem84], emitted over a broad temperature range ($\gtrsim 10$–100 MK). The weaker “Fe/Ni line" line complex at $\sim 8$ keV consists of higher-excitation lines ($1s^2 - 1snp$, $n \geqslant 3$) of and satellites, with a few-percent contribution from – lines. Line-to-continuum ratios can be derived from [*RHESSI*]{} spectra, and from these the iron abundance estimated using models for the thermal structure (isothermal or simple functions for the temperature distribution). This was done for [*RHESSI*]{} spectra taken during various phases of 27 flares between 2002 and 2005 [@phi06] using fluxes of the Fe line complex at 6.65 keV expressed as the equivalent width of nearby continuum. The observed equivalent widths were found to follow a dependence on $T_e$, derived from the energy dependence of the continuum emission, approximately equal to the theoretical dependence with an Fe abundance somewhat less than [@fel00]’s coronal value, $A({\rm Fe}) = 8.10$. In this work, this analysis is carried further. First, we have chosen only spectra during the gradual phases of flares with the [*RHESSI*]{} thin attenuators in place, and secondly we have used advances in the analysis software enabling [*RHESSI*]{} spectra to be better interpreted, including the use of the [chianti]{} code (v. 6) [@der97; @you03; @der09] with latest atomic data instead of the earlier [mekal]{} code [@mew85]. Also, instrumental effects can be adjusted in the fitting process. By these means, we derive an estimate of the Fe abundance with much reduced uncertainties compared with previous work.
SPECTRAL ANALYSIS AND OBSERVATIONS
==================================
The [*RHESSI*]{} Spectrometer
-----------------------------
The [*RHESSI*]{} spectrometer has been described by [@smi02] so only an outline is given here. Nine cryogenically cooled hyper-pure germanium detectors make up the spectrometer, each with a 1-cm-thick front segment which senses low-energy X-rays. Modulation collimators in front of each detector result in time-modulated counting rates as the spacecraft rotates, which are unscrambled with imaging software to form images. The X-ray emission is viewed through and partially absorbed by beryllium windows and aluminized Mylar insulation layers. To avoid detector saturation and to reduce pulse pile-up, sets of aluminum attenuators are moved in front of the detectors at increasing X-ray flux levels. Since instrumental effects like pile-up, slight changes in energy calibration at high count rates, and variable background rates are different for each detector, we chose to analyze spectra from the detector with best energy resolution rather than summing counts from multiple detectors to improve statistics. Following our earlier work [@phi06], we selected flare spectra during times of slowly decaying emission in the A1 state (thin attenuators in place) when the flare plasma is most nearly isothermal. Spectral fits were done over a range from $\sim 5.7$ keV (the count rates at lower energies are dominated by K-escape events in the A1 and A3 attenuator states) up to 20–40 keV, depending on the emission at higher energies relative to the background spectrum. Thick-target X-ray continuum from a power-law electron spectrum was included when necessary to achieve an acceptable fit to the data at the higher energies. The energy bins for the spectral fits were those $\frac{1}{3}$ keV-wide bins used by the on-board pulse-height analyzer.
Data Selection
--------------
Twenty flares having a steady decline of X-ray emission were observed by [*RHESSI*]{} between 2002 and 2005 in its A1 attenuator state during the $\sim 60$-minute solar-viewing part of its orbit. For two long-duration flares (on 2002 July 20/21 and July 26/27), the decline could be followed for up to four orbits. Table \[anal\_ints\] gives the dates and [*GOES*]{} classes with time intervals over which analysis of at least the Fe-line complex was possible, one interval per flare except for the 2002 July 20/21 and July 26/27 flares; numbers indicate flares, letters time intervals for the 2002 July 20/21 and 26/27 flares. Also listed are the number of spectra $N_{\rm full}$, and $N_{\rm cont}$ analyzed in each interval for each of the [chianti$\_$full]{} and [chianti$\_$cont]{} emission models, and details of the estimated Fe abundance which will be discussed in Section \[Fe\_abund\_anal\].
[rlccrcrccc]{}
1..& 2002 Mar 10 & 22:57 – 23:50 & M2.3& 9 & $0.54 \pm 0.04$ & $7.83 \pm 0.03$ & 186 & $7.91 \pm 0.03$ & $8.02 \pm 0.07$\
2..& 2002 Apr 15 & 00:15 – 00:40 & M3.7& 59 & $0.56 \pm 0.19$ & $7.85 \pm 0.13$ & 88 & $7.86 \pm 0.09$ & $8.02 \pm 0.10$\
3..& 2002 May 31 & 00:13 – 00:57 & M2.4& & indeterminate$^b$ & & 149 & $7.93 \pm 0.13$ & $8.04 \pm 0.14$\
4 ..& 2002 Jun 1 & 03:52 – 04:04 & M1.6& 17 & $0.93 \pm 0.55$ & $8.07 \pm 0.20$ & 36 & $8.01 \pm 0.09$ & indeterminate\
5a..& 2002 Jul 20/21 & 22:29 – 23:28 & X3.3& 21 & $0.52 \pm 0.07$ & $7.81 \pm 0.06$ & 90 & $7.80 \pm 0.08$ & $7.87 \pm 0.20$\
5b..& & 00:06 – 01:03 & & 205 & $0.53 \pm 0.22$ & $7.82 \pm 0.16$ & 172 & $7.89 \pm 0.12$ & $8.07 \pm 0.16$\
5c..& & 01:42 – 02:40 & & & indeterminate & & 172 & $7.87 \pm 0.20$ & indeterminate\
5d..& & 03:24 – 04:16 & & & indeterminate & & 86 & $8.03 \pm 0.21$ & indeterminate\
6a..& 2002 Jul 26/27 & 23:01 – 00:00 & M4.6& 185 & $0.67 \pm 0.12$ & $7.93 \pm 0.07$ & 188 & $7.90 \pm 0.07$ & $7.97 \pm 0.12$\
6b..& & 00:37 – 01:36 & & 59 & $0.66 \pm 0.09$ & $7.92 \pm 0.06$ & 59 & $7.99 \pm 0.09$ & $8.09 \pm 0.10$\
6c..& & 02:14 – 03:13 & & & indeterminate & & 118 & $8.04 \pm 0.13$ & $8.09 \pm 0.14$\
7..& 2002 Jul 29 & 10:50 – 11:26 & M4.7& 2 & $0.64 \pm 0.09$ & $7.91 \pm 0.05$ & 79 & $7.89 \pm 0.05$ & $7.91 \pm 0.07$\
8..& 2002 Oct 4 & 05:41 – 05:56 & M4.0& 33 & indeterminate & & 26 & $7.91 \pm 0.05$ & $8.03 \pm 0.12$\
9..& 2002 Dec 2 & 19:23 – 19:32 & C9.6& 23 & $0.73 \pm 0.23$ & $7.96 \pm 0.12$ & 22 & $8.00 \pm 0.03$ & $8.05 \pm 0.10$\
10..& 2002 Dec 17/18 & 23:35 – 01:01 & M1.6& 71 & $0.43 \pm 0.25$ & $7.73 \pm 0.20$ & 29 & $7.84 \pm 0.15$ & indeterminate\
11..& 2003 Apr 23 & 01:00 – 01:40 & M5.2& 43 & $0.73 \pm 0.27$ & $7.96 \pm 0.14$ & 101 & $7.99 \pm 0.05$ & $8.07 \pm 0.09$\
12..& 2003 May 29 & 01:10 – 01:42 & X1.1& & indeterminate & & 118 & $7.90 \pm 0.05$ & $7.99 \pm 0.09$\
13..& 2003 Aug 19 & 10:00 – 10:26 & M2.7& 121 & $0.85 \pm 0.37$ & $8.03 \pm 0.16$ & 106 & $7.94 \pm 0.05$ & $7.88 \pm 0.12$\
14..& 2003 Oct 22 & 20:16 – 20:37 & M9.9& & indeterminate & & 23 & $7.74 \pm 0.04$ & $7.83 \pm 0.08$\
15..& 2003 Oct 23 & 20:06 – 20:37 & X1.1& 37 & $0.58 \pm 0.14$ & $7.86 \pm 0.10$ & 39 & $7.94 \pm 0.04$ & $8.06 \pm 0.09$\
16..& 2003 Nov 2 & 18:37 – 18:59 & X8.3& 31 & $0.50 \pm 0.06$ & $7.80 \pm 0.05$ & 71 & $7.88 \pm 0.08$ & $8.03 \pm 0.12$\
17..& 2003 Nov 11 & 15:34 – 16:33 & C8.5& 21 & $0.64 \pm 0.17$ & $7.91 \pm 0.10$ & 85 & $8.00 \pm 0.04$ & $8.13 \pm 0.09$\
18..& 2004 Jan 5 & 04:05 – 04:52 & M6.9& 14 & $0.53 \pm 0.05$ & $7.82 \pm 0.04$ & 37 & $7.93 \pm 0.03$ & $8.08 \pm 0.02$\
19..& 2004 Jul 20 & 12:40 – 13:41 & M8.7& 41 & $0.60 \pm 0.15$ & $7.88 \pm 0.10$ & 91 & $7.92 \pm 0.06$ & $8.00 \pm 0.10$\
20..& 2005 Jan 16 & 01:29 – 02:27 & X2.6& 98 & $0.28 \pm 0.06$ & $7.55 \pm 0.08$ & 102 & $7.72 \pm 0.09$ & $7.90 \pm 0.18$\
Spectral Analysis
-----------------
For each time interval, data and detector response matrix (DRM) files were extracted and read by OSPEX (Object Spectral Executive), an object-oriented IDL program recently substantially modified. The non-solar background spectrum was determined from the night-time parts of the [*RHESSI*]{} orbit. An isothermal fitting function was chosen to model the continuum and line emission, with goodness of fit determined by the reduced chi-squared $\chi_{\rm red}^2$, calculated as the weighted sum of the squares of the differences between background-subtracted count rates in each energy bin and the predicted count rates computed by folding the assumed incident photon spectrum through the DRM. The weights were determined from the predicted rates assuming Poisson statistics with zero systematic uncertainties. A graphical user interface allowed least-squares spectra with normalized residuals to be viewed (see Figures \[count\_rate\_resids\_CHIANTI\_full\] and \[count\_rate\_resids\_CHIANTI\_cont\]). One form of the fitting function we chose, unavailable previously, is a thermal spectrum (vth) calculated from [chianti]{}, including all lines and free–free and free–bound continua. The abundances of individual elements, most especially Fe, may be varied independently by a factor $F$ from a particular set of abundances, chosen in our case to be the “coronal" set of [@fel92]. The ion fractions of [@bry09] were used. This value and $T_e$ and emission measure ($N_e^2 V$) were set as free parameters to be determined, as well as those describing any nonthermal continuum present. A further component of the fitting function (drm$\_$mod) allows for small modifications (gain and energy resolution) in the DRM. For this model fitting function ([chianti$\_$full]{}), there is a total of 8 free parameters. There were slight disagreements in the fits, particularly around the energy of the Fe/Ni line complex ($\sim 8$ keV), with the [chianti]{} spectrum underestimating the line flux, worsening the $\chi_{\rm red}^2$. Figure \[count\_rate\_resids\_CHIANTI\_full\] shows a [*RHESSI*]{} detector 4 A1 spectrum and fit with the [chianti$\_$full]{} function and normalized residuals. For this fit, 80 energy bins were used with 8 free parameters defining the model spectrum, giving $\chi_{\rm red}^2 = 1.07$. With 72 degrees of freedom, this indicates a probability of 32% of exceeding this value through random statistical fluctuations in the count rates. The mismatch at $\sim 8$ keV is unlikely to be due to the omission in [chianti]{} of $n>5$ lines of [@phi08]; more probably it is due to an instrumental line from the tungsten collimator grids not allowed for in the DRM. The fit gives an abundance factor $F=0.372$, or $A({\rm Fe}) =7.67$.
![(Upper panel) [*RHESSI*]{} detector 4 A1 background-subtracted count-rate spectrum (black histogram) for an interval during the decay of the 2003 October 23 flare (energy range is 5 to 40 keV). The least-squares fitted spectrum (red) consists of two components folded through the DRM: (a) a thermal spectrum (vth, in green) including all lines and continua in [chianti]{}; (b) a nonthermal photon spectrum (thick2, dark purple). The function (drm$\_$mod) allows the energy resolution and gain to vary to give the best fit to the Fe line complex. The energy range of the fit was 5.7–33 keV (vertical dotted and dashed lines). The reduced chi-squared $\chi_{\rm red}^2 = 1.07$. The pink histogram is the background spectrum. (Lower panel) Residuals normalized to the statistical $1 \sigma$ uncertainties in each energy bin. []{data-label="count_rate_resids_CHIANTI_full"}](f1.eps){width="12cm"}
A more general fitting function was also chosen that includes free–free and free–bound continua alone as calculated by [chianti]{} and line features with Gaussian profiles to fit the Fe and Fe/Ni line complexes. This fitting function ([chianti$\_$cont]{}) has a single temperature and emission measure describing the continuum, and the two line fluxes as free parameters. The line energies were kept fixed at their theoretical values (6.65 keV and 8 keV); there is only a slight temperature dependence of the line energies $\sim 0.05$ keV). As free–bound emission contributes to the continuum, there is some dependence on element abundances; as an initial estimate, we chose the [@fel92] abundance set for this calculation. As with the [chianti$\_$full]{} function, gain and energy resolution adjustments with drm$\_$mod and a nonthermal continuum were included. This gave a total of 9 free parameters. The observed and fitted spectra for the same interval and detector are shown in Figure \[count\_rate\_resids\_CHIANTI\_cont\], with normalized residuals. The value of $\chi_{\rm red}^2 = 0.80$ implies a 89% probability of exceeding this value; there is a fairly random distribution of residuals with energy over the fitted range (5.67–33 keV). Allowing the 8-keV line flux to be a free parameter removes the enhanced residuals at this energy.
As the Fe abundance will be derived from the line fluxes and the continuum values of $T_e$ and $N_e^2 V$, the dependence on element abundances of the continuum flux in the neighborhood of the Fe and Fe/Ni line complexes should be investigated, in particular the Fe abundance. Anticipating the discussion in Section \[Fe\_abund\_anal\], an Fe abundance of $A({\rm Fe}) = 7.91$ or Fe/H $=8.13\times 10^{-5}$ is derived, i.e. a factor 1.55 less than the [@fel92] coronal value. The effect on the total continuum of this abundance difference can be partly tested with [*RHESSI*]{} software, since a provision is made in the analysis software to adjust the abundances of Si, S, Ca, Fe, and Ni by different factors in the continuum function. We took spectra during the 2003 October 23 flare with the [chianti$\_$cont]{} model function but reducing [*all*]{} the Si, S, Ca, Fe, and Ni abundances by a factor 1.55. The abundances for these low-FIP elements would then be equal to the “hybrid" abundances by [@flu99]. (There is nevertheless evidence that a constant reduction of all low-FIP elements is not observed; estimates of the flare potassium abundance from the RESIK instrument [@syl10a] indicate $A({\rm K}) = 5.86 \pm 0.23$ may be [*enhanced*]{} over the coronal value, 5.67, of [@fel92].) For the interval shown in Figure \[count\_rate\_resids\_CHIANTI\_cont\], the temperature was practically unchanged at 1.28 keV = 14.8 MK but the emission measure was 16% higher ($0.57 \times 10^{49}$ cm$^{-3}$). This difference folds directly into the derived Fe abundance.
The contribution that Fe alone makes to the free–bound and total continuum is, however, much less than the total of Si, S, Ca, Fe, and Ni. The exact contribution was calculated at temperatures typical of those found in this analysis, the results being given in Table \[Fe\_contr\_cont\], where the percentage contribution of the Fe free–bound continuum at an energy of 10 keV (near the peak of the [*RHESSI*]{} count rate spectrum in the A1 state) is given to the total as a function of temperature (the contributions are similar at other energies). Thus, for $T_e=25$ MK, where the differences are most marked, the total continuum flux for a coronal Fe abundance ($A({\rm Fe}) = 8.10$) and an Fe abundance $A({\rm Fe}) = 7.91$ decreases by 6%, from 15 % to 21 %. This means that the derived emission measure will increase by 6% if the other elements (Si, S, Ca, and Ni) remain at their coronal [@fel92] abundances. At present, there is no recent detailed abundance analysis for these elements to confirm whether their abundances are the same as the [@fel92] abundances. However, it is worth pointing out that RESIK measurements of the continuum flux at somewhat lower energies (2.9–3.9 keV) [@phi10] are better described by coronal abundances [@fel92] than other abundance sets. Column 2 of Table \[Fe\_contr\_cont\] gives the percentage of the total free–bound continuum to the total continuum (free–free and free–bound) at 10 keV. The calculations in Table \[Fe\_contr\_cont\] are confirmed by more detailed calculations involving all significant elements (J. Sylwester, work in preparation).
[ccccc]{}
10 & 77 & 4 & 5 & 8\
12 & 73 & 6 & 8 & 12\
15 & 67 & 8 & 12 & 18\
20 & 58 & 9 & 15 & 21\
25 & 50 & 8 & 15 & 21\
![Same as Figure \[count\_rate\_resids\_CHIANTI\_full\] with the thermal spectrum used for that plot replaced with the thermal continuum given by [chianti]{} with coronal [@fel00] abundances and two lines with gaussian profiles at the mean energies of the Fe and Fe/Ni line complexes (6.65 and 8.0 keV respectively). The energy range is 5 to 40 keV. The value of $\chi_{\rm red}^2 = 0.80$. The mismatch in the residuals at $\sim 8$ keV has been significantly reduced. []{data-label="count_rate_resids_CHIANTI_cont"}](f2.eps){width="14cm"}
Because of the mismatch using the [chianti$\_$full]{} fitting function at $\sim 8$ keV, there were fewer spectral fits having small values of reduced $\chi_{\rm red}^2$ than with the [chianti$\_$cont]{} fitting function. Temperatures and emission measures from [chianti$\_$cont]{} and [chianti$\_$full]{} are compared in the plot shown in Figure \[compare\_T\_EM\] for time intervals during three flares for which $\chi_{\rm red}^2 < 1.5$. It is evident that temperatures from the [chianti$\_$full]{} model are smaller than those from [chianti$\_$cont]{} by $\sim 1$ MK and emission measures larger by a factor $\sim 2.5$ (0.4 in the logarithm). Although the specific reasons for these differences are unclear, they may be due to attempts in the fitting process with [chianti$\_$full]{} to correct for the $\sim 8$ keV mismatch with a continuum function that is slightly too steep at energies $\gtrsim 8$ keV.
![Temperatures (MK) and logarithms of emission measures (unit = $10^{49}$ cm$^{-3}$) from spectral fits to 3 flares using fitting function [chianti$\_$full]{} plotted against those from fitting function [chianti$\_$cont]{}. The three flare intervals are numbers 2, 3, and 6a in Table \[anal\_ints\], with symbols indicated in the legend. []{data-label="compare_T_EM"}](f3.eps){width="16cm"}
In our previous analysis, we compared spectral fits from seven of the nine [*RHESSI*]{} detectors suitable for spectral analysis in the low-energy region (detectors 2 and 7 have poorer energy resolution and higher energy threshold and so they were not used). Here, this comparison was done more systematically using a 30-minute time interval in the decline of the flare on 2002 July 26 with the [chianti$\_$cont]{} fitting function. From a detailed analysis of the temperature and emission measure over this period, the mean $1\sigma$ uncertainties (in MK) for the detectors 1, 3, 4, 5, 6, 8 and 9 were found to be respectively 1.20, 1.19, 1.05, 3.02, 1.12, 1.08, 1.16. They are thus smallest for detector 4 and largest for detector 5. This is also true for the uncertainties in the emission measure estimates. In light of these results, our choice of spectral fits from detector 4 in our earlier work appears to be vindicated, and so we chose to fit spectra from this detector.
Use and Validation of CHIANTI Spectra
-------------------------------------
The [*RHESSI*]{} analysis code now uses version 6 of the [chianti]{} atomic code and database [@der09] which includes the best atomic data available for the lines in the [*RHESSI*]{} range discussed here. In particular, an abundance of Fe is directly determined from comparison of the Fe line and Fe/Ni line complex fluxes with that from [chianti]{} using $T_e$ and emission measure from the continuum rather than with our earlier work [@phi06] in which the equivalent width of each line complex was determined from [*RHESSI*]{} spectra and compared with calculations based on the sum of individual lines in each complex by [@phi04]. Expressed as contribution functions $G_{\rm Fe}(T_e)$ and $G_{\rm Fe/Ni}(T_e)$, or the amount of emission from the Fe line complex (defined to be all lines in the 6.391–7.005 keV range) and Fe/Ni line complex (all lines in the 7.728–8.907 keV range) from a plasma with unit volume emission measure, there are differences of only a few per cent between the [@phi04] and [chianti]{} v. 6 calculations for the Fe line. There are more significant differences for the Fe/Ni line for $T_e \lesssim 12$ MK, for which the [chianti]{} v. 6 calculations are higher by amounts that depend on $T_e$. These are due to the addition of satellites near the $1s^2 - 1snp$ ($n=4$, 5) lines not included by [@phi04] or earlier versions of [chianti]{}. The [chianti]{} v. 6 calculations for $G_{\rm Fe/Ni}(T_e)$ are still incomplete in that $1s^2 - 1snp$ ($n > 5$) lines and the associated satellites are not included. However, based on work by [@phi08], this is unlikely to lead to an underestimate of $G_{\rm Fe/Ni}(T_e)$ by more than a few percent in the range of temperatures (approximately 10–22 MK) considered in this work.
Validation of [chianti]{} spectra is possible by comparing with solar flare spectra from the [*Solar Maximum Mission*]{} Bent Crystal Spectrometer for the Fe line complex. There are small differences apparent in the ratio of some of the intense satellites to the resonance line which provides the means of determining $T_e$ in high-resolution, crystal spectrometer data. The BCS spectra analyzed by [@lem84] give $T_e = 15.0 \pm 0.5$ MK, whereas a re-analysis with [chianti]{} v. 6 leads to a higher temperature, 16.5 MK. The difference is probably due to improved atomic data for both the satellites and the collisional rates for the resonance line. Rather large differences (up to 0.01 keV) are apparent in the energies of – satellites, though these are not likely to affect the total flux of the Fe line complex. There are no high-resolution flare spectra of the Fe/Ni line complex, so validation is not possible.
Fe ABUNDANCE ANALYSIS {#Fe_abund_anal}
=====================
Applying the [chianti$\_$full]{} fitting function to [*RHESSI*]{} spectra gives temperature, emission measure, and the abundance factor $F$, which is the factor applied to the baseline Fe and Ni abundance and is determined by the fluxes of the Fe and Fe/Ni line complexes. The baseline abundance set used was the coronal one of [@fel00] (for Fe this is $A({\rm Fe})_{FL} = 8.10$). For all the flares analyzed, the values of $F$ are practically constant with time and are unrelated to variations in $T_e$. Since the Fe and Fe/Ni line features in the energy range chosen are nearly entirely due to iron, the measured Fe abundance is $F$ multiplied by the coronal [@fel00] value, or in logarithmic terms, $A({\rm Fe})_{\rm meas} = {\rm log}_{10}F + A({\rm Fe})_{FL}$. Table \[anal\_ints\] gives the mean value of $F$ and standard deviation for each of the time periods listed with the number of spectra used to derive $F$ having $\chi_{\rm red}^2 < 1.5$. For a few flares there were too few good-quality spectra to give a reliable value of $F$ (marked “indeterminate"). For the remaining 18 sets of spectra, the mean value of $F$ is $0.55 \pm 0.08$ (s.d.), and so $A({\rm Fe})_{\rm meas} = 7.90 \pm 0.02$. This is several standard deviations less than $A({\rm Fe})_{FL}$, despite the fact that the value of $F$ is poorly determined for some flares. It is a factor 2.5 more than the photospheric Fe abundance of [@asp09]. To determine the abundance of Fe from the [chianti$\_$cont]{} model, we followed the procedure for analyzing spectra from the RESIK crystal spectrometer [@syl10a; @syl10b]. For a spectrum in a particular ($i$th) [*RHESSI*]{} time interval, the iron abundance is determined from the flux $\mathfrak{F}_i$ of either the Fe line or the Fe/Ni line complexes using
$$f_i ({\rm Fe}) = \frac{\mathfrak{F}_i}{G(T_i) EM_i}$$
\[eq\_for\_f\]
where the temperature $T_i$ and emission measure $EM_i$ are output from the $i$th best-fit model function. Thus, while $f_i$, like $F_i$ in the [chianti$\_$full]{} fitting model, is a factor multiplying the [@fel00] Fe abundance to give the measured Fe abundance, it is defined in terms of the measured fluxes of either the Fe or Fe/Ni line complex, the continuum temperature $T_i$ and emission measure $EM_i$, and the contribution function $G_{\rm Fe} (T_i)$ or $G_{\rm Fe/Ni} (T_i)$.
The estimated fluxes in the Fe line and Fe/Ni line complexes divided by the continuum emission measure when plotted against continuum temperature can be compared with the theoretical $G_{\rm Fe}(T_e)$ and $G_{\rm Fe/Ni}(T_e)$ functions calculated with $A({\rm Fe})_{FL} = 8.10$ and the photospheric abundance ($A({\rm Fe}) = 7.50$). This is done in Figure \[GofT\_20030819\] (left) for the 2003 August 19 flare. The uncertainties (from the OSPEX software) are larger for the Fe/Ni line complex, which for the temperatures analyzed here is about a factor 10–20 less intense than the Fe line complex, and are larger for lower-temperature spectra late in the flare when the emission was weaker.
Values of measured Fe abundances from Eq. (1) were obtained for each spectral interval in this and other flares, and histogram distributions found for the Fe and Fe/Ni line complexes. This is illustrated for the Fe line complex in Figure \[GofT\_20030819\] (right) where numbers of values falling into intervals of 0.05 in $A({\rm Fe})$ are plotted, together with a best-fit gaussian curve. From the best-fit curves, a mean value of $A({\rm Fe})$ and standard deviation can be derived. The values derived are listed for each of the time periods in the last two columns of Table \[anal\_ints\]. For some time periods, the Fe/Ni line complex was too weak to be measured so the Fe abundance could not be derived; the abundances are marked “indeterminate" when the standard deviation in $A({\rm Fe})$ was larger than 0.3 (corresponding to a factor 2 in the abundance estimate). Similar plots to Figures \[GofT\_20030819\] were done for the time ranges in Table \[anal\_ints\]. The total of all flares for the Fe line and Fe/Ni line complexes is shown in Figure \[line\_flux\_allflares\]. There is a clear trend for the values of flux divided by emission measure to be below the coronal $G(T_e)$ curves by a constant factor for the Fe line complex, suggesting (as with the [chianti$\_$full]{} analysis) that the Fe abundance is smaller than the [@fel00] value, but larger than the photospheric value. For the Fe/Ni line complex, the trend is similar at high temperatures ($T_e \gtrsim 16.5$ MK, but the points become steadily higher than the coronal abundance curve for decreasing temperatures. This may be because of the instrumental line at $\sim 8$ keV mentioned earlier.
Figure \[Fe\_abund\_hist\] shows the distribution of Fe abundance values for spectra during all flares lumped together for the Fe line and Fe/Ni line complexes, with best-fit gaussian curves. The mean value of $A({\rm Fe})$ is $7.91 \pm 0.10$ from the Fe line, $8.01 \pm 0.16$ from the Fe/Ni line complex. The larger Fe abundance and larger uncertainty from the Fe/Ni line reflects the departure of the points from the theoretical $G_{\rm Fe/Ni}(T_e)$ curve, so the Fe abundance from the Fe line complex is clearly the preferred value. We note that the continuum in this analysis has a small contribution from Fe emission, so there is a slight dependence on the coronal abundances used, which was the [@fel92] set. Use of an Fe abundance $A({\rm Fe}) = 7.91$ instead of $A({\rm Fe})_{FL} = 8.10$, if the elements Si, S, Ca, and Ni remain at their coronal values, should lead to emission measures that are $\sim 6$% larger. Eq. (1) indicates that the derived value of $f_i$ and therefore the Fe abundance using the [chianti$\_$cont]{} emission model with coronal abundances for the continuum will be too large by 6% (0.03 in the logarithm). The abundance variation from flare to flare, in spite of the factor-of-100 range in the [*GOES*]{} X-ray importance and [*RHESSI*]{} count rates, is very small and well within the standard deviations for each flare.
Discussion and Conclusions
==========================
Our analysis of [*RHESSI*]{} solar flare spectra gives estimates of the Fe abundance from two different emission models, [chianti$\_$full]{} (continuum and lines as given by the [chianti]{} atomic code) and [chianti$\_$cont]{} (continuum given by [chianti]{}, Fe and Fe/Ni line complexes at 6.65 keV and 8 keV separately fitted). Both methods give Fe abundance values that are constant to within the estimated uncertainties and higher than those derived from the photosphere or meteorites. The [chianti$\_$full]{} analysis was found to suffer from poor fits to the spectral region around the Fe/Ni line complex at 8 keV, leading to generally higher values of $\chi_{\rm red}^2$. Using only spectral fits with $\chi_{\rm red}^2 < 1.5$, we found from 18 time intervals that the mean $A({\rm Fe})$ to be $7.87 \pm 0.02$, the uncertainty being in the mean values for each of the 18 intervals. The [chianti$\_$cont]{} analysis has the advantage that the Fe and Fe/Ni line complexes can be fitted separately and $A({\rm Fe})$ can be derived for each. The possible presence of an instrumental line at $\sim 8$ keV adds to the observed Fe/Ni line complex and so the measured flux is an upper limit to the solar value. This probably explains the 23% difference between the values $A({\rm Fe}) = 7.91 \pm 0.10$ for the Fe line during 25 time intervals and $A({\rm Fe}) = 8.01 \pm 0.16$ for the Fe/Ni line during 21 time periods, though the Fe/Ni line complex is also a much weaker feature. When all the [chianti$\_$cont]{} observations are lumped together (Figures \[line\_flux\_allflares\] and \[Fe\_abund\_hist\]), it is clear, particularly from the Fe line, that the Fe abundance is constant to within estimated uncertainties from flare to flare. The distribution of all the estimates from the Fe line leads to an Fe abundance given by $A({\rm Fe}) = 7.91 \pm 0.10$, which is our preferred value.
Our value is higher by a factor $2.6 \pm 0.6$ than Fe abundance estimates from the photosphere (e.g. $A({\rm Fe}) = 7.50 \pm 0.04$: [@asp09]) and by a factor $2.9 \pm 0.7$ than those from meteorites ($7.45 \pm 0.01$: [@lod09]). However, it is lower than those given for coronal plasmas, which range from 7.65 [@par77; @flu99] to 8.19 [@whi00]; it is a factor $1.55 \pm 0.5$ lower than that given by [@fel00] and [@fel92], viz. $A({\rm Fe}) = 8.10$, which is used as the coronal abundance set in [chianti]{}. The preliminary results for flares observed with the XRS instrument on Mercury MESSENGER [@den08] give an Fe abundance that is a factor 2.3 times photospheric [@gre98], i.e. $A({\rm Fe}) = 7.9$, which is consistent with our preferred value. Our value is within $1\sigma$ of the Fe abundance of Table 2 of [@flu99], $A({\rm Fe}) = 7.83$, for their “hybrid" abundance model, taking account of the uncertainties in both our value and the hybrid model. Our value of $7.91 \pm 0.10$ is in very close agreement (0.02 less) with the energetic particle abundances reported by [@rea95] for gradual events, viz. $A({\rm Fe}) = 7.93 \pm 0.01$, suggesting a relationship of the fractionation process involved in the formation of hot flare plasmas and the acceleration of solar energetic particles. The constancy of the [*RHESSI*]{} Fe abundance estimates points to fractionation processes in flares, at least their declining stages, to be similar to those for the quiet Sun. It is rather against expectations, as discussed by [@asp09], who state that “the degree of chemical separation varies significantly, being more severe in regions of higher solar activity." Though our measurements are during the decay of flares, the flare plasma is unlikely to be mixed with other plasma after the impulsive stage when (as is widely accepted) chromospheric evaporation occurs.
If the Fe abundance derived here for flares is representative of the quiet solar corona and active-region levels, there are consequences for the radiation loss curve (Figure \[rad\_loss\]). Iron ions are an important emitter for solar plasmas in the temperature range spanning values of the quiet corona to the tens of MK of solar flares. A maximum in radiation loss at $\sim 1$ MK is due to emission lines of – between 171 Å and 284 Å, and a second maximum at $\sim 10$ MK is due to – X-ray lines. The precise abundance of iron is therefore important for studies of the energy balance in quiet coronal or flare loops, particularly for higher densities when radiation loss may dominate conduction losses [@kli08]. If the value obtained here, $A({\rm Fe}) = 7.91$, applies to all coronal plasmas with temperatures 1–20 MK, the radiation loss curve will be correspondingly modified. Figure \[rad\_loss\] shows the radiation loss for optically thin plasma with this temperature range and for Fe abundances with photospheric and coronal values including our own. The contribution from Fe ions and hydrogen alone is also shown, with maxima at $\sim 1$ MK and $\sim 10$ MK. (The maximum at $\sim 20000$ K is due to Ly-$\alpha$ emission of hydrogen.) The radiation loss is less for a photospheric Fe abundance than a coronal Fe abundance by a factor $\sim 3.4$ for quiet coronal loops ($T_e \sim 1$ MK) and a factor $\sim 3.0$ for flare loops with $T_e \sim 10$ MK. For $A({\rm Fe}) = 7.91$ obtained in this work, the radiation loss is less than that for coronal Fe abundance [@fel92] by a factor 0.8 ($T_e \sim 1$ MK) and 0.72 ($T_e \sim 10$ MK). The radiation loss curve approximated by [@kli08] with a piece-wise continuous temperature-dependent form assumed abundances that are twice those of [@mey85]; as [@mey85] gives $A({\rm Fe}) = 7.6$ for the corona, the [@kli08] value (7.9) is very nearly the one obtained here.
The various models advanced up to the mid-1990s for explaining the FIP effect have been reviewed by [@hen98]. Fractionation according to the first ionization potential is generally explained by the fact that low-FIP (FIP $\lesssim 10$ eV) elements are partly ionized in the photosphere but high-FIP elements are predominantly neutral. It is difficult to obtain from any of these early models a quantitative enhancement of low-FIP elements in coronal plasmas, however, and it appears to be a feature of the present work that for flares with a range of X-ray importance that the enhancement of Fe is a particular value, constant from flare to flare. The constancy of the Fe abundance in flare plasmas is also consistent with the analysis of 2795 RESIK spectra indicating a constant enhancement (of a factor 5 over the photospheric abundance) of potassium in flares [@syl10a] and an argon abundance that is also constant (in this case to within 20% of photospheric Ar abundance proxies [@syl10b]). The more recent FIP model of Laming [@lam04; @lam09] involves the ponderomotive force that arises when Alfvén waves pass from the chromosphere to the corona, and definite predictions about the enhancements of various elements can be made according to the dimensions of the coronal loop that the waves are incident on, its magnetic field, and the wave intensity. A particular example is given for a loop with length of 100,000 km and field strength of 7 G. [*RHESSI*]{} flare loops are likely to be much shorter, but it is interesting that for a fairly wide range of wave energy fluxes the coronal enhancement of Fe is between 2 and 3, as obtained in this work.
In a wider context, it has been found that the occurrence of giant planets around stars depends sensitively on metallicity, specifically the Fe abundance [@fis05; @val08]. This refers to Fe abundances in the stellar photospheres. However, it would be interesting to use the methods of this work to derive the coronal or flare abundance of Fe from the 6.65 keV line feature and compare with photospheric abundances to see if there are correlations of the difference with the probability of planet formation.
We thank A. K. Tolbert and R. A. Schwartz for their invaluable help in the data analysis and to A. Gopie for initial data reduction. J. Sylwester and B. Sylwester are thanked for the use of their methodology in the derivation of the Fe abundance in this work. K. J. H. P. acknowledges support from NASA through ADNET (under the SESAA-II contract, NNG06EB68C) for a visit to Goddard Space Flight Center in 2008 and support from a National Research Council Senior Research Associateship during the original conception of this work. [chianti]{} is a collaborative project involving the US Naval Research Laboratory, the Universities of Florence (Italy) and Cambridge (UK), and George Mason University (USA). We are grateful to the authors of the [chianti]{} code for continued help in adding data to the spectral regions discussed here.
Asplund, M., Grevesse, N., Sauval, A. J., & Scott, P. 2009, , 47, 481
Bryans, P., Landi, E., & Savin, D. W. 2009, , 691, 1540
Caffau, E., Ludwig, H.-G., Steffen, M., Freytag, B., & Bonifacio, P. 2011, Solar Phys., 268, 255
Dennis, B. R., & Starr, R. 2008, in preparation ([*RHESSI*]{} Science Nugget on http://sprg.ssl.berkeley.edu/$\sim$tohban/wiki/index.php/Flare Plasma Abundances - New X-ray Observations)
Dere, K. P., Landi, E., Mason, H. E., Monsignori Fossi, B. C., & Young, P. R. 1997, , 125, 149
Dere, K. P., Landi, E., Young, P. R., Del Zanna, G., Landini, M., & Mason, H. E. 2009, , 498, 915
Feldman, U., Doschek, G. A., Kreplin, R. W., & Mariska, J. T. 1980, , 241, 1175
Feldman, U. 1992, Phys. Scr., 46, 202
Feldman, U., & Laming, J. M. 2000, Phys. Scr., 61, 222
Fischer, D. A., & Valenti, J. 2005, , 622, 1102
Fludra, A, & Schmelz, J. T. 1999, , 348, 286
Grevesse, N., & Sauval, A. J. 1998, , 85, 161
Hénoux, J.-C. 1998, Space Sci. Rev., 85, 215
Klimchuk, J., Patsourakos, S., & Cargill, P. 2008, , 682, 1351
Laming, J. M. 2004, , 614, 1063
Laming, J. M. 2009, , 695, 954
Lemen, J. R., Phillips, K. J. H., Cowan, R. D., Hata, J., & Grant, I. P. 1984, , 135, 313
Lodders, K., Palme, H., & Gail, H.-P. 2009, in [*Landolt-Börnstein, New Series, Astronomy and Astrophysics*]{}, Springer Verlag, Berlin \[arXiv:0901.1149\]
Mewe, R., Lemen, J. R., Peres, G., Schrijver, J., & Serio, S. 1985, , 152, 229
Meyer, J.-P. 1985, , 57, 173
Parkinson, J. H. 1977, , 57, 185
Phillips, K. J. H. 2004, , 605, 921
Phillips, K. J. H., Chifor, C., & Dennis, B. R. 2006, , 647, 1480
Phillips, K. J. H. 2008, , 490, 823
Phillips, K. J. H., Sylwester, J., Sylwester, B., & Kuznetsov, V. D. 2010, ApJ, 711, 179
Reames, D. V. Adv. Space Res., 15 (7), 41
Smith, D. M., et al. 2002, , 210, 33
Sylwester, J., Lemen, J. R., & Mewe, R. 1984, , 310, 665
Sylwester, J., Sylwester, B., Phillips, K. J. H., & Kuznetsov, V. D. 2010, , 710, 804
Sylwester, J., Sylwester, B., Phillips, K. J. H., & Kuznetsov, V. D. 2010, , 720, 1721
Tanaka, K., Watanabe, T., Nishi, K., & Akita, K. 1982, , 254, L59
Valenti, J. A., & Fischer, D. A. 2008, in 14th Cambridge Workshop on Cool Stars, Stellar Systems, and the Sun, ASP Conference Series, 384, 292
White, S. M., Thomas, R. J., Brosius, J. W., & Kundu, M. R. 2000, , 534, L203
Young, P. R., DelZanna, G., Landi, E., Dere, K. P., Mason, H. E., Landini, M. 2003, , 144, 135
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we consider kink-antikink collisions for some classes of $(1,1)$-dimensional nonlinear models. We are particularly interested to investigate in which aspect the presence of a general kinetic content in the Lagrangian could be revealed in a collision process. We consider a particular class of models known as twin theories, where different models lead to same solutions for the equations of motion and same energy density profile. The theories can be distinguished in the level of linear stability of defect structure. We study a class of k-defect theories depending on a parameter $M$ which is the twin theory of the usual $\phi^4$ theory with standard dynamics. For $M\to\infty$ both models are characterized by the same potential. In the regime $1/M^2<<1$, we obtain analytically the spectrum of excitations around the kink solution. It is shown that with the increasing on the parameter $1/M^2$: i) the gap between the zero-mode and the first-excited mode increases and ii) the tendency of one-bounce collision between kink-antikink increases. We numerically investigate kink-antikink scattering, looking for the influence of the parameter changing for the thickness and number of two-bounce windows, and confronting the results with our analytical findings.'
author:
- 'A. R. Gomes$^{1}$, R. Menezes$^{2,3}$, K. Z. Nobrega ${^4}$, F. C. Simas$^{5}$'
title: 'Kink-antikink collisions for twin models'
---
Introduction
============
In this work we consider kink-antikink collisions for some classes of $(1,1)$-dimensional nonintegrable models. As one knows, the collision process for integrable systems has an intrinsic simplicity, with the solitary waves passing to each other with at most a phase shift. Despite this simplicity, analytical results for most integrable models is rather nontrivial. As an example, the analysis of kink-kink and kink-antikink scattering for integrable models using perturbative analysis can be found in Refs. [scat1,scat2]{}. For more recent results on this subject, see Refs. [scat3,scat4,scat5]{}. The effect of a phase shift of the collision process for an integrable model can be confronted to the richness of the collision process for nonintegrable models, where most of analysis must be done numerically. As already shown by Anninos et al [@aom] for the $\lambda\phi^4$ theory, for sufficiently small initial velocities, the kink and antikink capture one another and a trapped bion state is formed. On the other hand for larger velocities a simple collision occurs, and after the contact, the pair of defects retreats from each other. For intermediate initial velocities, the structure of the collision is fair more complex, with two-bounce windows appearing in an endless sequence between the larger bion region and the one-bounce ones. A self-similar structure is revealed if one zooms the initial approximation velocity close to the interface regions. Indeed, if one refines the initial velocity near to an interface between a two-bounce window and a bion window, one sees the appearance an endless sequence of three-bounce and bion windows. This refinement process continues with higher number of bounces being observed; however, a limit in this process is achieved due to losses by radiation. The higher number of bounces appears jointly with a higher number of internal mode oscillations. The different windows are related by a scaling relation between the window velocity thickness and the number of internal mode oscillations.
The intriguing character of these alternating regions conflicts with the naive expectation that bion states could not be formed for larger initial velocities than that verified for n-bounce collisions. A semi-phenomenological theory that accounts for the presence of two-bounce windows is due to Campbell et al [@csw]. In their work, the two-bounce behavior is described as a two-steps interacting process. In the first interaction the energy is transferred from the translational mode to an internal shape mode oscillation of the kink. In the second one, if the kink and antikink satisfy a resonance criteria [@csw; @pc], the energy in the vibrational mode is turned to the translational mode, and the pair $K\bar K$ is liberated from their mutual attraction. Quantitatively the simple relation must be obeyed [@csw] $$\label{T_omega}
\omega_1 T=2n\pi+\delta$$ where $\omega_1$ is the frequency of the internal mode, T is the time interval between the bounces, $n$ is an integer and $\delta$ is the phase shift between the incoming and outgoing kink. In this way, the higher is $\omega_1$, the smaller is the time interval between the bounces, which signals that the energy transference from translational to vibrational mode is more difficult to be achieved. The simple one bounce scattering process occurs for the initial kink velocity higher than a critical velocity $v_c$. A heuristic expression for the relation between $T$ and $v$ for the two-bounce windows is presented also in [@bk], namely $$T\propto \sqrt{v_c-v}, \,\,\,v<v_c.$$ The former two expressions can be used to understand quantitatively how the windows centers scale with $n$ [@bk], with the successive windows with even smaller thickness accumulating near $v=v_c$.
The escape of the pair kink-antikink from their mutual attraction is verified when the energy of the vibrational mode is less than the kinetic energy of the colliding kinks [@bk]. Usually one can use collective coordinates to obtain the $K\bar K$ attractive interaction potential $U_{K\bar K}$ as a function of the separation of the pair kink-antikink [kudr,sug]{}. The potential $U_{K\bar K}$ can be intuitively understood as the energy of the static field configuration consisting of a kink at $+Z$ and an antikink at $-Z$ [@cps]. On the other hand, if the time duration of internal oscillations is accompanied by a leaking of energy by radiation greater than the kinetic energy of the colliding kinks, a trapped bion $K\bar K$ state is formed [@bk].
The analysis of kink-antikink collisions can be found in the literature for several interesting examples of nonintegrable models. In addition to the already cited $\lambda\phi^4$ model [@aom], one can cite the modified sine Gordon [@pc] and the $\phi^6$ model [@dmrs]. Most of nonintegrable models have internal oscillatory modes, responsible for resonant scattering. However, there are exceptions, as the $\phi^6$ model where despite the absence of an internal oscillatory mode, resonant scattering was reported [@dmrs]. There, the potential for linear perturbations has a wide central well and the energy can be transferred from the translational mode to an extended meson state residing in this potential [@dmrs]. Indeed, the central well allow several discrete eigenvalues corresponding to meson-soliton bound states [@lohe]. This is contrary to what happens for the more usual models $\phi^4$ and sine-Gordon where mesons can pass through the kink and antikink without reflection, with only a phase shift [@lohe]. In the $\phi^6$ model the two-bounce windows satisfy the same relation given by eq. (\[T\_omega\]), but with $\omega_1$ as the frequency of the lowest collective mode. Some interesting discussions related to this topic can also be seen in Ref. [@weigel].
Our proposal here is to investigate in which measure the kink-antikink collision processes can be used to distinguish twin models, i.e. a class of topological defects with the same scalar field profiles and energy densities [@altw]. For some recent results on this subject, see [br, bhm,bllm]{}. We are particularly interested in k-defect theories, in part due to their use for explaining the accelerated expansion of the universe [k\_cosm]{}. Kink-antikink collision processes are useful for cosmology in theories with one extra dimension [@tm], as in the ekpyrotic proposal for brane collision [@ekpy]. Also, one can cite theories in the braneworld scenario with generalized dynamics [@bdglm; @blom]. The study of bubble cosmology has now being a subject of renewed interest since it can lead to possible ways to probe string landscape [@pol]. Cosmic bubble collisions in the regime of high nucleation rate can be studied ignoring the expansion and the curvature of the universe [@haw]. Despite the usual approach being lattice simulations in $(3,1)$ dimensions [@egl], in the special situation of bubbles with $SO(2,1)$ symmetry, a high speed collision of two bubbles is equivalent, depending on the vacuum configuration, to a $(1,1)$ dimensional $K\bar K$ collision [@lim]. This shows that the studies of $K\bar
K$ collision in twin theories can be useful for several cosmological scenarios.
In this work, as a starting point we will focus on $(1,1)$-dimensional scalar field theories in the Minkowski spacetime. Despite the simplicity of the proposal, we will show that our results lead to interesting insights about the influence of the k-dynamics on some characteristics of the collision, namely the presence of two-bounce windows. For this purpose, in the Sect. II we review the main first-order formalism for twin theories. Stability analysis is reviewed in Sect. III. Sect. IV specializes the discussion for the $\phi^4$ model and its twin theory. The discrete spectra of fluctuations is obtained for mass parameter $1/M^2\ll1$ and compared with known results for the $\phi^4$ theory. It is shown that an increasing of the parameter $1/M^2$ increases the gap between the zero- and excited modes. This is confronted with our numerical results, presented in Sect. V. Our main conclusions are presented in Sect. VI.
Twin theories
=============
We start revising some results of the general formalism of k-defects in $(1,1)$ dimensions [@bab; @asw; @blmo]. Consider a general Lagrangian density given by $$\mathcal{L}=\mathcal{L}(\phi,X)$$ with $$X=\frac12 \partial_\mu \phi \partial^\mu \phi.$$ A remark about notation: in this paper a subscript in $\mathcal{L}$ (and later on, in $W$) means partial derivative with respect to the argument. Then for instance $\mathcal{L}_X\equiv\partial\mathcal{L}/\partial X$, $\mathcal{L}_{X\phi}\equiv\partial^2\mathcal{L}/(\partial X\partial \phi)$ and so on. Now, the equation of motion is given by $$\partial_\mu \left(\mathcal{L}_X \partial^\mu \phi\right)=\mathcal{L}_\phi$$ or $$\label{eqm1}
G^{\mu\nu}\partial_\mu\partial_\nu \phi = - 2X \mathcal{L}_{X\phi}+ \mathcal{L}_{\phi}$$ with $$G^{\mu\nu}=\mathcal{L}_X \eta^{\mu\nu} +\mathcal{L}_{XX} \partial^\mu \phi
\partial^\nu \phi$$ The energy-momentum tensor is $$T^{\mu\nu}=\mathcal{L}_X \partial^\mu \phi \partial^\nu \phi- \eta^{\mu \nu}\mathcal{L}$$ and the corresponding energy is $$E=\int^{\infty}_{-\infty} T^{00} \, dx.$$ For static solutions, $\phi=\phi(x)$, and Eqs. can be rewritten as $$\label{eqmest}
\mathcal{L}_X A^2\phi^{\prime\prime} =- 2X\mathcal{L}_{X\phi}+ \mathcal{L}_\phi$$ with $$\label{Asq}
A^2=\frac{2 X \mathcal{L}_{XX}+\mathcal{L}_X}{\mathcal{L}_X}.$$
Here we will be interested in the corresponding model in the modified k-defect theory of a scalar field model with the standard Lagrangian density $$\label{L_st}
\mathcal{L}_{(S)}=X-V(\phi)$$ where $V(\phi)$ is the potential for the standard theory. The equation of motion for static solutions gives $$\phi^{\prime \prime }=V_{\phi}$$ and the energy density is $$T^{00}\equiv\rho_S(x)=\frac12{\phi^{\prime }}^2+V(\phi).$$ A very interesting proposal for a corresponding k-defect theory can be found in [@altw]: $$\label{L_kt}
\mathcal{L}_{(m)}=M^2-M^2\biggl(1+\frac{U(\phi)}{M^2} \biggr)\sqrt{1-\frac{2X}{M^2}}.$$ where $M$ is a mass scale. As pointed out in [@bdglm], the limit $1/M^2\to0$ turns this Lagrangian in a standard form with potential $U(\phi)$. In this way we can say $U(\phi)$ is the potential of the modified model. Stable static kinklike solutions must be pressureless ($T^{11}=0$) [bdglm,blmo]{}. In this way the equation of motion gives $$\label{phi'U}
{\phi^{\prime }}^2=2U(\phi)+\frac{U^2(\phi)}{M^2}$$ and the energy density is $$\rho_m(x)={\phi^{\prime }}^2.$$ If the potentials for the standard and k-defect theory are related by [altw]{} $$\label{VU_relation}
V(\phi)=U(\phi)+\frac12\frac{U(\phi)^2}{M^2}$$ then the models of Eqs. (\[L\_st\]) and (\[L\_kt\]) have the same stable defect structure $\phi(x)$ and energy density [@altw; @bdglm].
The first-order formalism for general Lagrangian density $\mathcal{L}(X,\phi) $ [@blm] is very useful to understand the connections between the potentials. Introducing a function $W(\phi)$ for the standard Lagrangian density, and for the potential $$V(\phi)=\frac12W_{\phi}^2$$ the solutions of the second-order equation of motion are also solutions of $$\label{eom_st}
\phi^{\prime }=W_{\phi}.$$ When applied to the k-defect theory, the corresponding potential is [bdglm]{} $$\label{U_M}
U(\phi)=-M^2+M^2\sqrt{1+\frac{W_{\phi}^2}{M^2}}$$
Stability analysis
==================
Here we will review the linear stability of the solutions. We consider $\phi=\bar\phi+\eta$, where $\bar\phi$ is the unperturbed solution and we suppose small perturbations $\eta$ around this solution. We can use Eq. to attain the first-order contribution in $\eta$ $$\label{eigenEq0}
\partial_\mu \!\!\left(\mathcal{L}_X \partial^\mu \eta +\mathcal{L}_{XX}
\partial^{\mu}\phi \partial_\alpha \phi \partial^\alpha\eta \right)\!=\! \left[\mathcal{L}_{\phi\phi}\!-\!\partial_\mu \!\left(\mathcal{L}_{\phi X}
\partial^\mu \phi\right)\right] \eta$$
We decompose the perturbations in terms of the modes $$\eta (x,t)=\sum_{n}a_{n}\cos (\omega _{n}\,t)\eta _{n}(x) \label{expan}$$where the $a_{n}$ are real coefficients. Eq. ([expan]{}) allow us to rewrite Eq. (\[eigenEq0\]) as [@blmo] $$-\left[ A^{2}\mathcal{L}_{X}\eta _{n}^{\prime }\right] ^{\prime }=[\mathcal{L}_{\phi \phi }+(\mathcal{L}_{\phi X}\phi ^{\prime })^{\prime }+\omega
_{n}^{2}\mathcal{L}_{X}]\eta _{n}, \label{eigenEq}$$and $A$ was defined in Eq. (\[Asq\]). This is a Sturm-Liouville equation, which means that the eigenfunctions $\eta _{n}$ satisfy a condition of orthonormality with weight function $\mathcal{L}_{X}$ [@blmo]: $$\int_{-\infty }^{\infty }dx\,\mathcal{L}_{X}\,\eta _{n}(x)\eta
_{m}(x)dx=\delta _{mn}, \label{etaeta}$$where the eigenfunctions $\eta _{n}$ obey appropriated boundary conditions or $\mathcal{L}_{X}\,$ converges to zero more rapidly than $\,\eta _{n}\left( x\right) $ at boundaries. The finiteness of Eq. (\[etaeta\]) must be analyzed in detail for every model to be studied.
[The eigenvalue problem (\[eigenEq\]) becomes more clear after a convenient change of variables (a procedure described in Ref. [@blm]) ]{}$$dx=Adz,~~\eta _{n}=\frac{u_{n}}{\sqrt{\mathcal{L}_{X}A}},$$[a Schrödinger-like equation can be obtained ]{}$$-\left( u_{n}\right) _{zz}+U_{sch}(z)u_{n}=\omega _{n}^{2}u_{n},
\label{sch_eq}$$[with [@blmo] ]{}$$U_{sch}(z)=\frac{({A\mathcal{L}_{X}})_{zz}^{\frac{1}{2}}}{({A\mathcal{L}_{X}})^{\frac{1}{2}}}-\frac{1}{{\mathcal{L}}_{X}}\biggl[\mathcal{L}_{\phi \phi }+\frac{1}{A}\biggl(\mathcal{L}_{\phi X}\frac{\phi _{z}}{A}\biggr)_{z}\biggr].$$[Note that Eq. (\[sch\_eq\]) is an eigenvalue equation where stable solutions correspond to ]{}$\omega _{n}^{2}\geq 0$[. The existence or not of tachyonic modes (]{}$\omega _{n}^{2}<0$[) depends on the model and must be analyzed separately for each case. The determination of the eigenvalues ]{}$\omega _{n}$ [ and stability analysis for the particular case of the ]{}$\phi ^{4}$ model and its twin counterpart is considered in the next section.
The $\protect\phi^4$ model and its twin counterpart
===================================================
In this section we will consider comparatively some properties of two specific twin models. The set of bound states obtained from the stability analysis will give informations that will be confronted later with the numerical analysis of the collisions. It is important to note that, for the models considered in this work, the absence of tachyonic modes is demonstrated in Sect. II of Ref. [@bdglm]. Also, numerical analysis discussing their eigenmodes can be found in two different ways in Refs. [altw,bdglm]{}.
We start considering a standard Lagrangian density with a $\phi^4$ potential, where $$W_{\phi}=1-\phi^2.$$ The kink-like solution of the first-order equation of motion (Eq. ([eom\_st]{})) is $$\phi_S(x)=\tanh(x-x_0)$$ where $x_0$ is a constant, identified as the center of the kink. Note that, as a consequence of the twin model construction, the former expression also corresponds to the solution $\phi_m(x)$ achieved for the modified k-defect theory. The first step in the investigation the kink-antikink collision process is to investigate the spectra of fluctuations for both models. The fluctuation modes around the kink are described as $\phi(x, t) = \phi_S(x) +
\sum_n a_n \eta_n(x) \cos(\omega_n t)$. For the standard Lagrangian, a Schrödinger-like equation is attained $$-\eta_n^{\prime \prime }(x)+V_q(x)\eta_n(x)=\omega_n^2\eta(x)$$ with the potential $$V_q(x)=W_{\phi\phi}^2+W_{\phi}W_{\phi\phi\phi}=2\big(3\tanh^2(x)-1\big).$$ For the twin k-defect model, a Schrödinger-like equation $$-u_{zz}+U_q(z)u=\omega^2u,$$ is possible after a change of variables [@blm]: $$dx=\biggl(1+\frac{W_{\phi}^2}{M^2}\biggr)^{-\frac12} dz,$$ $$\eta=\biggl(1+\frac{W_{\phi}^2}{M^2}\biggr)^{\frac14} u.$$ Analytic expressions for the potential $U_q(z)$ can be attained in the regime $1/M^2\ll1$. For the particular case $W_\phi=1-\phi^2$ considered, one gets [@bdglm] $$\label{Uqz}
U_q(z) = 4 - 6 \mathrm{sech}^2(z) + \frac1{M^2} [4 \mathrm{sech}^2(z)+14
\mathrm{sech}^4(z)-21 \mathrm{sech}^6(z)].$$ We see that for $1/M^2\to0$ we have identical potentials $V_{(q)}(z)=U_{(q)}(z)$ for the standard and k-defect theories, with the Schrödinger-like potential being a modified Pösh-Teller. This potential, analyzed in Sugyama [@sug] in the context of $K\bar K$ collisions, has two discrete eigenvalues $\omega_i^2$ with corresponding eingenfunctions $u_i(z)$: $$\omega_0^2=0, \,\,\, u_0(z)=\sqrt{\frac34}\mathrm{sech}(z)^2$$ and $$\omega_1^2=3, \,\,\,u_1(z)=\sqrt{\frac32}\tanh(z)\mathrm{sech}(z),$$ followed by a continuum of modes. The first excited state is an excitation trapped in the kink, crucial for the formation of two-bounce states [sug]{}.
Now, from Fig. \[fig\_sch\] we see that the increasing of $1/M^2$ leads to the potentials $V_q(z)$ and $U_q(z)$ to depart one from another, with the Schrödinger-like potential in the k-defect theory with a deeper and thinner potential. Nondegenerate perturbation theory at first-order in the parameter $1/M^2\ll 1$ can be used to attain the energy eigenvalues and eigenfunctions of the twin k-defect theory. Eq. (\[Uqz\]) can be written as $U_q(z)=V_q(z)+U_{pert}(z)/M^2$, where the perturbation potential is $$U_{pert}=4 \mathrm{sech}^2(z)+14 \mathrm{sech}^4(z)-21 \mathrm{sech}^6(z).$$ The ground state energy is invariant under such perturbation: $$\begin{aligned}
(\omega^{(1)}_0)^2 &=&\omega_0^2+\frac1{M^2}\int_{-\infty}^{\infty} dz |u_0(z)|^2 U_{pert}(z) = 0.
\label{omega0_pert}\end{aligned}$$ This means the presence of the translation mode for the corresponding twin model. The energy of the first-excited mode is changed to $$\begin{aligned}
(\omega^{(1)}_1)^2 &=&\omega_1^2+\frac1{M^2}\int_{-\infty}^{\infty} dz |u_1(z)|^2 U_{pert}(z)=3
+\frac85\frac1{M^2} \label{omega1_pert}.\end{aligned}$$
$1/M^2$ $(\omega^{(1)}_1)^2$ $(\omega^{(1)}_1)_{num}^2$ relative error
--------- ---------------------- ---------------------------- ----------------------
0 3 3.0000013 $-4.3\times 10^{-7}$
0.01 3.016 3.0159773 $7.5\times 10^{-6}$
0.1 3.16 3.1574820 $8.0 \times 10^{-4}$
: Energy of the first-excited state. In the second column we compare $(\omega^{(1)}_1)^2$, given by Eq. (\[omega1\_pert\]) obtained with time-independent perturbed theory. The third column shows the energy obtained numerically with a finite-element method. The fourth column shows the relative error between the results from second and third columns.
Table I shows, for some values of $1/M^2$, the evaluated eigenvalue of the internal mode from Eq. (\[omega1\_pert\]) and the corresponding numerical solution of the Schröedinger-like equation. We see from the Table that the relative error increases with $1/M^2$, with an acceptable value $\sim 10^{-5}$ for $1/M^2=0.01$. An increasing of $1/M^2$ by one order of magnitude increases the relative error $100$ times. Then, $1/M^2=0.1$ seems to be a too large value for the approximation to be valid. Also, from Eqs. (\[omega0\_pert\]) and (\[omega1\_pert\]), we see that there is a larger gap between the zero-mode and the first excited state with the increasing of the parameter $1/M^2$. This property is very important for the whole collision process, since it is in the core of the collision that occurs a transference of energy from translational mode to the vibrational one. If the gap increases, it turns more difficult the energy exchange between the two modes, meaning a disfavoring of the occurrence of bion states and n-bounce collisions with $n\ge 2$ and the enlargement of the range of velocities with one-bounce collisions.
Numerical Results
=================
Here we will describe our main results concerning to the number of bounces as a function of the initial kink velocity in a symmetric kink-antikink collision. We will consider the standard $\phi^4$ theory and twin theories with increasing parameters $1/M^2$. Our results will be confronted with the theoretical predictions attained previously in this paper.
$\protect\lambda\protect\phi^4$ theory
--------------------------------------
First of all we review the $\lambda\phi^4$ theory. The equation of motion is $$\ddot\phi-\phi^{\prime \prime }+V_\phi=0$$ where the dots and primes mean derivatives with respect to $t$ and $x$, respectively. We studied a symmetric $K\bar K$ collision, with an initial configuration where the pair $K\bar K$ is sufficiently separated for the free solution to be useful as an initial condition (kink with velocity $v_{in}$, antikink with velocity $-v_{in}$). This means to chose as the initial conditions $$\begin{aligned}
\phi(x,0)&=&\phi_K(x+x_0,v_{in},0)-\phi_K(x-x_0,-v_{in},0)-1 \\
\dot\phi(x,0)&=&\dot\phi_K(x+x_0,v_{in},0)-\dot\phi_K(x-x_0,-v_{in},0).\end{aligned}$$ We used a pseudospectral method on a grid with $2048$ nodes and periodic boundary conditions. We fixed $x_0=15$ as the initial kink position and we set the grid boundaries at $x_{max}=120$.
For the $\phi^4$ model we reproduced some results from the literature [aom]{} concerning to the appearance of two-bounce windows. For a particular initial velocity $v_{in}$, the structure of bounces can be easily verified with a plot of the scalar field at the center of mass $\phi(0,t)$ as a function of $t$. Some examples for three different initial velocities can be seen in Figs. \[phicm\], where we have a bion (\[phicm\]a), two-bounce (\[phicm\]b) and one-bounce (\[phicm\]c) collisions. As one knows, the dependence of the number of bounces with the modulus of the initial velocity $v_{in}$ of the pair kink-antikink is intricate. For low $v_{in}$ a bion state is formed, whereas for high $v_{in}$ one has a one-bounce scattering. For intermediate velocities, there appears two-bounce windows of variable size separated by regions of bion states.
\
We constructed a procedure to identify the number of bounces $N_b$ for a given collision process in a time interval $0<t<T=200s$. We defined a bounce as connected to the change of sign of the center-of-mass solution $\phi(0,t)$. In this way, Fig. \[phicm\]b and \[phicm\]c show collisions with $N_b=2 $ and $N_b=1$, respectively and \[phicm\]a shows a bion state. In our analysis, a too large value of $N_b$ will correspond to a bion state. Fig. \[bouncv\]a shows the behavior of $N_b$ as a function of $v_{in}$. Note from the figure the change of pattern for collisions around $v_{in}\sim
0.26$. The presence of higher number of bounces for $v_{in} \lesssim 0.26$ characterizes bion states, whereas there are intermediate regions with $N_b=2 $. States with $N_b=1$ appear for $v_{in} \gtrsim 0.26$. The structure of the two-bounce windows can be characterized by the integer $m$ labeling the number of oscillations in $\phi(0,t)$ between the bounces. For example, in the Fig. \[bouncv\]a the first two-bounce window corresponds to $m=1$ (see details in Ref. [@aom]). The two-bounce thickness decreases with $m$ accumulating around $v_{in}\simeq 0.26$, the limit above which the initial velocity is already sufficiently large for a one-bounce scattering. Table II shows some characteristics of the thickness of the first four two-bounce windows. This table shows how an increasing in $m$ reflects in the reducing of the two-bounce windows. This can be better seen in Fig. \[phitwin\_vm\] (see Ref. [@bk]).
m $v_1$ $v_2$ $\Delta v$ $v_{1T}$ $v_{2T}$ $\Delta {v_T}$
--- -------- -------- ------------ ---------- ---------- ---------------- --
1 0.1926 0.2034 0.0108 0.1757 0.1872 0.0115
2 0.2241 0.2288 0.0047 0.2095 0.2145 0.005
3 0.2372 0.2396 0.0024 0.2233 0.2259 0.0026
4 0.2440 0.2454 0.0014 0.2305 0.2320 0.0015
: Separation in velocities of the first four two-bounce windows. $m$ is a label corresponding to the particular two-bounce window (equal to the number of oscillations of $\protect\phi(0,t)$ between the two bounces). For the $\protect\phi^4$ model, the columns $v_1$ and $v_2$ correspond to the first and last points of the velocity of the window and $\Delta v$ the corresponding width (the numerical results are in agreement with Table II of [@cp]). For the twin k-defect model with $1/M^2=0.01$, columns to the right show similar informations identified by $v_{1T}$, $v_{2T}$ and $\Delta v_T$.
Also refining the input data of initial velocities around a border between a two-bounce window and bion, one sees the presence of a cascate of three-bounce windows, as can be seen in Fig. \[bouncv\]d. Note that the three-bounce windows accumulate around the border of the two-bounce windows, replicating the effect of reducing of thickness and separation of windows that occurs in Fig. \[bouncv\]a, and showing the well-known fractal pattern of $K\bar K$ collisions for the $\phi^4$ model.
twin theory
-----------
The twin theory in the regime $1/M^2\ll1$ has the corresponding equation of motion $$\label{eom_twin}
\ddot\phi-\phi^{\prime \prime }+\frac1{M^2}(\dot\phi^2\ddot\phi+\phi^{\prime
2}\phi^{\prime \prime }-2\dot\phi\phi^{\prime }\dot\phi^{\prime
})+U_\phi-\frac1{M^2}UU_\phi=0.$$ In the regime $1/M^2\ll1$, and neglecting terms of ${\mathcal{O}}(1/M^4)$ we have $$U(\phi)=\frac12W_\phi^2-\frac18 \frac{W_\phi^4}{M^2}.$$
Substituting this in Eq. (\[eom\_twin\]), one can easily verify that $\phi_K(x,t)=\tanh(\gamma(x-vt))$ is the solution for a free propagating kink. As expected this is the same solution already achieved for the $\phi^4$ model. This means that the initial conditions will be the same used for the $\phi^4$ model. Indeed, it is only the collision process that will distinguish both theories. Also, as used previously for the $\phi^4$ theory, this analytic solution can be used as a first test for the numerical solution for the twin theories. In Figs. \[phitwin\] we present some results for $\phi(0,t)$ for the twin theory with $1/M^2=0.01$. Comparing the figures with Fig. \[phicm\] for the $\phi^4$ theory we see that even for a small value of $1/M^2$ the behavior is altered sensibly. For example, for $v_{in}=0.2$ we have two-bounce collision with $m=1$ for the $\phi^4$ model (Fig. \[phicm\]b) and bion for the twin model (Fig. \[phitwin\]b). Also, for $v_{in}=0.18$ we have bion for the $\phi^4$ model (Fig. \[phicm\]a) and two-bounce collision with $m=1$ for the twin model (Fig. \[phitwin\]a). Our results of $N_b$ as a function of $v_{in}$ for $1/M^2=0.01$ are depicted in Fig. \[bouncv\]b. The same effect of appearance of two-bounce windows already known for the $\phi^4$ model is present for the twin model. In the twin model, however, the one-bounce collision occurs for $v_{in}\gtrsim 0.246$, smaller that $v_{in}\gtrsim 0.26$ for the $\phi^4$ theory. An increasing of $1/M^2$ shows the same pattern, as can be seen in Fig. \[bouncv\]c for $1/M^2=0.05$, where now the one-bounce collisions occur for $v_{in}\simeq 0.20$. This enlargement of the region where occurs the one-bounce window is directly related to the increasing of the gap between the translational and vibrational modes, as discussed in the final of the previous section.
Table II shows the separation in velocities of the first four two-bounce windows for both $\phi^4$ and the twin model for $1/M^2=0.01$. We see that the thickness of the windows are roughly the same, occurring for lower velocities in the twin model in comparison to the $\phi^4$ model. This can be better seen in Fig. \[phitwin\_vm\], where we also included the case where $1/M^2=0.05$. There one can also see that, similarly to the $\phi^4$ model, for the twin model the velocity thickness is reduced continuously with $m$, whereas $v_{in}$ grows and asymptotes to the minimum velocity $v_{in}^*$ for the occurrence of a one-bounce collision. Also note from the figure that the larger is $1/M^2$, the lower is $v_{in}^*$.
For the twin model we refined the input data of initial velocities around a border between a two-bounce window and bion, obtaining the results of Figs. \[bouncv\]e and \[bouncv\]f. There one can see the presence of a cascate of three-bounce windows, accumulating around the border of the two-bounce windows, replicates the effect of reducing of thickness and separation of windows that occurs in the corresponding Figs. \[bouncv\]b and \[bouncv\]c, and showing that there is a fractal pattern of $K\bar K$ collisions for the twin model in a similar way to the verified for $\phi^4$ model (compare with Figs. \[bouncv\]a and \[bouncv\]d).
Fig. \[bouncv\_twin\]a-b show the time $t$ of the first three bounces as a function of the initial velocity $v_{in}$. Our results with $1/M^2=0.01$ (Fig. \[bouncv\_twin\]b) clearly show that there is a displacement of the two-bounce windows for regions of lower velocities in comparison to the $\phi^4$ model (compare with Fig. [bouncv\_twin]{}a). This can be interpreted as a signal of a weaker $K\bar K$ interaction for the twin model, in comparison to the $\phi^4$ model.
We also investigated if the fundamental relation given by Eq. (\[T\_omega\]) holds in the modified twin model. In Fig. \[slope\_twin\] we plot the time between bounces versus $n=m+2$, where $m$ is the window number. The integer $n$ is constructed in order to achieve a phase shift $\delta$ between $0$ and $2\pi$. We note from the figure that the numerical points can be described with good approximation by a straight line. Table III compares the angular coefficients from the least-squares with the theoretical one $2\pi/\omega_1$, predicted by Eq. (\[T\_omega\]). From the Table we see that for $1/M^2=0$ the theoretical angular coefficient $2\pi/(\omega^{(1)}_1)$ is lower but comparable to the numerical value obtained by least-squares method, as already noted in Ref. [@aom]. Also Table III shows that the numerical angular coefficient [*grows*]{} with the increasing of $1/M^2$. This is an intriguing character, since from Eq. (\[T\_omega\]) it is expected a [*decreasing*]{} of the angular coefficient, as shown in the second row of Table III. Since we found no physical ground for such behavior, we must say this numerical result must be handled with care, for the following reasons: i) One must note that Fig. \[slope\_twin\] considers a quite large number of windows (up to $m\sim 35$). As the window number grows, also grows the numerical error of the simulations. To put in other words, a consistent numerical analysis of two-bounce windows is easier from the computational perspective. ii) Eq. (\[T\_omega\]) was proposed by Campbell et all in Ref. as a good simplification of an intrincate process of interaction between two extended objects. If we want to test a qualitative expression, we must not be so strict in quantitative agreement. iii) We cannot consider a too large value of $1/M^2$ as a valid approximation of a twin theory breaks. We could take $1/M^2=0.05$ as a higher bound for an analysis of a two-bounce windows, but this value could be even lower for larger number of bounces. In this paper, all the former numerical results and analysis concerning to two-bounce windows where shown to be compatible with the interpretation of Eq. (\[T\_omega\]). The aplicability of this equation can be confirmed also by Fig. \[slope\_twin\] and Table III in the sense that a straight line behavior was verified and the angular coefficients are roughly comparable, as already done in Ref. [@aom] for the $\phi^4$ model.
\[slope\_twin\]
$1/M^2$ $2\pi/(\omega^{(1)}_1)$ $a$ relative error
--------- ------------------------- ------ ----------------
0 3.63 3.65 $0.55 \% $
0.01 3.62 3.66 $1.1 \% $
0.05 3.58 3.69 $3.0 \% $
: Theoretical angular coefficient $2\pi/(\omega^{(1)}_1)$ versus angular coefficient $a$ obtained by least-squares method from Fig. \[slope\_twin\].
conclusions
===========
In this work we have studied kink-antikink collisions for twin theories. We were particularly interested in investigating in which aspect the presence of a general kinetic content (k-generalization) in the Lagrangian could be revealed in a collision process. Starting form a general Lagrangian $\mathcal{L}(X,\phi)$, and considering a convenient decomposition of the fluctuations, we analyzed the energy contribution of the fluctuations. After reviewing the first-order formalism of twin theories, we considered the $\phi^4$ theory and a class of twin theories depending on a mass parameter $M$. In the regime where $1/M^2\ll 1$, we obtained the spectra of excitations with two bound states: a zero-mode, responsible for the translation, and a vibrational one, crucial for the two-bounce collisions. We showed that the gap between the two bound states is larger for the twin model, and that it increases with $1/M^2$. A detailed numerical analysis reproduced some known results for $K\bar K$ collisions in the $\phi^4$ theory, used as a control model to be confronted with the results of a twin model of a general kinetic content. The numerical results corroborated the theoretical expectation that, in a collision process, the increasing of $1/M^2$ reduces the possibility of formation of a trapped $K\bar K$ bion state.
Acknowledgements
================
The authors thank FAPEMA, CAPES, CNPq and IFMA for financial support. The authors thank Herbert Weigel and R. Casana for clarifying several points of stability analysis. A. R. Gomes thanks A. S. Anjos and M. M. Ferreira Jr. for discussions.
[99]{} B. A. Malomed, Physica D: Nonlinear Phenomena 15, 385 (1985).
D. W. McLaughlin and A. C. Scott, Phys. Rev. A 18, 1652 (1978).
M. A. Amin, Phys. Rev. D87, 123505 (2013)
M. A. Amin, E. A. Lim, I. Sheng Yang, Phys.Rev.Lett. 111 (2013) 224101
M. A. Amin, E. A. Lim, I. Sheng Yang, Phys.Rev. D88 (2013) 105024.
P. Anninos, S. Oliveira, R. A. Matzner, Phys. Rev. D **44**, 1147 (1991).
D. K. Campbell, J. F. Shonfeld, C. A. Wingate, Physica D 9, 1 (1983).
M. Peyrard, D.K. Campbell, Physica D 9, 33 (1983).
T. Belova and A. Kudryavtsev, Physica D 32, 18 (1988).
A.E. Kudryavtsev, JETP Lett. 22, 82 (1975).
T. Sugiyama, Prog. Theor. Phys. 61, 1550 (1979).
D. K. Campbell, M. Peyrard, P. Sodano, Physics D 19, 165 (1986).
P. Dorey, K. Mersh, T. Romanczukiewicz, Y. Shnir, Phys.Rev.Lett. 107, 091602 (2011).
M. A. Lohe, Phys. Rev. D 20, 3120 (1979).
H. Weigel, J. Phys. Conf. Series, 482 (2014) 012045.
M. Andrews, M. Lewandowski, M. Trodden, and D. Wesley, Phys. Rev. D 82, 105006 (2010).
D. Bazeia, R. Menezes, Phys.Rev. D 84, 125018 (2011).
D. Bazeia, E. da Hora, R. Menezes, Phys.Rev. D85, 045005 (2012).
D. Bazeia, A.S. Lobão, L. Losano, R. Menezes, Eur. Phys. J. C74 (2014) 2755.
C. Armendariz-Picon, T. Damour, and V. F. Mukhanov, Phys. Lett. B 458, 209 (1999); J. Garriga and V. F. Mukhanov, Phys. Lett. B 458, 219 (1999); T. Chiba, T. Okabe, and M. Yamaguchi, Phys. Rev. D 62, 023511 (2000).
Y. I. Takamizu and K. I. Maeda, Phys. Rev. D 70, 123514 (2004); *ibid* Phys. Rev. D 73, 103508 (2006).
J. Khoury, B.A. Ovrut, P. J. Steinhardt, and N. Turok, Phys. Rev. D 64, 123522 (2001); J. Khoury, B.A. Ovrut, N. Seiberg, P. J. Steinhardt, and N. Turok, ibid. 65, 086007 (2002); J. Khoury, B.A. Ovrut, P. J. Steinhardt, and N. Turok, ibid. 66, 046005 (2002); A. J. Tolley and N. Turok, ibid. 66, 106005 (2002).
D. Bazeia, J. D. Dantas, A. R. Gomes, L. Losano and R. Menezes, Phys.Rev. D84, 045010 (2011).
D. Bazeia, A.S. Lobao, Jr., R. Menezes, Phys.Rev. D 86, 125021 (2012).
R. Bousso and J. Polchinski, J. High Energy Phys. 06, 006 (2000).
S. W. Hawking, I. G. Moss, J. M. Stewart, Phys. Rev. D 26, 2681 (1982).
R. Easther, J. T. Giblin Jr., Lam Hui, E. A. Lim, Phys. Rev. D 80, 123519 (2009).
J. T. Giblin Jr., Lam Hui, E. A. Lim, I. Sheng Yang, Phys. Rev. D 82, 045019 (2010).
E. Babichev, Phys. Rev. D 74, 085004 (2006).
C. Adam, J. Sanchez-Guillen, A. Wereszczynski, J. Phys. A 40, 13625 (2007).
D. Bazeia, L. Losano, R. Menezes and J. C. R. E. Oliveira, Eur. Phys. J. C 51, 953 (2007).
D. Bazeia, L. Losano and R. Menezes, Phys. Lett. B 668, 246 (2008).
D.K. Campbell, M. Peyrard, Physica D 18, 47 (1986).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have derived isotopic fractions of europium, samarium, and neodymium in two metal-poor giants with differing neutron-capture nucleosynthetic histories. These isotopic fractions were measured from new high resolution ($R\sim120,000$), high signal-to-noise (S/N$\,\sim\,$160–1000) spectra obtained with the 2dCoudé spectrograph of McDonald Observatory’s 2.7m Smith telescope. Synthetic spectra were generated using recent high-precision laboratory measurements of hyperfine and isotopic subcomponents of several transitions of these elements and matched quantitatively to the observed spectra. We interpret our isotopic fractions by the nucleosynthesis predictions of the stellar model, which reproduces nucleosynthesis from the physical conditions expected in low-mass, thermally-pulsing stars on the AGB, and the classical method, which approximates nucleosynthesis by a steady neutron flux impinging upon Fe-peak seed nuclei. Our Eu isotopic fraction in HD 175305 is consistent with an origin by the classical method and is consistent with either an $r$- or an origin by the stellar model. Our Sm isotopic fraction in HD 175305 suggests a predominantly origin, and our Sm isotopic fraction in HD 196944 is consistent with an origin. The Nd isotopic fractions, while consistent with either or origins, have very little ability to distinguish between *any* physical values for the isotopic fraction in either star. This study for the first time extends the origin of multiple rare earths in metal-poor stars from elemental abundances to the isotopic level, strengthening the interpretation for HD 175305 and the interpretation for HD 196944.'
author:
- 'Ian U. Roederer, James E. Lawler, Christopher Sneden, John J. Cowan, Jennifer S. Sobeck, Catherine A. Pilachowski'
title: 'Europium, Samarium, and Neodymium Isotopic Fractions in Metal-Poor Stars'
---
Introduction
============
A wealth of observational studies over the last decade have focused on measuring precise abundances of neutron ($n$)-capture elements in metal-poor stars. The heart of this enterprise is identifying the origin of the nuclides with $Z\,>\,30$ in our Galaxy. These nuclides are produced in stars through either the rapid ($r$)-process or the slow ($s$)-process. Rapid-process nucleosynthesis occurs in environments with high neutron fluxes and densities, allowing many neutrons to be captured by existing nuclei much more rapidly than the timescales for $\beta$-decay. Slow-process nucleosynthesis occurs where the neutron densities are lower, with $\beta$-decay timescales generally shorter than the times between events. To build heavy nuclei through the , a chain of stable or long-lived nuclei must exist from the existing seed nucleus to the final product. The and lead to the creation of different sets of heavy nuclei, some of which can only be created in one process or the other (“pure-$s$-” or “pure-$r$-nuclei”) and some of which are created by both processes [see, e.g., @cowan06b]. While the precise astrophysical site(s) of the have yet to be identified, nucleosynthesis takes place in low- and intermediate-mass stars on the asymptotic giant branch (AGB) [e.g., @busso99].
To illustrate the effects of both processes, in Figure \[nuclides\] we show a table of nuclides covering portions of the atomic number domain 60 $\leq$ $Z$ $\leq$ 63. The can add only one neutron before a $\beta$-decay occurs for an unstable nucleus. On the other hand, the quickly floods a target nucleus with many neutrons, overwhelming the $\beta$-decay rates. When neutrons are no longer being added, the nucleus will $\beta$-decay repeatedly until it reaches stability. For isotopes produced primarily in the (e.g., [$^{\rm 146}$Nd]{}, [$^{\rm 148,150}$Sm]{}) a stable isotope with one less nuclide always exists. For isotopes produced primarily or wholly by the (e.g., [$^{\rm 150}$Nd]{}, [$^{\rm 147,149,152,154}$Sm]{}), there is a clear path for $\beta$-decay from unstable nuclei to the lower right (in the table of nuclides) of the stable isotope, sometimes unreachable by the .
The nucleosynthetic signatures of the $s$- and $r$-processes stand in sharp contrast to one-another. The Solar System (S. S.) isotopic abundances are well-established from studies of CI meteorites, as summarized in, e.g., @anders89 and @lodders03. Two primary methods are used to decompose the S. S. isotopic abundances into their constituent $s$- and origins. In the “standard” or “classical” method, first articulated by @clayton61 and @seeger65 and with abundances later updated by, e.g., @kappeler89 [@burris00; @simmerer04], and @cowan06a, the is modeled as a slowly-varying function of neutron exposure. This method attempts to smoothly fit the “$\sigma N_{s}$” curve (the product of the cross-section and the S. S. $s$-process abundances) using lab measurements of $\sigma$ to determine $N_{s}$. In the “stellar” model of @arlandini99, isotopic abundances for nucleosynthesis in 1.5$-$3.0$M_{\sun}$ stars are predicted based upon nuclear physics cross sections and stellar model calculations for thermally-pulsing (TP) stars on the AGB. In both of these cases, the abundances are then derived as residuals between the total S. S. abundances and the $s$-process products. Both approaches have strengths and weaknesses. The classical method is model-independent, but it assumes an empirical, smoothly-varying relationship for the $\sigma N_{s}$ abundance curve (which relies on the measured S. S. abundances) and ignores details of the nuclear physics. The stellar model does not depend on knowledge of the S. S. abundance distribution, but is heavily dependent on nuclear physics laboratory measurements and complex stellar AGB model calculations.
These models can also be used to predict the relative amounts of $s$- or nucleosynthesis products in metal-poor stars. Several key rare earth elements are commonly used to assess $r$-/ abundance dominance in stars.[^1] Eu is commonly used as a surrogate for nucleosynthesis because it is so predominantly composed of material and has several easily-observed spectral lines. Similarly, Ba or La are commonly used as surrogates of nucleosynthesis. Thus, many studies have employed \[Ba/Eu\][^2] ratios in metal-poor stars to estimate the relative contributions from the $s$- and $r$-processes [e.g., @spite78; @gratton94; @burris00; @mashonkina03; @honda04b; @barklem05; @francois07], which in turn constrain models of chemical evolution in the early Galaxy [@travaglio99; @travaglio01; @cescutti06; @kratz07]. Recently, \[La/Eu\] ratios have been employed as an alternative to \[Ba/Eu\] to avoid difficulties in Ba abundance determinations [see, e.g., @simmerer04; @jonsell06; @winckler06].
Elemental abundances, the sums of abundances of their naturally-occurring constituent isotopes, are more easily measured in stars than the isotopic abundances; however, isotopic abundances should be more fundamental indicators of nucleosynthesis because they can directly confront and predictions without the smearing effect of multiple isotopes. Much recent evidence (e.g., the growing number of $r$-process-enhanced stars that conform to the scaled-S. S. elemental abundance distribution) supports the hypothesis of a universal mechanism for elements with $Z\geq56$. It is important to demonstrate that the elemental abundance pattern extends to the isotopic level. Such agreement would greatly strengthen the argument for a universal mechanism for the heavy elements.
Most isotope fractions, unlike elemental abundances, are very insensitive to the model atmosphere parameters. The wavelength of a spectral line is split by two effects, hyperfine structure (HFS) and isotope shifts, and the isotopic fractions can be measured by detailed comparisons of an observed absorption line profile to synthetic spectra of these line substructures. If this splitting is comparable to or greater than the combined effects of stellar thermal, macroturbulent, and rotational broadening and the broadening caused by the spectrograph slit, then it may be possible to measure the isotopic fractions. In principle, unlike with elemental abundances, any element with multiple naturally-occurring isotopes that are produced in different amounts by the $s$- and $r$-processes can be used to assess the relative $s$- and contributions to the stellar composition. In practice, the availability of quality atomic data has limited studies of isotopic fractions of rare earth elements in stars, but the relatively recent increase in laboratory spectral line substructure studies has led to the identification of a few lines of a few elements that might be analyzed at the isotopic level.
@cowley89 and @magain93a first suggested that the width of the Ba <span style="font-variant:small-caps;">ii</span> 4554Å line in stellar spectra would be sensitive to the isotopic fraction. @magain93a, @magain95, and @lambert02 all examined the Ba isotopic fraction in the metal-poor subgiant HD 140283. Their measured isotopic fractions had uncertainties too large to unambiguously discern the Ba nucleosynthetic origin in this star. A recent series of papers by @mashonkina99 [@mashonkina03] and @mashonkina06 have measured Ba isotopic fractions in thin and thick disk stars. Their approach is to set the Ba elemental abundance from lines not affected by HFS or isotope shifts and then to derive the isotopic fraction from the 4554Å line, whose shape is sensitive to both its elemental abundance and isotopic composition of Ba. @mashonkina06 clearly show that a higher fraction of Ba isotopes is present in thick disk stars than in S. S.material and thin disk stars, and they also show that a smooth trend of increasing contributions occurs with increasing metallicity. Since only the 4554Å line is useful for examining the Ba isotopic mix, and the profile of this line is known to suffer from non-LTE effects in metal-poor stars [e.g., @mashonkina99; @short06], we do not attempt to assess the Ba isotopic fractions in this study. The Mashonkina et al. studies of the Ba isotopic fraction do, however, employ non-LTE calculations.
Eu isotopic fractions in metal-poor stars were first reported by @sneden02, who found that the isotopic ratio in three -rich giants was consistent with the S. S. isotopic ratio. Subsequent studies by @aoki03a [@aoki03b] found that Eu isotopic ratios could marginally distinguish $s$- and contributions in stars that had previously been shown to exhibit such chemical signatures. @lundqvist07 have investigated the Sm isotopic fractions in a very metal-poor -enriched star, qualitatively showing that the isotopic mix is in agreement with an origin. No measurements of the Nd isotopic fraction have been made outside of the S. S., and no study has yet attempted to measure isotopic fractions of multiple elements in the same star.
In our study, we analyze the isotopic fractions of Eu, Sm, and Nd in two metal-poor giant stars. Our observations and methods of analysis are described in § \[observations\] and \[analysis\], respectively. We summarize the available atomic data for each species in § \[atomicdata\]. Our measurements of the elemental abundances and isotopic fractions are presented in § \[elabund\] and § \[isofrac\]. In § \[interpret\] we discuss the implications of our measurements in the context of using isotopic fractions of multiple species to constrain the $s$- and nucleosynthetic histories. In § \[eupb\] we use Eu and Pb abundances collected from the literature to assess the model predictions for the $^{151}$Eu isotope, since pure-$r$- and pure-$s$-process nucleosynthesis reactions produce such contrasting abundances of these elements. We also make suggestions for how future studies might utilize measurements of isotopic fractions to make similar comparisons for other isotopes in § \[future\]. We summarize our results in § \[conclusions\]. In the appendices, we include comments describing the nuances associated with analyzing each line of each species (Appendix \[comments\]), and we include our computations of the hyperfine structure subcomponents for both Sm (Appendix \[smappendix\]) and Nd (Appendix \[ndappendix\]).
Observations
============
Selection of Target Stars {#targets}
-------------------------
We choose two bright, metal-poor giants for our analysis, HD 175305 and HD 196944. The technical requirements of achieving high signal-to-noise (S/N) ratios at very high resolution on a telescope with a moderate aperture mandate that we select bright ($V\,\lesssim\,$8.5) targets. We select metal-poor giant stars for this analysis because (1) metal-poor stars present significantly less crowding of strong absorption features in the blue spectral region where the lines of interest are concentrated, (2) absorption lines of all species will exhibit greater line depth in giants than in dwarfs because of the decrease in the continuous opacity with decreasing electron pressure in giants, and (3) at a given metallicity, molecular formation will be less in giants than in dwarfs. To maximize the expected contrast in the isotopic fractions, we have deliberately selected stars with different chemical enrichment histories. Although these stars have been well-studied, no previous attempts have been made to examine their isotopic fractions.
HD 175305 is a metal-poor (\[Fe/H\] $=-1.6$), bright ($V$ $\sim\,7.2$) giant that was selected because it shows evidence of $r$-process enrichment (\[Eu/Fe\] $\sim\,+0.5$, \[Ba/Eu\] $<\,0.0$, \[La/Eu\] $\sim\,-0.4$; @burris00 [@fulbright00; @cowan05]). HD 175305 is not enriched in carbon (\[C/Fe\] $=+0.2$; @wallerstein79). Although only a handful of heavy $n$-capture abundances have been measured previously, the high \[Eu/Fe\] and low \[Ba/Eu\] ratios are suggestive of an origin. Combining @cowan05’s abundance measurements of osmium (Os, $Z\,=\,76$), iridium (Ir, $Z\,=\,77$), and platinum (Pt, $Z\,=\,78$) from UV spectra and @lawler07’s measurement of hafnium (Hf, $Z\,=\,72$) with additional heavy elemental abundances measured in our spectra will produce a more thorough understanding of this star’s enrichment history to complement our isotopic fraction measurements.
HD 196944 is a very metal-poor (\[Fe/H\] $=-2.4$), bright ($V$ $\sim\,8.4$) giant with clear signatures. This star has high C, N, and O abundances (\[C/Fe\] $\sim\,+1.3$, \[N/Fe\] $\sim\,+1.3$, \[O/Fe\] $\sim\,+1.4$; [@zacs98; @aoki02; @jonsell05]), high Ba and La abundances (\[Ba/Fe\] $\sim\,+1.1$, \[La/Fe\] $\sim\,+0.9$), a low Eu abundance (\[Eu/Fe\] $\sim\,+0.2$), and a large over-abundance of lead (\[Pb/Fe\] $\sim\,+2.0$; [@vaneck01; @aoki02; @vaneck03]). HD 196944 also exhibits radial velocity variations [@lucatello05], indicating that it is a member of a binary or multiple star system. These characteristics have led to the classification of HD 196944 as a carbon-enhanced metal-poor $s$-enriched (CEMP-$s$) star, as defined in @beers05. @gallino05 have shown that the abundance pattern of HD 196944 can be reproduced by assuming mass transfer of material enriched in elements from a companion star with an initial mass of 1.5$M_{\sun}$ that passed through the TP-AGB phase.
We also observed HD 122563 in our study; at $V$ $\sim\,6.2$, this star is the brightest very metal-poor (\[Fe/H\] $=-2.8$) star in the sky. Unfortunately the lines of Eu, Sm, and Nd proved to be too weak to provide any assessment of the isotopic fractions. This is unfortunate because @honda07, following studies of this star by @cowan05, @aoki05, and @honda06, have suggested that HD 122563 and HD 88609 may be the first two known representatives of a class of stars whose enrichment has been dominated by the weak , which produces a distinctly different abundance pattern than the main .[^3] Our analysis of HD 122563 did allow us to assess the location and strength of blending features for some of our lines of interest. It is possible that the Eu <span style="font-variant:small-caps;">ii</span> lines at 3819, 3907, and 4129Å, which were not covered in our spectrum, may give some hint to the Eu isotopic fraction.
Description of Observations {#moreobs}
---------------------------
For a set of lines of a given species with the same ionization state in a stellar atmosphere, the relative strength of each line depends on its transition probability, excitation potential, wavelength, and the continuous opacity of the atmosphere at that wavelength [see, e.g., @gray05 p. 389]. In metal-poor stars the continuous opacity (from H$^{-}$) increases slowly over the visible wavelength region as the wavelength increases; these two terms take opposite signs in calculations of the line strength and roughly offset each other. To first approximation then, the relative strengths of individual absorption lines are determined by the transition properties of the lines themselves. Figure 10 of @lawler06 illustrates this point for the case of Sm <span style="font-variant:small-caps;">ii</span> lines. For the species considered in our study, the first-ionized state is dominant in our stars, and most of the transitions have similar excitation potentials ($\sim$ 0.0–0.7eV), so the oscillator strengths most significantly affect the relative line strength. The log($gf$) values generally decrease when moving from the near-UV to the near-IR spectral regions, so the strongest lines in stellar spectra tend to lie in the near-UV and blue end of the visible spectrum. This is unfortunately where the most severe blending with other spectral features is likely to occur. The isotope shifts increase with increasing wavelength; this broadens the stellar line profiles, making it easier to measure the isotopic fraction. The decrease in line strength, decrease in blending features, and the increase in isotopic shifts at longer wavelengths compete against each other, producing a range in the visible spectrum from $\sim$ 4400$-$5100Å where isotopic fractions are most easily measured.
We acquired new observations of our target stars using the 2dCoudé cross-dispersed echelle spectrograph [@tull95] on the 2.7 m Harlan J. Smith Telescope at the W. J. McDonald Observatory. Observations of HD 122563 and HD 196944 were made on 2006 June 12–18, and observations of HD 175305 were made on 2006 August 08–12. Sets of 30-minute exposures were taken with a Tektronix $2048\times2048$ CCD (“TK3”). Around 5000Å, to sample the full width of a line profile ($\sim$0.2Å) with at least 10 points requires $\Delta\lambda\,\approx\,$0.02 per pixel and a two-pixel resolution of $R\,\sim\,$125,000. This requirement dictates the spectral resolution necessary to complete our analysis. Our data have FWHM resolving powers $R\,\sim\,120,000$. Lines in our Th–Ar comparison spectra can be fit with a pure Gaussian function, and no variation in the instrumental profile was found from night-to-night over the course of either observing run. Due to the limited wavelength coverage at this setup, we have observed each star with three grating settings to increase the wavelength coverage. Our spectra cover $4200 \leq \lambda \leq 6640$Åin HD 122563 and HD 196944 and $4120 \leq \lambda \leq 5890$Å in HD 175305; however, even when using multiple grating settings our wavelength coverage is still incomplete within these ranges.
Image processing, order extraction, and wavelength calibration were performed using standard IRAF[^4] tasks in the *echelle* and *onedspec* packages. Echelle orders were cross-correlated and co-added, and the continuum normalization was performed with the SPECTRE software package [@fitzpatrick87]. Our final S/N values (per pixel) range from $\sim 160$ at 4130Å to $\sim 330$ at 5100Å in HD 175305, $\sim 160$ at 4200Å to $\sim 400$ at 5000Å in HD 196944, and $\sim 430$ at 4200Å to $\sim 1000$ at 5000Å in HD 122563.
Atomic Data {#atomicdata}
===========
HFS Structure and Isotope Shifts {#hfsis}
--------------------------------
HFS is primarily caused by the magnetic interaction between an electron and the nucleus of an atom. This interaction occurs when the nucleus has a magnetic dipole moment that results from a non-zero angular momentum, which occurs when the nucleus has an odd number of protons and/or an odd number of neutrons. The strength of this interaction is characterized by the magnetic dipole “A” constant, which can be measured in laboratory studies. We also include the measured electric quadrupole interaction “B” constant in our HFS computations, but magnetic interaction is usually dominant. In stellar spectra, HFS is manifest as a broadening of a single line of a particular atom due to the energy shifts. Isotope shifts, on the other hand, are slight rearrangements of the energy levels of one isotope of a particular species relative to another isotope of the same species. The most useful lines of heavy atoms—such as those considered in this study—usually have isotope shifts dominated by the field shift. This shift is from the finite volume occupied by the nucleus of an atom. Proton charge is distributed within the nucleus, and not concentrated at the origin, hence the electric field inside the nucleus does not have a $1/r$ dependence. For electrons that have a significant probability of existing at small $r$, such as the 6$s$ electron in the ground state configuration of Eu <span style="font-variant:small-caps;">ii</span>, this can have a sizable effect. Both the reduced mass shifts and specific mass shifts of the upper and lower levels also contribute to the total isotope shifts of a line. Isotope shifts for our elements have been measured in laboratory experiments. In stellar spectra, the effect of isotope shifts is to broaden the absorption line profile when more than one isotope of the same atom are present in the stellar atmosphere.
Europium Atomic Data {#euatomic}
--------------------
There are two naturally-occurring isotopes of Eu: $^{151}$Eu (47.8% S. S.) and $^{153}$Eu (52.2%). Both Eu isotopes exhibit HFS. Both isotopes of Eu are produced in roughly equal amounts in the , which dominates over the as the origin of the S. S. Eu. Because there are only two Eu isotopes, it is straightforward to measure the isotopic fraction by measuring the fraction of one of these isotopes relative to the total amount of Eu. The fraction $f_{151}$ is defined as $$f_{151}\,=\,\frac{N(^{151}\rm Eu)}{N( \rm Eu)}.$$ In Table \[tab2\], we list the fraction of $^{151}$Eu that would be expected if all the Eu present in a given star were to have originated in only the or the for both the stellar model and classical method. The stellar model predicts a $f^{s}_{151}$ abundance that is consistent with two $(r+s)$-enriched stars [@aoki03b; @aoki06], but the classical method predicts a $f^{s}_{151}$ abundance that is nearly 30 times smaller. We also note that even the predictions for the pure- abundance of $^{151}$Eu vary greatly within different sets of classical method calculations [@arlandini99; @cowan06a], so this discrepancy is not limited to only the classical method versus stellar model predictions. Lengthy discussion of this discrepancy will be taken up further in § \[eupb\].
We adopt the Eu <span style="font-variant:small-caps;">ii</span> log($gf$) values of @lawler01b. The hyperfine and isotopic components are taken from @ivans06. The HFS of the three Eu <span style="font-variant:small-caps;">ii</span> lines that we consider is very wide, $\sim\,$0.2$-$0.3Å, which facilitates our ability to measure the isotopic fraction.
Samarium Atomic Data {#smatomic}
--------------------
There are seven naturally-occurring isotopes of Sm: $^{144}$Sm (3.1%S. S.), $^{147}$Sm (15.0%), $^{148}$Sm (11.2%), $^{149}$Sm (13.8%), $^{150}$Sm (7.4%), $^{152}$Sm (26.7%), and $^{154}$Sm (22.8%). The $^{144}$Sm isotope is exclusively produced by the $p$-process; fortunately for this analysis, this isotope only comprises 3.1% of the Sm in the S. S., so we can neglect its contribution to the isotopic mix. According to the stellar model and classical method, $^{148}$Sm and $^{150}$Sm are produced exclusively by the , while the dominates the production of the other four isotopes.
The two odd-$A$ isotopes exhibit HFS structure in addition to the energy shifts between all seven isotopes. The Sm isotope shifts are $\sim\,0.05-0.10$Å, which is slightly larger than the Sm HFS splittings, $\sim\,0.05$Å, yet both the Sm isotope shifts and HFS splittings are significantly smaller than the Eu HFS splittings. The two heaviest isotopes, $^{152}$Sm, and $^{154}$Sm, have large isotope shifts relative to the isotope shifts of the lighter five isotopes in many of the transitions we analyze here (cf. Figures 1 and 2 of @lundqvist07). Therefore, it makes sense to define the isotopic fraction in terms of these two heaviest isotopes. The quantity $f_{152+154}$ is defined by $$f_{152+154}\,=\,\frac{N(^{152} \rm Sm)+\it N \rm(^{154}{Sm})}{N(\rm Sm)}.$$ In Table \[tab2\], we also list the isotopic fraction of Sm that would be expected if all the Sm present in a particular star were to have originated in only the or the for both the stellar model and classical method. Note that this definition more easily distinguishes pure-$s$ and pure-$r$ isotopes of Sm than the $f_{\rm odd}$ definition suggested in § 7 of @lawler06. In addition, the pure-$s$ and pure-$r$ isotope fractions predicted by the stellar model and classical method do not differ significantly when expressed in terms of $f_{152+154}$.
We have computed the HFS structure patterns for 13 Sm lines in this wavelength range using the hyperfine A and B constants and isotope shifts of @masterman03, when available, supplemented with additional values from @lundqvist07 and others as described in Appendix \[smappendix\]. Three of these lines (4719, 5052, and 5103) do not have complete sets of hyperfine A and B constants reported in the literature. To gauge the impact of the HFS splitting on our stellar syntheses, we have compared two syntheses of lines with full sets of hyperfine A and B constants, one including the HFS splitting and isotope shifts and one including only the isotope shifts. Syntheses with a lower fraction of $f_{152+154}$ are slightly more broadened in the case of full HFS treatment than when only the isotope shifts are considered. The differences are small; the synthesis including HFS is $\approx$5$-$6% wider at FWHM than the synthesis with no HFS (when the percentage of isotopes exhibiting HFS is maximized in the synthesis). Nevertheless we recommend that lines lacking complete sets of hyperfine A and B constants should not be used to measure the isotopic fraction by fitting the shape of the line profile; however, they should still be useful for other methods of measuring the isotopic fraction and for elemental abundance analyses and are included in the discussion here for this reason. We adopt the Sm <span style="font-variant:small-caps;">ii</span> log($gf$) values of @lawler06.
Neodymium Atomic Data {#ndatomic}
---------------------
Nd, like Sm, has seven naturally-occurring isotopes in S. S. material: $^{142}$Nd (27.2%S. S.), $^{143}$Nd (12.2%), $^{144}$Nd (23.8%), $^{145}$Nd (8.3%), $^{146}$Nd (17.2%), $^{148}$Nd (5.7%), and $^{150}$Nd (5.6%). The lightest isotope is only produced in the , the two heaviest isotopes are almost exclusively produced in the , and the other four isotopes are produced in a combination of the two processes. The two odd-$A$ isotopes exhibit HFS structure, like Sm. @rosner05 have measured the isotope shifts and HFS A and B constants for Nd transitions between 4175 and 4650Å. For the lines examined in our study, the Nd <span style="font-variant:small-caps;">ii</span> HFS is comparable in scale to the Sm <span style="font-variant:small-caps;">ii</span> HFS, $\lesssim\,0.05$Å, while the Nd <span style="font-variant:small-caps;">ii</span> isotope shifts are relatively small, usually only $\lesssim\,0.025$Å. @denhartog03 measured log($gf$) values for many of these transitions and noted that several of the yellow-red lines (mostly redward of 5100Å) exhibited a doublet or triplet structure in their FTS profiles. We collect the Nd isotopes into fractions that might provide a good indication of pure-$s$- and pure-$r$-nucleosynthesis while making sensible choices about which sets of isotopes are shifted the most from one another. We adopt the Nd isotopic fraction defined by $$f_{142+144}\,=\,\frac{N(^{142}\rm Nd)+ \it N \rm (^{144}Nd)}{N(\rm Nd)}.$$ Predictions for this isotopic fraction are displayed in Table \[tab2\]. Note that the pure-$s$ and pure-$r$ predictions of the stellar model and classical method generally agree well for $f_{142+144}$. In Appendix \[ndappendix\] we present our calculations of the HFS structure patterns for 6 Nd lines in the wavelength region covered by our spectra.
Analysis
========
Radial Velocities {#radvel}
-----------------
To measure the radial velocities, we find the centroid of $\sim$17$-$19 Fe <span style="font-variant:small-caps;">i</span> lines and calibrate to the wavelengths in @learner88[^5], which are accurate to better than 0.3mÅ. We find our wavelength scale zeropoint from a small number of telluric lines in the reddest orders of our spectra. Heliocentric velocities for each observation are calculated using the IRAF *rvcorrect* task. For HD 175305 and HD 122563, we measure heliocentric radial velocities of $-184.2\,\pm\,0.3$and $-26.0\,\pm\,0.3$, respectively, which are both consistent with measurements made by previous studies to suggest no radial velocity variation (HD 175305: @carney03 [@nordstrom04]; HD 122563: @bond80 [@gratton94; @barbuy03; @honda04a; @aoki05; @aoki07]). For HD 196944, we measure a heliocentric radial velocity of $-166.4\,\pm\,0.3$. The scatter in our radial velocity measurements of this star from individual observations over the very small date range 2006 June 12–18 is consistent with a single value. @lucatello05 compiled recent radial velocity measurements for this star (ranging from $-174.76\,\pm\,0.36$, @aoki02 to $-168.49\,\pm\,0.11$, @lucatello05 [see additional references therein]) and concluded that it does exhibit radial velocity variations, consistent with their assertion that all carbon-enhanced, $s$-process-rich, very metal-poor stars are members of binary (or multiple) star systems. Our measurement supports the inclusion of HD 196944 in this category of stars.
Equivalent Width Measurements {#ewmeasure}
-----------------------------
Our high-resolution and high-S/N observations of these stars are superior to spectra analyzed in previous studies, so we re-derive the atmospheric parameters for these stars. We combine the linelists and adopt log($gf$) values from @fulbright00, @honda04a, and @ivans06. Equivalent widths of isolated lines are measured in IRAF by fitting Gaussian profiles or by direct integration. We later discard the equivalent width measurements of lines that are approaching saturation, as indicated by their position on an empirical curve-of-growth. In each of our stars, equivalent widths are measured for $\sim$40$-$50 Fe <span style="font-variant:small-caps;">i</span> lines, $\sim$8 Fe <span style="font-variant:small-caps;">ii</span> lines, $\sim$10 Ti <span style="font-variant:small-caps;">i</span> lines, and $\sim$7 Ti <span style="font-variant:small-caps;">ii</span> lines. A comparison of our equivalent width measurements with previous high-resolution studies of these reveals that there are no systematic differences in the measured equivalent widths between our study and their studies. We find mean offsets of only $\Delta\,$EW$\,\equiv$(previous study)$-$(our study)$\,=\,+0.5\,\pm\,1.2$mÅ for HD 122563 [@honda04a], $\Delta\,$EW$\,=\,-1.0\,\pm\,2.2$mÅ for HD 175305 [@fulbright00], and $\Delta\,$EW$\,=\,-0.8\,\pm\,2.3$mÅ for HD 196944 [@zacs98].
Atmospheric Parameters {#atmosphere}
----------------------
We use @kurucz93 model atmospheres without convective overshooting for HD 175305 and HD 196944 and with overshooting for HD 122563 because its metallicity is beyond the range of atmospheres provided without overshooting. Interpolation software for the Kurucz grids has been kindly provided by A. McWilliam and I. Ivans (2003, private communication).
We have obtained $V-J$, $V-H$, and $V-K$ colors for our stars from the SIMBAD[^6] and 2MASS[^7] [^8] [@skrutskie06] databases. We adopt zero reddening values for these stars as recommended by previous studies (HD 122563: @honda04b; HD 175305: @wallerstein79; HD 196944: @zacs98, @jonsell05). We derive effective temperatures for these stars using the empirical color-$T_{\rm eff}$ calibrations given by @ramirez05b, using interpolation software kindly provided by I. Ramírez[^9]. Their work recalibrated and extended the range of color and metallicity applicability of earlier work by @alonso96b [@alonso99b], who employed the infrared flux method to determine effective temperatures of F, G, and K dwarfs and giants. There is a rather small amount of scatter among giants in the temperature and metallicity ranges we are using (see @ramirez05b’s Figure 4). The uncertainties intrinsic to the scatter of these calibrations and in the scatter of the temperatures predicted by each individual color lead us to adopt $\pm\,100$K as the uncertainty in our effective temperatures.
To determine the remainder of the model atmosphere parameters we use the most recent version of the LTE spectrum analysis code MOOG [@sneden73]. We allow for slight ($\sim$20K) modifications to the photometric temperatures to ensure that the abundances derived from lines with both high and low excitation potentials agree in our final atmospheric model. We derive the surface gravity by requiring that the abundances derived from neutral and ionized lines of Fe and Ti agree; we adopt $\pm\,0.3$ as the uncertainty in log($g$). The microturbulence is measured by requiring that the abundances derived from strong and weak lines of each atomic species agree; we adopt $\pm\,0.3$as the uncertainty in our microturbulence. These parameters are varied iteratively until we arrive at our final values.
We employ synthetic spectra to measure the macroturbulent broadening in our stars by fitting the line profiles of $\sim$5$-$10 clean Fe <span style="font-variant:small-caps;">i</span> and Fe <span style="font-variant:small-caps;">ii</span> lines in each star. Uncertainties of 0.3$-$0.4 were found for the macroturbulent velocities. Our line profiles are well-fit by the convolution of the Gaussian (instrumental) and macroturbulent (stellar) broadening terms.
Our metallicity is defined to be the derived Fe <span style="font-variant:small-caps;">i</span> abundance, with an adopted uncertainty of $\pm\,0.15$ dex. This uncertainty represents the systematic uncertainties associated with determining $T_{\rm eff}$ ($\sim 0.09$–0.12 dex) and the intrinsic line-to-line scatter ($\sim 0.10$ dex). Each of these stars are known to be $\alpha$-enhanced (HD 175305: \[$\alpha$/Fe\] $=+0.3$, @fulbright00; HD 196944: \[$\alpha$/Fe\] $=+0.35$, @zacs98; HD 122563: \[$\alpha$/Fe\] $=+0.4$, @honda04b). We have increased the overall metal abundance in our model atmosphere, \[M/H\], by 0.2 dex relative to the Fe <span style="font-variant:small-caps;">i</span> abundance to account for the extra electron-donating $\alpha$-elements [see, e.g. @brown91; @sneden94][^10]. For elemental abundance ratios (e.g., \[X/Fe\]) measured in our study, we reference neutral species to the Fe <span style="font-variant:small-caps;">i</span> abundance and ionized species to the Fe <span style="font-variant:small-caps;">ii</span> abundance.
Our final model atmosphere parameters are listed in Table \[tab1\]. For comparison, we show the atmospheres derived in previous studies of these stars. Table \[tab1\] indicates that our atmospheric parameters are in good agreement with these studies.
Line Profile Analysis {#lineanalysis}
---------------------
To measure the isotopic fractions, the observed absorption lines are fit by synthetic spectra. Although our line lists are initially generated from the extensive lists of @kurucz95[^11], we employ experimental wavelengths and log($gf$) values for all lines of interest and blended features whenever possible. Our Eu, Sm, and Nd lines are synthesized by accounting for the individual hyperfine and isotopic components; the isotope fraction and elemental abundance are varied to provide a best fit to the observed profiles.
At the spectral resolution employed in this study, some of our unblended Fe <span style="font-variant:small-caps;">i</span> and <span style="font-variant:small-caps;">ii</span> lines exhibit a slight profile asymmetry. This observation is consistent with our assumption that the macroturbulent velocity arises, at least in part, from large-scale convective motions in the stellar photosphere. These asymmetries account for $\approx$ 1$-$2% of the total equivalent width of the line, and the typical bisector amplitude is $\approx$ 200$-$300ms$^{-1}$. The shapes of our line bisectors are consistent with those observed for K giants, e.g., Arcturus, by @dravins87. We have not explored any attempts to introduce 3D modeling of the stellar atmospheres into our analysis; such models are not yet available for metal-poor giant stars. In this study we restrict ourselves to the use of one-dimensional, plane-parallel model atmospheres computed for the LTE case.
In contrast to the challenge of detecting the presence of $^{6}$Li from the Li <span style="font-variant:small-caps;">i</span> 6707Å line profile, which likely requires 3D-NLTE modeling to account for the convective asymmetry that can mimic the presence of $^{6}$Li [@cayrel07], the change in the line profile that we analyze is very different. The presence of the $^{6}$Li component may account for only a few tenths of a percent of the total flux in the red wing of the $^{7}$Li line, which is a very subtle effect. The magnitude of the HFS for Eu causes gross changes in the shape of the entire line profile (see § \[euiso\]). Sm and Nd each have seven naturally-occuring isotopes, so the shape of the line profile will be asymmetric along all parts of the line and not just along the red wing where convective asymmetries are most prominent. Therefore our choice to compute 1D synthetic line profiles in LTE will not significantly affect our results for these species.
### Detailed Examination of the Absorption Line Profile {#linefits}
Isotope shifts and HFS alter the shape of the line profile, and the isotopic fraction can be assessed by examining the line profile from detailed comparisons between the observed and synthetic spectra. This method has been used previously to assess the Eu isotopic fraction by @sneden02 and @aoki03a [@aoki03b]. We quantify our measurements of the isotopic fractions using a $\chi^{2}$ algorithm (after, e.g., @smith01 [@aoki03a; @asplund06]), $$\chi^{2} = \left\langle
\frac{(O_{i}-C_{i})^{2}}{\sigma^{2}_{i}}\right\rangle .$$ Here $O-C$ is the difference between the observed and synthetic spectra at the $i$th spectrum point and $\sigma$ is defined as $\sigma_{i} = [$(S/N)$ \times (f_{i})^{1/2}]^{-1}$, where S/N is the approximate signal-to-noise ratio of the continuum and $f$ is the depth of the $i$th point relative to the continuum. Our best estimate of the isotopic fraction and abundance for each line is given by minimizing $\chi^{2}$. Uncertainties are estimated at 1, 2, and 3$\sigma$ confidence intervals from $\Delta\chi^{2}=\chi^{2}-\chi^{2}_{min}=1$, 4, and 9, respectively. We consider the number of degrees of freedom in the fit to be the number of points in the spectral window plus the continuum normalization, abundance, and wavelength offset, all of which were allowed to vary in our fits. Reduced $\chi^{2}$ values are $\sim 1$, differing by less than a factor of two, which is still larger than would be expected for a normal distribution of errors [e.g., @press92]. We attribute the scatter in the reduced $\chi^{2}$ values to our estimation of the S/N ratio. The *relative* $\chi^{2}$ values for a given line, represented by $\Delta\chi^{2}$, however, do exhibit well-determined minima. We fix the macroturbulence, as determined in § \[atmosphere\], but we also estimate the uncertainties in the isotopic fraction introduced by the uncertainty in the macroturbulence, $V_{\rm macro}\,\pm\,\Delta\,V_{\rm macro}$. Only our Eu and Sm lines with complete sets of hyperfine constants (see § \[atomicdata\]) can be assessed by this method.
### Set the Absolute Wavelength Scale from Nearby Wavelength Standards {#wavelengthstandard}
In principle, the absolute wavelengths of the observed and synthetic spectra can by matched using wavelength standard Fe <span style="font-variant:small-caps;">i</span> and Ti <span style="font-variant:small-caps;">ii</span> lines, and then the peak wavelength of the rare-earth line of interest can be measured. The isotopic fraction can be assessed because the isotope shifts—much more than the HFS—affect the peak absorption wavelength. @lundqvist07 have used this method to qualitatively demonstrate that the Sm isotopic mix in CS 31082-001 is consistent with an origin.
For our Sm lines, a small wavelength shift ($\sim$ 0.02$-$0.04Å) is the dominant effect of changing the isotopic fraction; for our Nd lines, the shift is even smaller ($\sim$ 0.01Å). In both cases, broadening of the line profile is a secondary effect. Therefore, given the small magnitudes of these shifts, it is critical that we match the wavelength scales of our synthetic and observed spectra with great accuracy. The absolute wavelengths of some of our Sm HFS components are accurate to $\pm$1mÅ, as determined by the FTS measurements of $^{154}$Sm by @lundqvist07 (see Appendix \[smappendix\]), while the absolute wavelengths of the Sm lines not included in that study are accurate to $\approx$ 2mÅ. Unlike Sm, however, the atomic data for Nd does not include FTS measurements of individual isotope positions, so our reported absolute wavelengths of Nd components are based solely on measurements of the Nd energy levels. To avoid mixing measurements of energy levels from different studies, we adopt the energy levels from @blaise84. They estimate an uncertainty of $\pm\,0.005$cm$^{-1}$ for each level. Doubling this uncertainty translates to an uncertainty of $\sim$ 1.5–2.0mÅ across the region where our Nd lines are located. We note that the relative HFS component wavelengths—measured from FTS spectra—are more accurate than this in all cases. The absolute wavelengths of the wavelength standard lines (Fe <span style="font-variant:small-caps;">i</span>: @learner88 and @nave94; Ti <span style="font-variant:small-caps;">ii</span>: @pickering01 [@pickering02]) have a reported absolute accuracy in their FTS measurements of $\sim\,\pm$0.001cm$^{-1}$ ($\lesssim$0.3mÅ) in our wavelength regions. The uncertainties from our wavelength scale and co-addition of individual orders are no larger than $\pm$0.9mÅ, and we have verified that, after the wavelength solution has been applied, the Th-Ar emission lines are correctly identified to within $\sim\,\pm\,$1mÅ. All of these uncertainties are sufficiently small to enable reliable measurement of the isotopic fraction.
In each echelle order containing a rare-earth line of interest, we locate 1–6 wavelength standard lines and match the wavelength scales of the observed and synthetic spectra at each of these lines, typically with a precision of $\sim$ 1mÅ per line. When multiple wavelength standards are used, the scatter in our measured wavelength offsets necessary to ensure the matching of each line is $\sim$ 5mÅ, much greater than anticipated. This scatter appears to be independent of the ionization state, species, excitation potential, and relative strength of the wavelength standards, each of which could be expected to introduce small systematic offsets among lines with these different characteristics (see, e.g., @dravins81 [@dravins86; @asplund00]. The isotopic fraction we would derive from our rare-earth lines of interest is therefore entirely dependent upon our choice of which wavelength standard line(s) we adopt and how many of these lines are present in the echelle order with the rare-earth line.
To further understand this matter, we have also attempted to match the observed and synthetic spectra for wavelength standards in the Solar spectrum. We compare both the @kurucz84 integrated-disk Solar spectrum, obtained with the FTS on the McMath Solar Telescope at Kitt Peak National Observatory, and daylight sky spectra, obtained using the same instrumental configuration on the McDonald Observatory 2dCoudé echelle spectrograph as our stellar observations and reduced in an identical fashion. It is well-known that astronomical echelle spectrographs and cameras can introduce a variety of non-linear distortions to the wavelength scale, whereas the spectrum obtained with the FTS should be immune to many of these distortions. We synthesize a spectral region from 4997–5010Å (one echelle order) using an interpolated @holweger74 empirical Solar model photosphere. We apply our matching technique to the five unblended Fe <span style="font-variant:small-caps;">i</span> wavelength standard lines in common to the two spectra. We measure the same line-to-line wavelength offsets for these lines in both spectra (agreement within our measurement precision, $\sim$ 1mÅ), suggesting that the wavelength scale of our 2dCoudé spectra is consistent with the Kitt Peak FTS spectra. Nevertheless, the scatter in the absolute wavelength scale as determined from each of these lines in the FTS spectrum is *still* $\sim$ 5mÅ, indicating that there is a more fundamental problem that does not result from our spectral reduction techniques or our ability to match the observed and synthetic spectra. Unable to further identify and quantify the source of this discrepancy, we are unable to proceed with confidence by this method of analysis.
### Set the Relative Wavelength Scale at the Point of Insensitivity to the Isotopic Mix {#wavelengthpoint}
We also develop one additional method of quantifying the isotopic fraction that does not rely on nearby wavelength standard lines or fits to the shape of the line profile. The reason for developing such a method is our desire to assess the isotopic mix from all lines with maximum isotopic shifts $\gtrsim 15$mÅ and from which an elemental abundance can be measured. This allows us to examine additional Sm lines that do not have full sets of measured hyperfine A and B constants as well as all of the Nd lines.
This method attempts to match the observed and synthetic spectrum using the point in the synthetic spectrum that is insensitive to the $s$- or mix. First, for each line, an isotopic fraction is assumed in the synthesis, and we always adopt $f=0.50$ by default. Then an elemental abundance is derived from a synthesis with this isotopic fraction. Using this elemental abundance, two new syntheses (one with a pure-$s$-process mix of isotopes and one with a pure-$r$-process mix) are created. The intersection of these two syntheses is defined to be the point of insensitivity to the isotopic fraction. The relative wavelength offset between the point of insensitivity and the observed spectrum is then adjusted until the point of insensitivity is coincident with the flux of the nearest pixel of the observed spectrum. Next, we adopt this relative wavelength offset and generate a set of synthetic spectra with a range of isotopic fractions but the same elemental abundance. A new isotopic fraction is then measured by minimizing the sum of the squares of the differences between the observed and synthetic spectra. This isotopic fraction is then used as the input to derive an elemental abundance, and the process is repeated until the elemental abundance and isotopic fraction have each converged. We show an example where the point of insensitivity has been matched to the observed spectrum in Figure \[4604shift\].
The uncertainty in the matching process is given by the width of one pixel; therefore the uncertainty in the derived isotopic fraction is estimated by shifting the synthetic spectrum relative to the observed spectrum by $\pm\,w_{\rm pixel}/2$. Even at a spectral resolution of $R~\sim~120,000$, the width of one pixel is still a substantial fraction of the width of the absorption line itself. This method also places a disproportionate amount of weight on the one or two flux measurements closest to the point of insensitivity, rather than taking account of the information from the full width of the line to match the observed and synthetic spectra. The precision achieved by this method is hardly adequate to clearly assess the isotopic mix—particularly for Nd—but it does provide a consistent set of measurements for all of our lines of Sm and Nd.
Neutron-Capture Elemental Abundances {#elabund}
====================================
Our mean elemental abundance of Eu in HD 175305, \[Eu/Fe\] $=+0.46\,\pm\,0.07$, is in good agreement with @fulbright00’s abundance, \[Eu/Fe\] $=+0.48\,\pm\,0.105$, and @cowan05’s abundance, \[Eu/Fe\] $=+0.55\,\pm\,0.10$. Our synthesis of the Eu <span style="font-variant:small-caps;">ii</span> 4205Å line in HD 122563 does permit measurement of the Eu elemental abundance, \[Eu/Fe\] $=-0.55\,\pm\,0.06$, assuming $^{12}$C/$^{13}$C$\,=\,5$ [@lambert77; @honda04b] and a Eu isotopic fraction of $f_{151}\,=\,0.5$. The Eu abundance changes by $\pm\,0.03$ dex when the assumed isotopic ratio is varied by $\pm\,0.2$ dex. This elemental abundance agrees very well with the value found by @honda06, \[Eu/Fe\] $=-0.52$, and with many previous studies of this star. The 4205Å line in HD 196944 is far too sensitive to the C abundance to derive a meaningful Eu elemental abundance.
We measure Sm elemental abundances of \[Sm/Fe\] $=-0.56\,\pm\,0.12$ from 1 line in HD 122563, \[Sm/Fe\] $=+0.37\,\pm\,0.11$ from 9 lines in HD 175305, and \[Sm/Fe\] $=+0.88\,\pm\,0.08$ from 6 lines in HD 196944. Previous studies of Sm elemental abundances in these stars have found \[Sm/Fe\] $=-0.40\,\pm\,0.17$ in HD 122563 [@honda06] and \[Sm/Fe\] $=+0.78\,\pm\,0.23$ in HD 196944 [@aoki02]; Sm has not been measured previously in HD 175305. Unlike the uncertainties in the isotopic fractions, the uncertainties in the elemental abundances are dominated by the systematic uncertainties in the model atmosphere parameters and are affected somewhat less by internal scatter from the log($gf$) values.
We measure Nd elemental abundances of \[Nd/Fe\] $=+0.22\,\pm\,0.13$ from 12 lines in HD 175305 and \[Nd/Fe\] $=+0.95\,\pm\,0.13$ from 11 lines in HD 196944. Previous studies of Nd elemental abundances in these stars have found \[Nd/Fe\] $=+0.34$ [@burris00] in HD 175305 and \[Nd/Fe\] $=+0.93\,\pm\,0.09$ [@vaneck01], \[Nd/Fe\] $=+0.94\,\pm\,0.17$ [@zacs98], and \[Nd/Fe\] $=+0.86\,\pm\,0.20$ [@aoki02] in HD 196944.
We have also measured elemental abundances for five additional heavy elements in HD 175305. These abundances are listed in Table \[tab3\] along with our Eu, Sm, and Nd abundances. The uncertainties listed in the table reflect the uncertainties in the atmospheric parameters, continuum placement, line-to-line scatter, and $\log(gf)$ values, all added in quadrature. In cases where only one line of a species was available for analysis, we substitute the uncertainty in the spectral synthesis fit to the observed spectrum in place of the line-to-line statistical scatter in this formula. In Figure \[175305ncap\] we plot our elemental abundances, combined with measurements from @fulbright00, @cowan05, and @lawler07, against the scaled S. S. abundance pattern of @simmerer04. While not a perfect fit, this figure suggests that the scaled S. S. abundance pattern provides a respectable fit of the observed abundances beyond Ba. HD 175305 may not be a star with pristine material, but the appears to have contributed the majority of the species present. We have also used the Keck I HIRES spectrum of HD 175305, obtained for the @cowan05 study, to determine an upper limit for Pb from the 4057Å line.
Neutron-Capture Isotopic Fractions {#isofrac}
==================================
We emphasize that measurements of isotopic fractions are highly insensitive to the choice of atmospheric parameters. Separate isotopic components of a line all share the same initial and final energy levels (since the HFS of these levels is negligible relative to the energy differences themselves), eliminating systematic errors associated with the excitation and ionization states of each isotope.
The Europium Isotopic Fraction {#euiso}
------------------------------
### In the Sun {#eusun}
For S. S. material, the isotopic abundances are precisely known from studies of terrestrial meteorites. The Eu <span style="font-variant:small-caps;">ii</span> isotopic fraction in the Solar spectrum was first studied by @hauge70 and @hauge72, who found $^{153}$Eu/$^{151}$Eu=(48$\pm$6)/(52$\mp$6) (or $f_{151}\,=\,0.52$ in our notation). @lawler01b reevaluated the Solar Eu isotopic fraction, finding $f_{151}\,=\,0.50\,\pm\,0.07$ from the 4129Å line. We do not repeat this measurement here.
### In Metal-Poor Stars {#eustars}
From the Eu <span style="font-variant:small-caps;">ii</span> 4129Å line in HD 175305 we measure an isotopic fraction of $f_{151}\,=\,0.50\,\pm\,0.04$, as shown in Figure \[fig4129iso\]. The $\Delta\,\chi^{2}$ values are shown in Figure \[4129chisq\] for different values of the isotopic fraction and macroturbulent velocity. The minimum $\Delta\,\chi^{2}$ values have well-defined minima, leading to the small uncertainties in the derived isotopic fraction. As noted in Appendix \[eulines\], we cannot trust our syntheses of the Eu <span style="font-variant:small-caps;">ii</span> lines at 4205 and 4435Åat the precision necessary to measure an isotopic fraction. For the sake of comparison, in Figures \[fig4205iso\] and \[fig4435iso\] we show our synthesis of these two lines in HD 175305 given our best estimate of the blending features, but we emphasize that these syntheses display the best-fit isotopic fraction derived from the 4129Å line.
The Eu <span style="font-variant:small-caps;">ii</span> 4129Å line was not covered in our spectra of either HD 196944 or HD 122563.
Although only two naturally-occurring isotopes of Eu exist, only a small spread is predicted between $f^{s}_{151}$ and $f^{r}_{151}$ (see panel **a** of Figure \[fig4129iso\]). It is difficult even with very high-resolution ($R\,\gtrsim\,100,000$) and high-S/N ($\sim$ few hundred) data on one or more lines to achieve sufficiently small uncertainties on isotopic abundances to determine the nucleosynthetic source of the Eu. @sneden02 [@aoki03a; @aoki03b] have successfully shown that the Eu $s$- and nucleosynthetic signatures can be distinguished at the isotopic level, using approximately three lines in each of seven stars. The measurements made from individual lines in each star in these studies have a large degree of scatter, causing the final Eu isotopic fraction to be mildly sensitive to the set of lines chosen. For example, we note that @aoki03a’s two measurements of $f_{151}$ from the 4435Å line appear systematically higher than the results obtained from the 4205Å line by $\approx$ 10$-$15%, which is roughly the same as the change in the Eu isotopic fractions that are trying to be distinguished. Although we do not measure an isotopic fraction from the 4435Å line in HD 175305, our syntheses hint at a similar result in the small region without obvious blends (4435.40$-$4435.50Å). Additional measurements of the Eu isotopic fraction from these lines would be helpful to confirm or refute these possible systematic trends. If such systematic trends exist, they likely result from the presence of blending features, and the measured isotopic fractions may be as sensitive to the blending features as to the Eu isotopes themselves. There is little hope of being able to distinguish Eu isotopic fractions with any greater precision than has already been achieved using these methods on this set of lines, including the Eu <span style="font-variant:small-caps;">ii</span> lines at 3819 and 3907Å.
The Samarium Isotopic Fraction {#smiso}
------------------------------
### In the Sun {#smsun}
@ekeland75 analyzed the Sm isotopic fraction in the Solar photosphere from two lines, 4467 and 4519Å, finding isotopic fractions consistent with the meteoritic values. The 4467Å line is broadened slightly by HFS, with almost zero broadening resulting from the isotope shifts, while the 4519Å line is broadened slightly by isotope shifts. High-precision single-frequency laser measurements of the isotope shifts for $^{147}$Sm and $^{149}$Sm have not been measured in the lab for the 4519Å line, therefore we do not attempt synthesis of this line.
We synthesize the 4467Å line using the @holweger74 model atmosphere and the @kurucz84 Solar spectrum. We fix the broadening from nearby Fe lines. The total spread in the HFS and isotope shifts for this line are $\lesssim$ 0.05Å and $\lesssim$ 0.01Å, respectively, and we cannot measure the isotopic fraction, either in terms of $f_{\rm odd}$ or $f_{152+154}$, from the line-profile-fitting method even when a high-quality Solar spectrum is used. @ekeland75 provide very little information about their method of measuring the Sm isotopic fraction, and their very small uncertainties ($f_{\rm odd}\,=\,0.28\,\pm\,0.07$) for a line with relatively small HFS and isotope shifts give us some reservation in accepting their result. All of the other Sm transitions listed in Appendix \[smlines\] are either far too weak or are severely-blended with strong transitions of other species to measure reliable isotopic fractions in the Solar spectrum, so we are unable to make any measurements of the Sm isotopic fraction in the Solar photosphere. We remind the reader that the blended lines may be reliable for measurements of the Sm elemental abundance, but we cannot trust our syntheses of the blending features at a precision necessary to measure isotopic fractions.
### In Metal-Poor Stars {#smstars}
Figures \[4424iso175\] and \[4424iso196\] show our syntheses of the Sm <span style="font-variant:small-caps;">ii</span> line at 4424Å in HD 175305 and HD 196944, respectively, and the results from our line profile fits. We also include the Cr abundance of the Cr <span style="font-variant:small-caps;">i</span> 4424.28Å line, which blends the blue wing of the Sm line, as a free parameter. In HD 122563, this Sm line has a relative depth of $\sim$2%. We do not claim measurements of either the elemental abundance or isotopic fraction of Sm from this line in HD 122563; it is clear that this line is simply too weak to warrant further analysis. In HD 175305, we measure a Sm isotopic fraction of $f_{152+154}\,=\,0.55\,\pm\,0.14$. In HD 196944, we measure a Sm isotopic fraction of $f_{152+154}\,=\,0.35\,^{+0.11}_{-0.20}$. In addition to showing the best fit curves and their uncertainties in Figures \[4424iso175\] and \[4424iso196\], we also show the syntheses for pure-$s$- and pure-$r$-process nucleosynthesis and the ($O-C$) curves for these syntheses, illustrating the contrast between the two extremes. In each of these figures, the pure-$s$- and pure-$r$-process syntheses in panel **b** are compared to the observed spectra in panel **a**, illustrating the degree to which the observed spectrum tends toward one synthesis or the other.
Figure \[4604iso\] shows our synthesis of the Sm <span style="font-variant:small-caps;">ii</span> line at 4604Å in HD 175305. We measure a Sm isotopic fraction of $f_{152+154}\,=\,0.50^{+0.09}_{-0.14}$. This line is weak, with a relative depth of only 5%. Figure \[4604chisq\] shows the $\Delta\,\chi^{2}$ values for our fit to this line; $\Delta\,\chi^{2}$ values from the Sm 4424Å line in both stars are similar to the 4604Å line.
If instead we match the observed and synthetic spectra at the point of insensitivity to the isotopic fraction, we find $f_{152+154}\,=\,0.41\,\pm\,0.32$ from the 4424Å line, $f_{152+154}\,=\,0.49\,\pm\,0.17$ from the 4604Å line, and $f_{152+154}\,=\,0.49\,\pm\,0.20$ from the 4719Å line in HD 175305. In HD 196944, we find $f_{152+154}\,=\,0.37\,\pm\,0.28$ from the 4424Å line. While we regard the line profile fits as the superior measurements of the isotopic fraction, the agreement between the sets of measurements from the two methods is reassuring.
The Neodymium Isotopic Fraction {#ndiso}
-------------------------------
### In the Sun {#ndsun}
We check our Nd isotopic fraction measurements against the Solar spectrum. We are aware of no previous attempts to examine the Nd isotopic mix in the Solar photosphere, but the individual S. S. isotopic abundances have been established from studies of terrestrial meteorites. The Nd <span style="font-variant:small-caps;">ii</span> line at 4446Å would appear to be the most reliable indicator of the Nd isotopic fraction due to its relative strength and freedom from blending features. The line profile shape does not change enough to allow us to apply our fitting algorithm to measure the isotopic fraction, and we can find no good candidate Fe <span style="font-variant:small-caps;">i</span> or Ti <span style="font-variant:small-caps;">ii</span> lines nearby to match the observed and synthetic spectra. By using the method of matching the observed and synthetic spectra at the point of insensitivity to the isotopic mix, we derive an isotopic fraction of $f_{142+144}\,=\,0.53^{+0.47}_{-0.53}$, which is unfortunately unable to exclude any possible combination of isotopes. Therefore, we are unable to provide independent confirmation of the S. S. Nd isotopic fraction from analysis of the Solar photosphere.
### In Metal-Poor Stars {#ndstars}
Because the isotope shifts of Nd are smaller than Sm and because we cannot match our observed and synthetic spectra with a high absolute accuracy, we can only rely on the third matching method to measure the isotopic fraction of the Nd lines in these stars. In HD 175305, we measure $f_{142+144}\,=\,0.25^{+0.75}_{-0.25}$ from the 4177Å line, $f_{142+144}\,=\,0.11^{+0.85}_{-0.11}$ from the 4446Å line, and $f_{142+144}\,=\,0.33^{+0.67}_{-0.33}$ from the 4567Å line. The mean isotopic fraction of these three lines is $f_{142+144}\,=\,0.21^{+0.56}_{-0.21}$. Combining these measurements together using the asymmetric uncertainties that reflect the fact that the isotopic fraction cannot be less than 0.0 or greater than 1.0 could introduce a bias into the weighted mean. Therefore we have estimated the uncertainties in these measurements also by assuming (non-physically) that the isotopic fraction can exceed these limits. We find uncertainties of $\pm\,1.10$, $\pm\,0.85$, and $\pm\,1.00$ on these three lines, respectively, resulting in a weighed uncertainty of $\pm\,0.56$ on the mean. Our final quoted uncertainties do impose the physical limits on the isotopic fractions. In HD 196944, we measure $f_{142+144}\,=\,0.36^{+0.64}_{-0.36}$ from the 4446Å line. The non-physical uncertainty on this measurement is $\pm\,0.68$, which provides a relative comparison to our ability to measure the 4446Å line in HD 175305. Unfortunately none of these individual measurements provides any real insight into the Nd isotopic fraction due to the size of the uncertainties. The large uncertainties also prevent us from pointing out even a contrast between the Nd isotopic mixes in the two stars, much less make any detailed assessment of the nucleosynthetic predictions.
Interpretation of the Nucleosynthetic Signatures {#interpret}
================================================
In HD 175305, our Sm isotopic fraction $f_{152+154}\,=\,0.51\,\pm\,0.08$, a weighted average of our measurements from profile fits to the 4424 and 4604Å lines, suggests an origin. It is tempting to surmise that our Nd isotopic fraction, $f_{142+144}\,=\,0.21^{+0.56}_{-0.21}$, derived from three lines, is suggestive of an origin, but the large uncertainty cannot exclude an origin. Our Eu isotopic fraction, $f_{151}\,=\,0.50\,\pm\,0.04$, derived from only the 4129Å line and shown to be a reasonable match for the 4205 and 4435Å lines, implies an origin by the @cowan06a predictions but excludes neither a pure-$s$- nor a pure-$r$-process origin by the @arlandini99 predictions. It is interesting to note that our isotopic fractions in HD 175305 suggest a nucleosynthetic history predominantly—but perhaps not purely—dominated by the , which also is suggested by the elemental abundance trends. The fact that both the elemental and the isotopic distributions in HD 175305 agree with the predicted elemental and isotopic distributions of material in the S. S. adds to the preponderance of evidence that supports the hypothesis of a universal mechanism for elements with $Z\geq56$.
In HD 196944, our measured Sm isotopic fraction, $f_{152+154}\,=\,0.35\,\pm\,0.14$, is suggestive of an origin. Our Nd isotopic fraction, $f_{142+144}\,=\,0.36^{+0.64}_{-0.36}$ can exclude no possible values of the isotopic mix. The interpretation for the Sm is consistent with the elemental abundance signatures and radial velocity variation described in § \[radvel\].
Our best measurements for each species are summarized in Figures \[summary1\] and \[summary2\] for both HD 175305 and HD 196944. In Figure \[summary1\], the differences between the pure-$r$- and pure-$s$-process predictions are scaled together to emphasize the relative precision with which the two processes can be distinguished by each species. In contrast, Figure \[summary2\] displays our measurements in an absolute sense to emphasize the relative precision with which each species was measured. Expressed this way, it is clear that Eu can be measured precisely but does not easily reveal its nucleosynthetic origin because of the small difference between $f^{r}_{151}$ and $f^{s}_{151}$. The uncertainties on each Nd line can hardly exclude any possible value of the isotopic fraction. Only the Sm isotopic fractions, which have uncertainties on the individual measurements that are smaller than the difference between $f^{r}_{152+154}$ and $f^{s}_{152+154}$, offer any opportunity to distinguish the two processes with reasonable precision.
Eu/Pb and $^{151}$Eu/Pb Abundance Ratios {#eupb}
========================================
In § \[euatomic\] we pointed out the great discrepancy between the stellar model of @arlandini99 and the classical method of @cowan06a in predicting the amount of $^{151}$Eu produced by the : $N_{s}\,=\,0.00304\,\pm\,0.00013$ and $N_{s}\,=\,0.0001$, respectively. The stellar model predicts that the should produce $\sim$ 30 times more $^{151}$Eu than the classical method does. For a certain amount of $^{151}$Eu produced by an nucleosynthesis event, the classical method would then predict that the isotopes near the termination of the $s$-process—such as Pb—should be overproduced by a factor $\sim$ 30 relative to the stellar model.
We choose Pb for this comparative analysis because it is so dominantly produced by the . In Pb-enriched metal-poor stars, such as HD 196944, the Pb overabundance often will reach \[Pb/Fe\] $\geq +2.0$, making it easy to detect at the Pb <span style="font-variant:small-caps;">i</span> 4057Å transition. Because Eu is produced more easily by the than by the (relatively speaking) and Pb is produced more easily by the than by the , the Eu/Pb and $^{151}$Eu/Pb abundance ratios in metal-poor stars should be extremely good discriminators of the predicted abundances of these species.
@arlandini99 do not predict pure-$r$-process Pb abundances due to the fact that Pb can be overproduced relative to the lighter-$s$-process nuclei in low metallicity AGB stars. In other words, at lower metallicity a greater number of neutrons is captured per Fe-peak seed than at higher metallicity, so the abundance distribution at low metallicity favors the heavest nuclei; see, e.g., @gallino98, @busso99, and @straniero06.
In Figure \[eupbplot\] we have compiled the derived Eu/Pb ratios (or upper limits) for 35 metal-poor stars from a variety of recent studies and for HD 175305. The Eu/Pb ratio shows no obvious trends when expressed as a function of \[Fe/H\], but we note that this sample of Eu and Pb abundances is far from complete. Adopting the original authors’ classifications for the enrichment patterns for these stars, this sample consists of 6 $r$-process-enriched stars (@cowan02 [@hill02; @aoki03a; @sneden03; @plez04; @ivans06; @frebel07]; A. Frebel, 2007, private communication), 15 $s$-process-enriched stars [@aoki01; @aoki02; @johnson02c; @cohen03; @barbuy05; @cohen06; @preston06], and 5 $(r+s)$-process-enriched stars [@ivans05; @aoki06; @jonsell06; @masseron06] from the Galactic halo population. The $r$-process-enhanced sample also includes 10 stars from the metal-poor globular clusters M13 and NGC6752 [@yong06]. The small number of $r$-process enriched stars results from the observational difficulty of detecting Pb in stars without enrichment. The Eu/Pb ratios of the $r$-process-enhanced halo stars and globular cluster stars all suggest that the predictions are correct. Despite the fact that $\log\,\varepsilon\,$(Eu/Pb)$_{s}$ differs by $\sim$ 0.6 dex between the stellar model and classical method predictions, the measured abundance spread in the $s$-process-enriched stars is still too great to unambiguously distinguish the two predictions from this information alone. The $s$- and $(r+s)$-enhanced classes of stars show little difference from one another when their Eu and Pb abundances are expressed this way.
In Figure \[151pbplot\] we have complied a subset of 7 of these stars with measured Eu isotopic fractions, and we plot $\log\,\varepsilon\,$($^{151}$Eu/Pb) as a function of metallicity. This subset includes 5 $r$-process-enhanced stars and 2 $(r+s)$-process-enhanced stars. Again, the $r$-process-enhanced halo stars cluster around the predictions. Here, though, the stellar model and classical method predictions for $\log\,\varepsilon\,$($^{151}$Eu/Pb) differ by $\sim$ 1.7 dex, and the two $r+s$ stars with measured Eu isotopic fractions clearly favor the @arlandini99 predicted $^{151}$Eu abundance. This is not unexpected, given that @aoki03a, who measured the Eu isotopic fraction in these two stars, reported that $f_{151}$ was consistent with the @arlandini99 predictions.
It is worth considering whether these $r+s$ stars may have enough enrichment to give them a $\log\,\varepsilon\,$($^{151}$Eu/Pb) abundance that lies between the @cowan06a pure-$s$ and pure-$r$ predictions, inadvertently suggesting the @arlandini99 value. To investigate this matter, in Figure \[bandplot\] we plot the $\log\,\varepsilon\,$(Ba/Nd) abundance ratios for the full sample of halo stars; @yong06 did not measure Ba or Nd abundances in their globular cluster stars. While there is still a relatively large amount of scatter among the $s$- and $(r+s)$-enriched stars, neither population appears distinct from the other, and the two $r+s$ stars shown in Figure \[151pbplot\] scatter above and below the predicted pure-$s$-process abundance value. We therefore conclude that these two stars are representative of predominantly-$s$-process-enriched material, and these observations thereby support the @arlandini99 $^{151}$Eu abundance prediction. We caution that this conclusion hinges on the Eu isotopic fraction measured in only two stars, and additional observations of the Eu isotopic fraction in $s$-process-enriched stars would be most helpful to further resolve the matter.
@straniero06 show that in low-metallicity stars in the TP-AGB phase the abundance ratios of nuclei near the first and second peaks are only weakly affected by the efficiency of the $^{13}$C pocket (and hence the neutron source for the ), whereas abundances of species near the third peak are strongly dependent on this parameter. Therefore the abundance predictions of the @arlandini99 stellar model, which was computed for $\sim$ Solar metallicity stars in the TP-AGB phase, should predict the lighter abundances at low metallicty better than it predicts the Pb abundance at low metallicity. In a low-metallicity nucleosynthesis event, we would expect the $^{151}$Eu/Pb ratio to decrease relative to a higher-metallicity case as a higher fraction of seed nuclei are converted to Pb. Therefore, the stellar model pure-$s$ Eu/Pb and $^{151}$Eu/Pb ratios should represent upper limits for these values. In both cases, decreasing these ratios would bring the stellar model predictions into better agreement with the classical method predictions.
Prospects for Future Studies of Neutron-Capture Isotopic Fractions {#future}
==================================================================
Previous studies have demonstrated that the Eu and Ba isotopic fractions can be measured, we have shown that Sm isotopic fractions can also be quantitatively measured, and we also suggest that Nd isotopic fractions can be measured, albeit with lower precision that the other rare-earth species. The measurements of each set of isotopes give an indication of the stellar nucleosynthetic history that is consistent with other indicators, and we now discuss the role that measurements of Ba, Nd, Sm, Eu, and other isotopic fractions might play in future studies.
Isotopic Fractions of Multiple Neutron-Capture Species {#multipleiso}
------------------------------------------------------
It is a success of both the stellar model and classical method that these models’ predictions of the Sm $f_{152+154}$ isotopic fraction are in good agreement. Suppose that both Sm isotopic fractions $f_{152+154}$ and $f_{\rm odd}$ could be measured from the same set of lines. This would enable a comparison of the models’ predicted abundances of $^{147}$Sm, which differs by a factor of $\sim$3 and affects the predicted $f_{\rm odd}$ values, whereas both models generally agree on $^{149}$Sm, $^{152}$Sm, and $^{154}$Sm. One could measure the isotopic fractions of Sm and Ba in an $r$-enriched star. If a convincing argument could be made that the Ba and Sm should have originated from the same nucleosynthetic event(s) (i.e., an extra contribution has not been made to the Ba abundances by another process), the isotopic fractions measured for both Ba and Sm should show the same relative proportion of isotopes. This scenario could permit a direct test to the predicted Ba isotopic fractions, which differ by 26% in their $r$-only predictions for $f_{\rm odd}$, with the most significant differences arising in the models’ predictions for $^{137}$Ba and $^{138}$Ba. Such tests would be greatly welcomed.
It is not obvious why Ba abundances are sometimes larger than the scaled S. S. $r$-process abundances (scaled to the heavier $n$-capture elements) in metal-poor stars with an $r$-only chemical signature and no evidence of binarity, such as HD 175305. (See also, e.g., Figure 15 of @honda04a or Figure 15 of @barklem05, where some of their $r$-enriched stars exhibit this behavior.) @asplund04 suggests that the 3D Ba abundance correction may be as much as $-$0.5 dex for some metal-poor dwarfs and subgiants. While this matter may indeed be a result of poor Ba abundance determinations due to 1D LTE modeling or to uncertainties in the S. S. abundance predictions, the search for an astrophysical explanation for this phenomenon should be explored further. Measurements of Ba and other isotopic fractions could be used to identify any correlations that may exist between super-$r$-process Ba abundances and other nucleosynthetic signatures.
The combination of Ba, Nd, Sm, and Eu isotopic fractions could provide more complete knowledge of abundances at the isotopic level over a range of $\sim$ 20 mass numbers ($A\,=\,$134$-$154). This information would complement elemental abundance measurements of lighter and heavier $n$-capture species to get a more complete picture of the nucleosynthesis, constrain the conditions (e.g., temperature, neutron density, etc.) that are required to produce the elements, and determine the actual path by identifying the individual isotopes that participate in this process (see @cowan04 for further discussion). Extended discussion of the relationship between neutron densities in the “main” and “weak” $r$-processes and the isotopic abundance signatures in this mass range is provided in @kratz07 and will not be repeated here. There is evidence that different mass regimes may be synthesized by different astrophysical processes in addition to the $s$- and $r$-processes, such as the light element primary process, or LEPP [@travaglio04]. While the LEPP would have the greatest effect on nucleosynthesis of lighter elements such as the Sr-Y-Zr group, knowledge of isotopic data for elements with very different $Z$ would still be extremely informative.
This study has made use of the most current and complete laboratory atomic data to compute the isotope shifts and HFS patterns for Sm and Nd. While we are pleased that the available atomic data permits measurement of the isotopic fractions of these two species, complete sets of HFS constants and isotope shifts have not been measured for many lines that lie in the yellow-red regions of the spectrum. These transitions are generally weaker in stellar spectra, but the isotope shifts may be large enough nonetheless to permit reliable measurements of the isotopic fraction even with weak lines. We encourage such efforts from the atomic physics community in the future. We note that these lines may be significantly stronger in more metal-rich stars or those which exhibit enhanced abundances for elements, such as Ba stars [e.g., @allen06a; @allen06b], where the blue lines we have used in our study are likely to be heavily blended with other atomic and molecular features; in these cases, lines in the yellow-red spectral regions may provide the only opportunity to establish the signatures at the isotopic level.
Individual Advantages and Disadvantages of Europium, Samarium, and Neodymium Isotopic Fractions {#euvssm}
-----------------------------------------------------------------------------------------------
One advantage of using measurements of the Eu isotopic fraction as a nucleosynthesis probe is the wide HFS structure that can be measured with only moderately-high spectral resolution ($R\,\sim$45–60,000). In $r$-enriched stars, the Eu lines are strong enough to facilitate measurement of the isotopic fraction even at very low metallicities. But, as we have argued in § \[euiso\], it is unlikely that Eu isotopes can be measured with the precision to unambiguously distinguish $s$- and nucleosynthesis using the lines in the near-UV and blue regions of the spectrum if indeed the @arlandini99 predictions are correct.
Sm presents a different set of circumstances. The HFS of Sm is not nearly as wide as Eu, and the lines with the widest structure are also very weak (or undetectable) and sometimes blended in very metal-poor stars. Though we are forced to parametrize the $s$- and nucleosynthetic predictions into some set of isotopes (e.g., $f_{\rm odd}$ or $f_{152+154}$) to facilitate distinguishing among seven isotopes simultaneously, the greater difference predicted between $f^{s}_{152+154}$ and $f^{r}_{152+154}$ allows one to cleanly distinguish these different nucleosynthetic signatures with reasonable measurement uncertainties. In addition, if Sm isotope fractions could be measured in a large sample of stars, it may be possible to observe the changing contributions of $s$- and nucleosynthesis in stars over the history of the galaxy (cf. Figure 6 of @mashonkina06, who performed a similar study using Ba isotopes). Therefore, while Sm lines that permit measurement of isotopic fractions may not be accessible in all metal-poor stars, we propose that in favorable cases Sm isotopes may give a clearer picture of the nucleosynthetic history than Eu isotopes can.
The Nd isotopic fraction is more difficult to measure reliably than the Sm isotopic fraction. The smaller isotope shifts of Nd translate directly into larger uncertainties in the measured isotopic fraction on any individual line. This necessitates that as many clean lines as possible be measured in order to reduce the uncertainties on the mean Nd isotopic fraction. Only then is it possible to unambiguously distinguish distinct chemical signatures in the Nd isotopes. Nd is produced more by the than the , though, so Nd lines may be easier to observe in stars with an chemical signature than either Eu or Sm.
Of course these suggestions only apply to the examination of the nucleosynthetic history at the isotopic level. At the elemental level Eu should continue to serve as an excellent barometer of nucleosynthesis, and Ba or La should continue to serve as barometers of nucleosynthesis.
Yields from Models of Low-Metallicity Stars on the AGB {#lowzagb}
-------------------------------------------------------
While the classical method of modeling the is, by definition, model-independent, it has been shown to predict an overproduction of $^{142}$Nd relative to other isotopes of Nd and pure-$s$ nuclei, e.g., as determined from studies of the Murchison meteorite [@zinner91; @guber97]. This isotope is only produced by the and contains 82 neutrons, one of the magic neutron numbers, making this nucleus relatively more stable than other nuclei with similar mass numbers and thus violating the classical method’s assumption of a smoothly-varying $\sigma N_{s}$ curve. The complex stellar model of @arlandini99 can avoid this difficulty near the magic neutron numbers.
Of the much theoretical and computational work that has been performed in recent years to better understand the structure and evolution of low-metallicity stars on the AGB, one goal has been to provide a complete set of chemical yields from H to Bi, at the termination of the . The first such set of yields was presented by @cristallo06, for 2M$_{\sun}$ stars on the AGB with $Z=1.5\times10^{-2}$, $1\times10^{-3}$, and $1\times10^{-4}$. Similar models surely will follow, and the observational verification for these models will rely heavily on isotopic abundances as well as elemental abundances, particularly for magic neutron number nuclei such as $^{142}$Nd. The interpretation of nucleosynthetic signatures of the first generations of stars will depend heavily on the reliability of these models. As these models mature and are (hopefully) shown to be reliable predictors of nucleosynthesis at low metallicity, they will provide insight into the structure and evolution of stars on the AGB that cannot be obtained from the classical ad hoc approach to understanding the .
The Role of Future Studies of Neutron-Capture Isotopic Fractions {#bigtelescopes}
----------------------------------------------------------------
We suggest that the features which are strongly blended with the lines in our stars (in particular, Ti, V, Fe, and Ni) may be somewhat diminished in very metal-poor, enhanced stars. The requirements of high spectral resolution and high S/N to clearly discern isotopic fractions are currently difficult to achieve for the numbers of such stars that have been detected by the HK Survey [@beers85; @beers92] and the Hamburg/ESO Survey [@wisotzki00; @christlieb03], owing to the faintness of these stars. From a technical outlook, observations of this nature will become more feasible with high-resolution echelle spectrographs on the next generation of 20$-$30m telescopes. Knowledge of the isotopic composition of these stars would complement the information gleaned from elemental abundances alone and further enhance our understanding of stellar nucleosynthesis and chemical evolution throughout the early history of our Galaxy.
Conclusions
===========
We have successfully measured the isotopic fractions of Eu and Sm in HD 175305, and we have assessed the Nd isotopic fraction as well. We find that the Sm isotopic fraction is suggestive of an origin. The Eu and Nd isotopic fractions are unable to distinguish between an $r$- or origin. Along with measurements of elemental abundances of species, our Sm isotopic fraction reinforces the assertion that HD 175305 has been enriched by nucleosynthesis. Both the elemental and isotopic abundance distributions in HD 175305 join the growing preponderance of evidence that supports the hypothesis of a universal mechanism for elements with $Z\geq56$.
We have measured the isotopic fraction of Sm in HD 196944, which suggests an origin. The Nd isotopic fraction is unable to distinguish between the two processes. The Sm isotopic fraction complements previous elemental abundance measurements and observed radial velocity variations of this star to confirm that the enrichment pattern of HD 196944 is consistent with transfer of material from a companion star in the AGB phase of stellar evolution.
We suggest that measurements of the Sm isotopic fraction will be more advantageous than measurements of the Eu isotopic fraction alone. The Sm isotopic fraction $f_{152+154}$ permits a clearer distinction of pure-$s$- and pure-$r$-process content than Eu isotopic fractions can. There often are fewer blending features in the spectral regions where the useful Sm lines are located, and there are more Sm lines available for analysis. Nd lines may offer more opportunities to measure an isotopic fraction in $s$-process-enriched stars than either Eu or Sm, although it is extremely challenging to discern the Nd isotopic fraction. We propose that measurements of isotopic fractions of multiple species in the same metal-poor star can allow direct quantification of the relative $s$- and contributions to material, enable comparison of these contributions as a function of increasing nuclear mass number, test yields of low-metallicity stars on the AGB, and provide tests of nucleosynthesis predictions. We have conducted one such test for $^{151}$Eu using Eu and Pb abundances collected from the literature, finding that the @arlandini99 stellar model predictions are a better match than the classical method predictions for $^{151}$Eu, although this conclusion hinges on measurements of the Eu isotopic fraction in only two $(r+s)$-enriched stars.
Based on our results, we argue that the rare earth $s$- and abundance patterns, which are observed in metal-poor stars at the elemental abundance level, are also present in the isotopic fractions. This implies that our understanding of nucleosynthesis is not wildly mistaken. This result has often been tacitly assumed in studies of nucleosynthesis and chemical abundance patterns in the Galactic halo, and we contend that increasing empirical evidence for multiple rare earth species now exists to support this hypothesis.
We are pleased to acknowledge the following individuals for their encouraging and helpful discussions: Carlos Allende Prieto, Jacob Bean, Anna Frebel, David Lambert, Martin Lundqvist, Rob Robinson, and Glenn Wahlgren. We thank the referee for providing a number of insightful comments on this manuscript. I. U. R. wishes to thank the McDonald Observatory staff for their hospitality and their dedication to the observatory, its visitors, and the West Texas community. This research has made use of the NASA Astrophysics Data System (ADS), NIST Atomic Spectra Database, Vienna Atomic Line Database (VALD; @kupka99), Two Micron All-Sky Survey (2MASS), and SIMBAD databases. The reliability and accessibility of these online databases is greatly appreciated. Funding for this project has been generously provided by the U. S. National Science Foundation (grant AST 03-07279 to J. J. C., grant AST 05-06324 to J. E. L., and grants AST 03-07495 and AST 06-07708 to C. S.) and by the Sigma Xi *Grants-in-Aid-of-Research* program.
Comments on Individual Transitions {#comments}
==================================
Here we describe the individual characteristics of our lines of interest as well as blending features. Where possible, we have tried to incorporate the highest quality atomic data for these blending features, including log($gf$) values calculated from laser-induced fluorescence measurements of the radiative level lifetimes and branching fractions measured from Fourier transform spectrometry. We alert the reader that many of the studies published before the early 1980s used less precise methods, and we employ those values with some caution.
Europium Lines {#eulines}
--------------
The Eu <span style="font-variant:small-caps;">ii</span> line at 4129Å is a clean, relatively unblended line in HD 175305; this spectral region was not observed in our other two stars. The only significant blend is a Fe <span style="font-variant:small-caps;">i</span> line at 4129.4611[^12]Å in the far blue wing; the wavelength of this line was taken from @nave94 and the log($gf$) value was taken from @fuhr05, who scaled the value found by @may74. We note that these values are not critical to determining the isotopic fraction of Eu.
The Eu <span style="font-variant:small-caps;">ii</span> line at 4205Å is strongly blended with CH (and, to a lesser extent, CN) features, two V <span style="font-variant:small-caps;">ii</span> lines, and a Y <span style="font-variant:small-caps;">ii</span> line. We adopt a CH linelist from the Kurucz database and a CN linelist from the Plez database[^13]. Precise laboratory wavelengths and log($gf$) measurements are not available for the V <span style="font-variant:small-caps;">ii</span> line at 4205.04Å; we adopt the V <span style="font-variant:small-caps;">ii</span> 4205.084Ålog($gf$) value from @biemont89. A log($gf$) value for the Y <span style="font-variant:small-caps;">ii</span> line at 4204.694Å can be calculated from the oscillator strength presented in @hannaford82. Performing a synthesis of this region of the Solar spectrum using an interpolated empirical @holweger74 model atmosphere, we find rather poor agreement with the @kurucz84 Solar flux spectrum, with the most significant discrepancies arising from the CH bands and the wavelength of the Y <span style="font-variant:small-caps;">ii</span> line. One would hope that in a very metal-poor star with sub-solar C abundance (HD 122563) these effects would be greatly diminished; the situation is much less hopeful for a C-enhanced star (HD 196944). Yet given the extent and severity of the blends of this Eu line, as well as its weakness in HD 122563, we are reluctant to trust any of the isotope fractions derived from this line. This is unfortunate because only this Eu line is strong enough to permit measurement of the elemental abundance in our observed spectra in HD 122563 and HD 196944.
The Eu <span style="font-variant:small-caps;">ii</span> line at 4435Å is strongly blended with a Ca <span style="font-variant:small-caps;">i</span> line at 4435.67Å and mildly blended with a Ni <span style="font-variant:small-caps;">i</span> line at 4435.33Å. In HD 122563 and HD 196944, this Eu line is only a small sliver on the blue wing of the Ca line, and therefore no reasonable assessment of the elemental abundance or isotopic fraction can be made. In HD 175305 the Ca line is saturated, and we cannot produce a satisfactory fit to the blended profile no matter what log($gf$) value we adopt for this line. We adjust the log($gf$) value of the Ni line to match the Solar spectrum, yet an additional adjustment of $+$0.65 dex to the Solar log($gf$) value is necessary to bring this line into rational agreement with our observed spectrum. Even then we are not satisfied with the fit of the blue wing of the Eu line. Consequently, because such a small region of the spectrum covered by the Eu line remains unblended, we also disregard this line in our measurement of the isotopic fraction.
Samarium Lines {#smlines}
--------------
The Sm <span style="font-variant:small-caps;">ii</span> line at 4424.34Å is relatively strong and is blended only slightly with a Cr <span style="font-variant:small-caps;">i</span> line at 4424.28Å. An unpublished log($gf$) value for this line, $-$0.63, was measured by @sobeck07. We measure Sm isotopic fractions from this line in HD 175305 and HD 196944; the line is too weak to measure well in HD 122563.
The Sm <span style="font-variant:small-caps;">ii</span> line at 4467.34Å is also relatively strong and blended slightly with a weak Fe <span style="font-variant:small-caps;">i</span> line at 4467.43Å and an unidentified (in, e.g., the @moore66 atlas) line at 4467.21Å. No experimental log($gf$) value is available for the Fe line, but we use an inverted Solar analysis to derive a log($gf$) of $-$2.92 for this line. The unidentified feature can be reasonably modeled as an Fe line. The isotope shifts of this line are almost negligible and the HFS is also small, but @ekeland75 used it in their analysis of the Sm isotopic fraction in the Sun This line is not covered in any of our stellar spectra, but we do use it to revisit the Sm isotopic fraction in the Solar photosphere.
The Sm <span style="font-variant:small-caps;">ii</span> 4591.81Å line lies near the edge of the blue wing of a strong Cr <span style="font-variant:small-caps;">i</span> line at 4592.05Å that is easily accounted for in the synthesis. Of greater concern, however, is an apparently unidentified blend at 4591.73Å in HD 175305. We do not measure an isotopic fraction from this line but suggest that it might be useful in future studies. In HD 122563 this line just falls off an echelle order, and in HD 196944 this line is too weak to observe.
The Sm <span style="font-variant:small-caps;">ii</span> 4593.53Å line is affected by a very weak Fe <span style="font-variant:small-caps;">i</span> line at 4593.53Å, whose log($gf$) value is given by @fuhr05, which has been scaled from @may74. We also notice a small, unidentified blend in the blue wing of this line at 4593.42Å; a close examination of @lundqvist07’s spectrum of CS 31082-001 in Figure 5 shows a similar feature. This Sm line is too weak to be observed in HD 122563, is observed but is too weak to assess the isotopic fraction in HD 196944, and, regrettably, in HD 175305 this line fell on a piece of the CCD where we encountered unrecoverable flat-fielding errors during the reduction process.
The Sm <span style="font-variant:small-caps;">ii</span> 4595.28Å line is severely blended with a strong Fe <span style="font-variant:small-caps;">i</span> line at 4595.3591Å. The log($gf$) value for this Fe line can be computed from the data in @obrian91, but we are hesitant to believe any Sm abundances derived from this line in any of our stars.
The Sm <span style="font-variant:small-caps;">ii</span> line at 4604.17Å is promising. A Fe <span style="font-variant:small-caps;">i</span> line at 4603.949Åaffects only the bluemost wing of this line, and a weak Fe <span style="font-variant:small-caps;">i</span> line blends with it at 4604.24Å; a log($gf$) value is available for the latter of these two Fe lines in @gurtovenko81 and the log($gf$) for the former is derived from an inverted Solar analysis ($-$3.15), which provides a far better fit than the log($gf$) rated an “E” in @fuhr05. We measure an isotopic fraction for this line in HD 175305, but the line is too weak in the other two stars.
The Sm <span style="font-variant:small-caps;">ii</span> 4693.63Å line is blended with a Ti <span style="font-variant:small-caps;">i</span> line at 4693.67Å. Using our Ti <span style="font-variant:small-caps;">i</span> abundance and the log($gf$) value for this line from @kuhne78, we cannot produce a satisfactory fit for the Sm line. An additional unidentified blend occurs at 4693.75Å in our spectrum of HD 175305. Our spectra of HD 122563 and HD 196944 do not cover this region. We do not make a measurement of the elemental abundance or isotopic fraction of Sm from this line in HD 175305, but we suggest that this line could be a candidate for future Sm isotopic analyses.
@lundqvist07 note that the Sm <span style="font-variant:small-caps;">ii</span> line at 4715.27Å is blended with a Ti <span style="font-variant:small-caps;">i</span> line at 4715.230Å. A log($gf$) value is available in @kuhne78. This line was not covered in any of our spectra, but isotopic analysis of this Sm line may be possible.
The Sm <span style="font-variant:small-caps;">ii</span> line at 4719.84Å is covered in all three of our spectra, but is only strong enough to measure in HD 175305 and HD 196944. Full sets of hyperfine A and B constants are not available for this line, so we compute the profile using only the isotope shifts measured by @lundqvist07. This line is blended with a weak La <span style="font-variant:small-caps;">ii</span> line at 4719.92Å; this La line was not included in the analysis of @lawler01a. While the Sm line is much stronger than the La line in both of our spectra, the uncertainty in the La log($gf$) value and the lack of HFS structure information for the Sm line lead us to disregard this line in our isotopic analysis.
The Sm <span style="font-variant:small-caps;">ii</span> line at 5052.75Å is not covered in any of our spectra. Full sets of hyperfine A and B constants are not available for this line, either, but we compute a synthesis using the isotope shifts only. This line is blended with a Ti <span style="font-variant:small-caps;">i</span> line at 5052.87Å and a Fe <span style="font-variant:small-caps;">i</span> line at 5052.9814Å. A log($gf$) value is listed for the Ti line in the catalog of @savanov90, and a wavelength for the Fe line is given in @nave94. We suggest that these blends may be able to be accounted for even without precise atomic data, and the isotope shifts alone could provide some indication of the Sm isotopic fraction in metal-poor stars, but measurement of an exact isotopic fraction is probably unwarranted here.
The weak Sm <span style="font-variant:small-caps;">ii</span> line at 5069.47Å is blended with a weak Fe <span style="font-variant:small-caps;">i</span> line at 5069.4233Å, which is blended with a weak Ti <span style="font-variant:small-caps;">i</span> line at 5069.35Å. The wavelength for this Fe line is listed in @nave94, but we cannot find an experimental log($gf$) value for this line, and the Ti blend prevents us from using an inverted Solar analysis to derive a log($gf$) value for either line. We observe the Sm line in our spectra of HD 122563 and HD 196944; however, in both cases the Sm line is weak and the observed spectrum is littered with small, undocumented blends in addition to the Fe line noted here. We recommend that any attempts to use this line for Sm isotopic analysis proceed with great caution.
The Sm <span style="font-variant:small-caps;">ii</span> line at 5103.09Å is weak and lies on the red edge of a strong Ni <span style="font-variant:small-caps;">i</span> line at 5102.964Å. A log($gf$) value for the Ni line is given in @doerr85, and we have set the wavelength of this line from analysis of the Solar spectrum. Full sets of hyperfine A and B constants are not available for this line, yet we compute a synthetic spectrum using only the isotope shifts. We can detect the Sm line in HD 175305, but it is so overwhelmed by the Ni line that we cannot trust even our derived elemental abundance. This Sm line is too weak to detect in HD 122563 and HD 196944.
Our redmost Sm <span style="font-variant:small-caps;">ii</span> line lies at 5104.48Å. It is blended by a Fe <span style="font-variant:small-caps;">i</span> line at 5104.4375Å. @nave94 quote a precise wavelength for this Fe line, and its log($gf$) value is listed in @fuhr05, who scaled @may74’s value. When we synthesize this line in HD 175305, we are forced to enhance the Sm abundance by more than a factor of 2. We have re-evaluated the log($gf$) value of the Fe blend by comparing with the Solar spectrum, and find that we can reasonably only increase @fuhr05’s log($gf$) value by $\approx$0.2 dex, which does little to provide a better match between our observed and synthetic spectra. In addition, there is a significant unidentified blend at 5104.59Å in the red wing of our Sm line. While it is clear that the Sm line is broadened by HFS and isotopic shifts, we unfortunately are unable to measure an isotopic fraction. This Sm line is too weak to detect in HD 122563 and is not covered in our spectrum of HD 196944.
Neodymium Lines {#ndlines}
---------------
The Nd <span style="font-variant:small-caps;">ii</span> line at 4177.32Åis relatively strong, and only its far red wing is blended with a strong Y <span style="font-variant:small-caps;">ii</span> and Fe <span style="font-variant:small-caps;">i</span> blend. A log($gf$) value can be calculated for the Y <span style="font-variant:small-caps;">ii</span> line at 4177.536Å from the $f$-value in @hannaford82, and a log($gf$) value for the Fe <span style="font-variant:small-caps;">i</span> line at 4177.5935Å is given in @obrian91. This blend is easily accounted for and has little effect on the derived Nd isotopic fraction. This Nd line was not covered in our spectra of HD 122563 and HD 196944, but we measure an isotopic fraction from this line in HD 175305.
The Nd <span style="font-variant:small-caps;">ii</span> line at 4232.38Å has two identified blends. The Hf <span style="font-variant:small-caps;">ii</span> line at 4232.386Å lies coincident with our Nd line; a log($gf$) value for this line was measured by @lundqvist06. A V <span style="font-variant:small-caps;">i</span> line at 4232.46Å also blends with our Nd line, and a log($gf$) value is provided by @martin88 in the NIST database. The uncertainties in the abundances and transition probabilities of these two lines, as well as their severity of blending with our Nd feature, lead us to disregard this line in our spectrum of HD 175305. This line was not covered in our spectra of HD 122563 and HD 196944. The isotope shifts of this Nd line are noticeable, though, so we report the hyperfine patterns for this line nonetheless.
The Nd <span style="font-variant:small-caps;">ii</span> line at 4314.51Å lies in a region where continuum placement is difficult due to the presence of the CH G-band. We have not attempted to measure the Nd isotopic fraction from this line in any of our stars, although the Nd line is visible in HD 175305 and HD 196944. The isotope shifts of this line are rather large ($\sim$ 0.035Å), so we report the hyperfine patterns for this line, in hopes that the Nd isotopic fraction could be discerned from this line in a favorable star. @denhartog03 did not measure a log($gf$) value for this line, but we derive log($gf$)$\,=\,-0.22$ from an inverted Solar analysis, assuming log$\,\varepsilon\,$(Nd)$_{\sun}\,=\,1.50$.
The Nd <span style="font-variant:small-caps;">ii</span> line at 4358.16Å is relatively strong and unblended in HD 175305, and we use it to measure an isotopic fraction. This line was not covered in our spectra of HD 122563 and HD 196944.
The Nd <span style="font-variant:small-caps;">ii</span> line at 4446.38Å has the largest isotope shifts of any of the Nd transitions for which we report an isotopic fraction. In HD 175305, this line is blended only with a weak Gd <span style="font-variant:small-caps;">ii</span> line at 4446.502Å in the red wing, and @denhartog06 provide a log($gf$) value for this line. In HD 196944, some unidentified blends are observed in both the red and blue wings of this line, but it does not appear that these blends affect our measurement of the isotope fraction. We measure an isotopic fraction from this line in each of these two stars. This line is observed in our high-S/N spectrum of HD 122563, but it has a continuum depth of only $\sim$ 1.5%, and we cannot measure an isotopic fraction from this line.
The Nd <span style="font-variant:small-caps;">ii</span> line at 4567.61Å has a depth of only 4% in HD 175305, and an isotopic fraction can only marginally be deduced from this line. We identify no blending features, although our observed spectrum clearly shows some weaker blends in the wings of this line. This line was not covered in our observed spectra of HD 122563 and HD 196944.
Hyperfine Component Data for Samarium {#smappendix}
=====================================
We present hyperfine and isotopic components for the 13 lines of Sm <span style="font-variant:small-caps;">ii</span> described in § \[smlines\] in Table \[tab4-stub\]. Hyperfine A (magnetic dipole) and B (electric quadrupole) constants, which govern the relative positions of the hyperfine components, along with isotope shifts, were taken first from @masterman03 if available. Constants for additional levels were taken from @dorschel81, @young87, and @villemoes95. We give preference to the hyperfine constants and isotope shifts determined from radio frequency measurements or with a single-frequency laser, as in @masterman03 and @beiersdorf95, supplemented with isotope shifts measured from FTS spectra in @lundqvist07 for lines not covered in earlier studies.
The relative strengths are calculated from the LS (Russell-Saunders) angular momentum coupling formulae (e.g., @condon53 [p. 238]), where we have replaced the orbital angular momentum quantum number ($L$) by the total electronic angular momentum ($J$), the electron spin ($S$) by the nuclear spin ($I$), and the total electronic angular momentum ($J$) by the total atomic angular momentum ($F$). Thus, $F=I+J$, and the energy of the $J$ level is split into a number of sublevels given by $F$, which runs from $|I-J|$ to $I+J$. The energy spacing between each component is proportional to $F$. The nucleus may also have an electric quadrupole moment, which produces an additional shift (but not splitting) in the energy levels. The strongest relative components are those for which $\Delta\, F\,=\,\Delta\,J$. We normalize the strengths for each line such that they sum to one. Absolute transition wavenumbers are from @lundqvist07 if the lines were included in their study, otherwise they are computed using the energy levels tabulated by @martin78. The center-of-gravity transition wavenumbers and air wavelengths are given in the first two columns, respectively. Component wavenumbers are converted to air wavelengths according to the formula given by @edlen53.
Hyperfine Component Data for Neodymium {#ndappendix}
======================================
We present hyperfine and isotopic components for the 6 lines of Nd <span style="font-variant:small-caps;">ii</span> described in § \[ndlines\] in Table \[tab5-stub\]. Hyperfine A and B constants and isotope shifts were taken from @rosner05. All calculations were performed analogously to the Sm case. Center-of-gravity transition wavenumbers for the lines of interest were taken from FTS measurements by @blaise84.
Allen, D. M., & Barbuy, B. 2006, , 454, 895
Allen, D. M., & Barbuy, B. 2006, , 454, 917
Alonso, A., Arribas, S., & Mart[í]{}nez-Roger, C. 1996, , 313, 873
Alonso, A., Arribas, S., & Mart[í]{}nez-Roger, C. 1999, , 140, 261
Anders, E., & Grevesse, N. 1989, , 53, 197
Aoki, W., et al. 2001, , 561, 346
Aoki, W., Ryan, S. G., Norris, J. E., Beers, T. C., Ando, H., & Tsangarides, S. 2002, , 580, 1149
Aoki, W., Honda, S., Beers, T. C., & Sneden, C. 2003a, , 586, 506
Aoki, W., et al. 2003b, , 592, L67
Aoki, W., et al. 2005, , 632, 611
Aoki, W., Bisterzo, S., Gallino, R., Beers, T. C., Norris, J. E., Ryan, S. G., & Tsangarides, S. 2006, , 650, L127
Aoki, W., et al. 2007, , 660, 747
Arlandini, C., K[ä]{}ppeler, F., Wisshak, K., Gallino, R., Lugaro, M., Busso, M., & Straniero, O. 1999, , 525, 886
Asplund, M., Nordlund, A., Trampedach, R., Allende Prieto, C., & Stein, R. F., 2000, , 359, 729
Asplund, M. 2004, , 75, 300
Asplund, M., Lambert, D. L., Nissen, P. E., Primas, F., & Smith, V. V. 2006, , 644, 229
Barbuy, B., et al. 2003, , 588, 1072
Barbuy, B., Spite, M., Spite, F., Hill, V., Cayrel, R., Plez, B., & Petitjean, P. 2005, , 429, 1031
Barklem, P. S., et al. 2005, , 439, 129
Beers, T. C., Preston, G. W., & Shectman, S. A. 1985, , 90, 2089
Beers, T. C., Preston, G. W., & Shectman, S. A. 1992, , 103, 1987
Beers, T. C., & Christlieb, N. 2005, , 43, 531
Beiersdorf, S., Heddrich, W., Kesper, K., Hühnermann, H., Möller, W., & Wagner, H.1995, J. Phys. G: Nucl. Part. Phys. 21, 215
Biémont, E., Grevesse, N., Faires, L. M., Marsden, G., & Lawler, J. E. 1989, , 209, 391
Bi[é]{}mont, E., Garnir, H. P., Palmeri, P., Li, Z. S., & Svanberg, S. 2000, , 312, 116
Blaise, J., Wyart, J. F., Djerad, M. T., & Ahmed, Z. B. 1984, Phys. Scr., 29, 119
Bond, H. E. 1980, , 44, 517
Brandt, H.-W., Mei[ß]{}ner, E., & Steudel, A. 1979, Z. Physik A, 291, 97 Brown, J. A., Wallerstein, G., Cunha, K., & Smith, V. V. 1991, , 249, L13
Burris, D. L., Pilachowski, C. A., Armandroff, T. E., Sneden, C., Cowan, J. J., & Roe, H. 2000, , 544, 302
Busso, M., Gallino, R., & Wasserburg, G. J. 1999, , 37, 239
Carney, B. W., Latham, D. W., Stefanik, R. P., Laird, J. B., & Morse, J. A. 2003, , 125, 293
Cayrel, R., et al. 2007, , 473, L37
Cescutti, G., Fran[c c]{}ois, P., Matteucci, F., Cayrel, R., & Spite, M. 2006, , 448, 557
Christlieb, N. 2003, Reviews in Modern Astronomy, 16, 191
Clayton, D. D., Fowler, W. A., Hull, T. E., & Zimmermann, B. A. 1961, Ann. Phys., 12, 331
Cohen, J. G., Christlieb, N., Qian, Y.-Z., & Wasserburg, G. J. 2003, , 588, 1082
Cohen, J. G., et al. 2006, , 132, 137
Condon, E. U., & Shortley, G. H. 1953, The Theory of Atomic Spectra (Cambridge: Cambridge Univ. Press), 238
Cowan, J. J., Pfeiffer, B., Kratz, K.-L., Thielemann, F.-K., Sneden, C., Burles, S., Tytler, D., & Beers, T. C. 1999, , 521, 194
Cowan, J. J., et al. 2002, , 572, 861
Cowan, J. J., & Thielemann, F.-K. 2004, Physics Today, 57, 47
Cowan, J. J., et al. 2005, , 627, 238
Cowan, J. J., Lawler, J. E., Sneden, C., Den Hartog, E. A., & Collier, J. 2006, in Proc. NASA LAW, ed. V. Kwong, (Washington: NASA), 82
Cowan, J. J., & Sneden, C. 2006, , 440, 1151
Cowley, C. R., & Corliss, C. H. 1983, , 203, 651
Cowley, C. R., & Frey, M. 1989, , 346, 1030
Cristallo, S. 2006, Ph.D. Thesis, Osservatorio Astronomico di Collurania
Den Hartog, E. A., Lawler, J. E., Sneden, C., & Cowan, J. J. 2003, , 148, 543
Den Hartog, E. A., Lawler, J. E., Sneden, C., & Cowan, J. J. 2006, , 167, 292
Doerr, A., & Kock, M. 1985, , 33, 307
Dörschel, K., Hühnermann, H, Knobl, E., Meier, Th., & Wagner, H. 1981, Z. Phys. A – Atoms and Nuclei, 302, 359
Dravins, D., Lindegren, L., & Nordlund, A. 1981, , 96, 345
Dravins, D., Larsson, B., & Nordlund, A. 1986, , 158, 83
Dravins, D. 1987, , 172, 211
Edvardsson, B., Andersen, J., Gustafsson, B., Lambert, D. L., Nissen, P. E., & Tomkin, J. 1993, , 275, 101
Edlén, B. 1953, J. Opt. Soc.Am., 43, 339
Ekeland, A., & Hauge, Ø. 1975, , 42, 17
Fitzpatrick, M. J., & Sneden, C. 1987, , 19, 1129
François, P., et al. 2007, , doi:10.1051/0004-6361
Frebel, A., Christlieb, N., Norris, J. E., Thom, C., Beers, T. C., & Rhee, J. 2007, , 660, L117
Fuhr, J. R., & Wiese, W. L.2005, J. Phys. Chem. Ref. Data, 35, 1669
Fulbright, J. P. 2000, , 120, 1841
Fulbright, J. P. 2002, , 123, 404
Gallino, R., Arlandini, C., Busso, M., Lugaro, M., Travaglio, C., Straniero, O., Chieffi, A., & Limongi, M. 1998, , 497, 388
Gallino, R., Delaude, D., Husti, L., Cristallo, S., Straniero, O., & Ryan, S. 2005, Nuclear Physics A, 758, 485
Gratton, R. G., & Sneden, C. 1994, , 287, 927
Gray, D. F. 2005, The Observation and Analysis of Stellar Photospheres, 3rd Edition.Cambridge, UK: Cambridge University Press
Grevesse, N., & Sauval, A. J. 2002, Adv. Space Res., 30, 3
Guber, K. H., Spencer, R. R., Koehler, P. E., & Winters, R. R. 1997, Phys. Rev. Lett., 78, 2704
Gurtovenko, E. A., & Kostik, R. I. 1981, , 46, 239
Hannaford, P., Lowe, R. M., Grevesse, N., Biemont, E., & Whaling, W. 1982, , 261, 736
Hauge, [Ö]{}. 1970, , 11, 17
Hauge, [Ø]{}. 1972, , 27, 286
Hill, V., et al. 2002, , 387, 560
Holweger, H., & Müller, E. A. 1974, , 39, 19
Honda, S., et al. 2004a, , 152, 113
Honda, S., Aoki, W., Kajino, T., Ando, H., Beers, T. C., Izumiura, H., Sadakane, K., & Takada-Hidai, M. 2004b, , 607, 474
Honda, S., Aoki, W., Ishimaru, Y., Wanajo, S., & Ryan, S. G. 2006, , 643, 1180
Honda, S., Aoki, W., Ishimaru, Y., & Wanajo, S. 2007, , 666, 1189
Ivans, I. I., Sneden, C., Gallino, R., Cowan, J. J., & Preston, G. W. 2005, , 627, L145
Ivans, I. I., Simmerer, J., Sneden, C., Lawler, J. E., Cowan, J. J., Gallino, R., & Bisterzo, S. 2006, , 645, 613
Johnson, J. A. 2002, , 139, 219
Johnson, J. A., & Bolte, M. 2002, , 579, 616
Johnson, J. A., & Bolte, M. 2002, , 579, L87
Johnson, J. A., Herwig, F., Beers, T. C., & Christlieb, N. 2007, , 658, 1203
Jonsell, K., Edvardsson, B., Gustafsson, B., Magain, P., Nissen, P. E., & Asplund, M. 2005, , 440, 321
Jonsell, K., Barklem, P. S., Gustafsson, B., Christlieb, N., Hill, V., Beers, T. C., & Holmberg, J. 2006, , 451, 651
Käppeler, F., Beer, H., & Wisshak, K. 1989, Rep. Prog. Phys., 52, 945
Klose, J. Z., Fuhr, J. R., & Wiese, W. L. 2002, J. Phys. Chem. Ref. Data, 31, 217
Kratz, K.-L., Farouqi, K., Pfeiffer, B., Truran, J. W., Sneden, C., & Cowan, J. J. 2007, , 662, 39
Kühne, M., Danzmann, K., & Kock, M. 1978, , 64, 111
Kupka, F., Piskunov, N., Ryabchikova, T. A., Stempels, H. C., & Weiss, W. W. 1999, , 138, 119
Kurucz, R. L., Rurenlid, I., Brault, J., & Testerman, L. 1984, Solar Flux Atlas from 296 to 1300 nm, National Solar Observatory Atlas No. 1 (Harvard University)
Kurucz, R. L. 1993, Kurucz CD-ROM 13, ATLAS9 Stellar Atmosphere Programs and 2 km/s Grid (Cambridge: SAO)
Kurucz, R. L., & Bell, B. 1995, CD-ROM, Cambridge, MA: Smithsonian Astrophysical Observatory
Lambert, D. L., & Sneden, C. 1977, , 215, 597
Lambert, D. L., & Allende Prieto, C. 2002, , 335, 325
Lawler, J. E., Bonvallet, G., & Sneden, C. 2001a, , 556, 452 Lawler, J. E., Wickliffe, M. E., den Hartog, E. A., & Sneden, C. 2001b, , 563, 1075
Lawler, J. E., Den Hartog, E. A., Sneden, C., & Cowan, J. J. 2006, , 162, 227
Lawler, J. E., Den Hartog, E. A., Labby, Z. E., Sneden, C., Cowan, J. J., & Ivans, I. I. 2007, , 169, 120
Learner, R. C. M., & Thorne, A. P. 1988, J. Opt. Soc. Am. B, 5, 2045
Lodders, K. 2003, , 591, 1220
Lucatello, S., Tsangarides, S., Beers, T. C., Carretta, E., Gratton, R. G., & Ryan, S. G. 2005, , 625, 825
Lundqvist, M., Nilsson, H., Wahlgren, G. M., Lundberg, H., Xu, H. L., Jang, Z.-K., & Leckrone, D. S. 2006, , 450, 407
Lundqvist, M., Wahlgren, G. M., & Hill, V. 2007, , 463, 693
Magain, P., & Zhao, G. 1993a, Origin and evolution of the elements: proceedings of a symposium in honor of H. Reeves, held in Paris, June 22-25, 1992. Edited by Prantzos, N., Vangioni-Flam, E., & Casse, M. Published by Cambridge University Press, Cambridge, England, 1993, 480
Magain, P., & Zhao, G. 1993b, , 268, L27
Magain, P. 1995, , 297, 686
Martin, W. C., Zalubas, R., & Hagan, L. 1978, NSRDS-NBS, Washington: National Bureau of Standards, U.S. Department of Commerce, |c1978, p. 174
Martin, G. A., Fuhr, J. R., & Wiese, W. L. 1988, J. Phys. Chem. Ref. Data, 17, Suppl. 3
Mashonkina, L., Gehren, T., & Bikmaev, I. 1999, , 343, 519
Mashonkina, L., Gehren, T., Travaglio, C., & Borkova, T. 2003, , 397, 275
Mashonkina, L. I., Kamaeva, L. A., Samotoev, V. A., & Sakhibullin, N. A. 2004, Astronomy Reports, 48, 185
Mashonkina, L., & Zhao, G. 2006, , 456, 313
Masseron, T., et al. 2006, , 455, 1059
Masterman, D., Rosner, S. D., Scholl, T. J., Sharikova, A., & Holt, R. A. 2003, Can. J. Phys., 81, 1389
May, M., Richter, J., & Wichelmann, J. 1974, , 18, 405
McWilliam, A. 1998, , 115, 1640
Meyer, B. S., & Clayton, D. D. 2000, , 92, 133
Moore, C. E., Minnaert, M. G. J., & Houtgast, J. 1966, National Bureau of Standards Monograph, Washington: US Government Printing Office (USGPO), 1966,
Nave, G., Johansson, S., Learner, R. C. M., Thorne, A. P., & Brault, J. W. 1994, , 94, 221
Nordstr[ö]{}m, B., et al. 2004, , 418, 989
Norlén, G. 1973, , 8, 249
O’Brian, T. R., Wickliffe, M. E., Lawler, J. E., Whaling, W., & Brault, J. W. 1991, J. Opt. Soc. Am. B, 8, 1185
Palmeri, P., Quinet, P., Wyart, J.-F., & Biémont, E. 2000, , 61, 323
Pickering, J. C., Thorne, A. P., & Perez, R. 2001, , 132, 403
Pickering, J. C., Thorne, A. P., & Perez, R. 2002, , 138, 247
Plez, B., et al. 2004, , 428, L9
Press, W. H., Teuk lsky, S. A., Vetterling, W. T., & Flannery, B. P. 1992, Numerical Recipes in FORTRAN: The Art of Scientific Computing, (2nd ed.; Cambridge: University Press), 655
Preston, G. W., Thompson, I. B., Sneden, C., Stachowski, G., & Shectman, S. A. 2006, , 132, 1714
Qian, Y.-Z., & Wasserburg, G. J. 2000, , 333, 77
Ram[í]{}rez, I., & Mel[é]{}ndez, J. 2005, , 626, 465
Rao, P. M., Ahmad, S. A., Venugopalan, A., & Saksena, G. D., 1990, Z. Phys. D – Atoms, Molecules, and Clusters, 15, 211
Rosman, K. J. R., & Taylor, P. D. P. 1998, Pure Appl. Chem., 70, 217
Rosner, S. D., Masterman, D., Scholl, T. J., & Holt, R. A. 2005, Can. J. Phys., 83, 841
Savanov, I. S. 1990, 3rd International Collogium of the Royal Netherlands Academy of Arts and Sciences, 165
Seeger, P. A., Fowler, W. A., & Clayton, D. D. 1965, , 11, 121
Short, C. I., & Hauschildt, P. H. 2006, , 641, 494
Simmerer, J., Sneden, C., Cowan, J. J., Collier, J., Woolf, V. M., & Lawler, J. E. 2004, , 617, 1091
Skrutskie, M. F., et al. 2006, , 131, 1163
Smith, V. V., Vargas-Ferro, O., Lambert, D. L., & Olgin, J. G. 2001, , 121, 453
Sneden, C. A. 1973, Ph.D. Thesis, Univ. Texas at Austin
Sneden, C., Kraft, R. P., Prosser, C. F., & Langer, G. E. 1991, , 102, 2001
Sneden, C., Kraft, R. P., Langer, G. E., Prosser, C. F., & Shetrone, M. D. 1994, , 107, 1773
Sneden, C., Cowan, J. J., Lawler, J. E., Burles, S., Beers, T. C., & Fuller, G. M. 2002, , 566, L25
Sneden, C., et al. 2003, , 591, 936
Sobeck, J. S., Lawler, J. E., & Sneden, C. 2007, , 667, 1267
Spite, M., & Spite, F. 1978, , 67, 23
Straniero, O., Gallino, R., & Cristallo, S. 2006, Nuclear Physics A, 777, 311
Travaglio, C., Galli, D., Gallino, R., Busso, M., Ferrini, F., & Straniero, O. 1999, , 521, 691
Travaglio, C., Galli, D., & Burkert, A. 2001, , 547, 217
Travaglio, C., Gallino, R., Arnone, E., Cowan, J., Jordan, F., & Sneden, C. 2004, , 601, 864
Tull, R. G., MacQueen, P. J., Sneden, C., & Lambert, D. L. 1995, , 107, 251
Van Eck, S., Goriely, S., Jorissen, A., & Plez, B. 2001, , 412, 793
Van Eck, S., Goriely, S., Jorissen, A., & Plez, B. 2003, , 404, 291
Villemoes, P., Wang, M., Arnesen, A., Weiler, C., & Wännström, A. 1995, , 51, 2838
Wallerstein, G., Pilachowski, C., Gerend, D., Baird, S., & Canterna, R. 1979, , 186, 691
Wanajo, S., & Ishimaru, Y. 2006, , 777, 676
Wasserburg, G. J., Busso, M., & Gallino, R. 1996, , 466, L109
Wasserburg, G. J., & Qian, Y.-Z. 2000, , 529, L21
Whaling, W., Anderson, W. H. C., Carle, M. T., Brault, J. W., & Zaren, H. A. 2002, J. Res. Natl. Inst. Stand. Technol., 107, 149
Wickliffe, M. E., Lawler, J. E., & Nave, G. 2000, , 66, 363
Winckler, N., Dababneh, S., Heil, M., K[ä]{}ppeler, F., Gallino, R., & Pignatari, M. 2006, , 647, 685
Wisotzki, L., Christlieb, N., Bade, N., Beckmann, V., K[ö]{}hler, T., Vanelle, C., & Reimers, D. 2000, , 358, 77
Wujec, T., & Weniger, S.1981, , 25, 167
Yong, D., Aoki, W., Lambert, D. L., & Paulson, D. B. 2006, , 639, 918
Young, L., Childs, W. J., Berry, H. G., Kurtz, C., & Dinneen, T. 1987, , 36, 2148
Začs, L., Nissen, P. E., & Schuster, W. J. 1998, , 337, 216
Zinner, E., Amari, S., & Lewis, R. S. 1991, , 382, L47
![ \[summary1\] A summary of our best isotopic fraction measurements for each species. The Nd isotopic fractions reported here were measured by the method of matching the observed and synthetic spectra at the point of insensitivity to the isotopic mix. The Sm and Eu isotopic fractions were measured by fitting the shape of the line profile. Squares represent measurements in HD 175305, while circles represent measurements in HD 196944. The dotted line and dashed line (colored blue and red, respectively, in the electronic edition) represent the pure-$s$- and pure-$r$-process predictions of @arlandini99. The differences between the pure-$s$- and pure-$r$-process predictions are scaled together to emphasize the relative precision with which the two processes can be distinguished by each species. \[See electronic edition for a color version of this figure.\] ](f12_color.eps)
![ \[summary2\] A summary of our best isotopic fraction measurements for each species. The data and symbols are the same as in Figure \[summary1\]. The measurements are cast here in an absolute sense, emphasizing the relative precision with which each species was measured. \[See electronic edition for a color version of this figure.\] ](f13_color.eps)
[^1]: The percentages for some of the commonly-studied rare earth elements, predicted by the stellar model and classical method, respectively, are: Ba (five naturally-occurring isotopes), 81.1% or 85.3%; La (one isotope), 62.1% or 75.4%; Nd (seven isotopes), 58.8% or 57.9%; Sm (seven isotopes), 30.7% or 33.1%; Eu (two isotopes), 5.8% or 2.3%.
[^2]: We adopt the usual spectroscopic notations that \[A/B\] $\equiv$ log$_{\rm 10}$(N$_{\rm A}$/N$_{\rm B}$)$_{\star}$ – log$_{\rm 10}$(N$_{\rm A}$/N$_{\rm B}$)$_{\sun}$, and that log $\varepsilon$(A) $\equiv$ log$_{\rm 10}$(N$_{\rm A}$/N$_{\rm H}$) $+$ 12.00, for elements A and B.
[^3]: Please see, e.g., @wasserburg96, @qian00, @wasserburg00, @johnson02b, @wanajo06, and @kratz07 for further discussion of these two processes.
[^4]: 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.
[^5]: @whaling02 propose a change in the @learner88 Fe <span style="font-variant:small-caps;">i</span> wavenumber scale, who calibrated to the absolute Ar <span style="font-variant:small-caps;">ii</span> wavenumber scale from @norlen73. The proposed change, a multiplicative correction of 1$+$67(8)$\times10^{-9}$, amounts to a correction of $\sim\,0.001-0.002$cm$^{-1}$ or $0.3-0.5$mÅ in the visible spectral region examined in our study. This correction is negligible for our purposes.
[^6]: This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
[^7]: This research has made use of the NASA/ IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
[^8]: This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation.
[^9]: Available online: <https://webspace.utexas.edu/ir68/teff/dataII.htm>
[^10]: We have performed a test to compare abundances measured from equivalent widths of Ti, Fe, and Eu using three different Kurucz model atmospheres. For a typical metal-poor star, we adopt $T_{\rm eff}=5000$K, $\log(g)=2.0$, and $v_{t}=2.0$kms$^{-1}$ in all three models. For atmosphere 1, \[Fe/H\] $=-2.0$ and \[$\alpha$/Fe\] $=+0.0$. For atmosphere 2, \[Fe/H\] $=-1.8$ and \[$\alpha$/Fe\] $=+0.0$. For atmosphere 3, \[Fe/H\] $=-2.0$ and \[$\alpha$/Fe\] $=+0.4$. Atmospheres 1 and 2 use the Kurucz grid of $\alpha$-normal atmospheres, while atmosphere 3 uses the Kurucz grid of $\alpha$-enhanced atmospheres. The derived abundances are equivalent to within 0.02 dex in all cases, and this systematic effect is completely negligible for purposes of our study.
[^11]: Available online: <http://kurucz.harvard.edu/linelists.html>
[^12]: For wavelengths of blended features, we quote the number of digits to the right of the decimal point to the precision given in the reference and the precision used in our syntheses.
[^13]: Available online: <ftp://saphir.dstu.univ-montp2.fr/GRAAL/plez/CNdata>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a technique for mapping the complete 3D spatial intensity profile of a laser beam from its fluorescence in an atomic vapour. We propagate shaped light through a rubidium vapour cell and record the resonant scattering from the side. From a single measurement we obtain a camera limited resolution of 200 x 200 transverse points and 659 longitudinal points. In constrast to invasive methods in which the camera is placed in the beam path, our method is capable of measuring patterns formed by counterpropagating laser beams. It has high resolution in all 3 dimensions, is fast and can be completely automated. The technique has applications in areas which require complex beam shapes, such as optical tweezers, atom trapping and pattern formation.'
address: 'SUPA, School of Physics and Astronomy, University of Glasgow, Glasgow G12 8QQ, UK'
author:
- 'Neal Radwell, Mordjane A. Boukhet and Sonja Franke-Arnold'
title: 3D Beam Reconstruction by Fluorescence Imaging
---
[10]{}
David G Grier. A revolution in optical manipulation. , 424(6950):810–816, 2003.
Rudolf Grimm, Matthias Weidem[ü]{}ller, and Yurii B Ovchinnikov. Optical dipole traps for neutral atoms. , 42:95–170, 2000.
Richard W Bowman and Miles J Padgett. Optical trapping and binding. , 76(2):026401, 2013.
J Arlt and MJ Padgett. Generation of a beam with a dark focus surrounded by regions of higher intensity: the optical bottle beam. , 25(4):191–193, 2000.
S. Franke-Arnold, J. Leach, M. J. Padgett, V. E. Lembessis, D. Ellinas, A. J. Wright, J. M. Girkin, P. Ohberg, and A. S. Arnold. Optical ferris wheel for ultracold atoms. , 15(14):8619–8625, Jul 2007.
Yaoju Zhang. Generation of three-dimensional dark spots with a perfect light shell with a radially polarized laguerre–gaussian beam. , 49(32):6217–6223, 2010.
Aidan S. Arnold. Extending dark optical trapping geometries. , 37(13):2505–2507, Jul 2012.
Roee Ozeri, Lev Khaykovich, and Nir Davidson. Long spin relaxation times in a single-beam blue-detuned optical trap. , 59(3):R1750, 1999.
Peng Xu, Xiaodong He, Jin Wang, and Mingsheng Zhan. Trapping a single atom in a blue detuned optical bottle beam trap. , 35(13):2164–2166, 2010.
Miles Padgett and Richard Bowman. Tweezers with a twist. , 5(6):343–348, 2011.
MP Lee, A Curran, GM Gibson, M Tassieri, NR Heckenberg, and MJ Padgett. Optical shield: measuring viscosity of turbid fluids using optical tweezers. , 20(11):12127–12132, 2012.
Richard Bowman, Graham Gibson, and Miles Padgett. Particle tracking stereomicroscopy in optical tweezers: controlof trap shape. , 18(11):11785–11790, 2010.
Graeme Whyte and Johannes Courtial. Experimental demonstration of holographic three-dimensional light shaping using a gerchberg–saxton algorithm. , 7(1):117, 2005.
J Leach, MR Dennis, J Courtial, and MJ Padgett. Vortex knots in light. , 7(1):55, 2005.
J Romero, J Leach, B Jack, MR Dennis, S Franke-Arnold, SM Barnett, and MJ Padgett. Entangled optical vortex links. , 106(10):100407, 2011.
J. Radon. Über die bestimmung von funktionen durch ihre integralwerte längs gewisser mannigfaltigkeiten. , 69:262–277, April 1917. In German. An English translation can be found in S. R. Deans: [ *The Radon Transform and Some of Its Applications*]{}.
R Bowman, V D’Ambrosio, E Rubino, O Jedrkiewicz, P Di Trapani, and MJ Padgett. Optimisation of a low cost slm for diffraction efficiency and ghost order suppression. , 199(1):149–158, 2011.
Jonathan Leach, Miles J Padgett, Stephen M Barnett, Sonja Franke-Arnold, and Johannes Courtial. Measuring the orbital angular momentum of a single photon. , 88(25; PART 1):257901–257901, 2002.
Introduction
============
Laser beams with increasingly intricate and complex profiles have become interesting for a range of applications. Particles from the micron size range down to single atoms can be trapped by the dipole forces produced from light beams [@grier2003revolution; @grimm2000optical]. These forces have been exploited in optical tweezers [@bowman2013optical], allowing micro manipulation of beads and biological matter which would otherwise be impossible. Atoms can also be trapped by the same forces and have the further advantage that their strong resonances allow tuning of the force as a function of the detuning.
With the advent of Spatial Light Modulators (SLMs) it has become possible to generate a wide variety of beam profiles, expanding the possibilities for trapping and guiding. Beam propagation and in particular focussing can result in an additional modification of the beam profile along the propagation axis. Complex 3D beam shapes have been proposed [@arlt2000generation; @franke-arnold2007; @zhang2010generation; @arnold2012extending] and implemented in atom trapping [@ozeri1999long; @xu2010trapping] and optical tweezers [@padgett2011tweezers; @lee2012optical; @bowman2010particle; @whyte2005experimental]. In order to verify the accuracy of the beam generation however, traditional methods would require stepping a camera in the beam path followed by reconstruction [@leach2005vortex; @romero2011entangled]. This method suffers from several drawbacks: The beam axis distance can be difficult to measure accurately, the process is manual and slow and it is clearly ineffective at imaging beam structures created by interference of counter-propagating laser beams. We present here an alternative method which avoids these issues.
{width="\columnwidth"}
Summary of Technique
====================
Our technique is based on tomographic reconstruction of images taken via fluorescence imaging. If a laser beam is shone through a fluorescent medium the amount of fluorescence depends on the intensity of the laser beam. In the case of an atomic gas the two-level photon scattering (fluorescence) rate is;
$$R_{\rm sc} =\frac{\Gamma}{2} \frac{(I/I_{\rm sat})}{1+(I/I_{\rm sat})+4(\Delta/\Gamma)^2},
\label{eqn:Rscatt}$$
where $\Gamma$ ($2\pi\times6.06$ MHz) is the inverse of the upper state lifetime, $I$ is the laser intensity, $I_{\rm sat}$ (1.6 mw/cm$^2$) is atom specific saturation intensity and $\Delta$ is the laser detuning. For laser beams well below saturation, as used in the reported experiments here, the scattering rate becomes directly proportional to the laser intensity. It is therefore possible to measure the intensity of a laser beam by shining a resonant laser through an atomic gas and detecting the emitted fluorescence from the side with a camera. This recorded fluorescence image contains the projection of the light profile onto e.g. the yz plane (see Fig. \[fig:Axes\]a) without any information on the x dependence, but the full beam profile can be recovered tomographically.
Tomography is the method to reconstruct 2 dimensional information from 2 or more 1 dimensional projections. A single column of pixels within the camera image provides the projection of the fluorescence along the x-direction (or in cylindrical coordinates at a particular angle $\theta$) at a particular z-position. If we were to rotate the camera around the z axis, we would have access to more projections at more angles. From this combined data we can tomographically reconstruct the 2D xy plane at the chosen z-position. By performing this analysis on all columns in each image, and [*stacking*]{} the retrieved profiles along the z-direction we recover the full 3 dimensional fluorescence distribution. For practical reasons, instead of rotating the camera around the beam, we instead rotate the beam itself with a fixed camera position.
Tomography
==========
The operation of projecting 2D data along an axis to produce a 1D profile (See Fig. \[fig:Axes\]b) is called a Radon transform [@Radon1], while the operation to return to the 2D data from the 1D profiles is the inverse Radon transform. There are several techniques to compute this inverse, the most common being the back projection method. This method comprises 4 simple steps. First, each 1D projection is filtered to correct for the oversampling of the central pixels. The ideal filter is simply a ramp filter in frequency space, however the filter may be optimised to enhance the important features in an image. The filtered 1D projections are then converted to a 2D image by simply copying the value in each column to every row (i.e. the values are “smeared” from 1D to 2D). Next each 2D image is rotated by the angle at which the projection was taken and finally all of these images are added together.
{width="\columnwidth"}
The reconstruction is performed in Matlab and runs extremely quickly. To test the reconstruction we calculate an ideal 2D profile, take 1D projections from this and then reconstruct the 2D profile from the projections. The test profile is shown in Fig. \[fig:Reconstruction\_Example\]a and a sample reconstruction in Fig. \[fig:Reconstruction\_Example\]b. The shown reconstruction is formed from 30 projections at 200 x 200 (x,y) resolution, and was generated in 0.15 seconds. The performance of the reconstruction has been analysed for different numbers of projections, with the results shown in Fig. \[fig:Reconstruction\_Example\]c. Increasing the number of projections increases the accuracy of the reconstruction at the cost of longer reconstruction times,where the time scales linearly with the number of projections (5ms per projection). We suggest 30 projections as a workable balance between speed and accuracy. For this proof of principle experiment however, we favour accuracy at the expense of speed and therefore the following results have been performed with a large number (116) of projections.
{width="8cm"}
Experimental Results
====================
To test our beam reconstruction we have set up an experiment as shown in Fig. \[fig:Setup\]. The experiment is composed of two main sections, beam creation and beam detection.
The laser source is an external cavity diode laser, tuned to $780$nm, resonant with the Doppler broadened $^{85}$Rb F=2 to F’=3 transition for maximum fluorescence. The mode is cleaned by a pinhole and then expanded. An SLM (Hamamatsu LCOS) is used to generate the shaped beams, allowing us to create arbitrary phase and intensity profiles. We refract the beam off an intensity modulated holographic grating [@bowman2011optimisation] and select the first order diffracted beam with an aperture placed in the Fourier plane of the SLM. The beam is then collimated again and sent to the beam detection section.
The beam is directed through a standard rubidium cell in the beam path. The fluorescence emitted from the side of the cell is imaged onto a camera (Prosilica GC660) by a single lens.
For convenience we rotate the beam by simply turning the pattern on the SLM, however a more general approach would be to use a series of two Dove prisms to rotate the beam, thereby allowing detection of arbitrary beams [@leach2002measuring].
{width="\columnwidth"}
The data shown is of a superposition of two Laguerre-Gaussian (LG) beams. A single LG beam is parameterised by the azimuthal and radial numbers $l$ and $p$ respectively. For simplicity we restrict our pattern to have p=0. A single LG beam has an intensity minimum in the centre, the size of which increases proportional to $\sqrt{l/2}$. A superposition of 2 beams, with different l numbers produces a ’petal pattern’ with a number of petals equal to the difference in l numbers. We use a superposition of l=3 and l=-3, resulting in the pattern shown in Fig. \[fig:Reconstruction\_Example\]a, giving 6 ’petals’. This pattern is then rotated and a video is recorded by the camera. Sample frames, separated by $90$ degrees, are shown in Fig. \[fig:Sample\_Frames\]a and b.
{width="\columnwidth"}
To perform the tomographic reconstruction from the experimental data it is necessary to know which angle the beam is rotated in each frame. This can be achieved either by rotating by a known amount from frame to frame, or in our case, by extracting the angles from the data. A sinogram can be created by stacking all of the 1D projections next to one another, thereby forming an image with $\theta$ along the x axis. An example of which is shown in Fig. \[fig:Sinogram\]a. This sinogram has an inherent periodicity which is related to the rotational symmetry of the beam, therefore by fitting a sine function to the x axis of Fig. \[fig:Sinogram\]a, we can infer the number of frames between 0 and 2$\pi$.
The result of the reconstruction for a single column of pixels is shown in Fig. \[fig:Sample\_Frames\]c. Reconstruction for every column results in a 3 dimensional matrix with values corresponding to the fluorescence of the atoms.
Visualisation
=============
{width="\columnwidth"}
The full 3D intensity profile can be visualised with our Matlab program which renders isosurfaces and plots these with a colour and transparency which is related to the intensity of the isosurface. An example plot with ten isosurfaces between 45 % and 80 % of the maximum intensity is shown in Fig. \[fig:3D\_Reconstruction\]. The visualisation technique has also been expanded to create fly-by videos as well as images and video capable of being displayed in 3D on compatible displays. We note that in fact the 3D intensity profile can also be observed directly by looking at the scattered light through an IR viewer.
Discussion
==========
The resolution of the 3D profiles is generally limted only by the camera. In order to reduce computation time, here we have used only 200x200x659 pixels of the 494x494x659 camera resolution. The accuracy of the reconstruction is limited by the number of projections and the noise. The number of projections increases the accuracy of the reconstruction at the expense of longer measurements and calculation time but is already very high for only 30 projections as demonstrated in Fig. \[fig:Reconstruction\_Example\]c . Regarding the noise, a single background measurement can be taken at the start of the experiment, and used to remove background light signals from all further measurements. This is particularly effective at improving the final signal to noise ratio, due to the high noise sensitivity of the back projection method used to reconstruct the beam.
We next consider high speed measurements or video framerates. Due to the 120 Hz framerate of our camera, a single measurement (at 30 projections) would take 250 ms, however our system is limited by the beam rotation speed which for our SLM is capped at 60 Hz, resulting in 30 projections in 500 ms. A setup involving rotating Dove prisms may be able to outperform the SLM. Either method could allow taking 3D measurements at a rate around 1 Hz, so that the limiting factor will still be the reconstruction time. A single z position at 30 projections has a reconstruction time of 150 ms, limiting the z resolution to $<$10 for 1Hz video. Alternatively, z resolution could be increased at the cost of transverse resolution, either by restricting the number of projections or pixels used.
Conclusion
==========
We have demonstrated a high resolution, simple, fast experimental method which maps the entire 3 dimensional structure of an arbitrary laser beam. As a proof of concept we have generated a superposition of co-propagating LG beams and reconstructed and plotted the results. The method benefits from a far higher resolution in the propagation direction than other invasive methods, without loss of resolution in the transverse dimension. The method is completely automated and capable of producing a single plane reconstruction in only 150 ms, limited by computation time.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the existence of stationary classical solutions of the incompressible Euler equation in the plane that approximate singular stationary solutions of this equation. The construction is performed by studying the asymptotics of equation $-{\varepsilon}^2 \Delta u^{\varepsilon}=(u^{\varepsilon}-q-\frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}})_+^p$ with Dirichlet boundary conditions and $q$ a given function. We also study the desingularization of pairs of vortices by minimal energy nodal solutions and the desingularization of rotating vortices.'
address:
- |
Laboratoire Jacques-Louis Lions\
Université Pierre et Marie Curie\
4 place Jussieu BC 187\
75252 Paris\
France
- |
Université catholique de Louvain\
Départment de Mathématique\
Chemin du Cyclotron 2\
1348 Louvain-la-Neuve\
Belgium
author:
- Didier Smets
- Jean Van Schaftingen
title: Desingularization of vortices for the Euler equation
---
Introduction
============
Singular solutions to the Euler equation
----------------------------------------
The incompressible Euler equations $$\left\{
\begin{aligned}
\nabla \cdot \mathbf{v} &= 0, \\
\mathbf{v}_t + \mathbf{v}\cdot \nabla \mathbf{v}&=-\nabla p,
\end{aligned}
\right.$$ describe the evolution of the velocity $\mathbf{v}$ and the pressure $p$ in an incompressible flow. In ${\mathbf{R}}^2$, the vorticity $\omega = \nabla \times \mathbf{v}=\partial_1 \mathbf{v}_2-\partial_2 \mathbf{v}_1$ of a solution of the Euler equations obey the transport equation $$\omega_t + \mathbf{v} \cdot \nabla \omega = 0$$ and the velocity field $\mathbf{v}$ can be recovered from the vorticity function $\omega$ through the Biot–Savart law $$\mathbf{v} = \omega * \frac{1}{2\pi} \frac{-x^\perp}{{\lvertx\rvert}^2},$$ where $x^\perp=(x_2, -x_1)$. Special singular solutions of the Euler equations are given by [^1] $$\omega = \sum_{i=1}^k \kappa_i \delta_{x_i(t)},$$ corresponding to $$\mathbf{v}(x)=-\sum_{i=1}^k \frac{\kappa_i}{2\pi} \frac{(x-x_i(t))^\perp}{{\lvertx-x_i(t)\rvert}^2},$$ and the positions of the vortices $x_i : {\mathbf{R}}\to {\mathbf{R}}^2$ satisfy $$\dot{x}_i(t)=-\sum_{\substack{j=1 \\ j \ne i}}^k \frac{\kappa_j}{2\pi} \frac{(x_i(t)-x_j(t))^\perp}{{\lvertx_i(t)-x_j(t)\rvert}^2}.$$ In terms of the Kirchhoff–Routh function $$\mathcal{W}(x_1, \dotsc, x_k)=\frac{1}{2} \sum_{i \ne j} \frac{\kappa_i\kappa_j}{2\pi} \log \frac{1}{{\lvertx_i-x_j\rvert}},$$ the positions obey Kirchhoff’s law $$\label{equationKirchhoff}
\kappa_i \dot{x}_i=(\nabla_{x_i} \mathcal{W})^\perp,$$ which is a Hamiltonian formulation of the dynamics of the vortices.
In simply-connected bounded domains $\Omega \subset {\mathbf{R}}^2$, similar singular solutions exist. If one requires for example that the normal component of $\mathbf{v}$ vanishes on the boundary, the associated Kirchoff–Routh function is then given by $$\label{eqKRDomainsHomog}
\mathcal{W}(x_1, \dotsc, x_k)=\frac{1}{2} \sum_{i \ne j} \kappa_i\kappa_j G(x_i, x_j)+\sum_{i=1}^k \frac{\kappa_i^2}{2}H(x_i, x_i),$$ where $G$ is the Green function of $-\Delta$ on $\Omega$ with Dirichlet boundary conditions and $H$ is its regular part.[^2] One can also prescribe a condition $v_n$ on the outward component of the velocity on the boundary. Since we are dealing with an incompressible flow, the boundary data should satisfy $\int_{\partial \Omega} v_n=0$. Let $\mathbf{v}_0$ be the unique harmonic field whose normal component on the boundary is $v_n$; i.e., $\mathbf{v}_0$ satisfies $$\left\{
\begin{aligned}
\nabla \cdot \mathbf{v}_0&=0, & & \text{in $\Omega$}, \\
\nabla \times \mathbf{v}_0&=0, & & \text{in $\Omega$}, \\
n \cdot \mathbf{v}_0&=v_n& & \text{on $\partial \Omega$},
\end{aligned}
\right.$$ where $\nabla \times (u, v)=\partial_1 v-\partial_2 u$ and $n$ is the outward normal, then the positions of the vortices are obtained by the modified law $$\dot{x}_i=(\nabla_{x_i} \mathcal{W})^\perp +\mathbf{v}_0.$$ Since $\Omega$ is simply-connected $\mathbf{v}_0$ can be written $\mathbf{v}_0=(\nabla \psi_0)^\perp$ where the stream function $\psi_0$ is characterized up to a constant by $$\label{eqpsi0}
\left\{
\begin{aligned}
-\Delta \psi_0&=0& &\text{in $\Omega$}, \\
-\frac{\partial \psi_0}{\partial \tau}&=v_n & & \text{on $\partial \Omega$},
\end{aligned}
\right.$$ where $\frac{\partial \psi_0}{\partial \tau}$ denotes the tangential derivative on $\partial \Omega$. The Kirchhoff–Routh function associated to the vortex dynamics becomes then $$\label{KRDomains}
\mathcal{W}(x_1, \dotsc, x_k)=\frac{1}{2} \sum_{i \ne j} \kappa_i\kappa_j G(x_i, x_j)+\sum_{i=1}^k \frac{\kappa_i^2}{2}H(x_i, x_i)+\sum_{i=1}^k \kappa_i \psi_0(x_i),$$ see C.C.Lin [@Lin1941] (who uses opposite sign conventions).
Desingularization of vortices
-----------------------------
One way to justify the weak formulation for point vortex solutions of the Euler equations is to approximate these solutions by classical solutions. This can actually be done, on finite time intervals, by considering regularized initial data for the vorticity (see e.g. C.Marchioro and M. Pulvirenti [@MarchioroPulvirenti1983]).
Critical points of the Kirchhoff–Routh function $\mathcal{W}$ give rise to stationary vortex points solutions of the Euler equations. As noted above, these weak stationary solutions can be approximated by classical solutions of the Euler equations. These do not need be stationary solutions though, and one can wish to approximate the stationary vortex-point solutions by stationary classical solutions. In the simplest case, corresponding to a single point vortex in a simply-connected domain, we obtain the following
\[thm:resu\] Let $\Omega \subset {\mathbf{R}}^2$ be a bounded simply-connected smooth[^3] domain and $v_n:\partial \Omega \to {\mathbf{R}}\in L^s(\partial \Omega)$ for some $s>1$ be such that $\int_{\partial \Omega} v_n = 0$. Let $\kappa >0$ be given. For ${\varepsilon}>0$ there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in $\Omega$ with outward boundary flux given by $v_n$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}) \subset B(x_{\varepsilon}, C{\varepsilon})$ for some $x_{\varepsilon}\in \Omega$ and $C>0$ not depending on ${\varepsilon}$. Moreover, as ${\varepsilon}\to 0$, $$\int_\Omega \omega_{\varepsilon}\to \kappa,$$ and $$\mathcal{W}(x^{\varepsilon}) \to \sup_{x \in \Omega} \mathcal{W}(x).$$
Other situations, corresponding to pairs of vortices of opposite signs, multiply-connected bounded domains or unbounded domains are discussed in Section \[sect:resu\].
We are aware essentially of two methods to construct stationary solutions of the Euler equations that we call the vorticity method and the stream-function method.
The vorticity method was introduced by V.Arnold (see [@ArnoldKhesin]\*[Chapter II §2]{}), and was implemented successfully by G.R.Burton [@Burton1988] and B.Turkington [@Turkington1983]. It roughly consists in maximizing the kinetic energy $$\frac{1}{2}\int_{\Omega} \int_{\Omega} \omega(x)G(x, y)\omega(y)\, dx\, dy+\int_{\Omega} \psi_0(x)\omega(x)\, dx+\frac{1}{2} \int_{\Omega} {\lvert\nabla \psi_0\rvert}^2,$$ under some constraints on the sublevel sets of $\omega$. The function $\omega$ is the vorticity of the flow and a stream function $\psi$ is the solution to $$\left\{
\begin{aligned}
-\Delta \psi &=\omega & & \text{in $\Omega$},\\
\psi&=\psi_0 & & \text{on $\partial \Omega$}.
\end{aligned}
\right.$$ Considering suitable families of constraints on the sublevel sets of $\omega$, one can obtain families of solutions converging to stationary vortex-point solutions. The differentiability of those solutions is not guaranteed (the solutions correspond to vortex patches of constant density).
The stream-function method starts from the observation that if $\psi$ satisfies $$-\Delta \psi=f(\psi),$$ for some arbitrary function $f \in C^1({\mathbf{R}})$, then $\mathbf{v}=(\nabla \psi)^\perp$ and $p=F(\psi)-\frac{1}{2}{\lvert\nabla \psi\rvert}^2$, with $F(s)=\int_0^s f$ form a stationary solution to the Euler equations. Moreover, the velocity $\mathbf{v}$ is irotational on the set where $f(\psi)=0$.
We now set $q=-\psi_0$ and $u=\psi-\psi_0$, so that $u=0$ on $\partial \Omega$ and $-\Delta u = f(u-q)$ in $\Omega$. If we assume that $\inf_{\Omega} q > 0$ and $f(t)=0$ when $t \le 0$, the vorticity set $\{ x {\: :\:}f(\psi(x))> 0 \}$ is bounded away from the boundary. When $f$ satisfies also some monotonicity and growth conditions, $\Omega={\mathbf{R}}^2_+$ and $q(x)=W x_1+d$ with $W > 0$ and $d>0$, J.Norbury [@Norbury1975] has shown the existence of solutions to $-\Delta u = \nu f(u-q)$, where $\nu > 0$ is a Lagrange multiplier a priori unknown by minimizing $\int_{\Omega} {\lvert\nabla u\rvert}^2$ under the constraint $$\int_{\Omega} F(u-q)=\mu$$ in $H^1_0(\Omega)$ when $\Omega$ is the half-plane ${\mathbf{R}}^2_+$. M.S.Berger and L.E.Fraenkel [@BergerFraenkel1980] have obtained corresponding results for a bounded domain $\Omega \subset {\mathbf{R}}^2$, and they began studying the asymptotics for variable $\mu$ and $q$, but the lack of information on $\nu$ remained an obstacle.
The unknown $\nu$ can be avoided by minimizing $\int_{\Omega} \frac{1}{2}{\lvert\nabla u\rvert}^2-\frac{1}{{\varepsilon}^2}F(u-q)$ under the natural constraint $\int_{\Omega} \frac{1}{2}{\lvert\nabla u\rvert}^2-\frac{1}{{\varepsilon}^2}uf(u-q)=0$. Yang Jianfu [@Yang1991] has used this approach in ${\mathbf{R}}^2_+$ with $q(x)=Wx_1+d$ and has studied the asymptotic behavior of the solution $u^{\varepsilon}$ when ${\varepsilon}\to 0$: If $$\begin{aligned}
A_{\varepsilon}&=\{ x \in {\mathbf{R}}^2_+ {\: :\:}f(u^{\varepsilon}-q) > 0\}, & \kappa_{\varepsilon}&=\frac{1}{{\varepsilon}^2} \int_{\Omega} f(u^{\varepsilon}-q),\end{aligned}$$ and $x^{\varepsilon}\in A^{\varepsilon}$, then $\operatorname{diam}A^{\varepsilon}\to 0$, $\operatorname{dist}(x^{\varepsilon}, \partial {\mathbf{R}}^2_+)\to 0$, and $$ \frac{u^{\varepsilon}}{\kappa^{\varepsilon}}-G(\cdot, x^{\varepsilon}) \to 0$$ in $W^{1, r}_{\mathrm{loc}}({\mathbf{R}}^2_+)$, for $r \in [1, 2)$. Li Gongbao, Yan Shusen and Yang Jianfu [@LiYanYang2005] obtained a similar result on bounded domains, with the additional information that $q(a^{\varepsilon}) \to \min_{\Omega} q$. These results are in striking contrast with the observation made at the beginning that the dynamics of the vortices is governed by the Kirchhoff–Routh function $\mathcal{W}$ defined by , which implies that stationary vortices should be localized around a critical point of $x \mapsto \frac{\kappa^2}{2} H(x, x)-\kappa q(x)$.
In fact, the results in [@Yang1991; @LiYanYang2005] do not answer the question about the desingularization of stationary vortex point solutions to the Euler equation. Indeed, in the case of bounded domains for example, their solutions satisfy ${\Vert\nabla u\Vert}_{\mathrm{L}^2}^2=O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr)$, so that testing the equation against the function $\min(u^{\varepsilon}, q)$ and using the fact that $q$ is harmonic and nonnegative, we have $$\kappa^{\varepsilon}\min_{\partial \Omega} q \le
\frac{1}{{\varepsilon}^2}\int_{\Omega} f(u^{\varepsilon}-q) =\int_{\Omega \setminus A^{\varepsilon}} {\lvert\nabla u^{\varepsilon}\rvert}^2=O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr),$$ i.e. $\kappa^{\varepsilon}\to 0$. In some sense, the family of solutions $u^{\varepsilon}$ provides a desingularization of point-vortex solutions with vanishing vorticity. The asymptotic position is consistent with the fact that when the vorticities tend to zero, the term $\sum_{i=1}^k \kappa_i \psi_0(x_i)$ becomes dominant in the Kirchhoff–Routh function .
In order to desingularize point-vortex solutions with non-vanishing vorticity, M.S. Berger and L.E.Fraenkel [@BergerFraenkel1980]\*[Remark 2]{} suggest that $q$ should grow like $\log \frac{1}{{\varepsilon}}$. This brings us to the study of the problem $$\label{problemPeps}
\left\{
\begin{aligned}
-\Delta u^{\varepsilon}&=\frac{1}{{\varepsilon}^2} f(u^{\varepsilon}- q^{\varepsilon}) & &\text{in $\Omega$, }\\
u^{\varepsilon}&= 0 & &\text{on $\partial \Omega$},
\end{aligned}
\right. \tag{\protect{$\mathcal{P}^{\varepsilon}$}}$$ where $q^{\varepsilon}=q+\frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}}$.
In Section \[sectionSingleVortex\], we study $(\mathcal{P}_{\varepsilon})$ in a bounded domain: we first construct solutions and then analyze their asymptotic behavior. Theorem \[thm:resu\] is an easy consequence of the results in Section \[sectionSingleVortex\]. In Section \[sectionmultiply\] we present and extension to multiply-connected domains, while in Section \[sectUnbounded\], we present an extension to unbounded domains which are a perturbation of a half-plane. In Section \[sectionVortexPair\] we modify slightly $(\mathcal{P}_{\varepsilon})$ in order to construct desingularized solutions for two point vortices of opposite signs.
As a final remark, our results seem connected with the work of M.del Pino, M.Kowalczyk, and M.Musso [@delPinoKowalczykMusso2005] on the equation $$-\Delta u={\varepsilon}^2 K(x)e^u$$ for which the energy concentrates in small balls around points $x_1^{\varepsilon}, \dotsc, x_k^{\varepsilon}$. These points tend to a critical point of the function $-\sum_{i=1}^k 2\log K(x_i)-8\pi H(x_i, x_i)-\sum_{i \ne j} 8\pi G(x_i, x_j)$. The connection is clear when one rewrites their equation as $-\Delta u=\frac{1}{{\varepsilon}^2}\exp(u+\log K-\frac{8\pi}{2\pi}\log \frac{1}{{\varepsilon}})$. Other related work include the study of the equation $-\Delta u = u^p$ as $p \to \infty$ by P.Esposito, M.Musso and A. Pistoia [@EspositoMussoPistoia2006; @EspositoMussoPistoia2007], and the recent work of T.Bartsch, A.Pistoia and T.Weth [@BartschPistoiaWeth] in which systems of three and four vortices are desingularized by studying the equation $-\Delta u= {\varepsilon}^2 \sinh u$. In all the references, whereas the vorticity concentrates at points, its support does not shrink as ${\varepsilon}\to 0$.
We also bring to the attention of the reader that there is a similar situation with similar results for three-dimensional axisymmetric incompressible inviscid flows by vorticity methods [@Burton1987; @FridemannTurkington1981] and stream-function methods [@BergerFraenkel1974; @AmbrosettiStruwe1989; @Yang1995]. However we are not aware of a counterpart of the present work for three-dimensional axisymmetric incompressible inviscid flows.
[**Acknowledgements.**]{} This work was initiated during a visit of the second author at Laboratoire Jacques-Louis Lions of Université Pierre & Marie Curie. The authors wish to thank Franck Sueur for fruitful remarks following a first version of the manuscript.
Single vortices in bounded domains {#sectionSingleVortex}
==================================
In this section, $\Omega\subset {\mathbf{R}}^2$ is a bounded simply-connected smooth domain, $f : {\mathbf{R}}\to {\mathbf{R}}$ is the real function defined by $f(s)=s_+^p$ for some $1<p<+\infty$ and where $s_+=\max(s, 0)$, $\kappa > 0$ is given as well as $q\in \mathrm{W}^{1, r}(\Omega)$ for some $r > 2$.[^4] We will consider solutions of the boundary value problem where ${\varepsilon}>0$ is a real parameter. The solutions we consider are the least energy solutions obtained by minimizing the energy functional $$\label{energyFunctional}
\mathcal{E}^{\varepsilon}(u)= \int_{\Omega} \Bigl(\frac{|\nabla u|^2}{2} -
\frac{1}{{\varepsilon}^2}F(u-q^{\varepsilon})\Bigr)$$ over the natural constraint given by the Nehari manifold $$\mathcal{N}^{\varepsilon}= \left\{ u\in H^1_0(\Omega)\setminus \{0\} \ : \ \langle
d\mathcal{E}^{\varepsilon}(u), u\rangle = 0\right\},$$ where $F(s)=\frac{1}{p+1}s_+^{p+1}$ is a primitive of $f$. It is standard to prove the (see e.g. [@Willem1996]\*[Theorem 2.18]{})
\[prop:2.1\] Assume that $q^{\varepsilon}\geq 0$ on $\Omega$, so that ${\mathcal{N}^{\varepsilon}}\neq \emptyset$, and define $$c^{\varepsilon}= \inf_{u\in \mathcal{N}^{\varepsilon}} \mathcal{E}^{\varepsilon}(u).$$ Then, there exists $u^{\varepsilon}\in \mathcal{N}^{\varepsilon}$ such that $\mathcal{E}^{\varepsilon}(u^{\varepsilon})=c^{\varepsilon}$, and $u^{\varepsilon}$ is a positive solution of ${(\mathcal{P}^{\varepsilon})}$.
Note that $q$ is bounded since $r>2$, and therefore $q^{\varepsilon}\geq 0$ provided ${\varepsilon}$ is sufficiently small.
Our focus is the asymptotics of $u^{\varepsilon}$ when ${\varepsilon}\to 0$. In order to describe the asymptotic behavior of $u^{\varepsilon}$, we introduce the limiting profile $U_\kappa : {\mathbf{R}}^2 \to {\mathbf{R}}$ defined as the unique radially symmetric solution of the problem $$\tag{\protect{$\mathcal{U}_\kappa$}}
\label{Ukappa}
\left\{
\begin{aligned}
&-\Delta U_\kappa = f(U_\kappa), \\
&\int_{{\mathbf{R}}^2} f(U_\kappa) =\kappa.
\end{aligned}
\right.$$ For every $\kappa>0$, there exists $\rho_\kappa>0$ such that $$U_\kappa(y)=
\left\{
\begin{aligned}
&V_{\rho_\kappa}(y)& &\text{if $y \in B(0, \rho_\kappa)$}, \\
&\frac{\kappa}{2\pi} \log \frac{\rho_\kappa}{{\lverty\rvert}} & &\text{if $y \in {\mathbf{R}}^2 \setminus B(0, \rho_\kappa)$},
\end{aligned}\right.$$ where $V_\rho : B(0,\rho) \to {\mathbf{R}}$ satisfies $$\left\{
\begin{aligned}
\displaystyle -\Delta V_\rho &= V_\rho^p & & \text{in $B(0, \rho)$}, \\
V_\rho &= 0 && \text{on $\partial B(0, \rho)$}.
\end{aligned}
\right.$$ One can show that $\kappa=\gamma \rho^{-\frac{2}{p-1}}$, for some constant $\gamma > 0$ depending on the value of $p$.
The Kirchhoff-Routh function $\mathcal{W}$ for one vortex of vorticity $\kappa$ is defined by $$\mathcal{W}(x)=\frac{\kappa^2}{2} H(x, x)-\kappa q(x).$$ Let us also define the quantity $$\mathcal{C} = \frac{\kappa^2}{4\pi} \log \rho_\kappa +
\int_{B(0, \rho_\kappa)}\Bigl(\frac{|\nabla U_{\rho_\kappa}|^2}{2} - \frac{U_{\rho_\kappa}^{p+1}}{p+1}\Bigr).$$ While the function $\mathcal{W}$ depends on $x \in \Omega$ and on $\kappa$, the quantity $\mathcal{C}$ only depends on $\kappa$ and on $p$.
We set $$\begin{aligned}\label{defiq}
A^{\varepsilon}&=\Big\{ x \in \Omega {\: :\:}u^{\varepsilon}(x)> q^{\varepsilon}(x)\Big\}, \\
\omega^{\varepsilon}&=\frac{1}{{\varepsilon}^2} f(u^{\varepsilon}-q^{\varepsilon}), \\
\kappa^{\varepsilon}&=\int_{\Omega} \omega^{\varepsilon}, \\
x^{\varepsilon}&=\frac{1}{\kappa^{\varepsilon}}\int_{\Omega} x \, \omega^{\varepsilon}(x)\, dx, \\
\rho^{\varepsilon}&=\rho_{\kappa^{\varepsilon}},
\end{aligned}$$ and respectively refer to these as the vorticity set, the vorticity, the total vorticity, the center of vorticity, and the vorticity radius.
We will prove
\[thm:K1\] As ${\varepsilon}\to 0$, we have $$u^{\varepsilon}=U_{\kappa^{\varepsilon}} \Big(\frac{\cdot-x^{\varepsilon}}{{\varepsilon}}\Big)+\kappa^{\varepsilon}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}\rho^{\varepsilon}}+ H(x^{\varepsilon}, \cdot)\Bigr)+o(1),$$ , where $$\kappa^{\varepsilon}=\kappa+\frac{2\pi}{\log \frac{1}{{\varepsilon}}}\Bigl(q(x^{\varepsilon})-\kappa H(x^{\varepsilon}, x^{\varepsilon}) -\frac{\kappa}{2\pi} \log \frac{1}{\rho_\kappa} \Bigr)+o({{\vert\!\log{\varepsilon}\vert}}^{-1}),$$ and $$\mathcal{W}(x^{\varepsilon}) \to \sup_{x \in \Omega} \mathcal{W}(x).$$ One also has $$B(x^{\varepsilon}, \Bar{r}^{\varepsilon}) \subset A^{\varepsilon}\subset B(x^{\varepsilon}, \mathring{r}^{\varepsilon}),$$ with $\Bar{r}^{\varepsilon}={\varepsilon}\rho_\kappa+o({\varepsilon})$ and $\mathring{r}^{\varepsilon}={\varepsilon}\rho_\kappa +o({\varepsilon})$. Finally, $$\mathcal{E}^{\varepsilon}(u^{\varepsilon})= \frac{\kappa^2}{4\pi}\log \frac{1}{{\varepsilon}}-\mathcal{W}(x^{\varepsilon})+\mathcal{C}+o(1).$$
Since $\mathcal{W}(x) \to -\infty$ as $x \to \partial \Omega$, by Theorem \[thm:K1\], up to a subsequence, $x^{\varepsilon}\to x^*\in \Omega$. Combined with standard elliptic estimates this yields the convergence $u^{\varepsilon}\to \kappa G(x_*, \cdot\, )$ in $\mathrm{W}^{1, p}_0(\Omega)$ for any $p<2$ and in $\mathcal{C}^k_{\mathrm{loc}}(\Omega\setminus \{x_*\})$ for any $k\in {\mathbf{N}}$. If $\partial \Omega$ is smooth enough, then one also has convergence in $\mathcal{C}^k_{\mathrm{loc}}(\Bar{\Omega}\setminus \{x_*\}\})$.
The proof of Theorem \[thm:K1\] is twofold. First, in Corollary \[cor:upper\], we prove a sharp upper bounds for the critical level $c^{\varepsilon}$. Then, in Proposition \[prop:1mai\] we show that any solution satisfying this upper bound needs to satisfy the asymptotic expansion.
Upper bounds on the energy {#upperBounds}
--------------------------
We will derive upper bounds for $c^{\varepsilon}$ by constructing elements of $\mathcal{N}^{\varepsilon}$ similar to the asymptotic expression of Theorem \[thm:K1\].
\[lemmaHatuNehari\] For every $\Hat{x} \in \Omega$, if ${\varepsilon}>0$ is small enough, there exists $$\Hat{\kappa}^{\varepsilon}=\kappa+\frac{2\pi}{\log \tfrac{1}{{\varepsilon}}}\Bigl( q(\Hat{x})-\kappa H(\Hat{x}, \Hat{x})+\dfrac{\kappa}{2\pi} \log \rho_\kappa \Bigr)+O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-2}\bigr),$$ such that, if $$\Hat{u}^{\varepsilon}(x)=U_{\Hat{\kappa}^{\varepsilon}}\Bigl(\frac{x-\Hat{x}}{{\varepsilon}}\Bigr)+\Hat{\kappa}^{\varepsilon}\Bigl( \frac{1}{2\pi} \log \frac{1}{{\varepsilon}\rho_{\Hat{\kappa}^{\varepsilon}}}+H(\Hat{x}, x) \Bigr),$$ then $$\Hat{u}^{\varepsilon}\in \mathcal{N}^{\varepsilon}.$$ Moreover, we have $$\Hat{A}^{\varepsilon}:=\Bigl\{ x {\: :\:}\Hat{u}^{\varepsilon}(x) > q(x)+\frac{\kappa}{2\pi} \log \frac{1}{\epsilon} \Bigr\} \subset B(\Hat{x}, \Hat{r}^{\varepsilon}),$$ with $\Hat{r}^{\varepsilon}=O({\varepsilon})$.
For $\sigma \in {\mathbf{R}}$, define $$\begin{aligned}
\Hat{\kappa}^{{\varepsilon}, \sigma}&=\frac{q^{\varepsilon}(\Hat{x})+\sigma}{\tfrac{1}{2\pi} \log \tfrac{1}{{\varepsilon}\rho_\kappa}+H(\hat{x}, \Hat{x})}, \\
\Hat{\rho}^{{\varepsilon}, \sigma}&=\rho_{\Hat{\kappa}^{{\varepsilon}, \sigma}}, \\
\Hat{u}^{{\varepsilon}, \sigma}(x)&=U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}\Bigl(\frac{x-\Hat{x}}{{\varepsilon}}\Bigr)+\Hat{\kappa}^{{\varepsilon}, \sigma} \Bigl( \frac{1}{2\pi} \log \frac{1}{{\varepsilon}\rho_{\Hat{\kappa}^{{\varepsilon}, \sigma}}}+H(\Hat{x}, x) \Bigr). \end{aligned}$$ First note that when ${\varepsilon}>0$ is sufficiently small, $\Hat{u}^{{\varepsilon}, \sigma}(x)=\hat{\kappa}_{\sigma, {\varepsilon}} G(\Hat{x}, x)$ in a neighborhood of $\partial \Omega$, so that $\Hat{u}^{{\varepsilon}, \sigma} \in
W^{1, 2}_0(\Omega)$ and we can define $$g^{\varepsilon}(\sigma)=\langle d \mathcal{E}^{\varepsilon}(\Hat{u}^{{\varepsilon}, \sigma}), \Hat{u}^{{\varepsilon}, \sigma} \rangle.$$ Among the terms involved in $g^{\varepsilon}(\sigma)$, we may already compute $$\begin{split}
\int_{\Omega} {\lvert\nabla \Hat{u}^{{\varepsilon}, \sigma}\rvert}^2
&=\int_{B(\Hat{x}, {\varepsilon}\rho_{\Hat{\kappa}^{{\varepsilon}, \sigma}})}\!\!\!\!\!\!\!\!\!\!\!\! {\lvert\nabla (U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}(\tfrac{\cdot-\Hat{x}}{{\varepsilon}}) +\Hat{\kappa}^{{\varepsilon}, \sigma} H(\Hat{x}, \cdot))\rvert}^2 \\
&\qquad\qquad+(\Hat{\kappa}^{{\varepsilon}, \sigma})^2\int_{\Omega \setminus B(\Hat{x}, \rho_{\Hat{\kappa}^{{\varepsilon}, \sigma}} {\varepsilon})}\!\!\!\!\!\!\!\!\!\!\!\! {\lvert\nabla G(\Hat{x}, \cdot)\rvert}^2 \\
&=\int_{B(0, \rho_{\Hat{\kappa}^{{\varepsilon}, \sigma}})}\!\!\!\!\!\!\!\!\!\!\!\! {\lvert\nabla U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}\rvert}^2+O({\varepsilon}) \\
&\qquad\qquad+(\Hat{\kappa}^{{\varepsilon}, \sigma})^2\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}\rho_{\Hat{\kappa}^{{\varepsilon}, \sigma}}}+H(\Hat{x}, \Hat{x})+O({\varepsilon}) \Bigr) \\
&=\int_{B(0, \rho_{\Hat{\kappa}^{{\varepsilon}, \sigma}})}\!\!\!\!\!\!\!\!\!\!\!\! {\lvert\nabla U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}\rvert}^2
+\Hat{\kappa}^{{\varepsilon}, \sigma} \bigl( q^{\varepsilon}(\Hat{x})+\sigma \bigr)+O({\varepsilon}).
\end{split}$$
In order to estimate the second term involved in $g^{\varepsilon}(\sigma)$, namely $\frac{1}{{\varepsilon}^2}\int_{\Omega}
f(\Hat{u}^{{\varepsilon}, \sigma}-q^{\varepsilon})\Hat{u}^{{\varepsilon}, \sigma}$, we first claim that $$\label{HatAepssigma}
\Hat{A}^{{\varepsilon}, \sigma}:=\bigl\{x \in \Omega {\: :\:}\Hat{u}^{{\varepsilon}, \sigma}(x) >
q^{\varepsilon}(x)\bigr\} \subset B(\Hat{x}, r^{\varepsilon}),$$ with $r^{\varepsilon}=O({\varepsilon})$. Indeed, let $x \in \Hat{A}^{{\varepsilon}, \sigma} \setminus
B(\Hat{x}, \Hat{\rho}^{{\varepsilon}, \sigma}{\varepsilon})$. One has, by definition of $\Hat{u}^{{\varepsilon}, \sigma}(x)$ and of $\Hat{\kappa}^{{\varepsilon}, \sigma}$, $$ \Hat{\kappa}^{{\varepsilon}, \sigma}\Bigl(\frac{1}{2\pi}\log
\frac{1}{{\varepsilon}}+\frac{1}{2\pi} \log \frac{{\varepsilon}}{{\lvertx-\Hat{x}\rvert}}+H(\Hat{x}, x)\Bigr)
> q(x)+\frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}},
$$ so that $$\label{ineqVorticitySetUpperFrac}
\frac{\dfrac{1}{2\pi} \log \dfrac{1}{{\varepsilon}}+\dfrac{1}{2\pi} \log
\dfrac{{\varepsilon}}{{\lvertx-\Hat{x}\rvert}}+ H(\Hat{x}, x)}{\dfrac{\kappa}{2\pi} \log \dfrac{1}{{\varepsilon}}+q(x)} \ge
\dfrac{\log \dfrac{1}{{\varepsilon}}+H(\Hat{x}, \Hat{x})}{\dfrac{\kappa}{2\pi} \log \dfrac{1}{{\varepsilon}}+q(\Hat{x})+\sigma}.$$ Since $q$ and $H(\Hat{x}, \cdot)$ are bounded functions, one obtains that $$\frac{1}{\kappa}+\frac{\log \frac{{\varepsilon}}{{\lvertx-\Hat{x}\rvert}}}{\kappa \log \dfrac{1}{{\varepsilon}}}\ge \frac{1}{\kappa}+O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr),$$ and the claim is proved. We deduce from , that for every $x \in \Hat{A}^{{\varepsilon}, \sigma}$ $$\Hat{u}^{{\varepsilon}, \sigma}(x)-q^{\varepsilon}(x)=U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}\Bigl(\frac{x-\Hat{x}}{{\varepsilon}}\Bigr)+\sigma+O({\varepsilon}).$$
We may now estimate $$\begin{split}
\frac{1}{{\varepsilon}^2}\int_{\Omega}& f(\Hat{u}^{{\varepsilon}, \sigma}-q^{\varepsilon})\Hat{u}^{{\varepsilon}, \sigma}
=\frac{1}{{\varepsilon}^2}\int_{\Hat{A}^{{\varepsilon}, \sigma}} f(\Hat{u}^{{\varepsilon}, \sigma}-q^{\varepsilon})\Hat{u}^{{\varepsilon}, \sigma}\\
&=\frac{1}{{\varepsilon}^2}\int_{\Hat{A}^{{\varepsilon}, \sigma}} f(\Hat{u}^{{\varepsilon}, \sigma}-q^{\varepsilon})U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}(\tfrac{\cdot-\Hat{x}}{{\varepsilon}}) \\
&\qquad\qquad+\frac{\Hat{\kappa}^{{\varepsilon}, \sigma}}{{\varepsilon}^2}\int_{\Hat{A}^{{\varepsilon}, \sigma}} f(\Hat{u}^{{\varepsilon}, \sigma}-q^{\varepsilon})\bigl(\tfrac{1}{2\pi}\log \tfrac{1}{{\varepsilon}\Hat{\rho}^{{\varepsilon}, \sigma}}+ H(\Hat{x}, \cdot)\bigr) \\
&=\int_{{\mathbf{R}}^2} f(U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}+\sigma)U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}+O({\varepsilon}) \\
&\qquad\qquad+ \Hat{\kappa}^{{\varepsilon}, \sigma}\bigl(\tfrac{1}{2\pi} \log \tfrac{1}{{\varepsilon}\rho_\kappa}+ H(\Hat{x}, \Hat{x})+O({\varepsilon})\bigr)\Bigl(\int_{{\mathbf{R}}^2} f(U_\kappa+\sigma)+O({\varepsilon})\Bigr)\\
&=\int_{{\mathbf{R}}^2} f(U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}+\sigma)U_{\Hat{\kappa}^{{\varepsilon}, \sigma}} \\
&\qquad\qquad+ \big(\tfrac{\kappa}{2\pi} \log
\tfrac{1}{{\varepsilon}}+q(\Hat{x})+\sigma\big)\int_{{\mathbf{R}}^2}f(U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}+\sigma)+O({\varepsilon}{{\vert\!\log{\varepsilon}\vert}}).
\end{split}$$ Summarizing, we have $$\begin{split}
g^{\varepsilon}(\sigma)&=\frac{\kappa}{2\pi}\log \frac{1}{{\varepsilon}} \Bigl(\Hat{\kappa}^{{\varepsilon}, \sigma} - \int_{{\mathbf{R}}^2} f(U_{\Hat{\kappa}^{{\varepsilon}, \sigma}}+\sigma)\Bigr)+O(1)\\
&=\frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}} \Bigl(\int_{{\mathbf{R}}^2} f(U_{\kappa})-f(U_{\kappa}+\sigma)\Bigr)+O(1).
\end{split}$$ Since $g^{\varepsilon}$ is continuous and $\sigma \cdot \Bigl(\int_{{\mathbf{R}}^2} f(U_\kappa)-f(U_\kappa+\sigma)\Bigr)<0$ when $\sigma \ne 0$, there exists $\sigma^{\varepsilon}$ such that $g(\sigma^{\varepsilon})=0$ and $\sigma^{\varepsilon}\to 0$ as ${\varepsilon}\to 0$. One then sets $\Hat{\kappa}^{\varepsilon}=\Hat{\kappa}^{{\varepsilon}, \sigma^{\varepsilon}}$.
\[lemmaEnergyHatu\] For every $\Hat{x} \in \Omega$, we have $$c^{\varepsilon}\le \frac{\kappa^2}{4\pi}\log \frac{1}{{\varepsilon}}
-\mathcal{W}(\Hat{x})+\mathcal{C}+o(1)\qquad\text{as }{\varepsilon}\to 0.$$
By Lemma \[lemmaHatuNehari\], $\Hat{u}^{\varepsilon}\in \mathcal{N}^{\varepsilon}$, so that $c^{\varepsilon}\leq \mathcal{E}^{\varepsilon}(\Hat{u}^{\varepsilon})$. We compute the energy of $\Hat{u}^{\varepsilon}$ as follows. First, $$\begin{split}
\int_{\Omega} {\lvert\nabla \Hat{u}^{\varepsilon}\rvert}^2
&=\int_{\Omega} \Hat{u}^{\varepsilon}\Delta \Hat{u}^{\varepsilon}\\
&= -\int_{{\mathbf{R}}^2} U_\kappa \Delta U_\kappa+(\Hat{\kappa}^{\varepsilon})^2\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}}+H(\Hat{x}, \Hat{x})\Bigr)+o(1)\\
&=\int_{{\mathbf{R}}^2} {\lvert\nabla (U_\kappa)_+\rvert}^2+\frac{\kappa^2}{2\pi} \log \frac{1}{{\varepsilon}} +2\kappa q(\Hat x)-\kappa^2 H(\Hat x, \Hat x)
+\frac{\kappa^2}{2\pi} \log{\rho_\kappa} +o(1).
\end{split}$$ Next, $$\begin{split}
\frac{1}{{\varepsilon}^2}\int_{\Omega} F(\Hat{u}^{\varepsilon}-q^{\varepsilon})
&=\frac{1}{{\varepsilon}^2}\int_{\Hat{A}^{\varepsilon}} F(\Hat{u}^{\varepsilon}-q^{\varepsilon})\\
&=\frac{1}{{\varepsilon}^2}\int_{\Hat{A}^{\varepsilon}} F(\Hat{u}^{\varepsilon}-q^{\varepsilon}(x^{\varepsilon}))+o(1)\\
&=\int_{{\mathbf{R}}^2} F(U_\rho)+o(1),
\end{split}$$ and the conclusion follows from the definitions of $\mathcal{W}$ and $\mathcal{C}$.
\[cor:upper\] We have $$c^{\varepsilon}\leq \frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}} -\sup_{x\in \Omega} \mathcal{W}(x) +\mathcal{C} + o(1).$$
Asymptotic behavior of solutions
--------------------------------
The main goal of this section is to prove
\[prop:1mai\] Let $(v^{\varepsilon})$ be a family of solutions to such that $v^{\varepsilon}\ne 0$ $$\label{assumptEnergyUpperbound}
\mathcal{E}^{\varepsilon}(v^{\varepsilon}) \le \frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}}+O(1),$$ as ${\varepsilon}\to 0$. Define the quantities $A^{\varepsilon}$, $\omega^{\varepsilon}$, $\kappa^{\varepsilon}$, $x^{\varepsilon}$ and $\rho^{\varepsilon}$ for $v^{\varepsilon}$ as in for $u^{\varepsilon}$. Then $$v^{\varepsilon}=U_{\kappa^{\varepsilon}} (\tfrac{\cdot-x^{\varepsilon}}{{\varepsilon}})+\kappa^{\varepsilon}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}\rho^{\varepsilon}}+ H(x^{\varepsilon}, \cdot)\Bigr)+o(1),$$ in $\mathrm{W}^{2, 1}_\mathrm{loc}(\Omega)$, in $\mathrm{W}^{1, 2}_0(\Omega)$, and in $\mathrm{L}^\infty(\Omega)$, where $$\kappa^{\varepsilon}=\kappa+\frac{2\pi}{\log \frac{1}{{\varepsilon}}}\Bigl(q(x^{\varepsilon})-\kappa H(x^{\varepsilon}, x^{\varepsilon}) -\frac{\kappa}{2\pi} \log \frac{1}{\rho_\kappa} \Bigr)+o({{\vert\!\log{\varepsilon}\vert}}^{-1}),$$ In particular, we have $$\mathcal{E}^{\varepsilon}(v^{\varepsilon})= \frac{\kappa^2}{4\pi}\log \frac{1}{{\varepsilon}}-\mathcal{W}(x^{\varepsilon})+\mathcal{C}+o(1)$$ and $$B(x^{\varepsilon}, \Bar{r}^{\varepsilon}) \subset A^{\varepsilon}\subset B(x^{\varepsilon}, \mathring{r}^{\varepsilon}),$$ with $\Bar{r}^{\varepsilon}={\varepsilon}\rho_\kappa+o({\varepsilon})$ and $\mathring{r}^{\varepsilon}={\varepsilon}\rho_\kappa +o({\varepsilon})$.
In other words, $v^{\varepsilon}$ satisfies the same asymptotics as the one stated in Theorem \[thm:K1\] for $u^{\varepsilon}$ except for the convergence of $x^{\varepsilon}$.
In the sequel, $v^{\varepsilon}$ denotes a family of nontrivial solutions to verifying . We divide the proof of Proposition \[prop:1mai\] into several steps.
### Step 1: First quantitative properties of the solutions
In this section, we derive various types of estimates for $v^{\varepsilon}$.
\[propositionEstimatesueps\] We have, as ${\varepsilon}\to 0$, $$\begin{gathered}
\label{ineqMuAeps}{\mathcal{L}^{2}}(A^{\varepsilon}) = O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr), \\
\label{ineqVortexEnergy} \int_{A^{\varepsilon}} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2 =O(1), \\
\label{ineqVortexPotential}\frac{1}{{\varepsilon}^2}\int_{A^{\varepsilon}} F(v^{\varepsilon}-q^{\varepsilon}) =O(1), \\
\label{eq:2etoiles}\int_{\Omega\setminus A^{\varepsilon}} |\nabla v^{\varepsilon}|^2 \leq \frac{\kappa^2}{2\pi} \log\frac{1}{{\varepsilon}} + O(1), \\
\label{ineqTotalVorticity}\int_{\Omega} \omega^{\varepsilon}\leq \kappa + O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr).\end{gathered}$$
First note that for ${\varepsilon}>0$ sufficiently small, $$\label{ineqEnergy}
\Bigl(\frac{1}{2}-\frac{1}{p+1}\Bigr)\int_\Omega |\nabla v^{\varepsilon}|^2 \leq \mathcal{E}^{\varepsilon}(v^{\varepsilon}).$$ Indeed, $$\mathcal{E}^{\varepsilon}(v^{\varepsilon}) = \frac{1}{2}\int_\Omega |\nabla v^{\varepsilon}|^2 -
\frac{1}{p+1}\int_\Omega
\frac{1}{{\varepsilon}^2}f(v^{\varepsilon}-q^{\varepsilon})(v^{\varepsilon}-q^{\varepsilon})_+,$$ and, by testing $(\mathcal{P}^{\varepsilon})$ against $v^{\varepsilon}$, $$0 = \frac{1}{p+1}\int_\Omega |\nabla v^{\varepsilon}|^2 -
\frac{1}{p+1}\int_\Omega
\frac{1}{{\varepsilon}^2}f(v^{\varepsilon}-q^{\varepsilon})v^{\varepsilon}.$$ Since $(v^{\varepsilon}-q^{\varepsilon})_+\leq v^{\varepsilon}$ when $q^{\varepsilon}\geq 0$, and hence when ${\varepsilon}$ is sufficiently small, follows by subtraction.
In order to obtain , first note that since $q$ is bounded from below, for ${\varepsilon}$ sufficiently small, $\inf_{\Omega} q_{\varepsilon}>
\frac{\kappa}{4\pi} \log \frac{1}{{\varepsilon}}$. By the Chebyshev and Poincaré inequalities, it follows that $${\mathcal{L}^{2}}(A^{\varepsilon}) \le \Bigl(\frac{1}{\inf_{\Omega} q^{\varepsilon}}\Bigr)^2 \int_{\Omega} {\lvertv^{\varepsilon}\rvert}^2 \le \frac{C}{{{\vert\!\log{\varepsilon}\vert}}^2} \int_{\Omega} {\lvert\nabla v^{\varepsilon}\rvert}^2\le \frac{C'}{{{\vert\!\log{\varepsilon}\vert}}},$$ where the last inequality is a consequence and .
We claim that $$\label{ineqomegaepsL1}
\int_{\Omega} \omega^{\varepsilon}\leq C.$$ By testing ${(\mathcal{P}^{\varepsilon})}$ against $\min(v^{\varepsilon}, q^{\varepsilon})$ we obtain $$\label{ineqVorticityEnergy}
\begin{split}
\int_{\Omega} \omega^{\varepsilon}= \int_{A^{\varepsilon}}
\frac{1}{{\varepsilon}^2}f(v^{\varepsilon}-q^{\varepsilon})& \leq \frac{1}{\inf_\Omega q^{\varepsilon}}\int_{A^{\varepsilon}}
\frac{q^{\varepsilon}}{{\varepsilon}^2}f(v^{\varepsilon}-q^{\varepsilon})\\
&=\frac{1}{\inf_\Omega q^{\varepsilon}}\int_{\Omega\setminus A^{\varepsilon}}
{\lvert\nabla v^{\varepsilon}\rvert}^2 + \frac{1}{\inf_\Omega q^{\varepsilon}}\int_{A^{\varepsilon}} \nabla v^{\varepsilon}\nabla q.
\end{split}$$ In view of , this yields $$\kappa^{\varepsilon}\leq C \frac{\mathcal{E}^{\varepsilon}(v^{\varepsilon}) +o(1)}{\log \frac{1}{{\varepsilon}}},$$ and the estimate follows from assumption .
Testing now ${(\mathcal{P}^{\varepsilon})}$ against $(v^{\varepsilon}-q^{\varepsilon})_+$, we obtain $$\label{eqNehariVortex}
\int_{A^{\varepsilon}} |\nabla (v^{\varepsilon}-q^{\varepsilon})|^2 = \int_{A^{\varepsilon}} \frac{1}{{\varepsilon}^2} (v^{\varepsilon}-q^{\varepsilon})_+^{p+1} -\int_{A^{\varepsilon}}\nabla (v^{\varepsilon}-q^{\varepsilon}) \nabla q.$$ The Gagliardo–Nirenberg inequality [@Nirenberg1959]\*[p.125]{} yields $$\label{ineqGN}
\int_{A^{\varepsilon}} \frac{1}{{\varepsilon}^2} (v^{\varepsilon}-q^{\varepsilon})_+^{p+1} \leq C \int_{A^{\varepsilon}}
\frac{1}{{\varepsilon}^2} (v^{\varepsilon}-q^{\varepsilon})_+^{p} \left(\int_{A^{\varepsilon}} |\nabla
(v^{\varepsilon}-q^{\varepsilon})|^2\right)^{\frac{1}{2}},$$ so that $$\begin{split}
\int_{A^{\varepsilon}} |\nabla (v^{\varepsilon}-q^{\varepsilon})|^2 &\leq C \bigl({\Vert\omega^{\varepsilon}\Vert}_{\mathrm{L}^1}+{\Vert\nabla q^{\varepsilon}\Vert}_{L^2(A^{\varepsilon})} \bigr) \Bigl(\int_{A^{\varepsilon}}
{\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2\Bigr)^{\frac{1}{2}} \\
&\leq C'\Bigl(\int_{A^{\varepsilon}}
{\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2\Bigr)^{\frac{1}{2}}.
\end{split}$$ Inequality can therefore be deduced from , and follows from . Finally, $$\begin{split}
\frac{1}{2}\int_{\Omega \setminus A^{\varepsilon}} {\lvert\nabla v^{\varepsilon}\rvert}^2&=\mathcal{E}^{\varepsilon}(v^{\varepsilon})+\frac{1}{{\varepsilon}^2}\int_{A^{\varepsilon}} F(v^{\varepsilon}-q^{\varepsilon})-\frac{1}{2} \int_{A^{\varepsilon}} {\lvert\nabla v^{\varepsilon}\rvert}^2 \\
&\le \frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}}+O(1),
\end{split}$$ so that holds, and inequality then follows from .
The use of the Gagliardo–Nirenberg inequality to obtain is the only step in our proof that requires $f$ to be a power-like nonlinearity.
### Step 2: Structure of the vorticity set
We now examine the vorticity set $A^{\varepsilon}$ further. Since $A^{\varepsilon}$ is open, it contains at most countably many connected components that we label $A^{\varepsilon}_i$, $i \in I^{\varepsilon}$. If $q$ were a harmonic function (e.g. if the only goal was to prove Theorem \[thm:resu\]), one would deduce from the fact that $u^{\varepsilon}$ is a minimal energy solution that $A^{\varepsilon}$ is connected whenever $q^{\varepsilon}\ge 0$ [@BergerFraenkel1974]\*[Theorem 3F]{}, [@Norbury1975]\*[Theorem 3.4]{}, [@AmbrosettiMancini1981]\*[Theorem 4]{}, [@Yang1991]\*[Theorem 1]{}, [@LiYanYang2005]\*[Proposition 3.1]{}; this would simplify considerably the analysis that we perform below.
First we have a control on the total area and on the diameter of each connected component.
\[lemmaAreaDiameter\] If ${\varepsilon}> 0$ is sufficiently small, we have $$\label{ineqVorticityAreaStrong}
{\mathcal{L}^{2}}(A^{\varepsilon}) \le C {\varepsilon}^2$$ and, for every $i \in I^{\varepsilon}$, $$\label{ineqVorticityDiameter}
\operatorname{diam}(A^{\varepsilon}_i) \le C {\varepsilon}.$$
Set $$w^{\varepsilon}=\frac{v^{\varepsilon}}{\min_{\partial A^{\varepsilon}}q^{\varepsilon}}.$$ Since $v^{\varepsilon}=q^{\varepsilon}$ on $\partial A^{\varepsilon}$, we have, by , $$\label{ineqCapacity}
\frac{2\pi}{\operatorname{cap}(A^{\varepsilon}, \Omega)}
\ge \frac{2\pi}{\displaystyle \int_{\Omega\setminus A^{\varepsilon}} {\lvert\nabla w^{\varepsilon}\rvert}^2}
\ge 2 \pi \frac{\frac{\kappa^2}{4\pi} \bigl(\log \frac{1}{{\varepsilon}}\bigr)^2+O({{\vert\!\log{\varepsilon}\vert}})}{\displaystyle \int_{\Omega\setminus A^{\varepsilon}}
{\lvert\nabla v^{\varepsilon}\rvert}^2}=\log \frac{1}{{\varepsilon}}+O(1).$$ By Proposition \[propositionCapacityArea\], it follows that $$\log \frac{{\mathcal{L}^{2}}(\Omega)}{{\mathcal{L}^{2}}(A^{\varepsilon})} \ge 2\log \frac{1}{{\varepsilon}}+O(1),$$ from which follows.
Similarly, we have $$\frac{2\pi}{\operatorname{cap}(A^{\varepsilon}_i, \Omega)} \geq \frac{2\pi}{\operatorname{cap}(A^{\varepsilon}, \Omega)} \geq \log \frac{1}{{\varepsilon}}+O(1).$$ It hence follows from Proposition \[propositionBoundDiameter\] and the boundedness of $\Omega$ that $$\log C\Bigl(1+\frac{1}{\operatorname{diam}(A_i^{\varepsilon})}\Bigr) \ge \log \frac{1}{{\varepsilon}}+O(1),$$ which implies .
\[lemmaVortexSplit\] There exist positive constants $\gamma$ and $c$ such that when ${\varepsilon}$ is small enough, for every $i \in I_{\varepsilon}$, if $$\label{eqSplitVortices}
\int_{A^{\varepsilon}_i} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2 > \gamma^2,$$ then $$\begin{gathered}
\label{ineqLowerBoundArea} {\mathcal{L}^{2}}(A_i^{\varepsilon})\ge c{\varepsilon}^2, \\
\label{ineqLowerBoundDiam} \operatorname{diam}(A^{\varepsilon}_i)\ge c{\varepsilon}, \\
\label{ineqLowerBoundDistance} \operatorname{dist}(A^{\varepsilon}_i, \partial \Omega)\ge c, \\
\label{ineqLowerBoundVortex} \int_{A^{\varepsilon}_i} \omega_{\varepsilon}\ge c,\end{gathered}$$ while if does not hold, then for every $s \ge 1$, $$\label{ineqfsVanishing} \int_{A^{\varepsilon}_i} f(v^{\varepsilon}-q^{\varepsilon})^s \le C {\Vert\nabla q\Vert}_{\mathrm{L}^r(A^{\varepsilon}_i)}^{sp} {\mathcal{L}^{2}}(A^{\varepsilon}_i)^{1+\frac{sp}{2}(1-\frac{2}{r})},$$ where $C>0$ only depends on $s \ge 1$.
Starting from , and applying the Sobolev and Cauchy–Schwarz inequalities we obtain $$\begin{gathered}
\label{ineqGradientVortices}
\int_{A^{\varepsilon}_i} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})_+\rvert}^2
= \int_{A^{\varepsilon}_i} \frac{f(v^{\varepsilon}-q^{\varepsilon})}{{\varepsilon}^2}(v^{\varepsilon}-q^{\varepsilon})_+-\int_{A^{\varepsilon}_i} \nabla q \cdot \nabla (v^{\varepsilon}-q^{\varepsilon})\\
\le C\frac{{\mathcal{L}^{2}}(A^{\varepsilon}_i)}{{\varepsilon}^2}
\Bigl(\int_{A^{\varepsilon}_i} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})_+\rvert}^2\Bigr)^{\frac{p+1}{2}}\\
+ {\Vert\nabla q\Vert}_{\mathrm{L}^2(A^{\varepsilon}_i)}{\Vert\nabla (v^{\varepsilon}-q^{\varepsilon})_+\Vert}_{\mathrm{L}^2(A^{\varepsilon}_i)}. \end{gathered}$$ By Lemma \[lemmaAreaDiameter\], we may choose $\gamma$ sufficiently small so that $$\gamma^{p-1}\le \frac{{\varepsilon}^2}{2C{\mathcal{L}^{2}}(A^{\varepsilon}_i)},$$ independently of ${\varepsilon}$, and therefore if does not hold we obtain $$\label{ineqVanVorticesuq}
\frac{1}{2}\int_{A^{\varepsilon}_i} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})_+\rvert}^2\le \int_{A^{\varepsilon}_i} {\lvert\nabla q\rvert}^2.$$ Applying successively Sobolev inequality, and Lemma \[lemmaAreaDiameter\], we conclude $$\begin{split}
\int_{A^{\varepsilon}_i} f(v^{\varepsilon}-q^{\varepsilon})^s
&\le C \Bigl( \int_{A^{\varepsilon}_i} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})_+\rvert}^2 \Bigr)^\frac{sp}{2} {\mathcal{L}^{2}}(A^{\varepsilon}_i)\\
&\le C'\Bigl( \int_{A^{\varepsilon}_i} {\lvert\nabla q\rvert}^2 \Bigr)^\frac{sp}{2} {\mathcal{L}^{2}}(A^{\varepsilon}_i)\\
&\le C'' {\Vert\nabla q\Vert}_{\mathrm{L}^r(A^{\varepsilon}_i)}^{sp}
{\mathcal{L}^{2}}(A^{\varepsilon}_i)^{1+\frac{sp}{2}(1-\frac{2}{r})}.
\end{split}$$
Assume now that holds. Combined with and , this yields $$\gamma^2 \le C \frac{{\mathcal{L}^{2}}(A^{\varepsilon}_i)}{{\varepsilon}^2}+C {\Vert\nabla q\Vert}_{\mathrm{L}^2(A^{\varepsilon}_i)}.$$ Since ${\Vert\nabla q\Vert}_{\mathrm{L}^2(A^{\varepsilon}_i)} \to 0$ as ${\varepsilon}\to 0$, one must have ${\mathcal{L}^{2}}(A^{\varepsilon}) \ge c {\varepsilon}^2$. The isodiametric inequality then yields .
Turning back to , and using Proposition \[propositionBoundDiameter\], we obtain $$\log C\Bigl(1+\frac{\operatorname{dist}(A_i^{\varepsilon}, \partial \Omega)}{{\varepsilon}}\Bigr) \ge \log \frac{1}{{\varepsilon}}+O(1),$$ from which follows.
Testing ${(\mathcal{P}^{\varepsilon})}$ against $(v^{\varepsilon}-q^{\varepsilon})_+ \chi_{A_{\varepsilon}^i}$, applying the Gagliardo–Nirenberg inequality and using then , we have $$\int_{A^{\varepsilon}_i} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2 \leq C(\int_{A^{\varepsilon}_i} \omega_{\varepsilon}+{\Vert\nabla q^{\varepsilon}\Vert}_{L^2(A^{\varepsilon}_i)})\Bigl(\int_{A^{\varepsilon}_i}
{\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2\Bigr)^{\frac{1}{2}}
\le C'\int_{A^{\varepsilon}_i} \omega_{\varepsilon},$$ (cf. the proof of Proposition \[propositionEstimatesueps\]) and the inequality follows.
In view of Lemma \[lemmaVortexSplit\], we can split the vortices in two classes: the vanishing vortices $$\begin{aligned}
\label{eqDefVeps} V^{\varepsilon}&=\bigcup \Bigl\{A_i^{\varepsilon}{\: :\:}\int_{A_i^{\varepsilon}} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2 \le \gamma^2\Bigr\}, \\
\intertext{and the essential vortices}
\label{eqDefEeps} E^{\varepsilon}&=\bigcup \Bigl\{A_i^{\varepsilon}{\: :\:}\int_{A_i^{\varepsilon}} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2 > \gamma^2\Bigr\}. \end{aligned}$$ In view of , $E^{\varepsilon}$ contains finitely many connected components. We can thus split $E^{\varepsilon}=\bigcup_{j=1}^{k^{\varepsilon}} E^{\varepsilon}_j$, where $E^{\varepsilon}_j$ are nonempty open sets which are not necessarily connected such that, up to a subsequence, $$\label{eqDistEepsi}
\frac{\operatorname{dist}(E^{\varepsilon}_i, E^{\varepsilon}_j)}{{\varepsilon}} \to \infty$$ as ${\varepsilon}\to 0$, and $$\label{ineqDiamEepsi} \Tilde{\rho}= \limsup_{{\varepsilon}\to 0} \frac{\operatorname{diam}(E^{\varepsilon}_i)}{{\varepsilon}} < \infty.$$ By definition of $E^{\varepsilon}$ and by , $k^{\varepsilon}$ is bounded as ${\varepsilon}\to 0$. Finally, $$\label{eqDistbord}
\liminf_{{\varepsilon}\to 0} \operatorname{dist}(E^{\varepsilon}_i, \partial \Omega)>0.$$
We set $$\begin{aligned}
\omega^{\varepsilon}_v&=\omega^{\varepsilon}{\chi_{V^{\varepsilon}}}, &
\omega^{\varepsilon}_i&=\omega^{\varepsilon}{\chi_{E^{\varepsilon}_i}}, &
\kappa^{\varepsilon}_i&=\int_{\Omega} \omega^{\varepsilon}_i. \end{aligned}$$ By , we have $$\label{ineqSumVortices}
\sum_{i=1}^{k^{\varepsilon}} \kappa^{\varepsilon}_i \le \kappa+O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr).$$
\[lemmaVanishingVorticity\] For every $s \ge 1$, we have $${\Vert\omega^{\varepsilon}_v\Vert}_{\mathrm{L}^s} = o\bigl({\varepsilon}^{p(1-\frac{2}{r})-2(1-\frac{1}{s})}\bigr).$$ In particular, if $\frac{1}{s} \ge 1-p(\frac{1}{2}-\frac{1}{r})$, then $\omega^{\varepsilon}_v \to 0$ in $\mathrm{L}^s(\Omega)$.
Set $$I_v^{\varepsilon}=\Bigl\{ i \in I^{\varepsilon}{\: :\:}\int_{A_i^{\varepsilon}} {\lvert\nabla (v^{\varepsilon}-q^{\varepsilon})\rvert}^2 \le \gamma^2\Bigr\}$$ We have, by Lemma \[lemmaVortexSplit\] and by , $$\begin{split}
\int_{\Omega} {\lvert\omega^{\varepsilon}_v\rvert}^s
&=\sum_{i \in I^{\varepsilon}_v}\int_{A^{\varepsilon}_i} {\lvert\omega^{\varepsilon}_v\rvert}^s \\
&\le C\frac{1}{{\varepsilon}^{2s}} \sum_{i \in I^{\varepsilon}_v} {\Vert\nabla q\Vert}_{\mathrm{L}^r(A^{\varepsilon}_i)}^{sp} {\mathcal{L}^{2}}(A^{\varepsilon}_i)^{1+\frac{sp}{2}(1-\frac{2}{r})}\\
&\le C{\mathcal{L}^{2}}(V^{\varepsilon}) \max_{i \in I^{\varepsilon}_v} {\Vert\nabla q\Vert}_{\mathrm{L}^r(A^{\varepsilon}_i)}^{sp} \frac{{\mathcal{L}^{2}}(A^{\varepsilon}_i)^{1+sp(\frac{1}{2}-\frac{1}{r})}}{{\varepsilon}^{2s}}\\
&\le C' {\Vert\nabla q\Vert}_{\mathrm{L}^r(V^{\varepsilon})}^{sp} {\varepsilon}^{sp(1-\frac{2}{r})-2(s-1)}. \qedhere
\end{split}$$
\[lemmaNonVanisingVortex\] For ${\varepsilon}> 0$ sufficiently small, $k_{\varepsilon}\ge 1$.
Assume by contradiction that there is a sequence $({\varepsilon}_n)$ such that ${\varepsilon}_n \to 0$ and $k_{{\varepsilon}_n} =0$. Take $s > 1$ such that $\frac{1}{s} \ge 1-p(\frac{1}{2}-\frac{1}{r})$. Since $\omega_{{\varepsilon}_n}=\omega_{{\varepsilon}_n}^v \to 0$ in $\mathrm{L}^s(\Omega)$ for some $s > 1$ by Lemma \[lemmaVanishingVorticity\]; by classical estimates, [@GilbargTrudinger2001]\*[Theorem 8.15]{} $v_{{\varepsilon}_n} \to 0$ in $L^\infty(\Omega)$. Therefore, when $n$ is large enough, one would have $\omega_{{\varepsilon}_n}=0$ and thus $v_{{\varepsilon}_n} = 0$.
### Step 3: Small scale asymptotics
We define $$x^{\varepsilon}_i=\frac{1}{\kappa^{\varepsilon}_i}\int_{\Omega}\omega^{\varepsilon}_i(x)x\, dx.$$ By and , $x^{\varepsilon}_i \in \Omega$ and $x^{\varepsilon}_i\ne x^{\varepsilon}_j$ when $i \ne j$ and ${\varepsilon}$ is small. We also define $$v^{\varepsilon}_i(y)=v^{\varepsilon}(x^{\varepsilon}_i+{\varepsilon}y)-q^{\varepsilon}(x^{\varepsilon}_i),$$ and $$q^{\varepsilon}_i(y)=q(x^{\varepsilon}_i+{\varepsilon}y)-q(x^{\varepsilon}_i).$$ By , for every $R>0$, $v^{\varepsilon}_i$ is well-defined in $B(0, R)$ when ${\varepsilon}$ is sufficiently small, and it satisfies there the equation $$\label{eqLimit}
-\Delta v^{\varepsilon}_i=f(v^{\varepsilon}_i-q^{\varepsilon}_i).$$
\[lemmaSmallScaleLocalEstimates\] For every $R > 0$ and $s\ge 1$, there exist ${\varepsilon}(R)>0$ and $C>0$ such that for $0<{\varepsilon}\leq {\varepsilon}(R)$ we have $$\label{ineqRenormEstimate}
{\Vertf(v^{\varepsilon}_i-q^{\varepsilon}_i)\Vert}_{\mathrm{L}^s(B(0, R))}\le C.$$ Moreover, for $2\Tilde{\rho} < {\lverty\rvert} < R$, we have $$\label{ineqvepsiDecay}
\Bigl\lvert v^{\varepsilon}_i(y)-\frac{\kappa^{\varepsilon}_i}{2\pi}\log \frac{1}{{\varepsilon}{\lverty\rvert}}+q^{\varepsilon}(x^{\varepsilon}_i)-\kappa^{\varepsilon}_i H(x^{\varepsilon}_i, x^{\varepsilon}_i)
-\sum_{j \ne i} \kappa_j^{\varepsilon}G(x^{\varepsilon}_i, x^{\varepsilon}_j)\Bigr\rvert \le \frac{\kappa}{2\pi} \log \frac{{\lverty\rvert}}{{\lverty\rvert}-\Tilde{\rho}}+o(1),$$ and $$\label{ineqNablavepsiDecay}
\Bigl\lvert \nabla v^{\varepsilon}_i(y)-\frac{\kappa}{2\pi} \frac{y}{{\lverty\rvert}^2}\Bigr\rvert \le \frac{\Bar{C}}{{\lverty\rvert}^3}+o(1).$$ as ${\varepsilon}\to 0$, where $\Bar{C}$ does not depend on $R$.
Consider $D^{{\varepsilon}, R}_i=\bigcup \bigl\{A^{\varepsilon}_j {\: :\:}A^{\varepsilon}_j \cap B(x^{\varepsilon}_i, {\varepsilon}R)\ne \emptyset \bigr\}$. By , ${\mathcal{L}^{2}}(D_i^{{\varepsilon}, R})=O({\varepsilon}^2)$ as ${\varepsilon}\to 0$, so that one obtains, by Sobolev’s inequality, $$\begin{gathered}
\int_{B(0, R)} f(v^{\varepsilon}_i-q^{\varepsilon}_i)^s \le
\frac{1}{{\varepsilon}^2}\int_{D_i^{{\varepsilon}, R}} f(v^{\varepsilon}-q^{\varepsilon})^s \\
\le C\frac{1}{{\varepsilon}^2} {\Vert\nabla(v^{\varepsilon}-q^{\varepsilon})_+\Vert}_{\mathrm{L}^2(A^{\varepsilon})}^{sp} {\mathcal{L}^{2}}(D_i^{{\varepsilon}, R}) = O(1),\end{gathered}$$ which proves .
We have $$\label{eqvepsiGomega}
v^{\varepsilon}_i(y)=\int_{\Omega} G(x^{\varepsilon}_i+{\varepsilon}y, z) \omega^{\varepsilon}(z)\, dz-q^{\varepsilon}(x^{\varepsilon}_i).$$ We first prove . By a classical estimate [@GilbargTrudinger2001]\*[Theorem 8.15]{}, $$\label{ineqGomegav}
\Bigl\lvert\int_{\Omega} G(x, z) \omega^{\varepsilon}_v(z)\, dz\Bigr\rvert
\le C {\Vert\omega^{\varepsilon}_v\Vert}_{\mathrm{L}^s}.$$ Since by Lemma \[lemmaVanishingVorticity\], $\omega^{\varepsilon}_v \to 0$ in $\mathrm{L}^s(\Omega)$ for some $s > 1$, we have $$\int_{\Omega} G(x^{\varepsilon}_i+{\varepsilon}y, z) \omega^{\varepsilon}_v(z)\, dz
\to 0$$ uniformly in $y$. We also have, since $\operatorname{diam}E^{\varepsilon}_j=O({\varepsilon})$, ${\lvertx^{\varepsilon}_i-x^{\varepsilon}_j\rvert}/{\varepsilon}\to \infty$, for $j \ne i$, and ${\lverty\rvert} \le R$, $$\int_{\Omega} G(x^{\varepsilon}_i+{\varepsilon}y, z) \omega^{\varepsilon}_j(z)\, dz
=\kappa_j^{\varepsilon}G(x^{\varepsilon}_i, x^{\varepsilon}_j)+o(1),$$ and $$\int_{\Omega} H(x^{\varepsilon}_i+{\varepsilon}y, z) \omega^{\varepsilon}_i(z)\, dz
=\kappa^{\varepsilon}_i H(x^{\varepsilon}_i, x^{\varepsilon}_i)+O({\varepsilon}).$$ Finally, we have $$\begin{split}
\int_{\Omega} \frac{1}{2\pi}\log \frac{1}{{\lvertx^{\varepsilon}_i+{\varepsilon}y-z\rvert}} \omega^{\varepsilon}_i(z)\, dz
&=\int_{E^{\varepsilon}_i} \frac{1}{2\pi} \log \frac{1}{{\lvertx^{\varepsilon}_i+{\varepsilon}y-z\rvert}} \omega^{\varepsilon}_i(z)\, dz\\
&=\frac{\kappa^{\varepsilon}_i}{2\pi}\log \frac{1}{{\varepsilon}{\lverty\rvert}}+\frac{1}{2\pi} \int_{E^{\varepsilon}_i} \log \frac{{\varepsilon}{\lverty\rvert}}{{\lvertx^{\varepsilon}_i+{\varepsilon}y-z\rvert}} \omega^{\varepsilon}_i(z)\, dz.
\end{split}$$ In view of , ${\lvertx^{\varepsilon}_i-z\rvert} \le (1+o(1))\Tilde{\rho}{\varepsilon}$ when $z\in {\rm supp}(\omega^{\varepsilon}_i)$ so that for sufficiently small ${\varepsilon}$ $$\left\lvert\int_{E^{\varepsilon}_i} \log \frac{{\lvert{\varepsilon}y\rvert}}{{\lvert{\varepsilon}y+x^{\varepsilon}_i-z\rvert}} \omega^{\varepsilon}_i(z)\, dz\right\rvert\le \kappa^{\varepsilon}_i \log\frac{{\lverty\rvert}}{{\lverty\rvert}-\Tilde{\rho}} + o(1).$$
We now prove . By Lemma \[lemmaVanishingVorticity\], ${\varepsilon}\omega^{\varepsilon}_v \to 0$ in $\mathrm{L}^s(\Omega)$ for $\frac{1}{s} \ge \frac{1}{2}-p(\frac{1}{2}-\frac{1}{r})$. Choosing $s > 2$, by and classical elliptic estimates, one obtains that $$\int_{\Omega} {\varepsilon}G(x, z) \omega^{\varepsilon}_v(z)\, dz \to 0$$ as a function of $x$ in $\mathrm{W}^{2, s}_{\mathrm{loc}}(\Omega)$ and thus in $C^1_{\mathrm{loc}}(\Omega)$. Therefore, $$\int_{\Omega} {\varepsilon}\nabla G(x_i^{\varepsilon}+{\varepsilon}y, z) \omega^{\varepsilon}_v(z)\, dz \to 0$$ uniformly in $y$ on compact subsets. One also has $$\int_{\Omega} {\varepsilon}\nabla G(x^{\varepsilon}_i+{\varepsilon}y, z) \omega^{\varepsilon}_j(z)\, dz
={\varepsilon}\kappa_j^{\varepsilon}\nabla G(x^{\varepsilon}_i, x^{\varepsilon}_j)+o(1)$$ and $$\int_{\Omega} {\varepsilon}\nabla H(x^{\varepsilon}_i+{\varepsilon}y, z) \omega^{\varepsilon}_j(z)\, dz
={\varepsilon}\kappa^{\varepsilon}_i \nabla H(x_i, x_j)+O({\varepsilon}^2).$$ Finally, recall that $\int_{\Omega} \omega^{\varepsilon}_i=\kappa^{\varepsilon}_i$ and $\int_{\Omega} (x^{\varepsilon}_i-z)\omega^{\varepsilon}_i(z)\, dz=0$, so that $$\begin{gathered}
\int_{\Omega} {\varepsilon}\frac{x^{\varepsilon}_i+{\varepsilon}y -z}{{\lvertx^{\varepsilon}_i+{\varepsilon}y-z\rvert}^2} \omega^{\varepsilon}_i(z)\, dz-\kappa^{\varepsilon}_i\frac{y}{{\lverty\rvert}^2}= \\
{\varepsilon}\int_{E^{\varepsilon}_i} \Bigl(\frac{x^{\varepsilon}_i+{\varepsilon}y -z}{{\lvertx^{\varepsilon}_i+{\varepsilon}y-z\rvert}^2}- \frac{{\varepsilon}y}{{\lvert{\varepsilon}y\rvert}^2}-L({\varepsilon}y) (x^{\varepsilon}_i-z) \Bigr)\omega^{\varepsilon}_i(z)\, dz,\end{gathered}$$ where $$L(a)h=\frac{{\lverta\rvert}^2h-2(a \cdot h)a}{{\lverta\rvert}^4}.$$ On the other hand, for $2 {\lverth\rvert} \le {\lverta\rvert}$, $$\Bigl \lvert \frac{a+h}{{\lverta+h\rvert}^2}-\frac{a}{{\lverta\rvert}^2} - L(a)h \Bigr \rvert
\le C \frac{{\lverth\rvert}^2}{{\lverta\rvert}^3},$$ so that, by , $$\begin{gathered}
\Bigl\lvert \int_{\Omega} {\varepsilon}\frac{x^{\varepsilon}_i+{\varepsilon}y -z}{{\lvertx^{\varepsilon}_i+{\varepsilon}y-z\rvert}^2} \omega^{\varepsilon}_i(z) \, dz-\kappa^{\varepsilon}_i\frac{y}{{\lverty\rvert}^2}\Bigr\rvert \\
\le \int_\Omega {\varepsilon}\frac{{\lvertx^{\varepsilon}_i-z\rvert}^2}{{\lvert{\varepsilon}y\rvert}^3}\omega_i^{\varepsilon}(z)\, dz
\le C {\varepsilon}\frac{(\operatorname{diam}E^{\varepsilon}_i)^2}{{{\lvert{\varepsilon}y\rvert}^3}}\le \frac{\Bar{C}}{{\lverty\rvert}^3},\end{gathered}$$ and the lemma is proved.
\[lemmaLocalAsymptotics\] When ${\varepsilon}$ is small, we have $k^{\varepsilon}=1$. Moreover, $$ \kappa^{\varepsilon}_1=\kappa+\frac{2\pi}{\log \frac{1}{{\varepsilon}}}\Bigl(q(x^{\varepsilon}_1)-\kappa H(x^{\varepsilon}_1, x^{\varepsilon}_1)-\frac{\kappa}{2\pi} \log \frac{1}{\rho_\kappa} \Bigr)+o({{\vert\!\log{\varepsilon}\vert}}^{-1})
$$ and $v^{\varepsilon}_1 \to U_\kappa$ in $\mathrm{W}^{3, r}_{\mathrm{loc}}({\mathbf{R}}^2)$ as ${\varepsilon}\to 0$.
Set $$w^{\varepsilon}_i(y)= v^{\varepsilon}_i(y)-\frac{\kappa^{\varepsilon}_i}{2\pi}\log \frac{1}{{\varepsilon}}+q^{\varepsilon}(x^{\varepsilon}_i)-\kappa^{\varepsilon}_i H(x^{\varepsilon}_i, x^{\varepsilon}_i)
-\sum_{j \ne i} \kappa_j^{\varepsilon}G(x^{\varepsilon}_i, x^{\varepsilon}_j),$$ so that in particular $$-\Delta w^{\varepsilon}_i=f(v^{\varepsilon}_i-q^{\varepsilon}_i).$$ By , and classical elliptic estimates [@GilbargTrudinger2001]\*[Theorem 9.11]{}, the sequence $(w^{\varepsilon}_i)$ is bounded in $\mathrm{W}^{2, s}_{\mathrm{loc}}({\mathbf{R}}^2)$ for every $s \ge 1$. By Rellich’s compactness theorem, it is compact in $\mathrm{W}^{1, t}_{\mathrm{loc}}({\mathbf{R}}^2)$ for every $1 \le t < \infty$, and therefore bounded on compact subsets. On the other hand, by construction, all the $v^{\varepsilon}_i+q^{\varepsilon}_i(x^{\varepsilon}_i)-q^{\varepsilon}_i$ take positive and negative value at a uniformly bounded distance from the origin, so that there exists a bounded sequence $\check{x}_i^{\varepsilon}$ such that $v^{\varepsilon}_i(\check{x}_i^{\varepsilon})=q^{\varepsilon}_i(\check{x}_i^{\varepsilon})-q^{\varepsilon}_i(x_i^{\varepsilon})$. Therefore, $v^{\varepsilon}_i(\check{x}_i^{\varepsilon})$ and $w^{\varepsilon}_i(\check{x}_i^{\varepsilon})$ remain bounded and we obtain that for each $i \in \{1, \dotsc, k^{\varepsilon}\}$ $$q^{\varepsilon}(x^{\varepsilon}_i)-\frac{\kappa^{\varepsilon}_i}{2\pi}\log \frac{1}{{\varepsilon}}-\kappa^{\varepsilon}_i H(x^{\varepsilon}_i, x^{\varepsilon}_i)
-\sum_{j \ne i} \kappa_j^{\varepsilon}G(x^{\varepsilon}_i, x^{\varepsilon}_j)=O(1).$$ This implies that $$\label{eqVorticitiesGreen}
\frac{\kappa^{\varepsilon}_i}{2\pi}\log \frac{1}{{\varepsilon}} + \sum_{\substack{ j \ne i}} \kappa_j^{\varepsilon}\log \frac{1}{{\lvertx^{\varepsilon}_i-x^{\varepsilon}_j\rvert}} = \frac{\kappa}{2\pi}\log \frac{1}{{\varepsilon}} +O(1),$$ and, in view of , that $$k_{\varepsilon}\frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}}\ge \sum_{1 \le i, j \le k_{\varepsilon}} \frac{\kappa^{\varepsilon}_i }{2\pi}\log \frac{1}{{\varepsilon}} +O(1)=k_{\varepsilon}\frac{\kappa}{2\pi}\log \frac{1}{{\varepsilon}}+\sum_{\substack{1 \le i, j \le k^{\varepsilon}\\ j \ne i}} \kappa^{\varepsilon}_j\log \frac{{\lvertx^{\varepsilon}_i-x^{\varepsilon}_j\rvert}}{{\varepsilon}}+O(1).$$ Therefore, $$\sum_{\substack{1 \le i, j \le k^{\varepsilon}\\ j \ne i}} \kappa^{\varepsilon}_j \log \frac{{\lvertx^{\varepsilon}_i-x^{\varepsilon}_j\rvert}}{{\varepsilon}} \le O(1),$$ and since ${\lvertx^{\varepsilon}_i-x^{\varepsilon}_j\rvert}/{\varepsilon}\to \infty $ as ${\varepsilon}\to 0$, we deduce by that $k^{\varepsilon}\le 1$ for ${\varepsilon}$ sufficiently small. By Lemma \[lemmaNonVanisingVortex\], $k^{\varepsilon}=1$. Going back to , we get $$\kappa^{\varepsilon}_1=\kappa+O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr).$$ Since $v_1^{\varepsilon}-q^{\varepsilon}_1$ is compact in $\mathrm{W}^{1, r}_{\mathrm{loc}}({\mathbf{R}}^2)$ and $f \in C^1({\mathbf{R}})$, the sequence $f(v_1^{\varepsilon}-q^{\varepsilon}_1)$ is compact in $\mathrm{W}^{1, r}_{\mathrm{loc}}({\mathbf{R}}^2)$. In view of , $v^{\varepsilon}_1$ is compact in $\mathrm{W}^{3, r}_{\mathrm{loc}}$. Let $v$ be one of its accumulation points. It satisfies $$-\Delta v=f(v)$$ and $$\int_{{\mathbf{R}}^2} f(v)=\kappa.$$ Moreover, letting ${\varepsilon}$ go to zero, by we obtain $$v(y)=\frac{\kappa}{2\pi} \log \frac{\Tilde{\rho}}{{\lverty\rvert}}+O\Bigl(\log\bigl(1+\frac{1}{{\lverty\rvert}} \bigr)\Bigr)$$ for some $\Tilde{\rho} \in {\mathbf{R}}$, and $$\nabla v(y)=\frac{\kappa}{2\pi}\frac{y}{{\lverty\rvert}^2}+O\Bigl(\frac{1}{{\lverty\rvert}^3}\Bigr).$$ By a symmetry result of L.A.Caffarelli and A.Friedman [@CaffarelliFriedman1980 Theorem 1] (see also [@Fraenkel2000 Theorem 4.2]), $v$ is radial, and therefore $$v(y)=\frac{\kappa}{2\pi}\log \frac{\rho_\kappa}{{\lverty\rvert}}$$ when ${\lverty\rvert} \ge \rho_\kappa$. Hence, $v=U_\kappa$. In view of , this yields $$\Bigl\lvert \frac{\kappa}{2\pi}\log \frac{\rho_\kappa}{{\lverty\rvert}}+q^{\varepsilon}(x^{\varepsilon}_1)-\frac{\kappa^{\varepsilon}_1}{2\pi}\log \frac{1}{{\varepsilon}{\lverty\rvert}}-\kappa^{\varepsilon}_1 H(x^{\varepsilon}_1, x^{\varepsilon}_1)
\Bigr\rvert \le \kappa \log \frac{{\lverty\rvert}}{{\lverty\rvert}-R}+o(1).$$ First fixing $y$, this implies that $$\frac{\kappa-\kappa^{\varepsilon}_1}{2\pi} \log \frac{1}{{\varepsilon}}=O(1),$$ and next we deduce that for every $2\Tilde{\rho}<{\lverty\rvert}<R$, $$\Bigl\lvert \frac{\kappa}{2\pi}\log \frac{\rho_\kappa}{{\varepsilon}}+q(x^{\varepsilon}_1)-\frac{\kappa^{\varepsilon}_1}{2\pi}\log \frac{1}{{\varepsilon}}-\kappa^{\varepsilon}_1 H(x^{\varepsilon}_1, x^{\varepsilon}_1)
\Bigr\rvert \le \kappa \log \frac{{\lverty\rvert}}{{\lverty\rvert}-\Tilde{\rho}}+o(1),$$ as ${\varepsilon}\to 0$. We obtain the required asymptotic development of $\kappa^{\varepsilon}_1$ by letting $R\to +\infty$ and choosing sufficiently large ${\lverty\rvert}$.
### Step 4: Global asymptotics
We are now going to prove that $v^{\varepsilon}$ is well approximated by $$\Tilde{v}^{\varepsilon}=U_{\kappa^{\varepsilon}_1}\Bigl(\frac{\cdot-x^{\varepsilon}_1}{{\varepsilon}}\Bigr)+\kappa^{\varepsilon}_1\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}\rho_{\kappa_1^{\varepsilon}}}+H(x^{\varepsilon}_1, \cdot)\Bigr).$$
\[propositionAsymptoticsW21\] We have $$v^{\varepsilon}=\Tilde{v}^{\varepsilon}+o(1)$$ in $\mathrm{W}^{2, 1}_{\mathrm{loc}}(\Omega)$, in $\mathrm{W}^{1, 2}_0(\Omega)$, and in $\mathrm{L}^\infty(\Omega)$.
Choose $r>\Tilde{\rho}$ so that $E^{\varepsilon}_1 \subset B(x^{\varepsilon}_1, {\varepsilon}r)$ when ${\varepsilon}$ is small. By Lemma \[lemmaLocalAsymptotics\], and the invariance of the $\dot{\mathrm{W}}^{2, 1}$ semi-norm by scaling, we have $$\int_{B(x^{\varepsilon}_1, 2{\varepsilon}r)} {\lvertD^2 v^{\varepsilon}-D^2 \Tilde{v}^{\varepsilon}\rvert} \to 0$$ as ${\varepsilon}\to 0$. Define $$\begin{aligned}
\Tilde{\omega}^{\varepsilon}_1(x)&=\frac{1}{{\varepsilon}^2}f(\Tilde{v}^{\varepsilon}-q^{\varepsilon}), \\
w_v^{\varepsilon}(x)&=\int_{\Omega} G(x, y) \omega^{\varepsilon}_v(y)\, dy, \\
w_r^{\varepsilon}(x)&=\int_{\Omega} H(x, y) \bigl(\omega^{\varepsilon}_1(y)-\Tilde{\omega}^{\varepsilon}_1(y)\bigr)\, dy, \\
w_s^{\varepsilon}(x)&=\int_{\Omega} \Gamma(x-y) \bigl(\omega^{\varepsilon}_1(y)-\Tilde{\omega}^{\varepsilon}_1(y)\bigr)\, dy,\end{aligned}$$ where $\Gamma (x)=\frac{1}{2\pi} \log \frac{1}{{\lvertx\rvert}}$, so that $v^{\varepsilon}- \Tilde{v}^ {\varepsilon}=w_v^{\varepsilon}+w_r^{\varepsilon}+w_s^{\varepsilon}$. Since by Lemma \[lemmaVanishingVorticity\], $\omega_v \to 0$ in $\mathrm{L}^s(\Omega)$ for some $s > 1$, we have, by elliptic estimates, $w^{\varepsilon}_v \to 0$ in $\mathrm{W}^{2, s}_{\mathrm{loc}}(\Omega)$. Next, since by $x^{\varepsilon}_1$ stays away from $\partial \Omega$ and $\omega^{\varepsilon}_1-\Tilde{\omega}^{\varepsilon}_1 \to 0$ in $\mathrm{L}^1(\Omega)$ by Lemma \[lemmaLocalAsymptotics\], we have $w_r^{\varepsilon}\to 0$ in $C^\infty_{\mathrm{loc}}(\Omega)$. Finally, we have $$D^2 w_s^{\varepsilon}(x)=\int_{\Omega} D^2\Gamma(x-y) \bigl(\omega^{\varepsilon}_1(y)-\Tilde{\omega}^{\varepsilon}_1(y)\bigr)\, dy.$$ Since $\int_{\Omega} \omega^{\varepsilon}_1=\int_{\Omega} \Tilde{\omega}^{\varepsilon}_1=\kappa^{\varepsilon}_1$, one also has $$D^2 w_s^{\varepsilon}(x)=\int_{B(x^{\varepsilon}_1, {\varepsilon}r)} \bigl(D^2\Gamma(x-y)-D^2\Gamma(x-x^{\varepsilon}_1)\bigr) \bigl(\omega^{\varepsilon}_1(y)-\Tilde{\omega}^{\varepsilon}_1(y)\bigr)\, dy.$$ For every $y \in B(x^{\varepsilon}_1, {\varepsilon}r)$ and $x \in \Omega \setminus B(x^{\varepsilon}_1, {\varepsilon}2r)$ $${\lvertD^2\Gamma(x-y)-D^2\Gamma(x-x^{\varepsilon}_1)\rvert} \le C \frac{{\lverty-x^{\varepsilon}_1\rvert}}{{\lvertx-x^{\varepsilon}_1\rvert}^3},$$ so that $${\lvertD^2 w_s^{\varepsilon}(x)\rvert} \le \frac{C {\varepsilon}}{{\lvertx^{\varepsilon}_1-x\rvert}^3}{\Vert\omega^{\varepsilon}_1-\Tilde{\omega}^{\varepsilon}_1\Vert}_{\mathrm{L}^1}.$$ Integrating the previous inequality we conclude $$\begin{gathered}
\int_{\Omega \setminus B(x^{\varepsilon}_1, {\varepsilon}2r)} {\lvertD^2 w_s^{\varepsilon}(x)\rvert} \le C{\varepsilon}{\Vert\omega^{\varepsilon}_1-\Tilde{\omega}^{\varepsilon}_1\Vert}_{\mathrm{L}^1}\int_{{\mathbf{R}}^2 \setminus B(x^{\varepsilon}_1, {\varepsilon}2r)}\frac{1}{{\lvertx^{\varepsilon}_1-x\rvert}^3}\, dx \\
=C{\Vert\omega^{\varepsilon}_1-\Tilde{\omega}^{\varepsilon}_1\Vert}_{\mathrm{L}^1}\frac{2\pi {\varepsilon}}{{\varepsilon}R}=o(1). \end{gathered}$$
The $\mathrm{W}^{2, 1}_\mathrm{loc}(\Omega)$ convergence implies the $\mathrm{W}^{1, 2}_\mathrm{loc}(\Omega)$ and the $\mathrm{L}^\infty_{\mathrm{loc}}(\Omega)$ convergences. One needs then to prove the convergence in a neighbourhood of the boundary. Consider $U \subset V \subset \Omega$ open bounded sets such that $\partial \Omega \subset \Bar{U}$, $\Bar{U} \subset V$ and $\operatorname{supp}\omega_{\varepsilon}\cap V = \emptyset$. One has $$\left\{
\begin{aligned}
-\Delta (v_{\varepsilon}-\Tilde{v}_{\varepsilon})&=\omega^{\varepsilon}_v && \text{in $U$},\\
v_{\varepsilon}-\Tilde{v}_{\varepsilon}&=0 && \text{on $\partial \Omega$}.
\end{aligned}
\right.$$ Since $v_{\varepsilon}-\Tilde{v}_{\varepsilon}\to 0$ in $\mathrm{W}^{1,2}(V \setminus U)$ and in $\mathrm{L}^\infty (V \setminus U)$ and $\omega^{\varepsilon}_v \to 0$ in $\mathrm{L}^s(\Omega)$ for some $s > 1$, one obtains by classical regularity estimates that $v_{\varepsilon}-\Tilde{v}_{\varepsilon}\to 0$ in $\mathrm{W}^{1,2}(U)$ and in $\mathrm{L}^\infty(U)$.
\[corollaryAsymptotic\] When ${\varepsilon}$ is small enough, $A^{\varepsilon}$ is connected, $x^{\varepsilon}_1=x^{\varepsilon}$, $\kappa^{\varepsilon}_1=\kappa^{\varepsilon}$, $\partial (A^{\varepsilon}_1-x^{\varepsilon}_1)/{\varepsilon}$ tends to $\partial B(0, \rho_\kappa)$ as a $C^2$ manifold. In particular, $-\Delta v^{\varepsilon}=0$ in $\Omega \setminus B(x^{\varepsilon}_1, 2{\varepsilon}\rho_\kappa)$ and $$\omega^{\varepsilon}=\Tilde{\omega}^{\varepsilon}+o(1)$$ in $\mathrm{L}^1(\Omega)$.
Assume that $y\in A^{\varepsilon}\setminus B(x_1^{\varepsilon}, {\varepsilon}\rho_{\kappa^{\varepsilon}})$. We have $$\label{ineqAepsBall}
q(y)+ \frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}} < v^{\varepsilon}(y) \le \frac{\kappa_1^{\varepsilon}}{2\pi} \log \frac{1}{{\lverty-x^{\varepsilon}_1\rvert}}+o(1),$$ uniformly in $y$, so that ${\lverty-x^{\varepsilon}_1\rvert}=O({\varepsilon})$. One obtains then in view of Proposition \[propositionAsymptoticsW21\] that $(A^{\varepsilon}_1-x^{\varepsilon}_1)/{\varepsilon}$ is connected when ${\varepsilon}$ is small and the required convergence of the boundary.
\[corEnergy\] We have $$\begin{split}
\mathcal{E}^{\varepsilon}(v^{\varepsilon})
=\frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}}-\mathcal{W}(x^{\varepsilon})+\mathcal{C}+o(1).
\end{split}$$
First we have in view of Proposition \[propositionAsymptoticsW21\] and Corollary \[corollaryAsymptotic\], $$\begin{split}
\int_{\Omega} {\lvert\nabla v^{\varepsilon}\rvert}^2
&=\int_{\Omega} v^{\varepsilon}\omega^{\varepsilon}\\
&=\int_{\Omega} \Tilde{v}^{\varepsilon}\omega^{\varepsilon}+o(1).
\end{split}$$ Since ${\Vert\Tilde{v}^{\varepsilon}-q^{\varepsilon}\Vert}_{\mathrm{L}^\infty}$ remains bounded as ${\varepsilon}\to 0$, we obtain, by Proposition \[propositionAsymptoticsW21\] $$\int_{\Omega} {\lvert\nabla v^{\varepsilon}\rvert}^2=\frac{1}{{\varepsilon}^2}\int_{\Omega} \Tilde{v}^{\varepsilon}f\bigl(\Tilde{v}^{\varepsilon}-q^{\varepsilon}(x^{\varepsilon})\bigr)+o(1).$$ Similarly, by Proposition \[propositionAsymptoticsW21\], $$\frac{1}{{\varepsilon}^2}\int_{\Omega} F(v^{\varepsilon}-q^{\varepsilon})=\frac{1}{{\varepsilon}^2}\int_{\Omega} F\bigl(\Tilde{v}^{\varepsilon}-q^{\varepsilon}(x^{\varepsilon})\bigr)+o(1).$$ It suffices then to compute $\mathcal{E}^{\varepsilon}(\Tilde{v}^{\varepsilon})$ as in the proof of Lemma \[lemmaEnergyHatu\].
### Conclusion
We are now in position to present the
It is a direct consequence of Lemma \[lemmaLocalAsymptotics\], Proposition \[propositionAsymptoticsW21\], Corollary \[corollaryAsymptotic\] and Corollary \[corEnergy\].
and the
It is a direct consequence of the upper estimate of Corollary \[cor:upper\] and the asymptotic properties obtained in Proposition \[prop:1mai\].
Single vortices in multiply connected domains {#sectionmultiply}
=============================================
In this section we assume that $\Omega \subset {\mathbf{R}}^2$ is a bounded smooth multiply-connected domain; it can be written as $$\Omega = \Omega_0 \setminus \bigcup_{h=1}^m \Omega_h,$$ where $\Omega_0, \dotsc, \Omega_m$ are bounded simply-connected domains with $\Bar{\Omega}_h \subset \Omega$ for every $h \in \{1, \dotsc, m\}$. In place of problem , we consider the problem of finding $u$ and $\lambda^{\varepsilon}_1, \dotsc, \lambda^{\varepsilon}_m$ such that $$\label{problemPepsstar}
\left\{
\begin{aligned}
-\Delta u^{\varepsilon}&=\frac{1}{{\varepsilon}^2} f(u^{\varepsilon}- q^{\varepsilon}) & &\text{in $\Omega$, }\\
u^{\varepsilon}&= 0 & &\text{on $\partial \Omega_0$},\\
u^{\varepsilon}&= \lambda^{\varepsilon}_h & &\text{on $\partial \Omega_h$},\\
\int_{\partial \Omega_h} \frac{\partial u}{\partial n}&=0 & & \text{for $h \in \{1, \dotsc, m\}$}.
\end{aligned}
\right. \tag{\protect{$\mathcal{P}^{\varepsilon}_*$}}$$
The natural space to deal with this problem is the space of functions that are constant on the complement of $\Omega$: $$H^1_*(\Omega)=\Bigl\{u \in H^1(\Omega) {\: :\:}\nabla u=0 \text{ in $\bigcup_{h=1}^m \Omega_h$}\Bigr\}.$$ It is standard to show that solutions of are critical points of the functional $\mathcal{E}^{\varepsilon}$ defined on $H^1_*(\Omega)$ by . We consider least energy solutions obtained by minimization of the functional on the Nehari manifold.
In order to state our result we also need the corresponding (appropriate) Green functions. Following C.C.Lin [@Lin1941; @Lin1943], we define $G_*$ as the solution of $$\left\{
\begin{aligned}
-\Delta G(\cdot, y)&=\delta_y & & \text{in $\Omega$,}\\
G(\cdot, y)&=0 & & \text{on $\partial \Omega_0$},\\
G&=\lambda_h & & \text{on $\partial \Omega_h$},\\
\int_{\partial \Omega_h} \frac{\partial G}{\partial n}&=0 & & \text{for $h \in \{1, \dotsc, m\}$}.\\
\end{aligned}\right.$$ Its regular part $H_*$ is defined by $$H_*(x,y)=G_*(x,y)-\frac{1}{2\pi} \log \frac{1}{{\lvertx-y\rvert}}.$$ P.Koebe [@Koebe1918]\*[§6]{} (see also [@Lin1943]\*[§9]{}), defined $G_*$ in terms of the Green function for the Dirichlet problem $G$ and the unique solutions $Z_k$ of $$\left\{
\begin{aligned}
-\Delta Z_k&=0 & & \text{in $\Omega$,}\\
Z_k&=0 & & \text{on $\partial \Omega_0$},\\
Z_k&=\delta_{kh} & & \text{on $\partial \Omega_h$ with $h \in \{1, \dotsc, m\}$.}\\
\end{aligned}\right.$$ Since the $Z_k$ are linearly independent, the matrix $(\omega_{kh})_{1 \le k, h \le n}$ defined by $$\omega_{kh}=\int_{\Omega} \nabla Z_k \cdot \nabla Z_h.$$ is invertible; let $(\omega^{kh})_{1 \le k, h \le n}$ denote its inverse. We have $$\label{eqGreenRelationship}
G_*(x,y)=G(x,y)+\sum_{k, h=1}^m Z_k(x)\omega^{kh}Z_h(y).$$
The Kirchhoff–Routh function in this context is defined by $$\mathcal{W}_*(x)=\frac{\kappa^2}{2}H_*(x, x)-\kappa q(x),$$ and the various quantities $A^{\varepsilon}, \omega^{\varepsilon}, \kappa^{\varepsilon}, x^{\varepsilon}, \rho^{\varepsilon}$ are still defined by .
Theorem \[thm:K1\] generalizes then to
\[thm:K1m\] As ${\varepsilon}\to 0$, we have $$u^{\varepsilon}=U_{\kappa^{\varepsilon}} \Big(\frac{\cdot-x^{\varepsilon}}{{\varepsilon}}\Big)+\kappa^{\varepsilon}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}\rho^{\varepsilon}}+ H_*(x^{\varepsilon}, \cdot)\Bigr)+o(1),$$ , where $$\kappa^{\varepsilon}=\kappa+\frac{2\pi}{\log \frac{1}{{\varepsilon}}}\Bigl(q(x^{\varepsilon})-\kappa H(x^{\varepsilon}, x^{\varepsilon}) -\frac{\kappa}{2\pi} \log \frac{1}{\rho_\kappa} \Bigr)+o({{\vert\!\log{\varepsilon}\vert}}^{-1}),$$ and $$\mathcal{W}_*(x^{\varepsilon}) \to \sup_{x \in \Omega} \mathcal{W}_*(x).$$ One also has $$B(x^{\varepsilon}, \Bar{r}^{\varepsilon}) \subset A^{\varepsilon}\subset B(x^{\varepsilon}, \mathring{r}^{\varepsilon}),$$ with $\Bar{r}^{\varepsilon}={\varepsilon}\rho_\kappa+o({\varepsilon})$ and $\mathring{r}^{\varepsilon}={\varepsilon}\rho_\kappa +o({\varepsilon})$. Finally, $$\mathcal{E}^{\varepsilon}(u^{\varepsilon})= \frac{\kappa^2}{4\pi}\log \frac{1}{{\varepsilon}}-\mathcal{W}_*(x^{\varepsilon})+\mathcal{C}+o(1).$$
The proof of Theorem \[thm:K1m\] follows almost the same lines as one of Theorem \[thm:K1\], so that we only mention the few adaptations. First, the functions $G$ and $H$ should be replaced by $G_*$ and $H_*$. In view of the regularity of $\Theta_h$ and of this does not bring any trouble in the upper estimate nor the small scale and global asymptotics.
Next, the proof of Theorem \[thm:K1\] relies on the Dirichlet boundary condition to estimate $\operatorname{cap}(A^{\varepsilon}, \Omega)$ in . Here, we define instead $$w^{\varepsilon}=\frac{v^{\varepsilon}-\max_{\partial \Omega} v^{\varepsilon}}{\min_{\partial A_{\varepsilon}} q^{\varepsilon}- \max_{\partial \Omega}v^{\varepsilon}}.$$ For every $h \in \{1, \dotsc, m\}$, let $\Theta_h \in \mathrm{H}^1_*(\Omega)$ be the unique solution of $$\left\{
\begin{aligned}
-\Delta \Theta_h&=0 & & \text{in $\Omega$,}\\
\Theta_h&=0 & & \text{on $\partial \Omega_0$},\\
\Theta_h&=\mu_{kh} & & \text{on $\partial \Omega_h$ and $h \in \{1, \dotsc, m\}$},\\
\int_{\partial \Omega_h} \frac{\partial \Theta_h}{\partial n}&=\delta_{hk} & & \text{for $k \in \{1, \dotsc, m\}$},
\end{aligned}\right.$$ where $\mu_{kh}$ are unknown constants that are part of the problem[^5]. By construction of $\Theta_k$, one has $$v^{\varepsilon}\vert_{\Omega_h}=\int_{\partial \Omega_h} v^{\varepsilon}\frac{\partial \Theta_h}{\partial n}
=\int_{\Omega} \nabla v^{\varepsilon}\cdot \nabla \Theta_h=\int_{\Omega} \omega_{\varepsilon}\Theta_h,$$ and hence, in view of , $${\Vertv^{\varepsilon}\Vert}_{\mathrm{L}^\infty(\partial \Omega)} \le \max_{h \in \{1, \dotsc, m\}} {\Vert\Theta_h\Vert}_{\mathrm{L}^\infty(\Omega)}\bigl(\kappa+O({{\vert\!\log{\varepsilon}\vert}}^{-1})\bigr),$$ Therefore, $$\frac{2\pi}{\operatorname{cap}(A^{\varepsilon}, \Omega)} \ge \frac{2\pi}{\int_\Omega {\lvert\nabla w^{\varepsilon}\rvert}^2}
\ge \log \frac{1}{{\varepsilon}}+O(1),$$ and one can continue as in the proof of Lemma \[lemmaAreaDiameter\].
Single vortices in unbounded domains {#sectUnbounded}
====================================
In this section, we assume that $\Omega \subset {\mathbf{R}}^2$ is an unbounded simply-connected domain whose boundary is bounded in one direction; to fix the ideas, $$]a_0, +\infty[\times {\mathbf{R}}\subset \Omega \subset ]a_1, +\infty[ \times {\mathbf{R}}.$$ Our goal is to carry out an analysis similar to that of the previous section.
We assume that $q \in \mathrm{W}^{1, 1}_\textrm{loc}(\Omega)$, $$\sup_{x \in \Omega} \int_{B(x, 1)} {\lvert\nabla q\rvert}^r < \infty$$ for some $r > 2$, and that $$q(x) \ge W(x_1-a_0)+d,$$ for some $W > 0$ and $d > 0$, where $x=(x_1, x_2)$. Since $\partial \Omega$ is bounded in the $x_1$ direction, this is equivalent with requiring that $$q(x) \ge W \operatorname{dist}(x, \partial \Omega)+d'.$$ The natural space for solutions is $$\mathrm{D}^{1, 2}_0(\Omega)= \{ u \in \mathrm{W}^{1, 1}_{\mathrm{loc}}(\Omega) {\: :\:}\int_{\Omega} {\lvert\nabla u\rvert}^2 < \infty\}.$$
The Nehari manifold $\mathcal{N}^{\varepsilon}$ and the infimum value $c^{\varepsilon}$ are defined as in Proposition \[prop:2.1\]. The existence of a minimizer $u^{\varepsilon}\in \mathcal{N}^{\varepsilon}$ as in Proposition \[prop:2.1\] such that $\mathcal{E}^{\varepsilon}(u^{\varepsilon})=c^{\varepsilon}$ is no longer direct nor true because of compactness issues.
In a first step, we derive upper bounds on $c^{\varepsilon}$. Next, we perform the a priori asymptotic analysis of solutions of ${(\mathcal{P}^{\varepsilon})}$ satisfying similar upper bounds. Finally, we prove existence results in appropriate cases of $\Omega$ and $q$.
Upper bound on the energy
-------------------------
\[propUnboundedUpper\] We have $$c^{\varepsilon}\leq \frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}} - \sup_{x\in \Omega} \mathcal{W}(x) +\mathcal{C} + o(1).$$
The proof goes as the proof of Corollary \[cor:upper\]. The main difference is that $q$ and $H(\Hat{x}, \cdot)$ are not bounded as in the proof of Lemma \[lemmaHatuNehari\]. However, since $\lim_{x \to \infty} \frac{1}{2\pi} \log \frac{1}{{\lvertx-\Hat{x}\rvert}} +H(x, \Hat{x})=0$, one still has, for every $x \in \Omega$, $$H(\Hat{x}, x) \le \frac{q(x)}{\kappa}+C,$$ whence, starting from , one obtains $$\frac{\dfrac{1}{2\pi} \log \dfrac{1}{{\varepsilon}}+\dfrac{1}{2\pi} \log \dfrac{{\varepsilon}}{{\lvertx-\Hat{x}\rvert}}+ q(x)+C}{\dfrac{\kappa}{2\pi} \log \dfrac{1}{{\varepsilon}}+\frac{q(x)}{\kappa}} \ge
\dfrac{\log \dfrac{1}{{\varepsilon}}+H(\Hat{x}, \Hat{x})}{\dfrac{\kappa}{2\pi} \log \dfrac{1}{{\varepsilon}}+q(\Hat{x})+\sigma}.$$ Since $q \ge 0$, it follows that $$\frac{1}{\kappa}+\frac{\dfrac{1}{2\pi} \log \dfrac{{\varepsilon}}{{\lvertx-\Hat{x}\rvert}}}{\kappa \log \dfrac{1}{{\varepsilon}}}\ge \frac{1}{\kappa}+O\bigl({{\vert\!\log{\varepsilon}\vert}}^{-1}\bigr),$$ and it suffices to continue as in Lemmas \[lemmaHatuNehari\] and \[lemmaEnergyHatu\].
Functional inequalities on the half-plane
-----------------------------------------
In order to perform the asymptotic analysis of the solutions and to study their existence, we first provide some useful functional type inequalities and convergence results on the half-plane ${\mathbf{R}}^2_+$ that will be used in the next section.
\[ineqUnboundedIneqW\] We have for $u\in \mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$, $${\mathcal{L}^{2}}\bigl(\{ x \in {\mathbf{R}}^2_+ {\: :\:}u(x) \ge Wx_1\}\bigr)\le C \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2,$$ and, for every $p>0$, $$\int_{{\mathbf{R}}^2_+} \bigl(u(x)-Wx_1\bigr)^p_+\, dx \le \frac{C}{W^2} \Bigl( \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2 \Bigr)^{1+\frac{p}{2}}.$$
A similar statement is proved by Yang Jianfu [@Yang1991 Lemma 4] with a different proof relying on an isometry between $\mathcal{D}^{1,2}_0({\mathbf{R}}^2_+)$ and the space of cylindrically symmetric elements of $\mathcal{D}^{1,2}_0({\mathbf{R}}^4)$ [@Yang1991 Lemma 1].
Define $A_u=\{ x \in {\mathbf{R}}^2_+ {\: :\:}u(x) \ge Wx_1\}$. First we have, by the Chebyshev and Hardy inequalities $${\mathcal{L}^{2}}(A_u ) \le \frac{1}{W^2} \int_{{\mathbf{R}}^2_+} \frac{{\lvertu(x)\rvert}^2}{{\lvertx_1\rvert}^2}\, dx \le \frac{4}{W^2} \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2.$$ By Sobolev’s inequality, it follows $$\begin{split}
\int_{{\mathbf{R}}^2_+} (u(x)-Wx_1)_+^p\, dx &=\int_{A_u} (u-Wx_1)^p\, dx \\
&\le C {\Vert\nabla(u-Wx_1)\Vert}^p_{2} {\mathcal{L}^{2}}(A_u) \\
&\le \frac{C'}{W^2} {\Vert\nabla u\Vert}_{2}^2({\Vert\nabla u\Vert}_2+W {\mathcal{L}^{2}}(A_u)^\frac{1}{2})^p \\
&\le \frac{C''}{W^2} \Bigl( \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2\Bigr)^{1+\frac{p}{2}}. \qedhere
\end{split}$$
As a consequence
\[lemmaUnboundedInequalityq\] We have for $u\in \mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$, $${\mathcal{L}^{2}}\bigl(\{ x \in {\mathbf{R}}^2_+ {\: :\:}u(x) \ge q(x)\}\bigr)\le C \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2,$$ and for every $p>0$ $$\int_{{\mathbf{R}}^2_+} (u-q)^p_+ \le C \Bigl( \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2 \Bigr)^{1+\frac{p}{2}}.$$
We also have a compactness theorem
\[lemmaUnboundedCompactness\] For every $p < \infty$ and $L>0$, the map $\Phi : \mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+) \to \mathrm{L}^p({\mathbf{R}}_+\times ]-L, L[) : u \mapsto (u-Wx_1)_+$ is completely continuous.
By Rellich’s Theorem, $u \mapsto \Phi(u)\chi_{]0, \lambda[\times ]-L, L[}$ is completely continuous for every $\lambda > 0$. On the other hand, $$\int_{]\lambda, +\infty[\times ]-L, L[ } \hspace{-2em}(u(x)-Wx_1)_+^p\, dx \le \frac{C}{\lambda}\int_{]\lambda, +\infty[\times ]-L, L[} \hspace{-2em}(u(x)-\tfrac{W}{2}x_1)_+^{p+1}\, dx \le \frac{C}{\lambda} {\Vert\nabla u\Vert}_2^{p+3},$$ therefore, on every bounded subset of $\mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$, $\Phi$ is a uniform limit of completely continuous maps. The conclusion follows.
\[campeones\] Let $(u_n) \subset \mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$. If $(u_n)$ is bounded in $\mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$ and $$\sup_{y\in {\mathbf{R}}} \int_{{\mathbf{R}}_+ \times ]y-1, y+1[} (u_n-Wx_1)_+^p \to 0,$$ then $$\int_{{\mathbf{R}}^+_2} (u_n-Wx_1)_+^s \to 0,$$ for every $s>0$.
This kind of result was first obtained by P.-L.Lions [@Lions1984 Lemma I.1]. The idea of our proof comes from V.Coti Zelati and P.Rabinowitz [@CotiZelatiRabinowitz1992].
By the Gagliardo–Nirenberg inequality [@Nirenberg1959 p.125], $$\begin{split}
\int_{{\mathbf{R}}_+\times ]y-1, y+1[}(u_n-Wx_1)_+^{p+2}
\le C\int_{{\mathbf{R}}_+\times ]y-1, y+1[}\hspace{-4em}& (u_n-Wx_1)_+^p \\
&\times \int_{{\mathbf{R}}_+ \times ]y-1, y+1[}\hspace{-4em} ({\lvert\nabla (u_n-Wx_1)_+\rvert}^2+{\lvert(u_n-Wx_1)_+\rvert}^2).
\end{split}$$ Integrating with respect to $y\in {\mathbf{R}}$, one obtains $$\begin{split}
\int_{{\mathbf{R}}^2_+}(u_n-Wx_1)_+^{p+2}
\le C\biggl(\sup_{y \in {\mathbf{R}}} \int_{{\mathbf{R}}_+\times ]y-1, y+1[}\hspace{-4em}& (u_n-Wx_1)_+^p\biggr)\\
&\times\int_{{\mathbf{R}}_+^2} \bigl({\lvert\nabla (u_n-Wx_1)_+\rvert}^2+{\lvert(u_n-Wx_1)_+\rvert}^2\bigr).
\end{split}$$ Since by Lemma \[ineqUnboundedIneqW\] $$\int_{{\mathbf{R}}_+^2} {\lvert\nabla (u_n-Wx_1)_+\rvert}^2+{\lvert(u_n-Wx_1)_+\rvert}^2
\le C \bigl({\Vert\nabla u_n\Vert}_{2}^2+{\Vert\nabla u_n\Vert}_{2}^4\bigr),$$ $(u_n-Wx_1) \to 0$ in $\mathrm{L}^{p+2}({\mathbf{R}}^2_+)$. By Lemma \[ineqUnboundedIneqW\], the general case $s \ne p+2$ follows by interpolation.
Asymptotic behavior of solutions {#sectionUnboundedAsymptotics}
--------------------------------
In this section, we assume that $(v^{\varepsilon})$ is a sequence of solutions to ${(\mathcal{P}^{\varepsilon})}$ satisfying . We shall prove
\[prop:1maiUnbounded\] Proposition \[prop:1mai\] holds under the assumptions on $\Omega$ and $q$ of this section.
### Step 1: First quantitative properties of the solutions
We first have the counterpart of Proposition \[propositionEstimatesueps\]
\[propositionUnboundedEstimatesueps\] The estimates , , , and hold for some constant $C$ independent of ${\varepsilon}$.
The proof of Proposition \[propositionEstimatesueps\] provides the estimates , , and without any modification. The inequality needs a little more work, since its proof in Proposition \[propositionEstimatesueps\] relies on the Poincaré inequality. In the present setting, we replace it by the Chebyshev inequality and Lemma \[lemmaUnboundedInequalityq\] $$\begin{split}
{\mathcal{L}^{2}}(\{ x \in \Omega {\: :\:}v^{\varepsilon}(x) \ge q(x)+\tfrac{\kappa}{2\pi} \log \tfrac{1}{{\varepsilon}}\})
&\le \frac{1}{(\frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}})^4} \int_{\Omega} (v^{\varepsilon}-q)_+^4 \\
&\le \frac{C}{{{\vert\!\log{\varepsilon}\vert}}^4} {{\vert\!\log{\varepsilon}\vert}}^3=C {{\vert\!\log{\varepsilon}\vert}}^{-1}. \qedhere
\end{split}$$
### Step 2: Structure of the vorticity set
As previously, we consider the connected components of $(A^{\varepsilon}_i)_{i \in I_{\varepsilon}}$ of $A_{\varepsilon}$.
\[lemmaUnboundedAreaDiameter\] If ${\varepsilon}> 0$ is sufficiently small, we have for every $i \in I^{\varepsilon}$, $$\label{ineqUnboundedVorticityDiameter}
\operatorname{diam}(A^{\varepsilon}_i) \le C {\varepsilon}\frac{\operatorname{dist}(A^{\varepsilon}_i, \partial \Omega)}{e^{2W \operatorname{dist}(A^{\varepsilon}, \partial\Omega)}}.$$ Moreover, if for every $x \in \Omega$, one defines $$A^{\varepsilon}_x=\bigcup \Bigl\{ A^{\varepsilon}_i {\: :\:}B(x, \tfrac{1}{2}\operatorname{dist}(x, \partial \Omega) +1) \cap A^{\varepsilon}_i \ne \emptyset \Bigr\},$$ then $${\mathcal{L}^{2}}(A^{\varepsilon}_x) \le C {\varepsilon}^2e^{-\mu \operatorname{dist}(x, \partial \Omega)}.$$
Let $$w=\frac{v^{\varepsilon}}{\min_{\partial A^{\varepsilon}_i}q^{\varepsilon}}.$$ Proceeding as in , we obtain, using once more Proposition \[propositionCapacityArea\] $$\frac{2\pi (\tfrac{\kappa}{2\pi}\log \tfrac{1}{{\varepsilon}}+W\operatorname{dist}(A^{\varepsilon}_i, \Omega)+d')^2}{\frac{\kappa^2}{2\pi} \log \frac{1}{{\varepsilon}}}\le \log \biggl(C\bigl(1+\operatorname{dist}(A^{\varepsilon}_i, \partial \Omega)\bigr)\Bigl(1+\frac{\operatorname{dist}(A^{\varepsilon}, \partial \Omega)}{\operatorname{diam}A^{\varepsilon}_i}\Bigr)\biggr).$$ Therefore, $$\frac{1}{{\varepsilon}}\le C\frac{1+\operatorname{dist}(A^{\varepsilon}, \partial \Omega)}{e^{2W(\operatorname{dist}(A_{\varepsilon}, \partial \Omega)-1)}} \Bigl(1+\frac{\operatorname{dist}(A^{\varepsilon}_i, \partial \Omega)}{\operatorname{diam}A^{\varepsilon}_i}\Bigr),$$ from which follows.
Consider now $A^{\varepsilon}_x$. By , $A^{\varepsilon}_x \subset B(x, \frac{2}{3} \operatorname{dist}(x, \partial \Omega)+1)$ when ${\varepsilon}$ is small enough, so that $$\frac{2\pi}{\operatorname{cap}_\Omega (A^{\varepsilon}_x)} \ge \frac{\bigl(\tfrac{\kappa}{2\pi}\log \tfrac{1}{{\varepsilon}} + \frac{W}{3}\operatorname{dist}(x, \partial \Omega)+d'\bigr)^2}{\frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}}}.$$ By Proposition \[propositionCapacityLocalArea\], we obtain $${\mathcal{L}^{2}}(A^{\varepsilon}_x) \le C\bigl(\operatorname{dist}(x, \partial \Omega)+1\bigr)^2 {\varepsilon}^2 e^{-\frac{4W}{3}\operatorname{dist}(x, \partial \Omega)} \le C {\varepsilon}^2 e^{-\mu\operatorname{dist}(x, \partial \Omega)}. \qedhere$$
A slightly more careful proof shows that one can take any $\mu < W/2$, provided $C$ is large enough.
The next Lemma, counterpart of Lemma \[lemmaVortexSplit\], insures that essential vortices are not too far from the boundary.
\[lemmaUnboundedVortexSplit\] There exists constants $\gamma, C, c>0$, such that, when ${\varepsilon}$ is small enough: If holds, we have , , , and $$\operatorname{dist}(A^{\varepsilon}_i, \partial \Omega)\le C,$$ while if does not hold, then holds.
The proof follows essentially the one of Lemma \[lemmaVortexSplit\]. The inequality follows immediately from and .
As in the case of a bounded domain, the vorticity set can be split into a vanishing vorticity set $V^{\varepsilon}$ and an essential one $E^{\varepsilon}$, defined by and . Since the gradient of $q$ is only locally integrable, Lemma \[lemmaVanishingVorticity\] only gives local information.
\[lemmaUnboundedVanishing\] For every $s \ge 1$, we have $$\sup_{x \in \Omega} {\Vert\omega^{\varepsilon}_v\Vert}_{\mathrm{L}^s(B(x, 1))} = o({\varepsilon}^{p(1-\frac{2}{r})-2(1-\frac{1}{s})}).$$ In particular, if $\frac{1}{s} \ge 1-p(\frac{1}{2}-\frac{1}{r})$, then $\omega^{\varepsilon}_v \to 0$ in $\mathrm{L}^s_{\mathrm{loc}}(\Omega)$.
### Step 3: Small scale asymptotics
For the small scale asymptotics, one first note that Lemma \[lemmaSmallScaleLocalEstimates\] still holds. Indeed, the only step that relied on the boundedness of $\Omega$ was . For every $\rho>0$, regularity estimates still yields for $x \in B(x^{\varepsilon}_i, \frac{1}{2})$ $$\Bigl\lvert\int_{B(x^{\varepsilon}_i, \rho)} G(x, y)\omega^{\varepsilon}_v(y)\, dy \Bigr\rvert\le C {\Vert\omega^{\varepsilon}_v\Vert}_{\mathrm{L}^s(B(x^{\varepsilon}_i, 2\rho))},$$ and the conclusion follows from Lemma \[lemmaUnboundedVanishing\]. On the other hand, since $\Omega$ is contained in a half-plane, by comparing its Green function by the Green function of a half-plane, we have $$G(x, y) \le \frac{1}{2\pi} \log \Bigl(1+\frac{C\bigl(1+ \operatorname{dist}(x, \partial \Omega)\bigr)}{{\lvertx-y\rvert}}\Bigr).$$ Since $\operatorname{dist}(x^{\varepsilon}_i, \partial \Omega)$ is bounded, we have, for every $x \in B(x^{\varepsilon}_i, 1)$, $$\int_{\Omega \setminus B(x^{\varepsilon}_i, \rho)} G(x, y)\omega^{\varepsilon}_v(y)\, dy \le \frac{\kappa^{\varepsilon}}{2\pi} \log \Bigl(1+\frac{C}{\rho}\Bigr) \to 0,$$ as $\rho \to \infty$, uniformly in ${\varepsilon}> 0$.
Lemma \[lemmaSmallScaleLocalEstimates\] being established, the proof of Lemma \[lemmaLocalAsymptotics\] also adapts straightforwardly.
### Step 4: Global asymptotics
For Proposition \[propositionAsymptoticsW21\], one obtains a little more than the $\mathrm{W}^{2, 1}_{\mathrm{loc}}(\Omega)$ convergence. Setting $\Omega_\delta=\{ x \in \Omega {\: :\:}\operatorname{dist}(x, \partial \Omega)> \delta\}$, one has
We have $$v^{\varepsilon}=\Tilde{v}^{\varepsilon}+o(1)$$ in $\mathrm{W}^{2, 1}_{\mathrm{loc}}(\Omega_\delta)$ for every $\delta > 0$, in $\mathrm{W}^{1, 2}_0(\Omega)$, and in $\mathrm{L}^\infty(\Omega)$.
One defines $\Tilde{\omega}^{\varepsilon}_1$ and $w^{\varepsilon}_v$, and $w^{\varepsilon}_s$ as in the proof of Proposition \[propositionAsymptoticsW21\]. One defines $$\begin{aligned}
w^{\varepsilon}_r(x)&=\int_{\Omega} \Bigl(H(x,y)-\frac{1}{4\pi}\log ({\lvertx-y\rvert}^2+4x_1y_1)\Bigr) \bigl(\omega^{\varepsilon}_1(y)-\Tilde{\omega}^{\varepsilon}_1(y)\bigr)\, dy,\\
w^{\varepsilon}_h(x)&=\int_{\Omega} \frac{1}{4\pi}\log ({\lvertx-y\rvert}^2+4x_1y_1) \bigl(\omega^{\varepsilon}_1(y)-\Tilde{\omega}^{\varepsilon}_1(y)\bigr)\, dy.\end{aligned}$$ Recalling that $0 < c \le \operatorname{dist}(x^{\varepsilon}_1, \partial \Omega) \le C$, one treats the terms $w^{\varepsilon}_v$, $w^{\varepsilon}_s$ and $w^{\varepsilon}_v$ as in the proof of Proposition \[propositionAsymptoticsW21\]; the term $w^{\varepsilon}_s$ is treated similarly to the term $w^{\varepsilon}_s$. The proof of the convergences up to the boundary follows then as in the proof of Proposition \[propositionAsymptoticsW21\].
For Corollary \[corollaryAsymptotic\], we have, instead of , $$q(y)+ \frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}} < v^{\varepsilon}(y^{\varepsilon}) \le \frac{\kappa_1^{\varepsilon}}{2\pi} \log \Bigl(1+\frac{C\operatorname{dist}(x^{\varepsilon}_1, \partial \Omega)}{{\lverty-x^{\varepsilon}_1\rvert}}\Bigr)+O(1).$$ The remaining part of the proof carries over identically since $\operatorname{dist}(x^{\varepsilon}_1, \partial \Omega)$ remains bounded as ${\varepsilon}\to 0$. Corollary \[corEnergy\] also follows without any modification.
Existence of solutions
----------------------
In this section we present sufficient conditions for the existence of a minimizer for $c^{\varepsilon}$.
Assume that $\Omega \subset ]a_0, +\infty[ \times {\mathbf{R}}$ is a Lipschitz domain, and that $$\label{condPerturbation}
\lim_{t \to +\infty} \inf \{ x_1 \in {\mathbf{R}}{\: :\:}\exists x_2 \in {\mathbf{R}}, (x_1, x_2) \in \Omega \text{ and } {\lvertx_2\rvert} \ge t\}=0.$$ Assume also that there exist $\Hat{W}, \Hat{d}>0$ such that and $$\lim_{t \to +\infty} \inf_{{\lvertx_2\rvert} > t} \frac{q(x)-\Hat{W}x_1-\Hat{d}}{1+{\lvertx_1\rvert}} \ge 0.$$ We define $$\Hat{\mathcal{E}}^{\varepsilon}(u)= \frac{1}{2} \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2-\frac{1}{{\varepsilon}^2}\int_{{\mathbf{R}}^2_+} F(u-\Hat{W}x_1-\Hat{d})$$ and the minimax level $$\Hat{c}^{\varepsilon}=\inf_{u \in \mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)}\max_{t>0} \Hat{\mathcal{E}}^{\varepsilon}(tu).$$
We first recall and investigate about the case where $q$ is affine and $\Omega$ is the half-plane. In this case, by definition, $c^{\varepsilon}=\Hat{c}^{\varepsilon}$.
\[thmYang\] If $\Omega={\mathbf{R}}^2_+$ and $q(x)=Wx_1+d$, then problem admits a solution $u \in \mathrm{D}^{1, 2}_0(\Omega)$.
The proof in [@Yang1991] allows to state that
\[prop:al\] The critical level $c^{\varepsilon}=\Hat{c}^{\varepsilon}$ depends continuously on $W$ and $d$.
We can assume without loss of generality that ${\varepsilon}=1$ and skip any reference to it. Given converging sequences $W_n \to W$ and $d_n \to d$, we set $$\mathcal{E}_n(u)=\frac{1}{2} \int_{{\mathbf{R}}^2_+} {\lvert\nabla u\rvert}^2-\int_{{\mathbf{R}}^2_+} F(u-W_nx_1-d_n)$$ By Theorem \[thmYang\], $\mathcal{E}$ and $\mathcal{E}_n$ possess (some) ground-states $u$ and $u_n$, for which we set $c_n=\mathcal{E}(u_n)$. There exist $\tau_n \to 1$ such that ${\langle d\mathcal{E}_n(\tau_n u) , \tau_n u \rangle}=0$. Therefore, $$c_n \le \mathcal{E}_n(\tau_n u) \to \mathcal{E} (u)=c.$$ This implies that $c$ is upper semi-continuous. In particular, since $$\Bigl(\frac{1}{2}-\frac{1}{p+1}\Bigr){\Vert\nabla u_n\Vert}^2\le \mathcal{E}_n(u_n)$$ the sequence $(u_n)$ is bounded in $\mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$. Choosing $\check{W}=\inf W_n>0$, we obtain by Proposition \[ineqUnboundedIneqW\] $$\begin{gathered}
\Bigl(\int_{{\mathbf{R}}^2_+} (u_n-\tfrac{1}{2}\check{W}x_1)^{p+1}_+\Bigr)^\frac{2}{p+3}
\le \int_{{\mathbf{R}}^2_+} {\lvert\nabla u_n\rvert}^2
\le \int_{{\mathbf{R}}^2_+} u_n f(u_n-W_nx_1-d_n) \\
\le \int_{{\mathbf{R}}^2_+} u_n f(u_n-\check{W}x_1)
\le C \int_{{\mathbf{R}}^2_+} (u_n-\tfrac{1}{2}\check{W}x_1)^{p+1}_+,\end{gathered}$$ so that $(u_n-\frac{1}{2} \check{W} x_1)_+ \not \to 0$ in $L^{p+1}({\mathbf{R}}^2_+)$. By Lemma \[campeones\], up to translation in the $x_2$ direction, we have $(u_n-\frac{1}{2} \check{W} x_1)_+ \not \to 0$ in $L^{p+1}({\mathbf{R}}_+\times ]-1, 1[)$. Hence, there exists $0\neq v\in \mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$ such that $u_n {\rightharpoonup}v$ in $\mathrm{D}^{1, 2}_0({\mathbf{R}}^2_+)$ and $u_n \to v$ almost everywhere and in $L^r_{\mathrm{loc}}({\mathbf{R}}^2_+)$ for $r \ge 1$. In particular, $d\mathcal{E}(v)=0$ and by Fatou’s Lemma, we have $$\begin{split}
c &\le \mathcal{E}(v)=\int_{{\mathbf{R}}^2_+} (W x_1+d)^p (v-Wx_1-d)+(\tfrac{1}{2}-\tfrac{1}{p+1})(v-Wx_1-d)_+^{p+1}\\
&\le \liminf_{n \to \infty}\int_{{\mathbf{R}}^2_+} (W_n x_1+d)^p (u_n-Wx_1-d)+(\tfrac{1}{2}-\tfrac{1}{p+1})(u_n-Wx_1-d)_+^{p+1}\\
&= \liminf_{n \to \infty} \mathcal{E}_n(u_n)=\liminf_{n \to \infty} c_n. \qedhere
\end{split}$$
\[propositionPS\] If $$c^{\varepsilon}< \Hat{c}^{\varepsilon}$$ then there exists $u_{\varepsilon}\in \mathrm{D}^{1, 2}_0(\Omega)$ such that $d\mathcal{E}^{\varepsilon}(u_{\varepsilon})=0$ and $\mathcal{E}^{\varepsilon}(u_{\varepsilon})=c^{\varepsilon}$.
We use the same strategy as P.Rabinowitz [@Rabinowitz1992] for the nonlinear Schrödinger equation on ${\mathbf{R}}^N$.
The minimization problem can be reformulated as a mountain-pass problem (see, e.g. [@Willem1996 Chapter 4]). By Ekeland’s variational principle, there exists a sequence $(u_n) \subset \mathrm{D}^{1, 2}_0(\Omega)$ such that $d\mathcal{E}^{\varepsilon}(u_n) \to 0$ and $\mathcal{E}^{\varepsilon}(u_n) \to c^{\varepsilon}$, see [@MawhinWillem Theorem 4.3] or [@Willem1996 Theorem 1.15]. We have $$\Bigl(\frac{1}{2}-\frac{1}{p+1}\Bigr){\Vert\nabla u\Vert}_{\mathrm{L}^2}^2\le \mathcal{E}^{\varepsilon}(u_n)-{\langle \mathcal{E}^{\varepsilon}(u_n) , u_n \rangle} \to c^{\varepsilon},$$ so that $(u_n)$ is bounded in $\mathrm{D}^{1, 2}_0(\Omega)$. There exists $u \in \mathrm{D}^{1, 2}_0(\Omega)$ such that, up to a subsequence, $u_n {\rightharpoonup}u$ $\mathrm{D}^{1, 2}_0(\Omega)$. By Rellich’s Theorem, for every $\varphi \in \mathrm{D}^{1, 2}_0(\Omega)$, ${\langle d\mathcal{E}^{\varepsilon}(u_n) , \varphi \rangle} \to {\langle d\mathcal{E}^{\varepsilon}(u) , \varphi \rangle}$, so that $d\mathcal{E}^{\varepsilon}(u)=0$. If $u \ne 0$, then $u \in \mathcal{N}^{\varepsilon}$ and by Fatou’s Lemma $$\begin{split}
\mathcal{E}^{\varepsilon}(u)&=\frac{1}{{\varepsilon}^2}\int_{\Omega} \frac{f(u-q^{\varepsilon})u}{2}-F(u-q^{\varepsilon})\\
&=\frac{1}{2{\varepsilon}^2}\int_{\Omega} q^{\varepsilon}(u-q^{\varepsilon})_+^{p}+(1-\tfrac{2}{p+1}) (u-q^{\varepsilon})_+^{p+1} \\
&\le \liminf_{n \to \infty} \frac{1}{2{\varepsilon}^2}\int_{\Omega} q^{\varepsilon}(u_n-q^{\varepsilon})_+^{p}+(1-\tfrac{2}{p+1})(u_n-q^{\varepsilon})_+^{p+1}\\
&= \liminf_{n \to \infty} \mathcal{E}^{\varepsilon}(u_n)-\tfrac{1}{2}{\langle d\mathcal{E}^{\varepsilon}(u_n) , u_n \rangle}=c^{\varepsilon},
\end{split}$$ so that $u$ fits the claim.
Otherwise, for any $\delta < \min( \Hat{W}, \Hat{d})$, let $R > 0$ be such that $$-\delta \le \inf \bigl\{ s \in {\mathbf{R}}{\: :\:}\exists r \in {\mathbf{R}}, (s, r) \in \Omega \text{ and } {\lverts\rvert}\ge R\bigr\},$$ and, $$\label{ineqqdelta}
q(x)\ge \Hat{q}_\delta(x):=(\Hat{W}-\delta)x_1+\Hat{d}-\delta \qquad \text{if ${\lvertx_2\rvert} \ge R$}.$$ We have, for $\Omega_R = \{ x \in \Omega {\: :\:}|x_2|\geq R\}$, and in view of Lemma \[lemmaUnboundedCompactness\], $$\label{ineqOmegaR}
\begin{split}
c^{\varepsilon}&=\lim_{n \to \infty}\mathcal{E}^{\varepsilon}(u_n)-{\langle d\mathcal{E}^{\varepsilon}(u_n) , u_n \rangle}\\
&\le \liminf_{n \to \infty} \frac{1}{{\varepsilon}^2}\int_{\Omega} u_n (u_n-q)_+^p
= \liminf_{n \to \infty} \frac{1}{{\varepsilon}^2}\int_{\Omega \setminus \Omega_R} u_n (u_n-q)_+^p \\
&\le C \liminf_{n \to \infty} \frac{1}{{\varepsilon}^2}\int_{\Omega_R} \Bigl(u_n-\frac{q}{1+\delta}\Bigr)_+^{p+1}
\le C \liminf_{n \to \infty} \frac{1}{{\varepsilon}^2}\int_{\Omega_R} (u_n-\Hat{q}_\delta)_+^{p+1}.
\end{split}$$ Let $\psi \in C^\infty({\mathbf{R}})$ such that $\operatorname{supp}\psi \subset [-2\delta, -\delta]$, $\psi (t)=0$ for $t \le -2 \delta$ and $\psi(t)=1$ for $t \ge -\delta$. We set $\varphi(x_1, x_2)=\psi(x_1)$. Note that $\operatorname{supp}\nabla \varphi \cap \Bar{\Omega}$ is compact, so that by Rellich’s Theorem, $$\int_{\Omega} {\lvert\nabla \varphi\rvert}^2{\lvertu_n\rvert}^2 \to 0,$$ and therefore, defining $v_n=\varphi u_n$, $$\int_{\Omega} {\lvert\nabla v_n\rvert}^2=\int_{\Omega} {\lvert\nabla u_n\rvert}^2+o(1).$$ For every $\tau > 0$ $$\begin{gathered}
\max_{\theta > 0} \mathcal{E}(\theta u_n)
\ge \mathcal{E}(\tau u_n)
=\Hat{\mathcal{E}}_{\delta}(\tau v_n)+\frac{\tau^2}{2}\int_{\Omega} {\lvert\nabla u_n\rvert}^2-{\lvert\nabla v_n\rvert}^2 \\
+\frac{1}{{\varepsilon}^2}\int_{\Omega} F(\tau v_n-\Hat{q}_\delta)-F(\tau u_n-q). \end{gathered}$$ Choose now $\tau_n$ such that $\Hat{\mathcal{E}}_\delta(\tau_n v_n)=\sup_{\tau > 0}\Hat{\mathcal{E}}_\delta(\tau v_n)$. If $\tau_n \ge 1$, we have, $$\begin{gathered}
\tau_n^2 \int_{\Omega} {\lvert\nabla v_n\rvert}^2 = \frac{1}{{\varepsilon}^2}\int_{\Omega} \tau_n v_n f(\tau_n v_n-\Hat{q}_\delta) \ge \tau_n^{p+1}\frac{1}{{\varepsilon}^2}\int_{\Omega} (v_n-\Hat{q}_\delta)_+^{p+1} \\
\ge \tau_n^{p+1}\frac{1}{{\varepsilon}^2}\int_{\Omega_R} (v_n-\Hat{q}_\delta)^{p+1}_+
= \tau_n^{p+1} \frac{1}{{\varepsilon}^2}\int_{\Omega_R} (u_n-\Hat{q}_\delta)^{p+1}_+,\end{gathered}$$ so that by we obtain $$\tau_n \le \max\Biggl(1, \biggl( \frac{\int_{\Omega} {\lvert\nabla v_n\rvert}^2}{\int_{\Omega_R} (u_n-\Hat{q}_\delta)^{p+1}_+} \biggr)^\frac{1}{p-1}\Biggr),$$ and the quantity on the right-hand side is bounded in view of . This implies that $\tau_n v_n {\rightharpoonup}0$ and $\tau_n u_n {\rightharpoonup}0$ in $D^{1, 2}(\Omega)$, and by Lemma \[lemmaUnboundedCompactness\], that $$\int_{\Omega \setminus \Omega_R} F(\tau_n v_n-\Hat{q}_\delta)-F(\tau_n u_n-q) \to 0, \qquad\text{as }n\to +\infty.$$ On the other hand, by , $\Hat{q}_\delta \le q$ in $\Omega \setminus \Omega_R$, and $$\int_{\Omega_R} F(\tau_n v_n-\Hat{q}_\delta)-F(\tau_n u_n-q)
= \int_{\Omega_R} F(\tau_n u_n-\Hat{q}_\delta)-F(\tau_n u_n-q)\ge 0.$$
Hence, $$\liminf_{n \to \infty} \mathcal{E}(u_n) \ge \liminf_{n \to \infty} \Hat{\mathcal{E}}_\delta(\tau_n v_n) \ge \Hat{c}_\delta := \inf_{v \in \mathrm{D}^{1, 2}_0(]-2\delta, +\infty[\times {\mathbf{R}})} \Hat{\mathcal{E}}_\delta(v),$$ and the conclusion follows from Proposition \[prop:al\], sending $\delta$ to zero.
From Proposition \[propositionPS\], we derive
\[theoremExistenceLevels\] If $$\sup_{x \in \Omega}\frac{\kappa^2}{2} H(x, x)-\kappa q(x) > \frac{\kappa^2}{4\pi} \Bigl(\log \frac{\kappa}{2\pi \Hat{W}}-1\Bigr)-\kappa \Hat{d},$$ then, if ${\varepsilon}$ is sufficiently small, there exists $u^{\varepsilon}\in \mathrm{D}^{1, 2}_0(\Omega)$ such that $d\mathcal{E}^{\varepsilon}(u^{\varepsilon})=0$ and $\mathcal{E}^{\varepsilon}(u^{\varepsilon})=c^{\varepsilon}$.
By Proposition \[propUnboundedUpper\], we have $$c^{\varepsilon}\le \frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}}-\sup_{x\in \Omega}\Bigl(\frac{\kappa^2}{2} H(x, x)-\kappa q(x)
\Bigr) +\mathcal{C} + o(1).$$ On the other hand, in view of Theorem \[thmYang\], $\Hat{\mathcal{E}}^{\varepsilon}$ possesses a ground-state whose energy is bounded by $\frac{\kappa^2}{4\pi}\log \frac{1}{{\varepsilon}}+O(1)$. It follows from Proposition \[prop:1maiUnbounded\] applied to these ground-states that $$\begin{split}
\Hat{c}^{\varepsilon}&=\frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}}-\sup_{x\in {\mathbf{R}}^2_+} \Bigl(\frac{\kappa^2}{4\pi} \log 2x_1-\kappa (\Hat{W}x_1+\Hat{d})\Bigr) +\mathcal{C} + o(1) \\
&=\frac{\kappa^2}{4\pi} \log \frac{1}{{\varepsilon}}- \Bigl( \frac{\kappa^2}{4\pi} \Bigl(\log \frac{\kappa}{2\pi \Hat{W}}-1\Bigr)-\kappa \Hat{d} \Bigr) +\mathcal{C} + o(1).
\end{split}$$ Therefore, when ${\varepsilon}$ is small enough, $c^{\varepsilon}< \Hat{c}^{\varepsilon}$, and the conclusion follows from Proposition \[propositionPS\].
Pair of vortices in bounded domains {#sectionVortexPair}
===================================
In this section, $\Omega\subset {\mathbf{R}}^2$, $f : {\mathbf{R}}\to {\mathbf{R}}$ and $q: \Omega \to {\mathbf{R}}$ are as in Section \[sectionSingleVortex\]. For ${{\boldsymbol{\varepsilon}}}=({\varepsilon}_+, {\varepsilon}_-) >0$, $\kappa_+>0$ and $\kappa_-<0$ given, and consider solutions of the boundary value problems $$\left\{
\begin{aligned}
-\Delta u^{{\boldsymbol{\varepsilon}}}&= \frac{1}{{\varepsilon}_+{}^2} f(u^{{\boldsymbol{\varepsilon}}}- q^{\varepsilon}_+) - \frac{1}{{\varepsilon}_-{}^2} f(q^{\varepsilon}_-
-u^{{\boldsymbol{\varepsilon}}}) &
& \text{in $\Omega$}, \\
u^{{\boldsymbol{\varepsilon}}}&= 0 & &\text{on $\partial \Omega$},
\end{aligned}
\right. \tag{\protect{$\mathcal{Q}^{{\boldsymbol{\varepsilon}}}$}}
\label{Qeps}$$ where $q^{\varepsilon}_\pm = q+ \frac{\kappa_\pm}{2\pi} \log \frac{1}{{\varepsilon}_\pm}$.
We consider are the least energy nodal solutions of obtained by minimizing the energy functional $$\mathcal{E}^{{\boldsymbol{\varepsilon}}}(u)= \int_{\Omega} \Bigl(\frac{|\nabla u|^2}{2} -
\frac{1}{{\varepsilon}_+{}^2}F(u-q^{\varepsilon}_+) -\frac{1}{{\varepsilon}_-{}^2}F(q^{\varepsilon}_- -u) \Bigr)
$$ over the natural constraint given by the nodal Nehari set $$\mathcal{M}^{{\boldsymbol{\varepsilon}}}= \left\{ u\in H^1_0(\Omega) \ : \ u_+\neq 0, u_- \ne 0, \ \langle
d\mathcal{E}^{{\boldsymbol{\varepsilon}}}(u), u_+\rangle = \langle
d\mathcal{E}^{{\boldsymbol{\varepsilon}}}(u), u_-\rangle =0\right\}.$$ It is a standard [@CastroCossioNeuberger; @BartschWethWillem; @BartschWeth2003; @BartschWeth2005] to prove the
Assume that $q^{\varepsilon}_+$ is positive on $\Omega$ and $q^{\varepsilon}_-$ is negative on $\Omega$, so that ${\mathcal{M}^{\varepsilon}}\neq \emptyset$, and define $$d^{{\boldsymbol{\varepsilon}}}= \inf_{u\in \mathcal{M}^{{\boldsymbol{\varepsilon}}}} \mathcal{E}^{{\boldsymbol{\varepsilon}}}(u).$$ There exists $u^{{\boldsymbol{\varepsilon}}}\in \mathcal{M}^{{\boldsymbol{\varepsilon}}}$ such that $\mathcal{E}^{{\boldsymbol{\varepsilon}}}(u^{{\boldsymbol{\varepsilon}}})=d^{{\boldsymbol{\varepsilon}}}$, and $u^{{\boldsymbol{\varepsilon}}}$ is a nonnegative solution of ${(\mathcal{Q}^{\varepsilon})}$.
Our focus is the asymptotics of $u^{{\boldsymbol{\varepsilon}}}$ for a sequence ${{\boldsymbol{\varepsilon}}}\to (0, 0)$. We assume that $0 < c < \frac{\log {\varepsilon}_+}{\log {\varepsilon}_-} < C < \infty$, and we will write ${{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}$ instead of $\log {\varepsilon}_+$ or $\log {\varepsilon}_-$ in asymptotic expansions.
We extend the definition of $U_\kappa$ given by for $\kappa < 0$ by $U_\kappa=-U_{-\kappa}$ and $\rho_{\kappa}=\rho_{-\kappa}$. One still has, when ${\lvertx\rvert}$ is large enough, $U_\kappa(x)=\frac{\kappa}{2\pi}\log \frac{\rho_\kappa}{{\lvertx\rvert}}$. We also set $$\mathcal{C}_\pm=
\frac{\kappa_\pm^2}{4\pi} \log \rho_{\kappa_\pm} +
\int_{B(0, \rho_{\kappa_\pm})}\Bigl(\frac{|\nabla U_{\rho_{\kappa_\pm}}|^2}{2} - \frac{U_{\rho_{\kappa_\pm}}^{p+1}}{p+1}\Bigr).$$ The Kirchhoff–Routh function $\mathcal{W}$ is defined for $(x_+, x_-)\in \Omega^2_*=\{ (y_+, y_-) \in \Omega {\: :\:}y_+ \ne y_-\}$ by $$\begin{split}
\mathcal{W}(x_+,x_-)= \,
&\frac{\kappa_+^2}{2}H(x_+, x_+) + \frac{\kappa_-^2}{2}H(x_-, x_-) +
\kappa_+\kappa_- G(x_+, x_-)\\
&
-\frac{\kappa_+}{2\pi} q(x_+) -\frac{\kappa_-}{2\pi} q(x_-).
\end{split}$$ We set $$\begin{aligned}\label{defiqNodal}
A^{{\boldsymbol{\varepsilon}}}_\pm &=\Big\{ x \in \Omega {\: :\:}\pm u^{{\boldsymbol{\varepsilon}}}(x)> \pm q^{{\boldsymbol{\varepsilon}}}(x) + \frac{\kappa_\pm}{2\pi} \log \frac{1}{{\varepsilon}_\pm}\Big\}, \\
\omega_\pm^{{\boldsymbol{\varepsilon}}}&=\pm \frac{1}{{\varepsilon}_\pm{}^2} f(\pm(u^{{\boldsymbol{\varepsilon}}}-q^{{\boldsymbol{\varepsilon}}})), \\
\kappa^{{\boldsymbol{\varepsilon}}}_\pm&=\int_{\Omega} \omega_\pm^{{\boldsymbol{\varepsilon}}}, \\
x^{{\boldsymbol{\varepsilon}}}&=\frac{1}{\kappa^{{\boldsymbol{\varepsilon}}}}\int_{\Omega} x \, \omega^{{\boldsymbol{\varepsilon}}}_\pm(x)\, dx, \\
\rho^{{\boldsymbol{\varepsilon}}}_\pm&=\rho_{\kappa^{{\boldsymbol{\varepsilon}}}_\pm}.
\end{aligned}$$
We will prove
\[thm:K3\] As ${{\boldsymbol{\varepsilon}}}\to 0$, we have $$\begin{split}
u^{{\boldsymbol{\varepsilon}}}= \, & U_{\kappa_+^{{\boldsymbol{\varepsilon}}}} \Big(\frac{\cdot-x^{{\boldsymbol{\varepsilon}}}}{{\varepsilon}_+}\Big)+\kappa_+^{{\boldsymbol{\varepsilon}}}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}_+ \rho_{+}^{{\boldsymbol{\varepsilon}}}}+ H(x^{{\boldsymbol{\varepsilon}}}_+, \cdot)\Bigr)\\
&+U_{\kappa_-^{{\boldsymbol{\varepsilon}}}} \Big(\frac{\cdot-x^{{\boldsymbol{\varepsilon}}}}{{\varepsilon}_-}\Big)+\kappa_-^{{\boldsymbol{\varepsilon}}}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}_- \rho_-^{{\boldsymbol{\varepsilon}}}}+ H(x^{{\boldsymbol{\varepsilon}}}_-, \cdot)\Bigr)+o(1),
\end{split}$$ in $\mathrm{W}^{2, 1}_\mathrm{loc}(\Omega)$, in $\mathrm{W}^{1, 2}_0(\Omega)$, and in $\mathrm{L}^\infty(\Omega)$, where $$\kappa^{{\boldsymbol{\varepsilon}}}=\kappa_\pm+\frac{2\pi}{\log \frac{1}{{\varepsilon}_\pm}}\Bigl(q(x^{{\boldsymbol{\varepsilon}}})-\kappa_\pm H(x^{{\boldsymbol{\varepsilon}}}, x^{{\boldsymbol{\varepsilon}}})-\kappa_\mp G(x_\pm, x_\mp) -\frac{\kappa}{2\pi} \log \frac{1}{\rho_{\kappa_\pm}} \Bigr)+o({{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}^{-1}),$$ and $$\mathcal{W}(x^{{\boldsymbol{\varepsilon}}}_+,x^{{\boldsymbol{\varepsilon}}}_-) \to \sup_{(x_+,x_-) \in \Omega^2_*} \mathcal{W}(x_+,x_-).$$ One also has $$B(x^{{\boldsymbol{\varepsilon}}}_\pm, \Bar{r}_\pm^{{\boldsymbol{\varepsilon}}}) \subset A_\pm^{{\boldsymbol{\varepsilon}}}\subset B(x_\pm^{{\boldsymbol{\varepsilon}}}, \mathring{r}_\pm^{{\boldsymbol{\varepsilon}}}),$$ with $\Bar{r}^{{\boldsymbol{\varepsilon}}}_\pm={\varepsilon}_\pm \rho_{\kappa_\pm}+o({\varepsilon}_\pm)$ and $\mathring{r}_\pm^{{\boldsymbol{\varepsilon}}}={\varepsilon}_\pm
\rho_{\kappa_\pm} +o({\varepsilon}_\pm)$. Finally, $$\mathcal{E}^{{\boldsymbol{\varepsilon}}}(u^{{\boldsymbol{\varepsilon}}})= \frac{\kappa^2_+}{4\pi} \log \frac{1}{{\varepsilon}_+}+\frac{\kappa^2_-}{4\pi} \log \frac{1}{{\varepsilon}_-}-\mathcal{W}(x^{{\boldsymbol{\varepsilon}}}_+,x^{{\boldsymbol{\varepsilon}}}_-)+\mathcal{C}_++\mathcal{C}_-+o(1).$$
Upper bounds on the energy {#upper-bounds-on-the-energy}
--------------------------
We compute upper bounds on $d^{\varepsilon}$ by constructing suitable elements in ${\mathcal{M}^{\varepsilon}}$.
\[lemNodalUpperBound\] For every $\Hat{x}_+, \Hat{x}_- \in \Omega$ such that $\Hat{x}_+\ne \Hat{x}_-$, there exists $$\Hat{\kappa}^{\pm}_{{{\boldsymbol{\varepsilon}}}}=\kappa_\pm+\frac{2\pi}{\log \dfrac{1}{{\varepsilon}_\pm}}\Bigl( q(\Hat{x}_\pm)-\kappa_\pm H(\Hat{x}_\pm, \Hat{x}_\pm)-\kappa_{\mp} G(\Hat{x}_\pm, \Hat{x}_\mp)+\dfrac{\kappa_\pm}{2\pi} \log \rho_{\kappa_\pm} \Bigr)+O\bigl({{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}^{-2}\bigr),$$ such that, if $$\begin{split}
\Hat{u}^{{\boldsymbol{\varepsilon}}}(x)
=&U_{\Hat{\kappa}_+^{{\boldsymbol{\varepsilon}}}}\Bigl(\frac{x-\Hat{x}_+}{{\varepsilon}_+}\Bigr)+\Hat{\kappa}^{{\boldsymbol{\varepsilon}}}_+
\Bigl( \frac{1}{2\pi} \log \frac{1}{{\varepsilon}_+\Hat{\rho}_+^{{\boldsymbol{\varepsilon}}}}+H(\Hat{x}_+, x) \Bigr)\\
&+U_{\Hat{\kappa}_-^{{\boldsymbol{\varepsilon}}}}\Bigl(\frac{x-\Hat{x}_-}{{\varepsilon}_-}\Bigr)+\Hat{\kappa}^{{\boldsymbol{\varepsilon}}}_-
\Bigl( \frac{1}{2\pi} \log \frac{1}{{\varepsilon}_-\Hat{\rho}_-^{{\boldsymbol{\varepsilon}}}}+H(\Hat{x}_-, x) \Bigr),
\end{split}$$ then $$\Hat{u}^{{\boldsymbol{\varepsilon}}}\in \mathcal{M}^{{\boldsymbol{\varepsilon}}}.$$ Moreover, $$\Hat{A}_\pm^{{\boldsymbol{\varepsilon}}}:=\bigl\{ x {\: :\:}\pm \Hat{u}^{{\boldsymbol{\varepsilon}}}(x) > \pm q^{{\boldsymbol{\varepsilon}}}_\pm(x) \bigr\} \subset B(\Hat{x}_\pm, \Hat{r}_\pm^{{\boldsymbol{\varepsilon}}}), \\
$$ with $\Hat{r}_\pm^{{\boldsymbol{\varepsilon}}}={\varepsilon}_\pm \rho_{\kappa_\pm}+o({{\boldsymbol{\varepsilon}}})$.
For every $\boldsymbol{\sigma}=(\sigma_+, \sigma_-) \in {\mathbf{R}}^2$, we define $$\begin{gathered}
\Hat{\kappa}^\pm_{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}=\frac{q^{{\boldsymbol{\varepsilon}}}_\pm(x_\pm)-\kappa_\mp G(x_\pm, x_\mp)+\sigma_\pm}{\frac{1}{2\pi} \log \frac{1}{{\varepsilon}_\pm \rho_{\kappa_\pm}}+H(x_\pm, x_\pm)},\\
\begin{split}
\Hat{u}_{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}} =&U_{\Hat{\kappa}_+^{{\boldsymbol{\varepsilon}}}}\Bigl(\frac{x-\Hat{x}_+}{{\varepsilon}_+}\Bigr)+\Hat{\kappa}^{\varepsilon}_+
\Bigl( \frac{1}{2\pi} \log \frac{1}{{\varepsilon}_+\Hat{\rho}_+^{{\boldsymbol{\varepsilon}}}}+H(\Hat{x}_+, x) \Bigr)\\
&+U_{\Hat{\kappa}_-^{{\boldsymbol{\varepsilon}}}}\Bigl(\frac{x-\Hat{x}_-}{{\varepsilon}_-}\Bigr)+\Hat{\kappa}^{\varepsilon}_-
\Bigl( \frac{1}{2\pi} \log \frac{1}{{\varepsilon}_-\Hat{\rho}_-^{{\boldsymbol{\varepsilon}}}}+H(\Hat{x}_-, x) \Bigr),
\end{split}\end{gathered}$$ and we set $$g^{{{\boldsymbol{\varepsilon}}}}_\pm(\boldsymbol{\sigma})=\langle d \mathcal{E}_{{{\boldsymbol{\varepsilon}}}} (\Hat{u}^{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}), \Hat{u}^{{{\boldsymbol{\varepsilon}}}, \sigma_\pm} \rangle.$$ We compute as in the proof of Lemma \[lemmaHatuNehari\], $$\label{eq:badaboum1}
\int_{\Omega} {\lvert\nabla u^{{{\boldsymbol{\varepsilon}}}, \sigma_\pm}\rvert}^2
=\int_{B(0, \rho_{\Hat{\kappa}_{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}})} {\lvert\nabla U_{\Hat{\kappa}_{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}}\rvert}^2+\Hat{\kappa}^{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}_{\pm} \Bigl(\frac{\kappa_\pm}{2\pi} \log \frac{1}{{\varepsilon}_\pm}+q(\Hat{x}_\pm)+\sigma_\pm\Bigr)+O({\lvert{{\boldsymbol{\varepsilon}}}\rvert}).$$ We also set $$\Hat{\omega}^{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}=\frac{1}{{\varepsilon}_+^2} f(\Hat{u}^{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}} - q^{{\boldsymbol{\varepsilon}}}_+) - \frac{1}{{\varepsilon}_-^2} f(q^{{\boldsymbol{\varepsilon}}}_- -\Hat{u}^{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}),$$ and we compute as in the proof of Lemma \[lemmaHatuNehari\] $$\begin{gathered}
\label{eq:badaboum2}
\frac{1}{{\varepsilon}_\pm^2}\int_{\Omega} \Hat{\omega}^{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}} \Hat{u}^{{{\boldsymbol{\varepsilon}}}, \boldsymbol{\sigma}}_\pm
=\int_{{\mathbf{R}}^2} F(U_{\kappa_\pm}+\sigma_\pm) \\+ (\tfrac{\kappa}{2\pi} \log \tfrac{1}{{\varepsilon}_\pm}+q(\Hat{x}_\pm)+\sigma_\pm)\int_{{\mathbf{R}}^2}f(U_{\kappa_\pm}+\sigma_\pm)+o(1). \end{gathered}$$ Combining and we obtain $$g^{{\boldsymbol{\varepsilon}}}_\pm(\boldsymbol{\sigma})=\frac{\kappa_\pm}{2\pi} \log \frac{1}{{\varepsilon}_\pm} \Bigl( \int_{{\mathbf{R}}^2} f(U_{\kappa_\pm})-f(U_{\kappa_\pm}+\sigma_\pm)\Bigr)+O(1).$$ By the Poincaré–Miranda Theorem (see e.g. [@Kulpa1997]), when ${\lvert{{\boldsymbol{\varepsilon}}}\rvert}$ is small, there exists $\boldsymbol{\sigma}_{{\boldsymbol{\varepsilon}}}$ such that $g^{{\boldsymbol{\varepsilon}}}(\boldsymbol{\sigma}_{{\boldsymbol{\varepsilon}}})=0$ and $\boldsymbol{\sigma}_{{\boldsymbol{\varepsilon}}}=o(1)$ as ${{\boldsymbol{\varepsilon}}}\to 0$.
Evaluating $\mathcal{E}_{{\boldsymbol{\varepsilon}}}(\Hat{u}_{{\boldsymbol{\varepsilon}}})$ yields
\[cor:Nodalupper\] As ${\lvert{{\boldsymbol{\varepsilon}}}\rvert} \to 0$, we have $$\begin{split}
d^{{\boldsymbol{\varepsilon}}}\leq\ &\frac{\kappa^2_+}{4\pi} \log \frac{1}{{\varepsilon}_+}+\frac{\kappa^2_-}{4\pi} \log \frac{1}{{\varepsilon}_-}-\mathcal{W}(x_+,x_-)+\mathcal{C}_++\mathcal{C}_-+o(1).
\end{split}$$
Asymptotic behavior of solutions {#asymptotic-behavior-of-solutions-1}
--------------------------------
We shall prove the counterpart of Proposition \[prop:1mai\]
\[prop:1maiNodal\] Let $(v^{{\boldsymbol{\varepsilon}}})$ be a family of solutions to such that $v^{{\boldsymbol{\varepsilon}}}_\pm \ne 0$ $$\label{assumptEnergyUpperboundNodal}
\mathcal{E}^{{\boldsymbol{\varepsilon}}}(v^{{\boldsymbol{\varepsilon}}}) \le \frac{\kappa^2_+}{4\pi} \log \frac{1}{{\varepsilon}_+}+\frac{\kappa^2_-}{4\pi} \log \frac{1}{{\varepsilon}_-}+O(1),$$ as ${{\boldsymbol{\varepsilon}}}\to 0$. Define the quantities $A_\pm^{{\boldsymbol{\varepsilon}}}$, $\omega_\pm^{{\boldsymbol{\varepsilon}}}$, $\kappa_\pm^{{\boldsymbol{\varepsilon}}}$, $x_\pm^{{\boldsymbol{\varepsilon}}}$ and $\rho_\pm^{{\boldsymbol{\varepsilon}}}$ for $v^{{\boldsymbol{\varepsilon}}}$ as in for $u^{{\boldsymbol{\varepsilon}}}$. Then $$\begin{split}
v^{{\boldsymbol{\varepsilon}}}= \, & U_{\kappa_+^{{\boldsymbol{\varepsilon}}}} \Big(\frac{\cdot-x^{{\boldsymbol{\varepsilon}}}}{{\varepsilon}_+}\Big)+\kappa_+^{{\boldsymbol{\varepsilon}}}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}_+ \rho_{+}^{{\boldsymbol{\varepsilon}}}}+ H(x^{{\boldsymbol{\varepsilon}}}_+, \cdot)\Bigr)\\
&+U_{\kappa_-^{{\boldsymbol{\varepsilon}}}} \Big(\frac{\cdot-x^{{\boldsymbol{\varepsilon}}}}{{\varepsilon}_-}\Big)+\kappa_-^{{\boldsymbol{\varepsilon}}}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}_- \rho_-^{{\boldsymbol{\varepsilon}}}}+ H(x^{{\boldsymbol{\varepsilon}}}_-, \cdot)\Bigr)+o(1),
\end{split}$$ , where $$\kappa^{{\boldsymbol{\varepsilon}}}=\kappa_\pm+\frac{2\pi}{\log \frac{1}{{{\boldsymbol{\varepsilon}}}_\pm}}\Bigl(q(x^{{\boldsymbol{\varepsilon}}})-\kappa_\pm H(x^{{\boldsymbol{\varepsilon}}}, x^{{\boldsymbol{\varepsilon}}})-\kappa_\mp G(x_\pm, x_\mp) -\frac{\kappa}{2\pi} \log \frac{1}{\rho_{\kappa_\pm}} \Bigr)+o({{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}^{-1}).$$ In particular, we have $$\mathcal{E}^{{\boldsymbol{\varepsilon}}}(v^{{\boldsymbol{\varepsilon}}})= \frac{\kappa^2_+}{4\pi} \log \frac{1}{{\varepsilon}_+}+\frac{\kappa^2_-}{4\pi} \log \frac{1}{{\varepsilon}_-}-\mathcal{W}(x_+,x_-)+\mathcal{C}_++\mathcal{C}_-+o(1).$$ and $$B(x^{{\boldsymbol{\varepsilon}}}_\pm, \Bar{r}_\pm^{{\boldsymbol{\varepsilon}}}) \subset A_\pm^{{\boldsymbol{\varepsilon}}}\subset B(x_\pm^{{\boldsymbol{\varepsilon}}}, \mathring{r}_\pm^{{\boldsymbol{\varepsilon}}}),$$ with $\Bar{r}^{{\boldsymbol{\varepsilon}}}_\pm={\varepsilon}_\pm \rho_{\kappa_\pm}+o({\varepsilon}_\pm)$ and $\mathring{r}_\pm^{{\boldsymbol{\varepsilon}}}={\varepsilon}_\pm
\rho_{\kappa_\pm} +o({\varepsilon}_\pm)$.
In other words, $v^{{\boldsymbol{\varepsilon}}}$ satisfies the same asymptotics as the one stated in Theorem \[thm:K1\] for $v^{{\boldsymbol{\varepsilon}}}$ except for the convergence of $x^{{\boldsymbol{\varepsilon}}}$.
### Step 1: First quantitative properties of the solutions
\[propositionNodalEstimatesueps\] We have, as ${\lvert{{\boldsymbol{\varepsilon}}}\rvert} \to 0$, $$\begin{gathered}
{\mathcal{L}^{2}}(A^{{\boldsymbol{\varepsilon}}}_\pm) =O\bigl({{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}^{-1}\bigr), \\
\int_{A^{{\boldsymbol{\varepsilon}}}_+} {\lvert\nabla (v^{{\boldsymbol{\varepsilon}}}-q^{{\boldsymbol{\varepsilon}}}_\pm)\rvert}^2 =O(1), \\
\frac{1}{{\varepsilon}_{\pm}^2}\int_{A^{{\boldsymbol{\varepsilon}}}_\pm} F(\pm(v^{{\boldsymbol{\varepsilon}}}-q^{{\boldsymbol{\varepsilon}}}_\pm)) =O(1),\\
\int_{\Omega\setminus A^{{\boldsymbol{\varepsilon}}}_\pm} {\lvert\nabla v^{{\boldsymbol{\varepsilon}}}_\pm\rvert}^2 \leq \frac{\kappa^2_\pm}{2\pi} \log\frac{1}{{{\boldsymbol{\varepsilon}}}_\pm} + O(1), \\
\pm \int_{\Omega} \omega^{{\boldsymbol{\varepsilon}}}_\pm \leq \pm \kappa_\pm + O\bigl({{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}^{-1}\bigr). \end{gathered}$$
First note that by Theorem \[thm:K1\], $$\mathcal{E}_{{\boldsymbol{\varepsilon}}}(v^{{\boldsymbol{\varepsilon}}}_\pm) \ge \frac{\kappa_+^2}{4\pi} \log \frac{1}{{\varepsilon}_\pm}+O(1).$$ By , this implies that $$\mathcal{E}_{{\boldsymbol{\varepsilon}}}(v^{{\boldsymbol{\varepsilon}}}_\pm) =\frac{\kappa_+^2}{4\pi} \log \frac{1}{{\varepsilon}_\pm}+O(1).$$ We are now in position to proceed as in the proof of Proposition \[propositionEstimatesueps\], testing $(\mathcal{Q}^{{\boldsymbol{\varepsilon}}})$ against $v^{{\boldsymbol{\varepsilon}}}_+$ and $v^{{\boldsymbol{\varepsilon}}}_-$ instead of $v^{{\boldsymbol{\varepsilon}}}$, then against $\min(v^{{\boldsymbol{\varepsilon}}}, q^{{\boldsymbol{\varepsilon}}}_+)$ and $\max(v^{{\boldsymbol{\varepsilon}}}, q^{{\boldsymbol{\varepsilon}}}_-)$ instead of $\min(v^{{\boldsymbol{\varepsilon}}}, q^{{\boldsymbol{\varepsilon}}})$, and finally against $(v^{{\boldsymbol{\varepsilon}}}-q^{{\boldsymbol{\varepsilon}}}_+)_+$ and $(q^{{\boldsymbol{\varepsilon}}}_--v^{{\boldsymbol{\varepsilon}}})_+$ instead of $(v^{{\boldsymbol{\varepsilon}}}-q^{{\boldsymbol{\varepsilon}}}_+)_+$. We skip the details.
### Step 2: Structure of the vorticity set
In this subsection we further describe the vorticity set $A^{{\boldsymbol{\varepsilon}}}=A^{{\boldsymbol{\varepsilon}}}_+\cup A^{{\boldsymbol{\varepsilon}}}_-$. Since it is an open set, it contains at most countably many connected components that we label $A^{{\boldsymbol{\varepsilon}}}_{\pm, i}$, $i \in I^{{\boldsymbol{\varepsilon}}}_\pm$. First we have a control on the total area and on the diameter of each connected component.
\[lemmaNodalAreaDiameter\] If ${\lvert{{\boldsymbol{\varepsilon}}}\rvert}$ is sufficiently small, we have $${\mathcal{L}^{2}}(A^{{\boldsymbol{\varepsilon}}}_\pm) \le C {\varepsilon}_\pm^2$$ and, for every $i \in I^{{\boldsymbol{\varepsilon}}}_\pm$, $$\label{ineqNodalVorticityDiameter}
\operatorname{diam}(A^{{\boldsymbol{\varepsilon}}}_{\pm, i}) \le C {\varepsilon}_\pm.$$
It suffices to repeat the arguments in the proof of Lemma \[lemmaNodalAreaDiameter\].
\[lemmaNodalVortexSplit\] There exists constants $\gamma, C, c>0$ such that, when ${\lvert{{\boldsymbol{\varepsilon}}}\rvert}$ is small enough, if $$\label{eqNodalSplitVortices}
\int_{A^{{\boldsymbol{\varepsilon}}}_{\pm, i}} {\lvert\nabla (v^{{\boldsymbol{\varepsilon}}}- q^{{\boldsymbol{\varepsilon}}}_\pm)\rvert}^2 > \gamma^2,$$ then for every $j \in I^{{\boldsymbol{\varepsilon}}}_\mp$, $$\begin{gathered}
\label{ineqNodalLowerBoundMeas} {\mathcal{L}^{2}}(A_{\pm, i}^{{\boldsymbol{\varepsilon}}})\ge c{\varepsilon}^2_\pm, \\
\label{ineqNodalLowerBoundDiam} \operatorname{diam}(A^{{\boldsymbol{\varepsilon}}}_{\pm, i})\ge c{\varepsilon}_\pm, \\
\label{ineqNodalLowerBoundBoundary} \operatorname{dist}(A^{{\boldsymbol{\varepsilon}}}_{\pm, i}, \partial \Omega)\ge c, \\
\label{ineqNodalSignDistance} \operatorname{dist}(A^{{\boldsymbol{\varepsilon}}}_{\pm, i}, A^{{\boldsymbol{\varepsilon}}}_{\mp, j})\ge c,\end{gathered}$$ while if does not hold, then $$\int_{A^{{\boldsymbol{\varepsilon}}}_{\pm, i}} {\lvert\omega^{{\boldsymbol{\varepsilon}}}\rvert}^s \le C {\Vert\nabla q\Vert}_{\mathrm{L}^r(A^{{\boldsymbol{\varepsilon}}}_{\pm, i})}^{sp} {\mathcal{L}^{2}}(A^{{\boldsymbol{\varepsilon}}}_{\pm,i})^{1+sp(\frac{1}{2}-\frac{1}{r})},$$ where $C$ only depends on $s \ge 1$.
The proof is very similar to the one of Lemma \[lemmaVortexSplit\] except for which remains to be proved. To that purpose, we consider the function $$\eta_{{\boldsymbol{\varepsilon}}}=\frac{\frac{v^{{\boldsymbol{\varepsilon}}}_+}{\kappa_+}+\frac{v^{{\boldsymbol{\varepsilon}}}_-}{\kappa_+}}{\log \frac{1}{{\varepsilon}_+{\varepsilon}_-}}.$$ We have $$\eta_{{\boldsymbol{\varepsilon}}}{\vert_{A^{{\boldsymbol{\varepsilon}}}_{\pm, i}}}=\frac{\log \frac{1}{{\varepsilon}_+}}{\log \frac{1}{{\varepsilon}_+{\varepsilon}_-}}+O\bigl({{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}^{-1}\bigr),$$ and $$\eta_{{\boldsymbol{\varepsilon}}}{\vert_{A^{{\boldsymbol{\varepsilon}}}_{\mp, j}}}=\frac{- \log \frac{1}{{\varepsilon}_-}}{\log \frac{1}{{\varepsilon}_+{\varepsilon}_-}}+O\bigl({{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}^{-1}\bigr).$$ Therefore, $$\frac{2\pi}{\operatorname{cap}(A^{{\boldsymbol{\varepsilon}}}_+, {\mathbf{R}}^2\setminus A^{{\boldsymbol{\varepsilon}}}_-)}\ge \log \frac{1}{{\varepsilon}_+ {\varepsilon}_-}+O(1).$$ Using Proposition \[propositionCapacityBoundDistance\] with $\Omega={\mathbf{R}}^2 \setminus \overline{A^{{\boldsymbol{\varepsilon}}}_{\pm, i}}$ and $K=\overline{A^{{\boldsymbol{\varepsilon}}}_{\mp, j}}$, and applying to $A^{{\boldsymbol{\varepsilon}}}_{\mp, j}$ and to $A^{{\boldsymbol{\varepsilon}}}_{\pm, i}$, we are led to $$\log \frac{1}{{\varepsilon}_+{\varepsilon}_-} \le \log C \Bigl( 1+\frac{\operatorname{dist}(A^{{\boldsymbol{\varepsilon}}}_{\pm, i}, A^{{\boldsymbol{\varepsilon}}}_{\mp, j})}{{\varepsilon}_\mp}\Bigr)\Bigl( 1+\frac{\operatorname{dist}(A^{{\boldsymbol{\varepsilon}}}_{\pm, i}, A^{{\boldsymbol{\varepsilon}}}_{\mp, j})}{{\varepsilon}_\pm}\Bigr)+O(1),$$ which can not hold if $\operatorname{dist}(A^{{\boldsymbol{\varepsilon}}}_{\pm, i}, A^{{\boldsymbol{\varepsilon}}}_{\mp, j}) \to 0$.
The vorticity set is split into four subsets: $$\begin{aligned}
V^{{\boldsymbol{\varepsilon}}}_\pm&=\bigcup \Bigl\{A_{\pm, i}^{{\boldsymbol{\varepsilon}}}{\: :\:}\int_{A_{\pm, i}^{{\boldsymbol{\varepsilon}}}} {\lvert\nabla (v^{{\boldsymbol{\varepsilon}}}- q^{{\boldsymbol{\varepsilon}}}_\pm\rvert}^2 \le \gamma^2\Bigr\}, \\
E^{{\boldsymbol{\varepsilon}}}_\pm&=\bigcup \Bigl\{A_{\pm, i}^{{\boldsymbol{\varepsilon}}}{\: :\:}\int_{A_{\pm, i}^{{\boldsymbol{\varepsilon}}}} {\lvert\nabla (v^{{\boldsymbol{\varepsilon}}}- q^{{\boldsymbol{\varepsilon}}}_\pm)\rvert}^2 > \gamma^2\Bigr\}. \end{aligned}$$ By Proposition \[propositionNodalEstimatesueps\], the sets $E^{{\boldsymbol{\varepsilon}}}_+$ and $E^{{\boldsymbol{\varepsilon}}}_-$ contain finitely many connected components, and by , , and , they can thus be split as $E^{{\boldsymbol{\varepsilon}}}_\pm=\bigcup_{j=1}^{k^{{\boldsymbol{\varepsilon}}}_\pm} E^{{\boldsymbol{\varepsilon}}}_{\pm, j}$, where $E^{{\boldsymbol{\varepsilon}}}_{\pm, j}$ are nonempty open sets such that $$\begin{gathered}
\frac{\operatorname{dist}(E^{{\boldsymbol{\varepsilon}}}_{\pm, i}, E^{{\boldsymbol{\varepsilon}}}_{\pm, j})}{{\varepsilon}_\pm} \to \infty,\\
\liminf_{{{\boldsymbol{\varepsilon}}}\to 0} \operatorname{dist}(E^{{\boldsymbol{\varepsilon}}}_{\pm, i}, E^{{\boldsymbol{\varepsilon}}}_{\mp, j}) >0,\\
\liminf_{{{\boldsymbol{\varepsilon}}}\to 0} \operatorname{dist}(E^{{\boldsymbol{\varepsilon}}}_{\pm, i}, \partial \Omega) >0, \\
\limsup_{{{\boldsymbol{\varepsilon}}}\to 0} \frac{\operatorname{diam}(E^{{\boldsymbol{\varepsilon}}}_{\pm, i})}{{\varepsilon}_\pm} < \infty,\end{gathered}$$ as ${{\boldsymbol{\varepsilon}}}\to 0$. By definition of $E^{{\boldsymbol{\varepsilon}}}$ and by , $k^{{\boldsymbol{\varepsilon}}}_+$ and $k^{{\boldsymbol{\varepsilon}}}_-$ remain bounded as ${{\boldsymbol{\varepsilon}}}\to 0$.
### Step 3: Small scale asymptotics
We set $$\begin{aligned}
\omega^{{\boldsymbol{\varepsilon}}}_{\pm, v}&=\omega^{{\boldsymbol{\varepsilon}}}{\chi_{V_{\pm}^{{\boldsymbol{\varepsilon}}}}}, &
\omega^{{\boldsymbol{\varepsilon}}}_{\pm, i}&=\omega^{{\boldsymbol{\varepsilon}}}{\chi_{E^{{\boldsymbol{\varepsilon}}}_{\pm, i}}}, \\
\kappa^{{\boldsymbol{\varepsilon}}}_{\pm, i}&=\int_{\Omega} \omega^{{\boldsymbol{\varepsilon}}}_{\pm, i}, &
x^{{\boldsymbol{\varepsilon}}}_{\pm, i}&=\frac{1}{\kappa^{{\boldsymbol{\varepsilon}}}_{\pm, i}}\displaystyle\int_{\Omega} x\omega^{{\boldsymbol{\varepsilon}}}_{\pm, i}(x)\, dx. \end{aligned}$$
Using the analogues of Lemma \[lemmaVanishingVorticity\] and Lemma \[lemmaSmallScaleLocalEstimates\], one obtains the analogue of Lemma \[lemmaLocalAsymptotics\].
\[lemmaNodalLocalAsymptotics\] When ${{\boldsymbol{\varepsilon}}}$ is small, we have $k^{{\boldsymbol{\varepsilon}}}_+=k^{{\boldsymbol{\varepsilon}}}_-=1$, and $$\begin{gathered}
\kappa^{{\boldsymbol{\varepsilon}}}_{\pm, 1}=\kappa_\pm +\frac{2\pi}{\log \frac{1}{{\varepsilon}_\pm}}\Bigl(q(x^{{\boldsymbol{\varepsilon}}}_\pm)-\kappa_\pm H(x^{{\boldsymbol{\varepsilon}}}_\pm, x^{{\boldsymbol{\varepsilon}}}_\pm)-\kappa_{\mp} G(x^{{\boldsymbol{\varepsilon}}}_{\pm}, x^{{\boldsymbol{\varepsilon}}}_{\mp})-\frac{\kappa_\pm}{2\pi} \log \frac{1}{\rho_{\kappa_\pm}} \Bigr) \\
+o({{{\vert\!\log\vert {\boldsymbol{\varepsilon}}\vert\vert}}}^{-1})\end{gathered}$$ and $v^{{\boldsymbol{\varepsilon}}}_{\pm} \to U_{\kappa_\pm}$ in $\mathrm{W}^{1, r}_{\mathrm{loc}}({\mathbf{R}}^2)$.
### Step 4: Global asymptotics
The counterpart of Proposition \[propositionAsymptoticsW21\] is now
\[propositionAsymptoticsW21Nodal\] We have $$\begin{split}
v^{{\boldsymbol{\varepsilon}}}= & \ U_{\kappa_{+,1}^{{\boldsymbol{\varepsilon}}}}\Bigl(\frac{\cdot-x^{{\boldsymbol{\varepsilon}}}_{+, 1}}{{\varepsilon}_+}\Bigr)+\kappa^{{\boldsymbol{\varepsilon}}}_{+, 1}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}_+ \rho_{\kappa_{+, 1}^{{\boldsymbol{\varepsilon}}}}}+H(x^{{\boldsymbol{\varepsilon}}}_{+, 1}, \cdot)\Bigr) \\
&+U_{\kappa_{-,1}^{{\boldsymbol{\varepsilon}}}}\Bigl(\frac{\cdot-x^{{\boldsymbol{\varepsilon}}}_{-, 1}}{{\varepsilon}_-}\Bigr)+\kappa^{{\boldsymbol{\varepsilon}}}_{-, 1}\Bigl(\frac{1}{2\pi} \log \frac{1}{{\varepsilon}_- \rho_{\kappa_{-, 1}^{{\boldsymbol{\varepsilon}}}}}+H(x^{{\boldsymbol{\varepsilon}}}_{-, 1}, \cdot)\Bigr)+o(1)
\end{split}$$ in $\mathrm{W}^{2, 1}_{\mathrm{loc}}(\Omega)$, in $\mathrm{W}^{1, 2}_0(\Omega)$, and in $\mathrm{L}^\infty(\Omega)$.
We have now all the ingredients to complete the
It follows from the combination of Lemma \[lemmaNodalLocalAsymptotics\], Proposition \[propositionAsymptoticsW21Nodal\] and the counterparts of Corollaries \[corollaryAsymptotic\] and \[corEnergy\].
Since the solutions have the upper bound Corollary \[cor:Nodalupper\], one can conclude from Proposition \[prop:1maiNodal\].
Desingularized solutions of the Euler equation {#sect:resu}
==============================================
Bounded domains
---------------
In bounded domains we shall successively consider stationary vortices, rotating vortices and stationary pairs of vortices.
### Stationary vortices in simply-connected bounded domains
Let us first deduce Theorem \[thm:resu\] from Theorem \[thm:K1\].
Take $q=-\psi_0$, where $\psi_0$ satisfies . One checks that $\psi_0 \in W^{1+\frac{1}{s}, s}(\Omega)$ so that $u \in W^{1, r}(\Omega)$ for every $r < \infty$. Define $\mathbf{v}_{\varepsilon}=(\nabla u_{\varepsilon})^\perp$ where $u_{\varepsilon}$ is given by Proposition \[prop:2.1\]. The conclusion then follows from Theorem \[thm:K1\].
We have constructed in Theorem \[thm:resu\] a family of solutions that concentrates around a global maximum of the Kirchhoff–Routh function $\mathcal{W}$; it is also possible to construct family of solutions that concentrate around a *local* maximum of $\mathcal{W}$:
\[thmLocalMinimum\] Let $\Omega \subset {\mathbf{R}}^2$ be a bounded simply-connected smooth domain and $v_n:\partial \Omega \to {\mathbf{R}}\in L^s(\partial \Omega)$ for some $s>1$ be such that $\int_{\partial \Omega} v_n = 0.$ Let $\kappa >0$ be given and let $\Hat{x} \in \Omega$ be a strict local minimizer of $\mathcal{W}$. For ${\varepsilon}>0$ there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in $\Omega$ with outward boundary flux given by $v_n$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}) \subset B(x_{\varepsilon}, C{\varepsilon})$ for some $x_{\varepsilon}\in \Omega$ and $C>0$ not depending on ${\varepsilon}$. Moreover, as ${\varepsilon}\to 0$, $$\int_\Omega \omega_{\varepsilon}\to \kappa$$ and $x_{\varepsilon}\to \Hat{x}$.
Assume that $\Hat{x}$ is the unique minimizer of $\mathcal{W}$ in $B(\Hat{x}, \rho)$. Define $q \in C^\infty(\Bar{\Omega})$ so that $q=-\psi_0$ in $B(\Hat{x}, \rho/2)$, where $\psi_0$ satisfies and for every $x \in \Omega$, $$\kappa q(x)-\frac{\kappa^2}{2} H(x, x) > \kappa q(x_*)-\frac{\kappa^2}{2} H(x_*, x_*).$$ We now apply Theorem \[thm:K1\] with $q$. By construction of $q$, we have $x_{\varepsilon}\to \Hat{x}$.
But then, one has, still by Theorem \[thm:K1\] $$u_{\varepsilon}(x) \ge \frac{\kappa}{2\pi} \log \frac{1}{{\lvertx_{\varepsilon}-x\rvert}}+O(1).$$ Therefore, when ${\varepsilon}$ is small enough, $u_{\varepsilon}\le -\psi_0+\frac{\kappa}{2\pi}\log \frac{1}{{\varepsilon}}$ and $u_{\varepsilon}\le q_{\varepsilon}$ in $\Omega \setminus B(x_{\varepsilon}, \rho/2)$. Therefore, for such ${\varepsilon}$, $u_{\varepsilon}$ solves $-{\varepsilon}^2\Delta u_{\varepsilon}=f(u_{\varepsilon}+\psi_0- \frac{\kappa}{2\pi} \log \frac{1}{{\varepsilon}})$ in $\Omega$. One can now take $\mathbf{v}_{\varepsilon}= (\nabla (u_{\varepsilon}+\psi_0))^\perp$ and show that this is a stationary solution to the Euler equation.
### Stationary vortices in multiply-connected bounded domains
If $\Omega$ is not simply connected then $\Omega = \Omega_0 \setminus \bigcup_{h=1}^m \Omega_h$, where $\Omega_0, \dotsc, \Omega_m$ are bounded simply connected domains, one can prescribe for $h \in \{1, \dotsc, m\}$, the circulations $\int_{\partial \Omega_h} \mathbf{v}\cdot \tau=\gamma_h$. In that case $\mathbf{v}_0$ is the unique harmonic field whose normal component on the boundary is $v_n$; i.e., $\mathbf{v}_0$ satisfies $$\left\{
\begin{aligned}
\nabla \cdot \mathbf{v}_0&=0, & & \text{in $\Omega$}, \\
\nabla \times \mathbf{v}_0&=0, & & \text{in $\Omega$}, \\
n \cdot \mathbf{v}_0&=v_n& & \text{on $\partial \Omega$},\\
\int_{\partial \Omega_h} \mathbf{v}_0 \cdot \tau &=\gamma_h & &\text{for $h \in \{1, \dotsc, m\}$}.
\end{aligned}
\right.$$ If $\int_{\partial \Omega_h} v_n=0$ for every $h \in \{1, \dotsc, m\}$, $\mathbf{v}_0=(\nabla \psi_0)^\perp$ where $$\label{eqpsi0NotConnected}
\left\{
\begin{aligned}
-\Delta \psi_0&=0& &\text{in $\Omega$}, \\
-\frac{\partial \psi_0}{\partial \tau}&=v_n & & \text{on $\partial \Omega$},\\
\int_{\partial \Omega_h} \frac{\partial \psi_0}{\partial n} & = \gamma_h & &\text{for $h \in \{1, \dotsc, m\}$}.\\
\end{aligned}
\right.$$ The Kirchhoff–Routh function associated to the vortex dynamics is then given by $$\mathcal{W}_*(x)=\frac{\kappa^2}{2}H_*(x, x)+\kappa \psi_0(x),$$ where one should recall that $\psi_0$ depends on $v_n$ and $\gamma_h$ for $h \in \{1, \dotsc, m\}$.
We have
\[thm:MultiplyConnected\] Let $\Omega \subset {\mathbf{R}}^2$ be a bounded smooth domain and $v_n:\partial \Omega \to {\mathbf{R}}\in L^s(\partial \Omega)$ for some $s>1$ be such that $\int_{\partial \Omega_h} v_n = 0$ for every $h \in \{0, \dotsc, m\}$. Let $\gamma_h \in {\mathbf{R}}$ for $h \in \{1, \dotsc, m\}$ and let $\kappa >0$ be given. For ${\varepsilon}>0$ there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in $\Omega$ with outward boundary flux given by $v_n$ and circulations given by $\gamma_h$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}) \subset B(x_{\varepsilon}, C{\varepsilon})$ for some $x_{\varepsilon}\in \Omega$ and $C>0$ not depending on ${\varepsilon}$. Moreover, as ${\varepsilon}\to 0$, $$\int_\Omega \omega_{\varepsilon}\to \kappa,$$ and $$\mathcal{W}_*(x^{\varepsilon}) \to \sup_{x \in \Omega} \mathcal{W}_*(x).$$
The proof is almost identical to the one of Theorem \[thm:resu\], it relies on Theorem \[thm:K1m\] instead of Theorem \[thm:K1\].
One could similarly prove a counterpart of Theorem \[thmLocalMinimum\] for multiply connected domains.
### Rotating vortices in a discs
If $\Omega$ is invariant under rotation, one can consider the Euler equation in a reference frame rotating with angular velocity $\alpha$: $$\left\{
\begin{aligned}
\nabla \cdot \mathbf{v} &= 0, \\
\mathbf{v}_t + \mathbf{v}\cdot \nabla \mathbf{v}&=-\nabla p+2\alpha \mathbf{v}^\perp-\alpha^2x.
\end{aligned}
\right.$$ The vorticity of $\mathbf{v}$ with respect to an inertial frame is $\nabla \times \mathbf{v}+2\alpha$. The movement of singular vortices is governed by Kirchhoff’s law , where $\mathcal{W}$ is replaced by $\mathcal{W}_\alpha(x)=\mathcal{W}(x)+\sum_i \alpha \frac{{\lvertx\rvert}^2}{2}$.
The stream-function method to construct stationary solutions in a rotating reference frame can be adapted to this situation. If $-\Delta \psi=f(\psi)-2\alpha$, setting $\mathbf{v}=(\nabla \psi)^\perp$ and $p= F(\psi)-\frac{\alpha^2}{2}{\lvertx\rvert}^2-\frac{1}{2}{\lvert\nabla \psi\rvert}^2$ yields a solution[^6]. In particular, the solution is irrotational outside on the set where $\psi=0$.
\[thmRotating\] Let $\rho > 0$, $\kappa >0$ and $\alpha > 0$. If $\kappa < 2\pi \alpha \rho^2$ For ${\varepsilon}>0$ there exist smooth rotating solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in $B(0, \rho)$ with angular velocity $\alpha$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon})$ is contained in a disc of radius $O({\varepsilon})$ around a point rotating on the circle of radius $\sqrt{\rho^2-\frac{\kappa}{2\pi\alpha}}$. Moreover, as ${\varepsilon}\to 0$, $$\int_\Omega \omega_{\varepsilon}\to \kappa.$$
Take $$q(x)=-\alpha \frac{{\lvertx\rvert}^2}{2}.$$ and apply Theorem \[thm:K1\]. One checks that $$\mathbf{v}_{\varepsilon}(x, t)=(\nabla u_{\varepsilon})^\perp (R(\alpha t) x),$$ where $R(\alpha t)$ denote the rotation of $\alpha t$, satisfies Euler equation. Since $$\mathcal{W}_\alpha(x)=\frac{\kappa^2}{4\pi}\log \frac{\rho^2-{\lvertx\rvert}^2}{\rho}+\frac{\kappa \alpha}{2} {\lvertx\rvert}^2,$$ attains its maximum on the circle of radius $\sqrt{\rho^2-\frac{\kappa}{2\pi\alpha}}$, one has the desired concentration result.
When $\kappa > 2\pi \alpha \rho^2$, the minimizer concentrates around $0$; one recovers thus stationary solutions as in Theorem \[thm:resu\].
### Stationary pairs of vortices in bounded domains
Let $\Omega \subset {\mathbf{R}}^2$ be a bounded simply-connected smooth domain and $v_n:\partial \Omega \to {\mathbf{R}}\in L^s(\partial \Omega)$ for some $s>2$ be such that $\int_C v_n = 0$ over each connected component $C$ of $\partial \Omega$. Let $\kappa_+ >0$ and $\kappa_- < 0$ be given. For ${\varepsilon}>0$ there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in $\Omega$ with outward boundary flux given by $v_n$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}^\pm) \subset B(x_{\varepsilon}^\pm, C{\varepsilon})$ for some $x^{\pm}_{\varepsilon}\in \Omega$ and $C>0$ not depending on ${\varepsilon}$. Moreover, as ${\varepsilon}\to 0$, $$\int_\Omega \omega_{\varepsilon}^\pm \to \kappa^\pm$$ and $$\mathcal{W}(x_{\varepsilon}^+, x_{\varepsilon}^-) \to \sup_{x^+, x^- \in \Omega} \mathcal{W}(x^+, x^-).$$
This follows from Theorem \[thm:K3\] in the same lines as Theorem \[thm:resu\].
There is also a counterpart of Theorem \[thmLocalMinimum\] for vortex pairs, concerning the existence of solutions near local maxima of the Kirchoff–Routh function and a counterpart of Theorem \[thm:MultiplyConnected\] for domains which are not simply connected.
One can also address the question of rotating vortex pairs. Combining the ingredients of the proof of Theorem \[thmRotating\], one can prove the existence of rotating vortex pairs of strength $\kappa_+ > 0$ and $\kappa_- > 0$ that concentrates around two antipodal rotating points at distance $\rho _+$ and $\rho _-$ which maximize the function $$\frac{\alpha \kappa_+}{2} \rho _+^2 +\frac{\alpha \kappa_-}{2}\rho _-^2+\frac{\kappa_+^2}{4\pi}\log (1-\rho _+^{2})+\frac{\kappa_-^2}{4\pi}\log (1-\rho _-^{2})+\frac{\kappa_+\kappa_-}{2\pi} \log \frac{1+\rho _+\rho _-}{\rho _++\rho _-}.$$ In contrast with Theorem \[thmRotating\], the pair of vortices obtained is always a nontrivial pair of rotating vortices for any $\alpha \ne 0$, $\kappa_+ > 0$ and $\kappa_- < 0$.
Unbounded domains
-----------------
We now consider the application of the results of Section \[sectUnbounded\] to the desingularization of vortices in unbounded domains.
### Translating vortex pair in the plane
We first consider the construction of a pair of vortices in ${\mathbf{R}}^2$. First recall that pair of vortices translating at velocity $\mathbf{W}$ in a flow with vanishing velocity at infinity is, up to a Galilean change of variables a pair of stationary vortices in a flow with velocity at infinity $-\mathbf{W}$. The stream-function of the corresponding irrotational flow is $\psi_0(x)=\mathbf{W}^\perp \cdot x$. Therefore, the positions of two vortices of opposite intensities $\kappa$ and $-\kappa$ in the moving reference frame is a critical point of the Kirchhoff–Routh $\mathcal{W}$ defined by $$\frac{-\kappa^2}{2\pi} \log \frac{1}{{\lvertx-y\rvert}} + \mathbf{W}^\perp \cdot x.$$
Let $W \ge 0$ and $\kappa \ge 0$, for every ${\varepsilon}> 0$ there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in ${\mathbf{R}}^2$ symmetric with respect to the $x_2$ axis and such that $\lim_{x_1 \to \infty} \mathbf{v}_{\varepsilon}(x)=(0, W)$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}) \cap {\mathbf{R}}^2_+ \subset B(\Bar{x}, C{\varepsilon})$, where $\Bar{x}= (\frac{\kappa}{4\pi W}, 0)$.
The problem can be reduced to finding a solution in ${\mathbf{R}}^2_+$ with vanishing flux on the boundary. The corresponding Kirchhoff–Routh function is $$\mathcal{W}(x)=\frac{\kappa^2}{4\pi} (\log 2x_1)-\kappa Wx_1$$ This follows from the existence result of Theorem \[thmYang\], the asymptotics of Proposition \[propUnboundedUpper\] and Proposition \[prop:1maiUnbounded\].
### Stationary vortex in the half-plane with non-vanishing flux
The method just used extends to non-vanishing flux boundary conditions:
Let $v_n \in L^1 ({\mathbf{R}})\cap L^s_{\mathrm{loc}}({\mathbf{R}})$ for $s > 1$. If $\int_{-\infty}^0 v_n=-\int_0^\infty v_n>0$. For every $W > 0$ and $\kappa > 0$, if $\kappa/W$ is small enough and if $\epsilon > 0$ sufficiently small there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in $\Omega$ with outward boundary flux given by $v_n$ and $\lim_{x_1 \to \infty} \mathbf{v}_{\varepsilon}(x)=(0, W)$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}) \subset B(x_{\varepsilon}, C{\varepsilon})$ for some $x_{\varepsilon}\in \Omega$ and $C>0$ not depending on ${\varepsilon}$, and $\int_{{\mathbf{R}}^2_+} \omega_{\varepsilon}\to \kappa$.
Define $\psi_0$ by $$\left\{
\begin{aligned}
-\Delta \psi_0 & = 0 & & \text{in ${\mathbf{R}}^2_+$}, \\
\partial_2 \psi_0 &=v_n& & \text{on $\partial {\mathbf{R}}^2_+$}, \\
\psi_0(0, x_2) &\to 0 & &\text{as ${\lvertx_2\rvert} \to \infty$},\\
\frac{\psi_0(x)}{x_1} &\to -W & & \text{as $x_1 \to \infty$}.
\end{aligned}
\right.$$ One checks that by our assumptions, $$\psi_0(0) > 0.$$ In order to apply Theorem \[theoremExistenceLevels\], we need to find $\Hat{x} \in \Omega$ such that $$\label{eqTrenchStrict}
\kappa \psi_0 (\Hat{x})+\frac{\kappa^2}{4\pi} \log 2 \Hat{x}_1 > \frac{\kappa^2}{4\pi} \Bigl(\log \frac{\kappa}{2\pi W}-1\Bigr).$$ One takes $\Hat{x}=(\frac{\kappa}{4\pi W}, 0)$. If $\kappa/W$ is small enough, one has $$\kappa \psi_0 (\Hat{x}) > 0,$$ and one checks that $$\kappa \psi_0 (\Hat{x})+\frac{\kappa^2}{4 \pi} \log 2 \Hat{x}_1 > \frac{\kappa^2}{4\pi} \log \frac{\kappa}{2\pi W} > \frac{\kappa^2}{4\pi} \Bigl(\log \frac{\kappa}{2\pi W}-1\Bigr).$$ The conclusion follows then from Theorem \[theoremExistenceLevels\].
### Stationary vortex in a perturbed half-plane
Instead of perturbing the boundary condition on the half-plane, one can instead perturb the geometry. The first situation is the situation in which one has for example enlarged a little bit the half-plane around $0$:
Assume that $\Omega$ is a simply-connected perturbation of ${\mathbf{R}}^2_+$ in the sense of . Let $\Bar{x} \in \partial \Omega$ be such that $x_1 > \Bar{x}_1$ for every $x \in \Omega$, $\partial \Omega$ is of class $C^2$ in a neighborhood of $\Bar{x}$, then for every $W > 0$, if $\kappa > 0$ is sufficiently small and if $\epsilon > 0$ sufficiently small there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in $\Omega$ with vanishing boundary flux and $\lim_{x_1 \to \infty} \mathbf{v}_{\varepsilon}(x)=(0, W)$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}) \subset B(x_{\varepsilon}, C{\varepsilon})$ for some $x_{\varepsilon}\in \Omega$ and $C>0$ not depending on ${\varepsilon}$ and $\int_{\Omega} \omega_{\varepsilon}\to \kappa$.
We are going to obtain the solutions by applying Theorem \[theoremExistenceLevels\] with $q=-\psi_0$. Let $\mathbf{v}_0$ be the irrotationnal stationary solution to the Euler equation with vanishing flux on $\partial \Omega$ and $\lim_{x_1 \to \infty} \mathbf{v}_0(x)=(0, W)$, i.e. $\mathbf{v}_0=\nabla \psi_0^\perp$ with $$\label{eqPsi0}
\left\{
\begin{aligned}
-\Delta \psi_0 &= 0, & &\text{in $\Omega$}, \\
\psi_0 &= 0 & &\text{on $\partial \Omega$}, \\
\tfrac{\psi_0(x)}{x_1} &\to -W &&\text{as $x \to \infty$}.
\end{aligned}
\right.$$ In order to apply Theorem \[theoremExistenceLevels\], we need to find $\Hat{x} \in \Omega$ such that the condition holds. First, by the strong maximum principle, one has $\partial_1 \psi(\Bar{x})>-W$, so that there exists $\gamma \in (0, W)$ such that in a neighborhood of $\Bar{x}$, $$\psi_0(x) > -\gamma (x_1-\Bar{x_1}).$$ If we consider the point $\Hat{x}=(\Bar{x}_1+\frac{\kappa}{4\pi W}, \Bar{x}_2)$, one has $$\kappa \psi_0(\Hat{x}) > -\gamma \frac{\kappa^2}{4\pi W}.$$ On the other hand, if $K$ denotes the curvature of $\partial \Omega$ at $\Bar{x}$, one has by Proposition \[propositionAsymptotH\], $$\frac{\kappa^2}{2} H (\Hat{x}, \Hat{x})=-\frac{\kappa^2}{4\pi} \log \frac{\kappa}{2\pi W}+O(\kappa^3).$$ Therefore, if $\kappa$ is small enough, one has , and one can then apply Theorem \[theoremExistenceLevels\] to obtain the conclusion.
### Translating vortex pair near a translating axisymmetric obstacle
We can also treat a situation in some sense opposite to the situation of the previous section. We obtain the desingularization of vortices on a set which is obtained by removing some part of the half-plane. By a Galilean change of variables and by extension by symmetry of the flow, this corresponds also physically to a rigid body in translation together with a pair of vortices. A similar problem was studied through the vorticity method by B.Turkington [@Turkington1983]
Let $D \subset {\mathbf{R}}^2$ be a compact simply-connected set with non-empty interior and symmetric with respect to the $x_1$ variable. Then for every $\kappa> 0$ and $W > 0$, if ${\varepsilon}> 0$ is sufficiently small there exist smooth stationary solutions $\mathbf{v}_{\varepsilon}$ of the Euler equation in ${\mathbf{R}}^2 \setminus D$ symmetric with respect to the $x_2$ axis, with vanishing boundary flux and such that $\lim_{x_1 \to \infty} \mathbf{v}_{\varepsilon}(x)=(0, W)$, corresponding to vorticities $\omega_{\varepsilon}$, such that ${\rm supp}(\omega_{\varepsilon}) \cap {\mathbf{R}}^2_+ \subset B(x_{\varepsilon}, C{\varepsilon})$ for some $x_{\varepsilon}\in \Omega \cap {\mathbf{R}}^2_+$ and $C>0$ not depending on ${\varepsilon}$ and $\int_{{\mathbf{R}}^2_+ \setminus D} \omega_{\varepsilon}\to \kappa$.
Set $\Omega={\mathbf{R}}^2_+ \setminus D$. We shall consider the case $W > 0$ and $\kappa > 0$, and we shall assume that $B(0, \rho) \subset D \subset B(0, R)$. We use again Theorem \[theoremExistenceLevels\] and therefore we shall prove that holds for some $\Hat{x} \in {\mathbf{R}}^2$ where $\psi_0$ solves . We shall take $\Hat{x}^\lambda=(\frac{\kappa}{4\pi W}, \lambda \frac{\kappa}{4\pi W})$ where $\lambda \in {\mathbf{R}}$. By the maximum principle on $\Omega$, one has for $x \in \Omega$, $$\psi_0(x)>-Wx_1+W \frac{x_1 \rho^2}{{\lvertx\rvert}^2}.$$ Hence, we have $$\kappa \psi_0(\Hat{x}^\lambda) \ge - \frac{\kappa^2 }{4\pi} + \frac{4\pi W^2}{1+\lambda^2},$$ with $c' > 0$. We also use the formula of the Green function $\Tilde{G}$ of ${\mathbf{R}}^2_+ \setminus B(0, R)$ used by B.Turkington [@Turkington1983 p.1047] $$\Tilde{G}(x, y)=\frac{1}{4\pi} \log \frac{1+\dfrac{4x_1y_1}{{\lvertx-y\rvert}^2}}{1+\dfrac{4R^2 x_1y_1 } {(x_1y_1+x_2 y_2-R^2)^2+(x_2y_1-x_1y_2)^2}}$$ Since $\Tilde{G}(x, y) \le G(x, y)$, one has therefore $$H(x,x)\ge \frac{1}{2\pi} \log 2 x_1 - \frac{1}{2\pi} \log \Bigl(1+ \frac{4R^2x_1^2}{({\lvertx\rvert}^2-R^2)^2}\Bigr),$$ whence $$\frac{\kappa^2}{2} H(\Hat{x}^\lambda, \Hat{x}^\lambda) \ge \frac{\kappa^2}{4\pi} \log \frac{\kappa}{2\pi W}+O(\lambda^{-4}).$$ One checks thus that for $\lambda$ sufficiently large, $$\kappa \psi_0 (\Hat{x}^\lambda)+\frac{\kappa^2}{4\pi} \log 2 \Hat{x}^\lambda_1 > \frac{\kappa^2}{4\pi} \Bigl(\log \frac{\kappa}{2\pi W}-1\Bigr).$$ and the conclusion thus follows from Theorem \[theoremExistenceLevels\].
Capacity estimates
==================
Let $\Omega \subset {\mathbf{R}}^2$ be open. The electrostatic capacity of a compact set $K \subset \Omega$ is $$\operatorname{cap}(K, \Omega)=\inf \Bigl\{ \int_{\Omega} {\lvert\nabla \varphi\rvert}^2 {\: :\:}\varphi \in C^\infty_c(\Omega)\text{ and } \varphi = 1 \text{ on $K$} \Bigr\}.$$ Let us first recall the following standard capacity estimate which was discovered by H.Poincaré [@Poincare p.17–22] and whose first complete proof was given by G.Szegő [@Szego1930].
\[propositionCapacityArea\] Let $\Omega \subset {\mathbf{R}}^2$ have finite measure. For every $K \subset \Omega$, $$\frac{4\pi}{\operatorname{cap}(K, \Omega)} \le \log \frac{{\mathcal{L}^{2}}(\Omega)}{{\mathcal{L}^{2}}(K)}.$$
One shows by the Pólya–Szegő inequality (for a modern treatment, see e.g. [@Kawohl1985], [@LiebLoss2001] or [@BrockSolynin2000]) that $$\operatorname{cap}(K, \Omega) \ge \operatorname{cap}(\overline{B(0, \rho)}, B(0,R)$$ if $\rho$ and $R$ are chosen so that ${\mathcal{L}^{2}}(B(0,\rho))={\mathcal{L}^{2}}(K)$ and ${\mathcal{L}^{2}}(B(0,R))={\mathcal{L}^{2}}(\Omega)$. One can then compute explicitly the right-hand-side to reach the conclusion.
When $\mathcal{L}^2(\Omega)=+\infty$, Proposition \[propositionCapacityArea\] loses its interest. However, one still has:
\[propositionCapacityLocalArea\] Let $K \subset {\mathbf{R}}^2_+$, we have $$\frac{4\pi}{\operatorname{cap}(K, \Omega)} \le \log \frac{8\pi \sup_{x \in K} {\lvertx\rvert}^2}{{\mathcal{L}^{2}}(K)}.$$
Set $a=\sup_{x \in K} {\lvertx\rvert}^2=1$ and define the conformal transformation $$\psi(z)= \frac{z-a}{z+a}.$$ We have $\psi({\mathbf{R}}^2_+)=B(0, 1)$. By the previous Lemma, we have $$\frac{4\pi}{\operatorname{cap}(\psi(K), B(0, 1))} \le \log \frac{2\pi}{{\mathcal{L}^{2}}(\psi(K))}.$$ The conclusion comes from $${\mathcal{L}^{2}}(\psi(K))=\int_K {\lvert\psi'\rvert}^2\ge \frac{{\mathcal{L}^{2}}(K)}{4a^2}. \qedhere$$
Another question about estimates of the capacity is whether one can estimate the diameter of $K$, instead of its area, by its capacity. This is possible if one assumes moreover that $K$ is connected. L.E.Fraenkel [@Fraenkel1981] has obtained in this direction the inequality $$\frac{2\pi}{\operatorname{cap}(K, \Omega)}\le \log C \frac{\operatorname{diam}K}{\sqrt{\mathcal{L}^2(\Omega)}}.$$ We improve this estimate so that it holds on unbounded sets and it takes into account the distance from the boundary.
\[propositionBoundDiameter\] Let $\Omega$ be such that ${\mathbf{R}}^2\setminus \Omega$ is connected and contains a ball of radius $\rho$ and $K \subset \Omega$ be compact. Then, $$\frac{2\pi}{\operatorname{cap}(K, \Omega)}\le \log 16\Bigl(1+ \frac{\operatorname{dist}(K, \partial \Omega)}{2\rho}\Bigr)\Bigl(1+ \frac{2\operatorname{dist}(K, \partial \Omega)}{\operatorname{diam}(K)}\Bigr).$$
Since $K$ is compact, up to translations and rotations we can assume that $0 \in K$ and $\operatorname{dist}(K, \partial \Omega)=\operatorname{dist}(0, \partial \Omega)$. Let $A^*$ and $\Omega^*$ be the sets obtained by circular symmetrization around $0$ introduced by V.Wolontis [@Wolontis1952 III.1] (see also J.Sarvas [@Sarvas1972]). We have $$\begin{gathered}
\operatorname{cap}(A^*, \Omega^*) \le \operatorname{cap}(A, \Omega),\\
[-\operatorname{diam}(A)/2, 0] \subset A^*,\end{gathered}$$ and, since ${\mathbf{R}}^2 \setminus \Omega^*$ contains a ball of radius $\rho$, $$[\operatorname{dist}(A, \partial \Omega), \operatorname{dist}(A, \partial \Omega)+2\rho] \subset {\mathbf{R}}^2\setminus \Omega^*.$$ We have thus $$\operatorname{cap}(A, \Omega) \ge \operatorname{cap}([-\operatorname{diam}(A)/2, 0], {\mathbf{R}}^2 \setminus [\operatorname{dist}(A, \partial \Omega), \operatorname{dist}(A, \partial \Omega)+2\rho].$$ Now, identifying ${\mathbf{R}}^2$ with ${\mathbf{C}}$, there exists a Möbius transformations that brings the points $-\operatorname{diam}(A)/2$, $0$, $\operatorname{dist}(A, \partial \Omega)$ and $\operatorname{dist}(A, \partial \Omega)+2\rho$ to $-1$, $0$, $s$ and $\infty$ with $$s=\frac{(2\rho+\operatorname{dist}(K, \partial \Omega)+\frac{1}{2}\operatorname{diam}(K))\operatorname{dist}(K, \partial \Omega)}{\rho \operatorname{diam}(K)},$$ from which we deduce that $$\operatorname{cap}(A, \Omega) \ge \operatorname{cap}([-1, 0], {\mathbf{C}}\setminus [s, +\infty[).$$ The conclusion comes from the next lemma.
As in L.E.Fraenkel’s proof [@Fraenkel1981], we use
Let $s>0$. We have $$\frac{2\pi}{\operatorname{cap}([-1, 0], {\mathbf{R}}^2 \setminus [s, \infty))}\le \log 16(1+s).$$
We have the formula [@Vuorinen1988 5.60 (1)] $$\operatorname{cap}([-1, 0], {\mathbf{R}}^2 \setminus [s, \infty))=2 \frac {\mathcal{K}(\sqrt{1/(1+s)})}{\mathcal{K}(\sqrt{s/(1+s)})},$$ where $\mathcal{K}$ is the complete elliptic integral of the first kind $$\mathcal{K}(\gamma)=\int_0^{\frac{\pi}{2}} \frac{1}{\sqrt{1-\gamma^2 (\sin \theta)^2}}\,d\theta.$$ Since (see [@AndersonVamanamurthyVuorinen1997]) $$\frac{\mathcal{K}(\gamma)}{\mathcal{K}(\sqrt{1-\gamma^2})} > \frac{\pi}{2\log\Bigl(2
\dfrac{1+\sqrt{1-\gamma^2}}{\gamma}\Bigr) }$$ We have then $$\operatorname{cap}([-1, 0], {\mathbf{R}}^2 \setminus [s, \infty))> \frac{\pi}{\log 2(\sqrt{s}+\sqrt{1+s})}> \frac{\pi}{\log 4\sqrt{1+s}}=\frac{2\pi}{\log 16(1+s)}.\qedhere$$
We also have an estimate in the case where the inner radius $\rho$ of ${\mathbf{R}}^2 \setminus \Omega$ is replaced by the connectedness and the measure of ${\mathbf{R}}^2 \setminus \Omega$.
\[propositionCapacityBoundDistance\] Let $\Omega$ be such that ${\mathbf{R}}^2\setminus \Omega$ is connected and has finite measure and $K \subset \Omega$ be compact. We have $$\frac{2\pi}{\operatorname{cap}(K, \Omega)}\le \log 16\Bigl(1+ \frac{\pi \operatorname{dist}(K, \partial \Omega)\operatorname{diam}({\mathbf{R}}^2 \setminus \Omega)}{2{\mathcal{L}^{2}}({\mathbf{R}}^2 \setminus \Omega)}\Bigr)\Bigl(1+ \frac{2\operatorname{dist}(K, \partial \Omega)}{\operatorname{diam}(K)}\Bigr)$$
One begins as in the proof of the previous proposition. We have then that $$[\operatorname{dist}(K, \partial \Omega), \operatorname{dist}(K, \partial \Omega)+\frac{2{\mathcal{L}^{2}}({\mathbf{R}}^2 \setminus \Omega)}{\pi \operatorname{diam}({\mathbf{R}}^2 \setminus \Omega)}] \subset {\mathbf{R}}^2\setminus \Omega^*.$$ And one continues as previously.
Green function asymptotics
==========================
This appendix is devoted to the study of the asymptotic expansion of Green’s function near a point of the boundary:
\[propositionAsymptotH\] Let $\Omega \subset {\mathbf{R}}^2$ and assume that $\Omega$ is of class $C^2$ around $0$ and that the tangent to $\partial \Omega$ is perpendicular to $x_1$. One has then as ${\varepsilon}\to 0$, $$G ({\varepsilon}x, {\varepsilon}y)=\frac{1}{4\pi} \log \frac{{\lvertx-y\rvert}^2+4x_1y_1}{{\lvertx-y\rvert}^2}-
{\varepsilon}\frac{K}{2\pi} \frac{x_1 {\lverty\rvert}^2+y_1 {\lvertx\rvert}^2}{{\lvertx-y\rvert}^2+4x_1y_1}+o({\varepsilon}).$$ uniformly on compact subsets of ${\mathbf{R}}^2_+ \times {\mathbf{R}}^2_+$, where $K$ is the curvature of $\partial \Omega$ at $0$. In particular, $$H ({\varepsilon}x, {\varepsilon}x)=\frac{1}{2\pi} \log 2{\varepsilon}x_1-{\varepsilon}\frac{K{\lvertx\rvert}^2}{4\pi x_1} +o({\varepsilon}).$$
Define $$w_{{\varepsilon}, y}(x) = \frac{1}{{\varepsilon}} \Bigl(\frac{1}{4\pi} \log \frac{{\lvertx-y\rvert}^2+4x_1y_1}{{\lvertx-y\rvert}^2}-G({\varepsilon}x, {\varepsilon}y)\Bigr).$$ This function is defined for every $x, y \in \Omega^{\varepsilon}=\{ z \in {\mathbf{R}}^2 {\: :\:}{\varepsilon}z \in \Omega \}$. Moreover, $w_{{\varepsilon}, y}$ satisfies $$\left\{\begin{aligned}
-\Delta w_{{\varepsilon}, y} &= 0 &&\text{in $\Omega$},\\
w_{{\varepsilon}, y} &= \frac{1}{4\pi{\varepsilon}} \log \frac{{\lvertx-y\rvert}^2+4x_1y_1}{{\lvertx-y\rvert}^2} && \text{on $\partial \Omega$}.
\end{aligned}\right.$$
By construction, $w_{{\varepsilon}, y}$ is a bounded function. We first claim that $w_{{\varepsilon}, y}$ is bounded uniformly in $L^\infty(\Omega_{\varepsilon})$ as ${\varepsilon}\to 0$ and $y$ stays in a compact subset of ${\mathbf{R}}^2$. Indeed, since $\Omega$ is $C^2$ around $0$, there exists $r > 0$ such that if $z \in \partial \Omega \cap B(0, r)$, ${\lvertz_1\rvert} \le C {\lvertz_2\rvert}^2$. One has thus, for $x \in \partial \Omega_{\varepsilon}\cap B(0, \frac{r}{{\varepsilon}})$, ${\lvertx_1\rvert} \le C {\varepsilon}{\lvertx_2\rvert}^2$, and therefore, when ${\varepsilon}$ is small enough $${\lvertw_{{\varepsilon}, y}(x)\rvert} \le \frac{C'}{{\varepsilon}} \frac{{\varepsilon}y_1 {\lvertx_2\rvert}^2}{{\lvertx-y\rvert}^2}$$ On the other hand, if $x \in \partial \Omega_{\varepsilon}\setminus B(0, \frac{r}{{\varepsilon}})$, then if ${\varepsilon}$ is small enough, $x \in \partial \Omega_{\varepsilon}\cap B(0, \frac{r}{2{\varepsilon}})$ so that $x_1 \le 2 {\lvertx-y\rvert}$ and ${\lvertx-y\rvert} \ge \frac{r}{2{\varepsilon}}$, and $${\lvertw_{{\varepsilon}, y}\rvert}(x) \le C {\varepsilon}.$$
Since, $\Omega$ is of class $C^2$, there exists a function $f : I \subset {\mathbf{R}}\to {\mathbf{R}}$ such that $\partial \Omega \cap B(0, r')= \{(f(t), t) \in \Omega {\: :\:}t \in I \}$. One has thus, using the Taylor expansion of $f$ and recalling that $f(0)=0$ and $f'(0)=0$, $$w_{{\varepsilon}, y}(x)=\frac{1}{4\pi {\varepsilon}} \log \Bigl( 1+ \frac{4 y_1 {\varepsilon}^{-1}f({\varepsilon}x_2)}{( {\varepsilon}^{-1}f({\varepsilon}x_2)-y_1)^2+(x_2-y_2)^2}\Bigr).$$
Therefore, by classical regularity estimates, $w_{{\varepsilon}, y}$ converges uniformly with respect to compact subsets of ${\mathbf{R}}^2_+ \times {\mathbf{R}}^2_+$ to the unique bounded solution of $$\left\{
\begin{aligned}
-\Delta w_y &= 0 &&\text{in ${\mathbf{R}}^2_+$}, \\
w_y&=\frac{f''(0)}{2\pi}\frac{y_1 x_2^2}{y_1^2+(x_2-y_2)^2}&&\text{on $\partial {\mathbf{R}}^2_+$}.
\end{aligned}
\right.$$ One can check that $$w_y(x)=\frac{f''(0)}{2\pi}\frac{y_1 (x_1^2+x_2^2)+x_1(y_1^2+y_2^2)}{(x_1+y_1)^2+(x_2-y_2)^2}.$$ The announced expressions for $G({\varepsilon}x, {\varepsilon}y)$ and $H({\varepsilon}x, {\varepsilon}x)$ follow.
[^1]: One needs to give a meaning to the equation in this case, since the velocity field generated by a vortex point is singular precisely on that vortex point. It consists in considering that each vortex point is transported only by the velocity field created by the other vortex points (see e.g. S.Schochet for details and further discussion).
[^2]: The function $x \mapsto H(x, x)$ is called the *Robin function* of $\Omega$.
[^3]: Here and in the sequel, smooth means Lipschitz and is sufficient for our goals.
[^4]: Notice that for the proof of Theorem \[thm:resu\] we only require a harmonic function $q$ but the proofs of Theorems \[thmLocalMinimum\] and \[thmRotating\] require more general $q$.
[^5]: This solution can be found by minimizing the functional $u \mapsto \frac{1}{2}\int_{\Omega} {\lvert\nabla u\rvert}^2+u\vert_{\partial \Omega_k}$ over $\mathrm{H}^1_*(\Omega)$. (A similar problem appears in [@BBH Chapter I, (3)])
[^6]: With the same velocity field, choosing as pressure $p=F(\psi)-2\alpha \psi-\frac{1}{2}{\lvert\nabla \psi\rvert}^2$ would of course give a solution to the Euler equation in a Galilean frame.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We establish an objective scheme to determine the macroscopicity of quantum mechanical superposition tests, which is based on the Bayesian hypothesis falsification of macrorealistic modifications of quantum theory. The measure uses the raw data gathered in an experiment, taking into account all measurement uncertainties, and can be used to directly assess any conceivable quantum test. We determine the resulting macroscopicity for three recent tests of quantum physics: double-well interference of Bose-Einstein condensates, Leggett-Garg tests with atomic random walks, and entanglement generation and read-out of nanomechanical oscillators.'
author:
- Björn Schrinski
- Stefan Nimmrichter
- 'Benjamin A. Stickler'
- Klaus Hornberger
title: Macroscopicity of quantum mechanical superposition tests via hypothesis falsification
---
Introduction
============
Any experiment witnessing or exploiting quantum coherent phenomena may be viewed as a test of whether quantum theory is complete at a fundamental level. While quantum mechanics is supported by all empirical observations up to date, all these observations are equally compatible with a number of alternative theories restoring macroscopic realism and resolving the measurement problem [@Leggett2002; @Bassi2013].
In recent years, various experiments demonstrated quantum superpositions or entanglement with mechanical objects of increasingly high masses and particle number, involving ever larger spatial delocalizations and coherence times. They include setups as diverse as counter-propagating superconducting loop currents [@friedman2000; @vanDerWal2000], large path-separation atom interferometers [@peters1999; @dimopoulos2007], high-mass molecular near-field interferometers [@Gerlich2011; @eibenberger2013], trapped and freely-falling Bose-Einstein condensates [@Schmiedmayer2013; @Kovachy2015], de-localized states and Leggett-Garg tests in optical lattices [@Alberti2009; @Robens2015], entangled ion chains [@jurcevic2014; @islam2015], and nanomechanical oscillators [@Riedinger2018; @ockeloen2018; @Marinkovi2018]. While all these experiments establish variants of a Schr[ö]{}dinger-cat-like state, an obvious question is the degree of macroscopicity (or ‘cattiness’) reached.
There are many ways to assess the macroscopicity of a Schr[ö]{}dinger cat realized in a quantum experiment [@frowis2018]. Most measures quantify the complexity of the quantum state based on information- or resource-theoretic concepts [@Korsbakken2007; @Marquardt2008; @Froewis2012; @yadin2016general; @Yadin2018], or introduce suitable distance measures in Hilbert space [@Bjoerk2004; @Lee2011]. While such abstract state vector ranking schemes may be used to compare experimental setups of similar kind, none can cover the entire variety of present-day superposition experiments [@friedman2000; @vanDerWal2000; @peters1999; @dimopoulos2007; @Gerlich2011; @eibenberger2013; @Schmiedmayer2013; @Kovachy2015; @Alberti2009; @Robens2015; @jurcevic2014; @islam2015; @Riedinger2018; @ockeloen2018; @Marinkovi2018].
A viable alternative is to regard a Schr[ö]{}dinger cat as more macroscopic than others if its demonstration is more at odds with the classical expectations shaped by our every-day experiences. In Ref. [@Nimmrichter2013] this was cast into a macroscopicity measure by quantifying the extent to which a superposition experiment rules out a wide and natural class of objective modifications of quantum theory that predict classical behavior on the macroscale, so-called *classicalizing modifications*. Recent tests of nonlocality and macrorealism, demonstrating the violation of Bell and Leggett-Garg inequalities at unprecedented mass and time scales, call for a generalization of this measure for arbitrary quantum tests.
![Scheme to compare the macroscopicity of two different quantum superposition tests: The experiments deliver raw data sets $D_1$ and $D_2$, which may be of arbitrary type and structure. They can be used to rule out modifications of standard quantum theory which classicalize the dynamics. Combining the data with the theoretical expectation yields a probability distribution for the classicalization timescale $\tau_e$, given the modification parameters $\sigma$ and the background information $I$. A quantum experiment is considered more macroscopic if the data rule out greater values of $\tau_e$, as inferred from the $5\,\%$ quantile $\tau_{\rm m}$.[]{data-label="fig:scheme"}](fig1.pdf){width="39.00000%"}
In this article, we present the most general framework for assigning the macroscopicity reached in quantum mechanical superposition experiments, based on non-informative Bayesian hypothesis testing, see Fig. \[fig:scheme\]. As the natural generalization of the measure presented in [@Nimmrichter2013], it relies only on the empirical evidence (i.e. the raw measurement outcomes) delivered by a given superposition test. It thus accounts for the measurement imperfections independently of the chosen experimental figure of merit.
This measure of macroscopicity can be applied to assess any mechanical superposition experiment. It is unbiased by construction and it accounts naturally for experimental uncertainties and statistical fluctuations. These advantages come at the expense of a certain theoretical effort required for calculating the macroscopicity of a given experiment. Specifically, the time evolution of the quantum system must be calculated in presence of classicalizing modifications to obtain the probability distribution for all possible measurement outcomes. In the second part of this article we demonstrate how this task is accomplished for three superposition tests at the cutting-edge of quantum physics: double-well interference of number-squeezed Bose-Einstein condensates (BECs) [@Schmiedmayer2013], Leggett-Garg inequality tests with atomic quantum random walks [@Robens2015], and generation and witnessing of entanglement between two spatially separated nanomechanical oscillators [@Riedinger2018].
{width="99.00000%"}
{width="99.00000%"}
{width="99.00000%"}
Macroscopicity of three recent superposition tests {#sec:neue2}
==================================================
Before presenting the formal framework of the proposed measure of macroscopicity, we illustrate its application to three recent superposition tests [@Schmiedmayer2013; @Robens2015; @Riedinger2018]. As a common theme, these experiments use derived quantities, such as visibilities, correlation functions, and entanglement witnesses, to certify the quantumness of their observations. One important advantage of the Bayesian approach advocated here is that it is independent of such data processing (and thus of secondary observables) and based exclusively on likelihoods associated with elementary measurement events. A theoretical derivation of the likelihoods required to assess the three mentioned experiments is presented in Secs. \[sec:3\]–\[sec:5\].
The measure uses the experimental data $D$ to determine the posterior probability distribution $p(\tau_e|D,\sigma, I)$ of classicalization timescales $\tau_e$, given the modification parameters $\sigma$, and any background information $I$ required to model the experiment. To ensure that each experiment is rated without bias, the least informative prior is used for Bayesian updating to yield the final posterior distribution. Figures \[fig:2\]–\[fig:4\] show how disparate experimental measurement protocols and data sets [@Schmiedmayer2013; @Robens2015; @Riedinger2018] yield comparable posterior distributions, narrowly peaked around a definite modification timescale. As an increasing number of data-points is included in the Bayesian updating procedure, the distributions shift to higher modification time scales, while their widths decrease. The lowest five percent quantile $\tau_{\rm m}(\sigma)$ of the posterior distribution determines the macroscopicity as $$\mu_{\rm m} = \max_\sigma \left [ \log_{10}\left(\frac{\tau_{\rm m}(\sigma)}{1\,{\rm s}}\right) \right ].$$ The value $\mu_{\rm m}$ thus quantifies the degree to which the quantum measurement data rules out a wide and natural class of classicalizing modifications of quantum theory.
The resulting macroscopicities of the experiments are: $\mu_{\rm m}=8.5$ for the BEC interferometer [@Schmiedmayer2013], $\mu_{\rm m}=7.1$ for the atomic Leggett-Garg test [@Robens2015], and $\mu_{\rm m}=7.8$ for the entangled nanobeams. That the BEC and the atomic random walk experiments exhibit comparable macroscopicities is due to the fact that they both witness single atom interference at a similar product of squared mass and coherence time. The macroscopicity associated with the entangled nanobeam experiment is roughly on the same order of magnitude on the logarithmic scale, despite the high mass and the large separation between the two beams and as well as coherence times of microseconds. This surprising result can be explained by the fact that the probed superposition state is delocalized merely by a few femtometers, and thus probes quantum theory only on sub-atomic scales.
Comparison of the three experiments also reveals that the convergence rate of the posterior distribution can vary strongly. In case of the Leggett-Garg test with an atomic quantum random walk [@Robens2015], the data set consists of 627 walks which all end in one of five final lattice sites. Since the likelihood of two of those outcomes is independent of the modification they include no information for the hypothesis test, which slows the convergence of the Bayesian updating procedure. In contrast, the double-well BEC-interferometer [@Schmiedmayer2013] provides a distribution of measurement outcomes over a practically continuous range of values, so that each experimental run yields a high degree of information gain, implying that 1457 measured population imbalances lead to a relatively narrow posterior distribution. In the case of nanobeams only two of four possible coincidence outcomes have different likelihoods, and thus several thousand repetitions of the measurement protocol are required to make the posterior converge.
Macroscopicity via hypothesis falsification {#sec:2}
===========================================
Empirical measure of macroscopicity {#sec:2A}
-----------------------------------
Classicalizing modifications of quantum theory propose an alternative (stochastic) evolution equation for the wavefunction. The observable consequences of these alternative theories are then encoded in the dynamics of the state operator $\rho_t$, which evolves according to a modified von Neumann equation $$\begin{aligned}
\label{eq:modvonneum}
\partial_t \rho_t=\mathcal{L}\rho_t+\frac{1}{\tau_e}\mathcal{M}_{\sigma}\rho_t.\end{aligned}$$ Here ${\cal L}\rho_t$ denotes the time evolution according to standard quantum theory (including possible decoherence) and ${\cal M}_\sigma \rho_t / \tau_e$ describes the effect of the proposed modification, characterized by the time scale $\tau_e$ and the set of modification parameters $\sigma$.
Indeed, a wide class of modification theories are compatible with all observations up to date, and they restore realism on the macroscale. This class can be parametrized by imposing a few natural consistency requirements, such as Galilean and scale invariance and exchange symmetry [@Nimmrichter2013]. The parameters $\sigma=(\sigma_q,\sigma_s)$ then specify the length- and momentum-scale on which the modification acts by means of the distribution function $g_\sigma(q,s)$ with zero mean and widths $\sigma_q,\sigma_s$, $$\begin{aligned}
\mathcal{M}_{\sigma}\rho_t=&\int d^3{\bf q}d^3{\bf s}\,g_\sigma(q,s)\left [
{\sf L}({\bf q},{\bf s})\rho_t{\sf L}^{\dagger}({\bf q},{\bf s}) \vphantom{\frac{1}{2}}\right. \nonumber \\
& \left. -\frac{1}{2}\left\{{\sf L}^{\dagger}({\bf q},{\bf s}){\sf L}({\bf q},{\bf s}),\rho_t
\right\}\right].
\label{eq:MIMpointparticle}\end{aligned}$$ The Lindblad operators in second quantization, $$\begin{aligned}
\label{eq:lindblad}
{\sf L}({\bf q},{\bf s})=\sum_\alpha\frac{m_\alpha}{m_e}\int d^3{\bf p}e^{i{\bf p}\cdot m_e{\bf s}/m_\alpha\hbar}{\sf c}_\alpha^{\dagger}({\bf p}){\sf c}_\alpha({\bf p}-{\bf q})\,,\end{aligned}$$ induce displacements in phase-space by means of the annihilation operator ${\sf c}_\alpha({\bf p})$ for momentum ${\bf p}$. They involve a sum over the different particle species $\alpha$ with mass $m_\alpha$, whose ratio over the electron mass $m_e$ effectively amplifies the strength of the modification for heavy particles, ensuring that macrorealism is restored [@Nimmrichter2013]. We take $g_\sigma$ to be Gaussian in the following. The modification then reduces to the model of continuous spontaneous localization (CSL) [@Bassi2013] for fixed $\sigma_q$ and $\sigma_s=0$. As explained in Ref. [@Nimmrichter2013], the bounds $\hbar/\sigma_q\gtrsim 10\,$fm and $\sigma_s\lesssim 20\,$pm ensure that the modification does not drive the system into the regime of relativistic quantum mechanics. In what follows, we will define the empirical measure of macroscopicity as the extent to which a quantum experiment rules out such classicalizing modifications.
Since the modified evolution predicts deviations from standard quantum mechanics at some scale these modification theories are empirically falsifiable. Thus, any quantum experiment gathering measurement data $D$ can be considered as testing the hypothesis $H_{\tau_e^*}$:
> *Given a classicalizing modification with parameters $\sigma$, the dynamics of the system state $\rho_t$ are determined by Eq. with a modification time scale $ \tau_e \leq \tau_e^*$.*
Note that greater values of $\tau_e$ imply weaker modifications.
The empirical data $D$ determine the Bayesian probability $P(H_{\tau_e^*} | D,\sigma, I)$ that $H_{\tau_e^*}$ is true, given the background information $I$. The latter includes all knowledge required for describing the experiment, such as the Hamiltonian, environmental decoherence processes, and the measurement protocol.
In order to compare $H_{\tau_e^*}$ with the complementary hypothesis $\overline{H}_{\tau_e^*}$, one defines the odds ratio [@von2014bayesian] $$\label{eq:oddsratio1}
o(\tau_e^* | D,\sigma, I) = \frac{P(H_{\tau_e^*} | D,\sigma, I)}{P(\overline{H}_{\tau_e^*} | D,\sigma, I)}.$$ If the data implies that the odds ratio is less than a certain maximally acceptable value $o_{\rm m}$ we can favor $\overline{H}_{\tau_e^*}$ over $H_{\tau_e^*}$. Modifications of quantum theory with $\tau_e \leq \tau_e^*$ are then ruled out by the data at odds $o_{\rm m}$ .
In order to evaluate the odds ratio we use Bayes’ theorem and exploit that for the hypothesis test to be unbiased by earlier experiments, $H_{\tau_e^*}$ and $\overline{H}_{\tau_e^*}$ must be [*a priori*]{} equally probable. Further using that the hypothesis $H_{\tau_e^*}$ implies $\tau_e \leq \tau_e^*$ yields $$\begin{aligned}
o(\tau_e^*|D,\sigma,I)=&\frac{\displaystyle \int_0^{\tau_e^*}d\tau_e\,P(D|\tau_e,\sigma,I)p(\tau_e|\sigma,I)}{\displaystyle \int_{\tau_e^*}^{\infty}d\tau_e\,P(D|\tau_e,\sigma,I)p(\tau_e|\sigma,I)},
\label{eq:oddsratio}\end{aligned}$$ where $p(\tau_e|\sigma,I)$ is the prior distribution of $\tau_e$, whose choice will be discussed in Sec. \[subsec:32\]. The probabilities $P(D |\tau_e,\sigma,I)$ are independent of the hypothesis $H_{\tau_e^*}$; they can be calculated for any experiment by solving the modified evolution equation with classicalization time scale $\tau_e$ and parameters $\sigma$.
The data $D$ is usually gathered in $N$ consecutive independent runs, $D = \{D_1,D_2,\ldots,D_{N}\}$, where $D_k$ denotes the set of (possibly correlated) measurement outcomes of round $k$. The likelihood for the entire data set $D$ is then given by $$\begin{aligned}
P(D|\tau_e,\sigma,I)=\prod_{k}P(D_k|\tau_e,\sigma,I).
\label{eq:likelihoodproduct}\end{aligned}$$ Every additional experimental run thus refines the posterior probability density, according to Bayes’ theorem $$\begin{aligned}
\label{eq:posterior}
p(\tau_e|D,\sigma,I) = \frac{P(D|\tau_e,\sigma,I)p(\tau_e|\sigma,I)}{P(D|\sigma,I)},\end{aligned}$$ where the normalization constant $P(D|\sigma,I)$ plays no role for the odds ratio. For sufficiently large data sets the posterior turns independent of the prior distribution $p(\tau_e|\sigma, I)$ under very general conditions [@schwartz1965bayes; @ghosh2003springer].
For what follows, we choose the threshold odds $o_{\rm m} = 1:19$, corresponding to the posterior probability $$\label{eq:taum}
P(\tau_e \leq \tau_{\rm m}|D,\sigma,I) \equiv \int_0^{\tau_{\rm m}} d \tau_e p(\tau_e \vert D,\sigma,I) = 5\,\%.$$ This determines the greatest excluded modification time scale $\tau_{\rm m}$ (at odds $o_{\rm m}$) so that for all $\tau_e^* \leq \tau_{\rm m}$ the odds ratio is smaller than $o_{\rm m}$ for given modification parameters $\sigma$.
Given the greatest excluded modification time scale $\tau_{\rm m}(\sigma)$ as a function of the modification parameters $\sigma$, one defines the empirical measure of macroscopicity as $$\mu_{\rm m} = \max_\sigma \left [ \log_{10}\left(\frac{\tau_{\rm m}(\sigma)}{1\,{\rm s}}\right) \right ],
\label{eq:Macroscopicity}$$ where $\tau_{\rm m}(\sigma)$ \[Eq. \] is the extent to which the measurement data $D$ of a given quantum experiment rules out the class of modifications . The value of $\mu_{\rm m}$ thus ranks superposition experiments against each other according to the degree to which they are at odds with our classical expectation.
We emphasize that this definition must only be used for experiments that undeniably show genuine quantum signatures. It cannot be used to *certify* whether a given experiment observes a superposition state. This is due to the fact that the absence of modification-induced heating and momentum diffusion can be observed also in classical experiments. Even though quantum coherence plays no role in such setups, they can serve to exclude combinations of classicalization timescales and modification parameters [@Laloe2014; @Nimmrichter2014; @carlesso2016experimental; @li2016discriminating; @goldwater2016testing; @vinante2017; @schrinski2017collapse; @adler2018bulk; @bahrami2018testing].
Even in genuine quantum superposition experiments the observed absence of modification-induced heating may dominate the range of excluded modification parameters. In this case it is necessary to recombine the observables in such a way that they separate into a subset of random variables $d$ providing information about quantum coherence and a subset $d_{\rm heat}$ yielding only information about the energy gain. (For example, in the case of the double-well BEC interference experiment, where one measures the particle numbers in the two different wells, their difference shows interference based on quantum coherence, while their sum constraints particle loss due to heating.) For a fair assessment of the macroscopicity, the likelihood $P(d,D_{\rm heat}|\tau_e,\sigma, I)$ must be conditioned on the realized data $D_{\rm heat}$ restricting modification induced heating, $$\label{eq:10cl}
P(d|\tau_e,\sigma, I, D_{\rm heat}) = \frac{P(d,D_{\rm heat}|\tau_e,\sigma, I)}{ P(D_{\rm heat}|\tau_e,\sigma, I)}\,$$ with $P(D_{\rm heat}|\tau_e,\sigma, I)=\sum_d P(d,D_{\rm heat}|\tau_e,\sigma, I)$. This way the witnessed lack of heating is effectively added to the background information $I$. (It also shows how to formally take into account the observation that the experiment could be executed at all, i.e. that the setup did not disintegrate due to modification-induced heating.) In Sect. \[sec:3\] we demonstrate how the conditioning on quantum observables works in practice by means of a nontrivial example.
Jeffreys’ prior {#subsec:32}
---------------
If the data set is not sufficiently large, the measure will in general depend on the prior distribution chosen to evaluate the odds ratio . It is therefore necessary to specify which prior distribution $p(\tau_e | \sigma,I)$ must be used to calculate the macroscopicity .
In order to ensure that the macroscopicity $\mu_{\rm m}$ does not have a bias towards a selected class of quantum superposition tests, the prior must be chosen in the most uninformative way, i.e. without including any [*a priori*]{} believes. For instance, this implies that it must not play a role whether we use the time scale $\tau_e$ or the rate $1/\tau_e$ to parametrize the class of modifications, which already excludes a uniform or piecewise-constant prior. Therefore, the natural choice is Jeffreys’ prior [@jeffreys1998theory]. Given the likelihood $P(d|\tau_e, \sigma,I)$ associated with a random variable $d$, it is defined as the square root of the Fisher information, $$\begin{aligned}
p(\tau_e|\sigma,I)=\sqrt{ \left \langle \left ( \frac{\partial}{\partial \tau_e}\log[P(d|\tau_e,\sigma,I)] \right )^2\right \rangle_d}\,.
\label{eq:JeffreysPrior}\end{aligned}$$ The ensemble average $\langle \cdot \rangle_d$ is performed over the entire range of possible measurement outcomes $d$ with Probability $P(d |\tau_e,\sigma,I)$.
This prior coincides with the so-called reference prior, so that it maximizes the Kullback-Leibler-divergence between prior and posterior and thus the average information gain in the Bayesian updating process [@bernardo1979reference; @ghosh2011objective]. In this sense, Jeffreys’ prior can be considered as the least informative prior [@berger2009formal]. In addition, it is invariant under re-parametrizations of the model [@jeffreys1998theory], implying that it is irrelevant whether we use the timescale $\tau_e$ or the rate $\lambda=1/\tau_e$ (as employed in the model of Continuous Spontaneous Localization [@Bassi2013]) or any other power of $\tau_e$ as the fundamental parameter of our model. We demonstrate in App.\[app:beweis\] that for all practical purposes Eq. yields a normalizable posterior distribution because the master equation and thus the likelihood $P(d|\tau_e,\sigma, I)$ are smooth functions of $\tau_e$.
If different measurement protocols are implemented, it is not obvious which likelihood should be used for calculating the prior since the Bayesian updating is independent of the order of measurement runs. A natural choice is to use the least favorable measurement protocol, i.e. the one yielding the lowest macroscopicity if no updating is performed. However, in experiments with sufficiently large data sets the posterior will in any case become independent of the prior distribution.
General scheme for assigning macroscopicities
---------------------------------------------
The formal framework of how to assess the macroscopicity of arbitrary quantum mechanical superposition tests is now complete:
1. Determine the Hamiltonian, environmental decoherence channels, and quantum measurement protocol, and use these to calculate the likelihood $P(d|\tau_e, \sigma,I)$ in presence of the modification . If appropriate use Eq. to focus on data demonstrating quantum coherence.
2. Calculate Jeffreys’ prior . In case that more than one quantum measurement protocol was experimentally implemented, use the one yielding the smallest prior macroscopicity.
3. Determine the posterior distribution via Bayesian updating to extract $\tau_{\rm m}(\sigma)$ via .
4. Find the maximum of the function $\tau_{\rm m}(\sigma)$, which determines the macroscopicity .
This recipe prescribes how to calculate the macroscopicity based on the empirical evidence of a quantum experiment. It formalizes and generalizes the notion of macroscopicity introduced in Ref. [@Nimmrichter2013]. The approximate expressions derived in Ref. [@Nimmrichter2013] intrinsically assume that imperfections of a given experiment yield a definite value of $\tau_e<\infty$, corresponding to a delta-peaked posterior distribution. The Bayesian framework put forward here extends this to measurement schemes and data sets yielding a finite posterior distribution $p(\tau_e|D,\sigma,I)$. It is thus the natural extension for noisy data and arbitrary measurement strategies.
In practice, the most complicated part of the above scheme is calculating the likelihoods in step 1. This requires finding an appropriate and quantitative description of the quantum dynamics in presence of decoherence and the modification. Note that the macroscopicity is underestimated if relevant decoherence channels are neglected in the calculation of the likelihoods. The remainder of this article demonstrates how the likelihoods can be calculated for the three superposition tests discussed in Sec. \[sec:neue2\].
Ramsey interferometry with a number-squeezed BEC {#sec:3}
================================================
Experimental Setting and Basics
-------------------------------
In the experiment reported in Ref. [@Schmiedmayer2013] a ${}^{87}$Rb BEC is trapped in a double-well potential and made to interfere, see Fig. \[fig:2\](a). The two involved modes $a,b$ form an effective two-level system described by the annihilation operators ${\sf c}_{a}$, ${\sf c}_{b}$. The state of the BEC can thus be represented by a collective pseudospin, defined by means of the (dimensionless) quasi angular momentum operators [@Arecchi1972] $$\begin{aligned}
{\sf J}_x&=\frac{1}{2}\left({\sf c}^{\dagger}_a{\sf c}_b+{\sf c}^{\dagger}_b{\sf c}_a\right)\nonumber\\
{\sf J}_y&=\frac{1}{2i}\left({\sf c}^{\dagger}_a{\sf c}_b-{\sf c}^{\dagger}_b{\sf c}_a\right)\nonumber\\
{\sf J}_z&=\frac{1}{2}\left({\sf c}^{\dagger}_a{\sf c}_a - {\sf c}^{\dagger}_b{\sf c}_b\right).
\label{eq:AngularMomentumOperators}\end{aligned}$$ They fulfill the angular momentum commutation relations $[{\sf J}_{\lambda},{\sf J}_{\mu}]=i\epsilon_{\lambda,\mu,\nu}{\sf J}_{\nu}$. The simultaneous eigenstates of ${\sf J}^2$ with eigenvalue $J(J+1)$ and ${\sf J}_z$ with eigenvalue $m$ is denoted by $\left|J,m\right\rangle$ (Dicke state), where $J=N/2$.
{width="99.00000%"}
The product of $N$ bosons being in a superposition state (coherent spin state; CSS) can be represented on a generalized Bloch sphere (see Fig. \[fig:blochspheres\]), whose polar angle $\theta$ indicates the relative population in $a$ and $b$, while the azimuth $\phi$ is the relative phase of the superposition state. Such a product state $\vert \theta, \phi \rangle$ can be expanded in terms of Dicke states as $$\begin{aligned}
\left|\theta,\phi\right\rangle \equiv &\frac{1}{\sqrt{(2J)!}}\left(\cos\left(\frac{\theta}{2}\right){\sf c}^{\dagger}_a+e^{i\phi}\sin\left(\frac{\theta}{2}\right){\sf c}^{\dagger}_b\right)^{2J} \vert {\rm vac} \rangle \nonumber\\
=&\sum_{m=-J}^J{2J \choose J+m}\cos\left(\frac{\theta}{2}\right)^{J-m}\sin\left(\frac{\theta}{2}\right)^{J+m}\nonumber\\
&\times e^{-i(J+m)\phi}\left|J,m\right\rangle.\end{aligned}$$ It has minimal and symmetric uncertainties, e.g. $\Delta {\sf J}_z^2=\Delta{\sf J}_y^2=|\left\langle{\sf J}_x\right\rangle/2|=J/2$ for $\theta = \pi/2$ and $\phi = 0$.
Applying a nonlinear squeezing operator turns the CSS into a squeezed spin state (SSS) [@Kitagawa1993; @Ma2011], which can be useful for metrology [@toth2012; @toth2014quantum; @hosten2016quantum] or robust against dephasing processes [@Javanainen1997PhaseDispersion; @Schmiedmayer2013]. In addition, it has been demonstrated that the [*depth of entanglement*]{} increases with squeezing [@sorensen2001many; @Sorensen2001; @Toth2014], as quantified by the squeezing parameter $\xi^2 = 2(\Delta {\sf J}_{\rm min})^2 / J$. We note that according to the information-theoretic measure from Ref. [@Froewis2012] already the existence of such a state yields a large macroscopicity since squeezing increases the quantum Fisher information.
In terms of the depth of entanglement [@Sorensen2001; @Toth2014] the non-classicality of SSS lies between a product state (CSS) and the maximally entangled NOON-state $\left|\psi\right\rangle\propto\left|N,0\right\rangle+\left|0,N\right\rangle$, a superposition of all particles being either in mode $a$ or mode $b$. Applying the modification on this NOON state yields a decoherence rate proportional to $N^2$, while that of a product state is proportional to $N$. It is thus easy to see that a NOON-state with stable phase could serve to exclude a large range of classicalization time scales [@bilardello2017collapse], but they have not been generated experimentally thus far. In contrast, the modification-induced dynamics of SSS, which are frequently realized in experiments, is much more intricate, as discussed in the following.
The free time evolution ${\cal L}\rho = -i [{\sf H},\rho]/\hbar$ of the BEC is characterized by the energy difference $\epsilon$ between the two modes and by the interaction between the particles. Approximating the latter to leading order in ${\sf J}_z$, yields the Hamiltonian [@Javanainen1997PhaseDispersion] $${\sf H}=\epsilon {\sf J}_z +\hbar\zeta{\sf J}^2_z,
\label{eq:FreeDephasingHamiltonian}$$ where $\zeta=d\tilde{\mu}/d (\hbar m)|_{m=0}$ is the change of chemical potential with the occupation difference $m$. Thus, the first term of the Hamiltonian describes rotations around the $z$-axis with angular frequency $\epsilon/\hbar$ on the Bloch sphere, while the second term leads to dispersion.
The experiment starts with the BEC in the state $|\theta=\pi/2,\phi=0\rangle$, which is then squeezed in $z$-direction and freely evolved for up to 20ms. Finally, a $\pi/2$-rotation around the $x$-axis converts the phase distribution into mode occupation differences, which are read-out by time-of-flight measurements, see Fig.\[fig:2\].
The likelihood required for the hypothesis test is the probability of observing a number difference of $m$ between the two modes, $$\begin{aligned}
P(m | \tau_e,\sigma,I)= \sum_{J = 0}^{\infty} \left\langle J, m\right|e^{-i\pi{\sf J}_x/2}\rho_t e^{i\pi{\sf J}_x/2}\left| J, m\right\rangle\,,
\label{eq:PiHalfPulse}\end{aligned}$$ where the sum over $J$ accounts for the possibility of modification-induced particle loss from the BEC during the experiment [@Laloe2014]. The modification parameters $\tau_e$ and $\sigma$ enter through the modified time evolution of the state $\rho_t$, which will be discussed next.
Double-well potential: phase flips
----------------------------------
Expanding the momentum annihilation operators ${\sf c}({\bf p})$ in Eq. in the single-particle eigenmodes in presence of the potential, and neglecting particle loss for the moment (${\mathsf c}_a^\dagger{\mathsf c}_a+{\mathsf c}_b^\dagger{\mathsf c}_b=2J$), yields $$\begin{aligned}
\mathcal{M}_\sigma\rho=&\frac{4 m_{\rm Rb}^2}{\tau_e m_e^2}\int d^3{\bf q}\,f_\sigma(q)\nonumber\\
&\times\left[
{\sf A}({\bf q})\rho{\sf A}^{\dagger}({\bf q})-\frac{1}{2}\{{\sf A}^{\dagger}({\bf q}){\sf A}({\bf q}),\rho\}
\right].\end{aligned}$$ Here, we used that spatial displacements are negligible on the length scale of the experiment and thus $f_\sigma(q) = \int ds g_\sigma(q,s)$ depends only on $\sigma_q$. The Lindblad operators are given by $$\begin{aligned}
{\sf A}({\bf q})=a_x({\bf q}){\sf J}_x+a_z({\bf q}){\sf J}_z\,,\end{aligned}$$ with $$\begin{aligned}
a_x({\bf q})=&\left\langle \psi_a\right|{\sf W}({\bf q})\left|\psi_b\right\rangle
\nonumber\\
a_z({\bf q})=&\left\langle \psi_a\right|{\sf W}({\bf q})\left|\psi_a\right\rangle\sin\left(\frac{\Delta_x q_x}{2\hbar}\right).
\label{eq:FlipCoefficients}\end{aligned}$$ Here, $\left|\psi_a\right\rangle$ and $\left|\psi_b\right\rangle$ are the single-atom eigenstates of the two level system with real wavefunctions $\psi_b({\mathbf r})=\psi_a({\mathbf r}-\Delta_x {\mathbf e}_x)\in \mathbb{R}$ and ${\sf W}({\bf q})=\exp(i {\bf q}\cdot \boldsymbol{\mathsf{r}}/\hbar)$ is the momentum transfer operator.
The first part of the Lindblad operator describes rotations around the $x$-axis, or spin-flips, while the second one induces rotations around the $z$-axis, or phase-flips. Such flip operators are frequently used to describe disturbance channels in collective spin states [@wang2010sudden; @Ma2011]. Since the spatial overlap between the two modes is negligible, $a_x({\bf q})\ll a_z({\bf q})$, the spin-flip contribution will be neglected in the following, implying that $\langle {\sf J}^2_z \rangle_t$ remains constant.
The expectation value of the perpendicular spin components decays as $\left\langle{\sf J}_y\right\rangle_t=e^{-\Gamma_{\rm P} t/2}\left\langle{\sf J}_y\right\rangle_{{\rm f}, t}$ with phase-flip rate (or dephasing rate) $$\begin{aligned}
\quad\Gamma_{\rm P}=\frac{4 m_{\rm Rb}^2}{\tau_e m_e^2}\int d^3{\bf q}\,f_\sigma(q)|a_z({\bf q})|^2.
\label{eq:SqueezingTEFirstMoments}\end{aligned}$$ Here $\left\langle{\sf J}_y\right\rangle_{{\rm f},t}$ denotes the free time evolution of the expectation value due to Eq. ; the same relation holds for $\langle {\sf J}_x \rangle_t$. Note that the phase-flip decay rate $\Gamma_{\rm P}$ is independent of the degree of squeezing.
The phase-flip operators induce diffusion in the azimuthal plane of the Bloch sphere. The second moment of ${\sf J}_y$ thus evolves as $$\begin{aligned}
\left\langle{\sf J}^2_y\right\rangle_t=\frac{1}{2} \left\langle{\sf J}^2_x + {\sf J}^2_y\right\rangle_{{\rm f},t}
-
\frac{e^{-2\Gamma_{\rm P} t}}{2} \left\langle{\sf J}^2_x - {\sf J}^2_y\right\rangle_{{\rm f},t},
\label{eq:SqueezingTESecondMoments}\end{aligned}$$ and similar for ${\sf J}^2_x$. For sufficiently large $N$ the squeezing loss rate is again independent of the initial squeezing since $\left\langle{\sf J}^2_x\right\rangle_{{\rm f},t}\approx J^2$ (as long as oversqueezing is avoided).
Equations and show that squeezing has no direct implications for the sensitivity on modification-induced decoherence. In contrast to what might be expected intuitively, an increased [*depth of entanglement*]{} does therefore not improve substantially the macroscopicity of experiments that measure only the first two moments of the collective spin observables.
Continuum approximation
-----------------------
In order to calculate the likelihood , we will utilize a continuum approximation on the tangent plane of the Bloch sphere, replacing the discrete probability $P(m|\tau_e,\sigma,I)$ by the continuous probability density $p(m|\tau_e,\sigma,I)$ for real $m$. For this sake, we use that the initial state is aligned with the $x$-axis, $\langle {\sf J}_x \rangle \approx J$, so that $$[{\sf J}_y, {\sf J}_z] \approx i J\,,$$ which is approximately constant (and not operator valued). Thus we locally replace the sphere by its flat tangent plane and may interpret ${\sf J}_y$ as a position and ${\sf J}_z$ as a momentum operator, see Fig.\[fig:blochspheres\]. The Wigner function of the initial state is then approximated by a Gaussian distribution, $$\begin{aligned}
\label{eq:ini}
w_0(j_y,j_z) = &\frac{1}{\sqrt{4\pi^2\left\langle{\sf J}^2_y\right\rangle_0\left\langle{\sf J}^2_z\right\rangle_0}}\nonumber\\
&\times \exp\left[
-\frac{1}{2}\frac{j_y^2}{\left\langle{\sf J}^2_y\right\rangle_0}-\frac{1}{2} \frac{j_z^2}{\left\langle{\sf J}^2_z\right\rangle_0}
\right],\end{aligned}$$ where $(j_y,j_z) \in \mathbb{R}^2$ are continuous variables in the flat tangent plane.
The time evolution of the initial state contains the free rotation and dispersion described by Eq. , as well as modification-induced dephasing. Representing the dynamics in quantum phase space, the quadratic term in the Hamiltonian induces shearing in $j_y$, while the linear term leads to a translation in $j_y$ with constant velocity. The phase flips induce diffusion in $j_y$, which increases the variance linearly with time. The corresponding time evolved state can thus be written as $$\begin{aligned}
\label{eq:wignert}
w_t(j_y,j_z) = &\frac{1}{\sqrt{4\pi^2\left(\left\langle{\sf J}^2_y\right\rangle_0+\Gamma_{\rm P} J^2 t\right)\left\langle{\sf J}^2_z\right\rangle_0}}\nonumber\\
&\times \exp\left[
-\frac{1}{2}\frac{(j_y-\epsilon t/\hbar - 2\zeta j_z t)^2}{\left\langle{\sf J}^2_y\right\rangle_0+\Gamma_{\rm P} J^2 t}-\frac{1}{2} \frac{j_z^2}{\left\langle{\sf J}^2_z\right\rangle_0}
\right],\end{aligned}$$ implying that the marginal distribution of $j_z$ remains unaffected by the dynamics.
In order to calculate the likelihood $P_J(m|\tau_e,\sigma,I)=\langle J, m | e^{-i\pi{\sf J}_x/2}\rho_t e^{i\pi{\sf J}_x/2}\left| J, m\right\rangle$ at fixed $J$, we first perform the $\pi/2$-rotation around the $x$-axis, which exchanges $j_y$ and $j_z$ in Eq. . The resulting distribution is then integrated over $j_y$, and $j_z$ is wrapped back onto the sphere by using $\sin(j_z) = m/J$ and the summation $\int dj_y\sum_k w_t(j_z+2\pi k,j_y)$. This way one obtains the continuous probability density approximating $P_J$,
$$\begin{aligned}
p_J(m|\tau_e,\sigma,I)=&\frac{\Theta(J^2-m^2)}{2\pi \sqrt{J^2-m^2}}\left[
\vartheta_3\left(\frac{\arcsin(m/J)-\epsilon t/\hbar}{2},g(t)\right) + \vartheta_3\left(\frac{\pi-\arcsin(m/J)-\epsilon t/\hbar}{2},g(t)\right)
\right],
\label{eq:PhaseFlipDistributionApprox}\end{aligned}$$
where $\Theta(x)$ is the Heaviside function, $\vartheta_3$ denotes the Jacobi-theta functions of the third kind $$\begin{aligned}
\vartheta_3(u,q)=\sum_{n=-\infty}^{\infty}q^{n^2}e^{2inu}\,,\end{aligned}$$ and the dependence on the initial state is expressed by $$\begin{aligned}
\label{eq:gdouble}
g(t)=\exp\left[-\frac{\left\langle{\sf J}^2_y\right\rangle_0}{2J^2}-\frac{\Gamma_{\rm P}t}{2}-2\zeta^2t^2\left\langle{\sf J}^2_z\right\rangle_0\right]\,.\end{aligned}$$
This analytic result captures the generic dephasing effect of random phase flips on a two-mode BEC. The comparison of Eq. with exact numerical calculations shows very good agreement, as demonstrated in Fig.\[fig:blochspheres\].
At this stage it might be tempting to use Eq. for Bayesian updating to calculate the macroscopicity. However, since the spatial distance between the two wells of the potential is not much greater than the extension of the modes, the resulting maximizing modification parameters $\sigma$ imply a moderate heating of the BEC. This must be taken into account for a consistent description. A brief discussion of the role of spin flips in single-well potentials will prepare this.
Single-well potentials: spin flips
----------------------------------
The dynamics of a BEC in the two lowest eigenstates of a single-well potential, as studied in Ref. [@Schmiedmayer2014], is strongly affected by spin flips. This marked difference to the double well is due to the spatial overlap between the two modes, see Eq. . The resulting Lindblad operators do not commute with ${\sf J}_z $, but induce additional diffusion in $z$-direction. In combination with the Hamiltonian this leads to an enhanced dispersion.
If the free rotation frequency $\epsilon/\hbar$ exceeds the spin-flip diffusion rate $$\begin{aligned}
\quad\Gamma_{\rm S}=\frac{4m_{\rm Rb}^2}{\tau_e m_e^2}\int d^3{\bf q}\,f_\sigma(q)|a_x({\bf q})|^2\,,\end{aligned}$$ the average gain in the second moment of $\sf J_z$ can be easily calculated. For times much greater than the rotation period one obtains $$\begin{aligned}
\left\langle{\sf J}^2_z\right\rangle_t\approx
\frac{\left\langle{\sf J}^2_z\right\rangle_0+J^2}{3}
+\frac{2\left\langle{\sf J}^2_z\right\rangle_0-J^2}{3}
e^{-3\Gamma_{\rm S} t/2}\,.
\label{eq:SpinFlipVarianceBroadening}\end{aligned}$$ For single wells, spin flips will typically dominate the influence of the modification, and phase flips can safely be neglected.
Expanding Eq. for small $\Gamma_{\rm S} t$ and exploiting that $J^2\gg\left\langle {\sf J}^2_z\right\rangle$, yields in the continuum approximation (see App. \[app:a\]) $$\begin{aligned}
\Delta j_y^2(t)\approx\Delta j_y^2(0)+4\zeta^2 J^2 t^2\left[\left\langle{\sf J}^2_z\right\rangle_0+\frac{\Gamma_{\rm S}J^2 t}{6}\right].
\label{eq:AngleVariance}\end{aligned}$$ Thus the random spin flips enhance dispersion so that the variance of $j_y$ increases with $t^3$. This results in the probability distribution with $$\begin{aligned}
g(t)=\exp\left[-\frac{\left\langle{\sf J}^2_y\right\rangle_0}{2J^2}-2\zeta^2t^2
\left(\left\langle{\sf J}^2_z\right\rangle_0
+\frac{\Gamma_{\rm S}J^2t}{6}
\right)\right].
\label{eq:SpinFlipGaussApprox}\end{aligned}$$
In single-well BEC interferometers the modification thus strongly influences the final occupation difference, rendering them attractive for future superposition tests. As explained next, diffusion in the orthogonal $z$-direction is also caused by modification-induced particle loss. The above results can be directly transferred.
Heating-induced particle loss
-----------------------------
In order to include modification-induced particle loss from the BEC, we assume that atoms leaving the two ground modes will never return. This assumption is well justified for a large modification parameter $\sigma_q$, where the particles have a negligible probability of being scattered back to the two lowest modes.
In this simplified scenario their populations decay exponentially, $$\begin{aligned}
\left\langle{\sf c}^{\dagger}_a{\sf c}_a\right\rangle_t
=e^{-\Gamma_a t}
\left\langle{\sf c}^{\dagger}_a{\sf c}_a\right\rangle_0,\quad
\left\langle{\sf c}^{\dagger}_b{\sf c}_b\right\rangle_t
=e^{-\Gamma_b t}
\left\langle{\sf c}^{\dagger}_b{\sf c}_b\right\rangle_0,
\label{eq:GSpopulation}\end{aligned}$$ with loss rates $$\begin{aligned}
\Gamma_{a,b}=&\frac{m_{\rm Rb}^2}{\tau_e m_e^2}\int d^3{\bf q}\,f_\sigma(q)\left[1-\left|\left\langle \psi_{a,b}\right|{\sf W}({\bf q})\left|\psi_{a,b}\right\rangle\right|^2\right].\end{aligned}$$ The radius of the generalized Bloch sphere thus decreases with time, and for $\Gamma_a \neq \Gamma_b$ the state is shifted towards one of the poles.
Also the coherences decay exponentially, $$\begin{aligned}
\left\langle{\sf c}^{\dagger}_a{\sf c}_b\right\rangle_t
=e^{-\Gamma_{\rm C} t}
\left\langle{\sf c}^{\dagger}_a{\sf c}_b\right\rangle_0,\quad
\langle{\sf c}^{\dagger}_b{\sf c}_a\rangle_t
=e^{-\Gamma_{\rm C} t}
\langle{\sf c}^{\dagger}_b{\sf c}_a\rangle_0\,,
\label{eq:Coherences}\end{aligned}$$ with $$\begin{aligned}
\Gamma_{\rm C}=\frac{m_{\rm Rb}^2}{\tau_e m_e^2}\int d^3 {\bf q}\, f_\sigma(q)
\left[
1-\left\langle \psi_a\right|{\sf W}({\bf q})\left|\psi_a\right\rangle
\left\langle \psi_b\right|{\sf W}^{\dagger}({\bf q})\left|\psi_b\right\rangle
\right]\,.
\label{eq:CoherenceDecay}\end{aligned}$$
In order to evaluate the effect of particle loss on the likelihood (\[eq:PhaseFlipDistributionApprox\]) we use the result of Ref. [@Ma2011] to determine how the variance of $\sf J_{\bf n}$, i.e. the angular momentum component in direction ${\bf n}$, changes due to particle loss. Using $J_0,J\gg 1$ one obtains $$\begin{aligned}
\label{eq:rescale}
\frac{\left\langle{\sf J}^2_{\bf n}\right\rangle_{J}}{J^2}\approx \frac{\left\langle{\sf J}^2_{\bf n}\right\rangle_{J_0}}{J_0^2}
+\frac{J_0-J}{2J_0J},\end{aligned}$$ where $J$ ($J_0$) is the current (initial) collective spin after the loss of $2(J_0-J)$ particles, and angular brackets $\left\langle \dots\right\rangle_{J}$ denote expectation values after tracing out the lost particles. The second term shows that the rescaled second moment $\left\langle{\sf J}^2_{\bf n}\right\rangle_{J}/{J^2}$ increases due to the particle loss.
Combining Eq. with Eq. (\[eq:GSpopulation\]), using that in the double-well $\Gamma_a=\Gamma_b \equiv \Gamma_{\rm L}$, expanding the result to linear order in $\Gamma_{\rm L}t$, and finally repeating the steps carried out in the previous section to account for simultaneous shearing and diffusion, yields the distribution (\[eq:PhaseFlipDistributionApprox\]) with $$\begin{aligned}
g(t)=&\exp\left[-\frac{\left\langle{\sf J}^2_y\right\rangle_0}{2J_0^2}
-\frac{\Gamma_{\rm P}t}{2}-\frac{\Gamma_{\rm L}t}{4J_0}
\right.\nonumber\\
&\left.-2\zeta^2t^2
\left(\left\langle{\sf J}^2_z\right\rangle_0
+\frac{\Gamma_{\rm L} J_0 t}{6}
\right)\right].
\label{eq:ContinuousDoubleWellGauss}\end{aligned}$$
The enhancement of the dispersion looks similar to the single-well case , but it is weaker by the (significant) factor $1/J_0$. Note that the dispersion rate $\zeta$ decreases with decreasing $J_0$, and the linear approximation of the chemical potential leading to the free Hamiltonian (\[eq:FreeDephasingHamiltonian\]) will fail if too many particles are lost.
The distribution of the remaining particles turns out to be binomial [@schrinski2017sensing] given that $\Gamma_a=\Gamma_b\equiv \Gamma_{\rm L}(\tau_e,\sigma,I)$. The probability density for $m \in \mathbb{R}$, i.e. the continuous approximation of Eq. , therefore takes the final form $$\begin{aligned}
\label{eq:37}
p(m|\tau_e,\sigma,I) = & \sum_{J=0}^{J_0}{{J_0}\choose{J}}\left(1-e^{-\Gamma_{\rm L}t}\right)^{J}\left (e^{-\Gamma_{\rm L}t}\right)^{J_0-J} \nonumber \\
& \times p_J(m|\tau_e,\sigma,I),\end{aligned}$$ where $p_J(m|\tau_e,\sigma,I)$ is given by Eqs. and and $p_0(m|\tau_e,\sigma,I) = \delta(m)$. This equation can now be used for the Bayesian updating procedure and for evaluating the macroscopicity .
Experimental parameters
-----------------------
The BEC reported in Ref. [@Schmiedmayer2013] consists of $N=2J_0\approx1200$ $^{87}$Rb atoms in a double-well configuration with a spatial separation of $\Delta_x\approx2\,\mu{\rm m}$ in $x$-direction and an initial number squeezing of $\Delta {\mathsf J}_z^2=0.41^2 J_0/2$. The trapping frequencies are $\omega_x/2\pi=1.44\,{\rm kHz}$, $\omega_y/2\pi=1.84\,{\rm kHz}$ and $\omega_z/2\pi=13.2\,{\rm Hz}$, so that the motion in $z$-direction is quasi-free. The two lowest energy levels of this potential have a gap of $\epsilon/\hbar=2.19\,{\rm kHz}$ and the first order corrections of the chemical potential are characterized by $\zeta=4\,$Hz.
Approximating the ground states harmonically with the widths $\sigma_{x,y}=\sqrt{\hbar/2m_{\rm Rb}\omega_{x,y}}$ yields the phase-flip and loss rates $$\begin{aligned}
\Gamma_{\rm P}= & \frac{m_{\rm Rb}^2}{\tau_e m_e^2} \frac{1-\exp[-\Delta_x^2\sigma_q^2/(4\sigma_q^2\sigma_x^2+2\hbar^2)]}{\sqrt{(1+2\sigma_q^2\sigma_x^2/\hbar^2)(1+2\sigma_q^2\sigma_y^2/\hbar^2)}} \\
\Gamma_{\rm L}= &\frac{m_{\rm Rb}^2}{\tau_e m_e^2}\left(1-\frac{1}{\sqrt{(1+2\sigma_q^2\sigma_x^2/\hbar^2)(1+2\sigma_q^2\sigma_y^2/\hbar^2)}}\right).
\label{eq:Gl}\end{aligned}$$
For the experimental parameters given above, the particle loss rate $\Gamma_{\rm L}$ cannot be neglected compared to the phase-flip rate $\Gamma_{\rm P}$ in the entire parameter regime of $\sigma$. This is due to the fact that the widths of the ground state modes $\sigma_{x,y}$ are comparable to the spatial separation of the wells $\Delta_x$. Consequently, it cannot be excluded that the observed lack of particle loss due to modification-induced heating may significantly affect the hypothesis test, even though confirming the conservation of particle number does not verify quantum coherence.
As a remedy, we condition the likelihood on the observed particle number, as explained at the end of Sect. \[sec:2A\]. This makes the overall atom number part of the experimental background information, and we can separately assess the modification-induced loss of interference visibility *given that* a certain particle number was detected. The conditioned likelihood is obtained by dividing the likelihood by the probability $$\begin{aligned}
P(D_{\rm heat}|\tau_e,\sigma, I) = & \sum_{J=\lfloor 0.9J_0\rfloor}^{J_0}{{J_0}\choose{J}}\left(e^{\Gamma_{\rm L}t}-1\right)^{J}e^{-J_0\Gamma_{\rm L}t}\,\end{aligned}$$ that not more than 10% of the particles are lost, $D_{\rm heat}:=\{J\geq 0.9J_0\}$. This threshold value is taken as a conservative estimate given that the number of the trapped particles fluctuates by at most 10% between the individual experimental runs.
All information is now available to perform the Bayesian hypothesis test, as described in Sect. \[sec:2\] using the 1438 data points presented in Fig. \[fig:2\](b). Numerical maximization of $\tau_{\rm m}(\sigma)$ yields a macroscopicity value of $\mu_{\rm m}=8.5$. The maximum of $\tau_{\rm m}(\sigma)$ is attained for the modification parameter $\sigma_q\simeq\hbar/0.77\,{\rm mm}$. As one would expect, this roughly corresponds to the parameter value where the phase-flip rate is maximized (at $\Gamma_{\rm P}=1.7/\tau_e$), implying that dephasing is most pronounced. The corresponding particle loss rate is an order of magnitude lower ($\Gamma_{\rm L}=0.11/\tau_e$).
The macroscopicity attained in the double-well BEC interferometer is comparable to the value expected for an atom interferometer operating single Rubidium atoms on the same timescale. For instance, using the estimate in [@Nimmrichter2013] with an interference visibility $f=0.2$ after $t=20\,$ms, one would also obtain $\mu=8.5$. This close match might be expected for an unsqueezed BEC, where all atoms are uncorrelated. That the number squeezed BEC discussed here does not reach an appreciably higher macroscopicity, despite its large depth of entanglement, can be attributed to the fact that single-particle observables are measured. They are not sensitive to many-particle correlations that are potentially destroyed by the classicalizing modification. In contrast, if the modification had induced spin flips, as in a single-well interferometer scenario [@Schmiedmayer2014], the resulting destruction of number-squeezing could be observed due to the interplay between the modification effect and the intrinsic dispersion caused by atom-atom interactions, see Eq. .
Leggett-Garg test with an atomic quantum random walk {#sec:4}
====================================================
Setup
-----
Reference [@Robens2015] describes a test of the Leggett-Garg inequality with single atoms performing a quantum random walk in an optical lattice formed by two circularly polarized laser beams. The form of the lattice potential depends on the hyperfine state of the atoms, so that by preparing single $^{133}$Cs atoms in a superposition of two hyperfine states and displacing the two lattices in opposite directions, one can prepare the atom in a superposition of left- and right-directed movements. We denote the displacement length of a single step by $d$, and the associated time required to displace the lattices by $T_{\rm d}$.
The quantum random walk (Fig.\[fig:3\]) is performed by first applying a $\pi/2$-pulse over the duration $T_{\rm r}$, which prepares the atom in a superposition of the hyperfine states and then transforming this into a spatial superposition by displacing the lattices for the duration $T_{\rm d}$. This scheme is iterated four times and finally a position measurement of the atom is performed, collapsing its position into a definite lattice site. Since no $\pi/2$-pulse is applied after the fourth step, atoms which do not end up in the same hyperfine state are excluded by the measurement protocol. This means that all paths which contribute to the interference must recombine after the third step.
Representing the two-level internal degree of freedom by a spinor, the action of a single step in the quantum random walk is given by the unitary operator $$\begin{aligned}
\label{eq:trafo}
{\sf S}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
{\sf U}_d & -{\sf U}_d \\
{\sf U}_{d}^\dagger & {\sf U}_{d}^\dagger
\end{pmatrix},\end{aligned}$$ with the translation operator ${\sf U}_d=\exp (-i{\sf p}d/\hbar )$. A straight-forward calculation shows that in addition to the classical random-walk trajectories, involving no coherences, there are only two classes of trajectories contributing to the interference pattern, see Fig. \[fig:QRWSkizze\]: (i) the atomic wavefunction is split and recombines immediately in the following step; (ii) the atomic wavefunction is split in the first step, then both parts are displaced either to the left or the right in the second step, and they recombine in the third step. To model the experimental outcome, one has to determine the likelihood $$\label{eq:qrwlike}
P(\ell | \tau_e,\sigma,I) = {\rm tr}_{\rm spin} \left(\langle \ell | \rho | \ell \rangle\right ),$$ where $\ell\in \{-2,-1,0,1,2\}$ labels the lattice sites that can be reached in four steps and $\rho$ is the final state evolved under influence of the modification with parameters $\tau_e$ and $\sigma$.
Impact of the modification
--------------------------
![(a) Examples of the two classes of coherently split trajectories contributing to the quantum random walk: (i) the atomic wavefunction splits in the first or second step and recombines afterwards; (ii) the atomic wavefunction splits in the first step, then both parts are move one step in parallel, and recombine in the third step. (b) Quantum-to-classical transition of the quantum random walk with decreasing classicalization timescale $\tau_e$. The diagrams depict the final-site probabilities for modification parameters $\hbar/\sigma_q=d/10$ and $\tau_e m_e^2/m_{\rm Cs}^2=1\,\mu{\rm s},50\,\mu{\rm s},100\,\mu{\rm s},10\,{\rm ms}$ from left to right.[]{data-label="fig:QRWSkizze"}](fig6.pdf){width="45.00000%"}
Since the separation between neighboring lattice sides is $d = 433$nm, spatial displacements can be neglected in the modification , i.e. we can set $\sigma_s =0$. The influence of the modification on a superposition of momentum states can be calculated by drawing on the results in Ref. [@schrinski2017sensing], where the momentum superposition of a non-interacting BEC in the limit of a high number of atoms was approximated by a macroscopic wave function (obeying the single particle Schrödinger equation). One can directly carry over these results to the present case of a single Cesium atom. As a result, the likelihood can be calculated with the help of the dimensionless coherence reduction factor $$\begin{aligned}
\label{eq:redfac}
R(t) = & \exp\left[
-\frac{2 T_{\rm d}m_{\rm Cs}^2}{\tau_e m_e^2} \left (1 - \frac{\sqrt{\pi} \hbar}{\sqrt{2} d \sigma_q}\operatorname{erf}\left (\frac{d \sigma_q}{\sqrt{2} \hbar } \right ) \right ) \right ] \nonumber \\
& \times \exp\left[
-\frac{t m_{\rm Cs}^2}{\tau_e m_e^2} \left (1
- \exp \left ( -\frac{d^2 \sigma_q^2}{2 \hbar^2} \right ) \right ) \right],\end{aligned}$$ where $t$ is the time over which the superposition state is maintained at a constant distance of $d$. Thus, in the case of the path (i) $t = T_{\rm r}$, and in case of path (ii) $t = T_{\rm d} + 2 T_{\rm r}$.
Initializing the random walk in the upper hyperfine state, one can identify all contributing trajectories by applying Eq. four times. After weighting these with the appropriate reduction factors , the trace finally yields the probability distribution[^1]
\[eq:QRW4\] $$\begin{aligned}
P(-2|\tau_e,\sigma,I)=&\frac{1}{16}, \\
P(-1|\tau_e,\sigma,I)= & \frac{1}{4} + \frac{1}{4}R(T_{\rm r})+\frac{1}{8}R(T_{\rm d} + 2T_{\rm r})\\
P(0|\tau_e,\sigma,I)=&\frac{3}{8}- \frac{1}{4}R(T_{\rm r}),\\
P(1|\tau_e,\sigma,I)=& \frac{1}{4}- \frac{1}{8}R(T_{\rm d} + 2T_{\rm r}),\\
P(2|\tau_e,\sigma,I)= & \frac{1}{16}.\end{aligned}$$
These results reflect what is to be expected from a classicalizing modification applied to the quantum random walk: The classical random walk probabilities are retrieved in the limit $\tau_e \to 0$, where $R(t) = 0$, while the opposite limit $\tau_e \to \infty$, i.e. $R(t) = 1$, yields the ideal quantum random walk probabilities. The gradual transition between classical and quantum behavior is depicted in Fig. \[fig:QRWSkizze\].
In the Leggett-Garg test of Ref. [@Robens2015] additional measurement results were postselected conditioned on whether the walker moves in the first step to the left or to the right. In this case the random walk effectively starts one step later, and thus only trajectories of type (i) contribute to the interference. The resulting probabilities can be determined as above,
\[eq:QRW3\] $$\begin{aligned}
P_{\rm L}(-2|\tau_e,\sigma,I)& = & P_{\rm R}(2|\tau_e,\sigma,I)=\frac{1}{8}, \\
P_{\rm L}(-1|\tau_e,\sigma,I)& = & P_{\rm R}(1|\tau_e,\sigma,I)= \frac{3}{8}+ \frac{1}{4} R(T_{\rm r}), \\
P_{\rm L}(0|\tau_e,\sigma,I)&= & P_{\rm R}(0|\tau_e,\sigma,I)= \frac{3}{8} - \frac{1}{4} R(T_{\rm r}), \\
P_{\rm L}(1|\tau_e,\sigma,I)&= &P_{\rm R}(-1|\tau_e,\sigma,I)= \frac{1}{8},\\
P_{\rm L}(2|\tau_e,\sigma,I) &= &P_{\rm R}(-2|\tau_e,\sigma,I)=0.\end{aligned}$$
The subscripts L or R denote that the first step was performed to the left or right.
For completeness, we note that the Leggett-Garg inequality studied in [@Robens2015] reads as $$\begin{aligned}
\sum_{\ell=-2}^2 {\rm sgn}(\ell)\left (P(\ell) - \frac{1}{2} \left [ P_{\rm L}(\ell)+P_{\rm R}(\ell) \right ] \right )
\leq 0,\end{aligned}$$ where we dropped the parameters $\tau_e,\sigma,I$ for brevity. This Leggett-Garg inequality can be rewritten in terms of the modification parameters through the reduction factor by inserting Eqs. (\[eq:QRW4\]) and (\[eq:QRW3\]), $$\begin{aligned}
R(T_{\rm r}) + R(T_{\rm d} + 2 T_{\rm r}) \le 0.
\label{eq:LGTInequalityMIM}\end{aligned}$$ This inequality is always violated unless $\tau_e $ vanishes, but the left-hand side approaches zero exponentially with decreasing $\tau_e$. Note that our assessment of macroscopicity is not based on such a derived quantity, but on the raw data of detection clicks.
Experimental parameters
-----------------------
In the experiment the displacement and resting time are $T_{\rm d}=21\,\mu{\rm s}$ and $T_{\rm r}=5\,\mu{\rm s}$ and the distance between each lattice site is $d=433\,{\rm nm}$. Maximizing the effect of the modification we note that the reduction factor decreases with increasing $\sigma_q$ and that the five percent quantile $\tau_{\rm m}(\sigma)$ saturates for $\hbar/\sigma_q \ll d$. To assess the macroscopicity, we take the value $\hbar/\sigma_q \approx d/10$, where $\tau_{\rm m}(\sigma)$ already takes the saturated value, yielding $\mu_{\rm m}=7.1$.
Finally, since we neglected possible effects of modification-induced heating so far, we have to verify that this is justified here, i.e. at the stated value of $\sigma_q$ and for the relevant range of classicalization time scales $\tau_{e }$. This can be done conservatively by calculating the heating rate with the 5% quantile of the Jeffreys’ prior ($\tau_{e} = 16.75\,\mu{\rm s}$). It serves as an upper bound (see Fig. \[fig:3\]) due to Bayesian updating. The resulting temperature increase of $\Delta T\approx 5.6\,\mu{\rm K}$ over the duration of the whole experiment is moderate, amounting to 1/16 of the potential depth. It thus renders particle loss negligible, so that no explicit conditioning on a likelihood which accounts for heating is required to arrive at and .
In summary, the macroscopicity of the atomic Leggett-Garg test is dominated by the timescale on which the experiment was performed, i.e. the ramp- and waiting-time between random walk steps. Since only neighboring trajectories contribute to interference, the relevant length scale of the superposition state is given by the lattice spacing $d$ rather than by the spatial extension of the final state. This could be enhanced by implementing a $\pi/2$-pulse after the fourth step, or by performing more steps, so that also trajectories separated by more distant sites contribute to the interference pattern.
Mechanical entanglement of photonic crystals {#sec:5}
============================================
Measurement protocol
--------------------
The observation of entanglement between two nanomechanical oscillators reported in Ref. [@Riedinger2018] is based on a coincidence measurement of Stokes- and anti-Stokes photons created in photonic crystal nanobeams placed in the two arms of a Mach-Zehnder interferometer, see Fig. \[fig:4\]. In the first step (pump), a photon is sent through the entrance beam splitter, excites a single phonon in one of the two nanobeams, thereby creating entanglement in their mechanical excitation. The Stokes-scattered photon is detected behind the exit beam splitter. In the second step (read), a further photon enters the interferometer through the entrance beam splitter, leading to stimulated emission in the photonic crystal. The resulting anti-Stokes scattered photon, which serves to read out the entanglement, is also detected behind the exit beam splitter.
We denote the measurement outcomes of the Stokes and the anti-Stokes photon detectors by $\pm_{1,2}$, where $+$ ($-$) refers to the upper (lower) detector behind the exit beam splitter and the index refers to the pump and read photon, respectively. The likelihood for a certain coincidence measurement is $$\label{eq:likenano}
P(\pm_1,\pm_2|\tau_e,\sigma,I) = {\rm tr} \left (| \pm_1, \pm_2 \rangle\langle \pm_1, \pm_2| \rho_{\rm fin} \right )$$ where $\rho_{\rm fin}$ is the total final state of both oscillators and both photons. The modification parameters $\tau_e$ and $\sigma$ only enter through their influence on the dynamics of the nanomechanical oscillators.
In each nanobeam a single mechanical mode contributes to the measurement signal of the experiment. Even though the pump photon can excite this mode only once, we will in the following allow for arbitrary phonon occupations $|k, \ell \rangle$ of the two oscillators to account for modification-induced heating.
Given that the two relevant oscillator modes are initially in the ground state, the total wave function of the system after the pump photon traversed the exit beam splitter reads $$\begin{aligned}
\left|\psi\right\rangle_{t=0} =&
\frac{1}{2}\left[
\left|+\right\rangle_1\left(\left|1,0\right\rangle+e^{i\phi}\left|0,1\right\rangle\right)\right.\nonumber\\
&\left.+\left|-\right\rangle_1\left(\left|1,0\right\rangle-e^{i\phi}\left|0,1\right\rangle\right)
\right] | {\rm vac} \rangle_2,
\label{eq:MHOInitialStateSP} \end{aligned}$$ where $\phi$ is the initial relative phase. The state now evolves freely according to the modified master equation into the mixed state $\rho_t$ until the read photon passes the interferometer.
The measurement with the read photon can be described through application of the read operator ${\sf R}$, as $\rho_{\rm fin}={\sf R} \rho_t {\sf R}^\dagger/\mathcal{N}$. Here, the factor $\mathcal{N}={\rm tr}({\sf R}^\dagger{\sf R} \rho_t )$ accounts for the conditioning on coincident detections of Stokes and anti-Stokes photons. The read operator ${\sf R}$ first annihilates a phonon in one of the two oscillators and simultaneously creates a read photon in the corresponding interferometer arm, with the relative phase $\theta$ between the two arms determined by the experimental setup. In a second step, the thus created photon traverses again the beam splitter, yielding in total $$\begin{aligned}
\label{eq:read}
{\sf R} |\pm\rangle_1|k, \ell \rangle |{\rm vac}\rangle_2 = & \frac{|\pm\rangle_1}{\sqrt{2k+2\ell}} \left [\sqrt{k} |k-1,\ell \rangle \left ( |+\rangle_2 + |-\rangle_2 \right ) \right. \nonumber \\
& \left. + e^{i \theta} \sqrt{\ell} |k,\ell-1\rangle \left ( |+\rangle_2 - |-\rangle_2 \right ) \right ]\end{aligned}$$ for $(k,\ell) \neq (0,0)$. By in addition setting ${\sf R}|\pm\rangle_1 |0, 0 \rangle |{\rm vac}\rangle_2=0 $ we account for the fact that the phonon ground state (which may be populated by modification-induced transitions) cannot lead to a coincidence detection involving an anti-Stokes photon.
The probability can be written as due to a generalized measurement, $P(\pm_1,\pm_2|\tau_e,\sigma,I) = {\rm tr} ({\sf F}_{\pm_2} \rho^{(\pm_1)}_t )/\mathcal{N}$. Here, the oscillator state $$\begin{aligned}
\label{eq:rhopm}
\rho^{(\pm_1)}_t=\langle \pm|_1\langle{\rm vac}|_2\rho_t|{\rm vac}\rangle_2| \pm\rangle_1\end{aligned}$$ is conditioned on the detection of the Stokes photon, and ${\sf F}_{\pm_2} = {\rm tr}_1 (\langle {\rm vac} |_2 {\sf R}^\dagger | \pm \rangle_2 \langle \pm |_2 {\sf R} | {\rm vac} \rangle_2)$ describes the measurement of the anti-Stokes photon, $$\begin{aligned}
{\sf F}_{\pm_2} = & \frac{1}{2}\left(\sum_{k=1,\ell=0}^{\infty}\frac{k}{k+\ell}\left|k,\ell\right\rangle\left\langle k,\ell\right|+\sum_{k=0,\ell=1}^{\infty}\frac{\ell}{k+\ell}\left|k,\ell\right\rangle\left\langle k,\ell\right|\right.\nonumber \\
&\left.+\sum_{k=1,\ell=0}^{\infty}e^{i\theta}\frac{\sqrt{k(\ell+1)}}{k+\ell}\left|k,\ell\right\rangle\left\langle k-1,\ell+1\right|\right.\nonumber\\
&\left.+\sum_{k=0,\ell=1}^{\infty}e^{-i\theta}\frac{\sqrt{(k+1)\ell}}{k+\ell}\left|k,\ell\right\rangle\left\langle k+1,\ell-1\right|\right) \,.
\label{eq:MHORiedingerProjectorPositive}\end{aligned}$$
To prepare the calculation of the likelihoods, we now determine the influence of the modification on the initial oscillator state .
Impact of the modification
--------------------------
To handle the elastic deformation of a single nanomechanical beam, we first note that all atoms in the solid can be treated as distinguishable. One can therefore use the Lindblad operators in first quantization, $$\begin{aligned}
\label{eq:lindblad2}
{\sf L}({\bf q},{\bf s})
=&\sum_n\frac{m_n}{m_e}\exp\left[-i\frac{\textbf{\textsf{r}}_n\cdot{\bf q}-\textbf{\textsf{p}}_n\cdot{\bf s}}{\hbar}\right]\,.\end{aligned}$$ To express this in terms of the mode variables, we expand the position operator $\textbf{\textsf{r}}_n$ of each individual atom around its equilibrium position ${\bf r}_n^{(0)}$, $$\begin{aligned}
\textbf{\textsf{r}}_n={\bf r}_n^{(0)}+{\bf w}({\bf r}_n^{(0)}){\sf Q}\,,\end{aligned}$$ in terms of the classical mode function [@madelung2012introduction; @fetter2003theoretical] of the relevant displacement mode ${\bf w}({\bf r})$ and its operator-valued amplitude ${\sf Q}$. The latter can also be written using the mode creation and annihilation operators ${\sf a}^\dagger$ and ${\sf a}$, $$\begin{aligned}
\label{eq:displfield}
{\sf Q}
=\sqrt{\frac{\hbar}{2\varrho V_{\rm m}\omega}}\left({\sf a}+{\sf a}^{\dagger}\right),\end{aligned}$$ where $\varrho$ is the mass density of the material, $\omega$ the mechanical frequency, and $V_{\rm m}$ the mode volume, see App. \[app:b\].
Accordingly, the momentum operator in takes the form $$\begin{aligned}
\label{eq:momfield}
\textbf{\textsf{p}}_n =\frac{m_n}{\varrho V_{\rm m}}{\bf w}({\bf r}^{(0)}_n){\sf P}= i \sqrt{\frac{\hbar \omega_k m_n^2}{2\varrho V_{\rm m}}}{\bf w}({\bf r}^{(0)}_n)\left({\sf a}^{\dagger}-{\sf a}\right).\end{aligned}$$ This equation implies that the modification-induced spatial displacement $\mathbf{s}$ in scales with the mass of the atom divided by the effective mass of the mechanical mode, which is on the order of the nanobeam mass. The spatial displacement is therefore negligible for all scenarios that lead to observable decoherence, allowing us to approximate the Lindblad operators as $$\begin{aligned}
&{\sf L}({\bf q})
\simeq\sum_n\frac{m_n}{m_e}\exp\left[-\frac{i}{\hbar}\left({\bf r}_n^{(0)}+\sum_k{\bf w}_k({\bf r}^{(0)}_n){\sf Q}_k\right)\cdot{\bf q}\right]
\nonumber\\
&= \frac{1}{m_e}\int d^3{\bf r}\,\varrho({\bf r}) \exp\left[-\frac{i}{\hbar}\left({\bf r}+\sum_k{\bf w}_k({\bf r}){\sf Q}_k\right)\cdot{\bf q}\right],
\label{eq:HarmOscFullLindbladOp}\end{aligned}$$ where $k$ is a mode index, and $\varrho({\bf r})=\sum_n m_n\delta({\bf r}-{\bf r}^{(0)}_n)$ denotes the mass density of the oscillator. The latter can be replaced by a continuous, homogeneous mass density provided the characteristic length scale $\hbar/\sigma_q$ is much greater than the lattice spacing of the crystal structure.
The Lindblad operators may be expanded to first order in the relevant mode amplitude ${\sf Q}$ as long as $\sigma_q \ll \sqrt{2 \varrho V_{\rm m} \omega\hbar}$. This decouples the different modes and we have $$\begin{aligned}
{\sf L}({\bf q})=-\frac{i}{\hbar}\left[\widetilde{{\bf w}}_{\varrho}({\bf q})\cdot{\bf q}\right]{\sf Q},
\label{eq:HarmOscDiffLO}\end{aligned}$$ where we introduced $$\begin{aligned}
\widetilde{{\bf w}}_{\varrho}({\bf q})&=
\frac{1}{m_e}\int d^3{\bf r}\,\varrho({\bf r}){\bf w}({\bf r})e^{-i{\bf r}\cdot{\bf q}/\hbar}.\end{aligned}$$
The total master equation including the free harmonic Hamiltonian and the Lindblad operators of both oscillators can be solved analytically with the help of the characteristic function $$\begin{aligned}
\chi_t({\bf Q},{\bf P})=\int d^2{\bf Q}'\,e^{i{\bf P}\cdot{\bf Q}'/\hbar}\left\langle {\bf Q}'+\frac{{\bf Q}}{2}\right|\rho_t\left|{\bf Q}'-\frac{{\bf Q}}{2}\right\rangle,\end{aligned}$$ where ${\bf Q} = (Q_1,Q_2)$ and ${\bf P} = (P_1,P_2)$ are the joint position and momentum coordinates of both oscillators. The evolution equation for the characteristic function reads $$\begin{aligned}
\partial_t\chi_t({\bf Q},{\bf P})=&
\left(- \frac{1}{\varrho V_{\rm m}}{\bf P}\cdot \nabla_{\bf Q}+\varrho V_{\rm m}{\bf Q}\cdot{\rm \Omega}^2\nabla_{\bf P}-\frac{U(\sigma) {\bf Q}^2}{\tau_e}\right)\nonumber\\
&\times\chi_t({\bf Q},{\bf P}),
\label{eq:HarmOscDiffCharTE}\end{aligned}$$ where ${\rm \Omega}$ is the diagonal matrix containing the two slightly detuned frequencies of both oscillators and $$\begin{aligned}
U(\sigma)=&\frac{1}{2\hbar^2}\int d^3 {\bf q}\, f_\sigma(q) \left|\widetilde{{\bf w}}_{\varrho}({\bf q})\cdot {\bf q}\right|^2 .
\label{eq:HarmOscDiffFT}\end{aligned}$$ Here we exploited that the separation of the two oscillators is much greater than $\hbar/\sigma_q$.
The time evolved characteristic function is given by $$\begin{aligned}
\label{eq:soldifflim}
\chi_t({\bf Q},{\bf P})=&
\exp\left[
-\frac{U(\sigma)}{\tau_e}\int_0^t dt' {\bf Q}^2_{t'} \right]\chi_0({\bf Q}_t,{\bf P}_t),\end{aligned}$$ with $$\begin{aligned}
{\bf Q}_t=&\cos({\rm \Omega t}){\bf Q}+\frac{1}{\varrho V_{\rm m}}{\rm \Omega}^{-1}\sin({\rm \Omega t}){\bf P}\nonumber\\
{\bf P}_t=&\cos({\rm \Omega t}){\bf P}-\varrho V_{\rm m}{\rm \Omega}\sin({\rm \Omega t}){\bf Q}.
\label{eq:MarmoOscFreeQP}\end{aligned}$$
Calculating the initial characteristic function of the state and evaluating for a given mode function ${\bf w}({\bf r})$ allows one to determine analytically the likelihoods .
Particle loss
-------------
For increasing $\sigma_q$ the energy gain induced by momentum translations due to the Lindblad operators can exceed the binding energy of the silicon atoms in the crystal. Thus, the modification may induce particle loss already deep in the diffusive regime. The solution of the mode dynamics cannot capture this because the mode expansion assumes the atoms to reside in infinitely extended harmonic potentials. Due to the finiteness of the real binding potential there is a critical momentum transfer $q_c$ beyond which the sole effect of the modification is a reduction of the atom number in the crystal.
To account for this particle loss, we split Eq. into a part $\mathcal{M}_\sigma^{<}$ with momentum transfers $|{\bf q}|<q_c$ that will most likely leave the atoms in the crystal, and into the part $\mathcal{M}_\sigma^{>}$ with $|{\bf q}|>q_c$ removing them into the vacuum, $$\begin{aligned}
\label{eq:LindbladSum}
\mathcal{M}_\sigma\rho_t&=\int\displaylimits_{q<q_C} d^3 {\bf q}\,f_\sigma(q)\left[{\sf L}({\mathbf q})\rho {\sf L}^\dagger({\mathbf q})-\frac{1}{2}\left\{{\sf L}^\dagger({\mathbf q}){\sf L}({\mathbf q}),\rho\right\}\right]\nonumber\\
&+\int\displaylimits_{q>q_C} d^3 {\bf q}\,f_\sigma(q)\left[{\sf L}({\mathbf q})\rho {\sf L}^\dagger({\mathbf q})-\frac{1}{2}\left\{{\sf L}^\dagger({\mathbf q}){\sf L}({\mathbf q}),\rho\right\}\right].\end{aligned}$$ A Dyson expansion shows that the final state can be written as a sum $$\begin{aligned}
\rho_t=\exp\left[\frac{t}{i\hbar}\mathcal{H}+\frac{t}{\tau_e}\mathcal{M}_\sigma^{<}\right]\rho_0+\widetilde{\rho}\end{aligned}$$ where only the first term is consistent with the coincidence measurement . Its reduced trace can be absorbed in the normalization $\mathcal{N}$ reflecting the conditioning on the coincidence measurements.
The time evolution under the modification $\mathcal{M}_\sigma^{<}/{\tau_e}$ can now be treated as in the previous section, yielding Eq. with $U(\sigma)$ replaced by $$\begin{aligned}
U_{<}(\sigma)=&\frac{1}{2\hbar^2}\int_{q<q_c} d^3 {\bf q}\, f_\sigma(q) \left|\widetilde{{\bf w}}_{\varrho}({\bf q})\cdot {\bf q}\right|^2 .\end{aligned}$$
Experimentally achieved macroscopicity
--------------------------------------
The two oscillators in Ref. [@Riedinger2018] are characterized by the effective mass $\varrho V_{\rm m} \approx 9\times10^{-17}$kg [@PrivateCommRiedinger] and the mechanical frequency $\omega \approx 2\pi\times 5\,{\rm GHz}$. The exact displacement field depends on the precise geometry of the photonic crystal, and is only numerically accessible. Since the details of the mode function are expected to be of minor relevance we approximate the shape of the oscillator by an elastic silicon cuboid containing only those atoms of the nanobeam that contribute to the elastic deformation. The dimension of this cuboid is set by the effective mass and frequency of the oscillator, yielding for its ground mode $L_x\times L_y\times L_z \approx 0.31\,\mu{\rm m}\times0.31\,\mu{\rm m}\times0.84\,\mu{\rm m}$, using the speed of sound $v = 8433$m/s and density $\varrho = 2300$kg/m$^3$ of silicon.
The resulting displacement field of the simplest longitudinal mode has the form $$\begin{aligned}
\label{eq:SineMode}
{\bf w}({\bf r})={\bf e}_z \sin\left(\frac{\pi z}{L_z}\right),\end{aligned}$$ for $-L_z/2 \leq z \leq L_z/2$. This can now be used to calculate the Lindblad operators .
The likelihood can be calculated with the characteristic function of the state as a phase space integral $$\begin{aligned}
P(\pm_1,\pm_2|\tau_e,\sigma,I)=\int d^2{\bf Q}d^2{\bf P}
\chi^{\pm_1}_t({\bf Q},{\bf P})\eta^{\pm_2}({\bf Q},{\bf P}),
\label{eq:MHOProbabilities}\end{aligned}$$ where $\eta^{\pm_2}({\bf Q},{\bf P})$ is the characteristic function of the operator .
![ The maximally excluded time parameter $\tau_{\rm m}$ as defined by the five percent quantile obtained via Bayesian updating with Eq. . The local maximum to the right is assumed for values of $\hbar/\sigma_q$ roughly equal to the spatial extension of the crystal mode $L_{x,z}$. The global maximum is achieved at $\hbar/\sigma_q\simeq\sqrt{\hbar^2/2m_{\rm Si}E_{\rm b}}$ where the momentum transfers become sufficiently strong to remove particles from the crystal. The fading of the graph indicates where the analytical descriptions derived in App. \[app:b\] fail: First, when $\hbar/\sigma_q$ is on the order of several Ångström so that the mass density can no longer be approximated as continuous, and second, when $\hbar/\sigma_q$ is on the order of femtometers where the diffusive regime ceases to be valid.[]{data-label="fig:BeyondContinuum"}](fig7.pdf){width="45.00000%"}
This expression can now be simplified by noting that the oscillator frequency is large on the timescale of the experiment, $\omega t \gg 1$, so that the time-averaged phase space coordinates can be used in the exponent of , $$\begin{aligned}
\chi_t({\bf Q},{\bf P}) \approx &
\exp\left[
-\frac{U_<(\sigma) t}{2 \tau_e} \left ( {\bf Q}^2 + \varrho^2 V_{\rm m}^2 (\Omega^{-1} {\bf P})^2 \right )\right]
\nonumber\\
&\times\chi_0({\bf Q}_t,{\bf P}_t)\,.\end{aligned}$$ Moreover, the modification cannot create coherences between the oscillator states. In Eq. one can therefore keep only the diagonal terms and the initial coherences between ground state and first excited states, $$\begin{aligned}
{\sf F}_{\pm_2} = \frac{1}{2} \left (\mathbb{1} - \left|0,0\right\rangle\left\langle 0,0\right|+e^{i\theta} \left|1,0\right\rangle\left\langle 0,1\right| + e^{-i\theta}\left|0,1\right\rangle\left\langle 1,0 \right|\right) .
\label{eq:MHORiedingerProjectorPositive2}\end{aligned}$$ The corresponding characteristic function is given in App. \[app:b\], together with the characteristic function of the state .
The integral Eq. yields the likelihood in its final form, $$\begin{aligned}
\label{eq:Ppm222}
&P(\pm_1,\pm_2|\tau_e,\sigma,I)=\nonumber\\
&\frac{1}{4\mathcal{N}}
+\frac{(\pm_1)(\pm_2) 4\cos (\theta- \Delta \Omega t )-2\xi t/\tau_{ e}-\xi^2 t^2/\tau^2_{ e}}{\mathcal{N}(2+\xi t/\tau_{ e})^4},\end{aligned}$$ where $\Delta \Omega = 2\pi \times 45$MHz is the frequency mismatch between the oscillators, and we defined the dimensionless parameter $\xi = 2 U_<(\sigma) \hbar/\varrho V_{\rm m} \omega$ characterizing the sensitivity of the relevant nanobeam mode to the modification parameter $\sigma_q$. The geometric factor $U$, as defined in Eq. , is evaluated in App. \[app:b\].
The phase- or time-sweep measurement protocols performed in [@Riedinger2018] are described by varying $\theta$ and $t$, respectively. The (unreported) initial phase is deduced to be $ \phi\approx1.8\, {\rm rad}-\Delta\Omega\times123\,{\rm ns}$ by optimization. In order to obtain the achieved macroscopicity, we perform Bayesian updating to determine the posterior and maximize over $\sigma_q$. The resulting $\tau_{\rm m}$ is plotted in Fig. \[fig:BeyondContinuum\] for $q_c = \sqrt{2 m_{\rm Si} E_b}$ with $E_{\rm b}=4.6\,$eV [@Farid1991]. It exhibits a global maximum of $\tau_{\rm m}=6.6\times 10^{7}\,$s at $\hbar/\sigma_q\simeq\sqrt{\hbar^2/2m_{\rm Si}E_{\rm b}}$, yielding a macroscopicity value of $\mu_{\rm m}=7.8$.
Given the relatively high mass of the nanomechanical oscillators and the fairly long coherence time achieved, one might expect the entangled nanobeams to be characterized by a higher degree of macroscopicity. That this is not the case can be traced back to the fact that the superposition state is delocalized only on the scale of femtometers. For such small spatial delocalizations, the sole influence of the modification is to add momentum diffusion to the nanobeam dynamics, leading to weakest possible form of spatial decoherence.
Conclusion
===========
The empirical measure discussed in this article serves to quantify the macroscopicity reached in quantum mechanical superposition experiments by the degree to which they rule out classicalizing modifications of quantum theory. We showed how the framework of Bayesian hypothesis testing allows one to assess diverse experiments based on their raw data, thus accounting appropriately for all measurement uncertainties. The fact that measurement errors are fundamentally unavoidable, ensures that the macroscopicity $\mu_{\rm m}$ will always converge to a finite value, even if quantum mechanics holds on all scales. For sufficiently large data sets, when statistical errors tend to be negligible, the here presented measure will approach the one given in Ref. [@Nimmrichter2013] for interferometric superposition tests. Equation is thus the natural generalization of the latter.
A great benefit of the formalism is that it allows one to straightforwardly combine independent parts of an experiment, e.g. quantum random walks of different lengths (Sec.\[sec:4\]) or different measurement protocols for entangled nanobeams (Sec.\[sec:5\]). Moreover, the Bayesian updating process naturally allows for correlated observables to be taken into account, as for instance the total atom number and the population imbalance in BEC interferometers (see Sec.\[sec:3\]). Finally, the use of Jeffreys’ prior ensures that the macroscopicity measure is solely determined by the experimental data at hand, irrespective of prior believes. In particular, using this least informative prior prevents the macroscopicity measure to favor any one type of quantum test against others. We showed that Jeffreys’ prior exists for all physically relevant situations, where the likelihood is a smooth function of the modification parameters.
These advantages come at the cost that the required likelihoods are in general considerably more difficult to determine than e.g. specific coherences of the statistical operator. It requires one to capture appropriately how the relevant quantum degrees of freedom are affected by the master equation describing the impact of the modification on the many-particle system state. We explained in Secs. \[sec:3\]-\[sec:5\] how this works in practice for three rather different quantum superposition tests.
We reemphasize that a naive application of the macroscopicity measure may yield a finite value even for experiments demonstrating no quantum superposition, because already the absence of observed heating can constrain the classicalization parameters. To be on the safe side, one must identify those observations that yield information only about modification-induced heating and use this data to condition the likelihoods as described at the end of Sect. \[sec:2A\]. In most quantum tests this is not necessary because the conditioning is already implemented in the measurement protocol.
The measure of macroscopicity put forward in this article can be used for any superposition test, provided a mechanical degree of freedom is involved, be it the electronic excitation of an atom or the motion of a kilogram-scale mirror. As such it does not apply to quantum tests involving only spins or photons. It seems natural to generalize the macroscopicity measure to pure photon experiments by drawing on a minimal class of classicalizing modifications of QED, but it is still an open problem how to get hold of the latter. Beyond the assessment of macroscopicity, the Bayesian hypothesis testing presented in Sec.\[sec:2\], can also be used for a proper statistical description of tests of specific modification models, e.g. the various extensions of the Continuous Spontaneous Localization model [@Bassi2013], but also of environmental decoherence mechanisms.
Finally, it goes without saying that the macroscopicity $\mu_{\rm m}$ attributed to a given superposition test serves to highlight a single aspect of the experiment, albeit an important one. It must not be taken as a proxy for the overall significance of an experimental finding.
We thank Andrea Alberti, Tarik Berrada, and Ralf Riedinger for helpful comments on their experiments, and the authors of Ref. [@Riedinger2018] for providing us with the unpublished raw data reported in Fig. \[fig:4\]. BS thanks Gilles Kratzer for helpful discussions on the topic of Bayesian statistics. This work was funded by Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 298796255.
Integrability of the posterior distribution {#app:beweis}
===========================================
To see that the Jeffreys’ prior always yields a normalizable posterior distribution , we first consider the limit $\tau_e \to \infty$, where the modification becomes arbitrarily weak. In this case the general solution of the master equation can be expanded to first order in $1/\tau_e$ by its Dyson series. Calculating the likelihood then yields $$\begin{aligned}
\label{eq:exp}
P(d|\tau_e,\sigma,I) \simeq P_\infty(d |I)+\frac{1}{\tau_e} q(d|\sigma,I) \quad \text{for} \quad \tau_e \to \infty,\end{aligned}$$ where $q$ is independent of $\tau_e$. Inserting the expansion into Jeffreys’ prior (\[eq:JeffreysPrior\]) yields $$\begin{aligned}
p(\tau_e | \sigma, I ) \stackrel{\tau_e\to\infty}{\sim} \begin{cases}
\tau_e^{-3/2} & \exists\, d_0 :P_\infty(d_0|I)=0, \\
\tau_e^{-2} & \, \text{else},
\end{cases}\end{aligned}$$ implying that the posterior decays at least as $\tau_e^{-3/2}$ for $\tau_e \to \infty$.
Second, for $\tau_e \to 0$, where modification-induced decoherence and heating get stronger and stronger, we use that the likelihood $P(d|\tau_e,\sigma,I)$ will continuously approach some limiting classical probability, $$P(d|\tau_e,\sigma,I) \simeq {P}_0(d|\sigma, I) + \tau_e^{\alpha} \tilde{q}(d|\sigma,I) \quad \text{for} \quad \tau_e \to 0,$$ where $\alpha>0$ may depend on $d$. Using this to evaluate Jeffreys’ prior yields that $$p(\tau_e | \sigma, I ) \stackrel{\tau_e\to 0}{\sim} \begin{cases}
\tau_e^{-(1 - \alpha_{\rm min}/2)} & \exists\, d_0 :P_0(d_0|\sigma,I)=0, \\
\tau_e^{-(1 - \alpha_{\rm min})} & \, \text{else},
\end{cases}
\label{eq:LimitTauToZero}$$ where $\alpha_{\rm min}>0$ is the minimal $\alpha$. Physically speaking, this means that no quantum superposition test will support a classical model of infinitely strong heating. Equation implies that the posterior always diverges weaker than $1/\tau_e$ for $\tau_e \to 0$.
Finally, to rule out that the posterior diverges at a finite $\tau_e \in (0,\infty)$, we note that the likelihood $P(d|\tau_e,\sigma,I)$ stays non-negative for all $\tau_e$. Thus, whenever it vanishes for some value of $\tau_e$, its first derivative must also be zero and its second derivative must be non-negative. Application of L’Hospital’s rule then shows that the posterior stays finite for all intermediate values of $\tau_e$. This completes the argument why the choice of Jeffrey’s prior always leads to a normalizable posterior and thus yields a well-defined value of macroscopicity .
Simultaneous shearing and diffusion of number squeezed BECs {#app:a}
===========================================================
For simultaneous phase diffusion and shearing the time evolution of the tangent space Wigner function $w_t(j_y,j_z)$ is given by the equation $$\begin{aligned}
\partial_t w_t(j_y,j_z)=-2\zeta j_z\partial_{j_y}w_t(j_y,j_z)+
\frac{\Gamma_{\rm P}}{2}\partial^2_{j_y} w_t(j_y,j_z)\,,\end{aligned}$$ which is solved by . If diffusion takes place perpendicular to the shearing, the time evolution is given by the equation $$\begin{aligned}
\partial_t w_t(j_y,j_z)=-2\zeta j_z\partial_{j_y}w_t(j_y,j_z)+
\frac{\Gamma_{\rm S}}{4}\partial^2_{j_z} w_t(j_y,j_z)\,.
\label{eq:ShearingDiffusion}\end{aligned}$$ Its general solution is $$\begin{aligned}
w_t(j_y,j_z)=\frac{1}{4\pi^2}\int dp_ydp_zdj_y'dj_z'w_0(j_y',j_z')e^{i(j_y-2\zeta j_z t)p_y-izp_z-ip_yj_y'+ip_z j_z'}\exp\left[-J^2\Gamma_{\rm S}\left(\frac{ p_z^2}{4}t-\frac{ \zeta p_yp_z}{2}t^2-\frac{\zeta^2 p_y^2}{3}t^3\right)\right]\,.\end{aligned}$$ We take the initial distribution $w_0(j_y,j_z)$ to be a Gaussian with width $\sigma_y$ and $\sigma_z$. Integrating $j_z$ preserves the Gaussian form, yielding the marginal distribution $$\begin{aligned}
p_t(j_y) = \int dz\,w_t(j_y,j_z)=\frac{1}{\sqrt{2\pi\sigma^2_y(t)}}\exp\left[\frac{j_y^2}{2\sigma^2_y(t)}\right],\end{aligned}$$ with variance $$\begin{aligned}
\sigma^2_y(t)=\sigma^2_y+4\zeta^2t^2\left(\sigma_z^2+\frac{J^2\Gamma_{\rm S}t}{6}\right).\end{aligned}$$
Calculational details for the entangled nanobeams experiment {#app:b}
============================================================
Normalization of displacement fields
------------------------------------
The equation of motion of a classical displacement field in an isotropic elastic medium can be derived from the Lagrangian density [@fetter2003theoretical] $$\begin{aligned}
\mathcal{L}=\frac{\varrho}{2}\dot{\bf u}^2 -V\left[{\bf u}({\bf r},t)\right] =\frac{\varrho}{2}\dot{\bf u}^2-
\frac{1}{4}\sum_{klmn=1}^3\left(\frac{\lambda}{2}\delta_{kl}\delta_{mn}+\mu \delta_{km}\delta_{ln}\right)\left(\frac{\partial u_k}{\partial x_l}+\frac{\partial u_l}{\partial x_k}\right)\left(\frac{\partial u_m}{\partial x_n}+\frac{\partial u_n}{\partial x_m}\right),\end{aligned}$$ where $\lambda$ and $\mu$ are the Lamé coefficients. Thus, the dynamics of ${\bf u}({\bf r},t)$ are given by $$\begin{aligned}
\varrho\frac{\partial^2{\bf u}}{\partial t^2} = \mu\nabla^2{\bf u}+(\lambda+\mu)\nabla(\nabla\cdot{\bf u}).\end{aligned}$$ This equation can be solved by introducing the mode functions ${\bf u}_k({\bf r},t)$ as the eigenfunctions of the differential operator on the left hand side with eigenvalues $-\omega_k^2 \varrho$. The total displacement field can then be written as $$\begin{aligned}
{\bf u}({\bf r},t)=\sum_k\sqrt{\frac{\hbar}{2 \rho V_k \omega_k}}{\bf w}_k({\bf r})\left(e^{-i\omega_k t}a_k+e^{i\omega_k t}a_k^*\right),
\label{eq:ClassicalDisplacementMode}\end{aligned}$$ so that its mean energy is $$\begin{aligned}
\left\langle E\right\rangle_t=\left\langle \int d^3{\bf r} \left[
\frac{\varrho}{2}\dot{{\bf u}}^2({\bf r},t)+V\left[{\bf u}({\bf r},t)\right]
\right]\right\rangle_t
=\sum_k\frac{\hbar\omega_k}{V_k}\int d^3{\bf r}\, {\bf w}_k^2({\bm r})a_k^*a_k.
\label{eq:MHOmeanenergy}\end{aligned}$$ Demanding that $\left\langle E\right\rangle_t=\sum_k\hbar\omega_k a_k^*a_k$ yields the normalization condition $\int d^3{\bm r}\, {\bm w}_k^2({\bm r})=V_k$.
Characteristic functions of mechanical oscillator states
--------------------------------------------------------
The characteristic function of the initial oscillator state in for $\phi=0$ can be calculated as $$\begin{aligned}
\chi^{\pm_1}({\bf Q},{\bf P})
=&\frac{1}{2}\exp\left[-\frac{1}{4\hbar\varrho V_{\rm m}}\left({\bf P}\cdot\Omega^{-1}{\bf P}+\varrho^2 V_{\rm m}^2{\bf Q}\cdot\Omega{\bf Q}\right)\right]
\nonumber\\
&\times\left(
1-\frac{1}{4\hbar\varrho V_{\rm m}}\left[\sum_{\lambda = 1,2}(\pm_1)^{\lambda}\left(\Omega^{-1/2}{\bf P}\right)_{\lambda}\right]^2
-\frac{\varrho V_{\rm m}}{4\hbar}\left[\sum_{\lambda = 1,2}(\pm_1)^{\lambda}\left(\Omega^{1/2}{\bf Q}\right)_{\lambda}\right]^2
\right)\,,\end{aligned}$$ where ${\rm \Omega}={\rm diag}(\Omega_1,\Omega_2)$. In a similar fashion, one obtains the characteristic symbols of the effect as $$\begin{aligned}
\eta^{\pm_2}({\bf Q},{\bf P})
=&\frac{1}{2}\delta({\bf Q})\delta({\bf P})
-\left[1\pm_2\cos\theta\left(\frac{P_1P_2}{\hbar\varrho V_{\rm m}\sqrt{\Omega_1\Omega_2}}+Q_1Q_2\frac{\varrho V_{\rm m}\sqrt{\Omega_1\Omega_2}}{\hbar}\right)\pm_2\sin\theta\left(P_1Q_2\sqrt{\frac{\Omega_2}{\hbar^2\Omega_1}}-P_2Q_1\sqrt{\frac{\Omega_1}{\hbar^2\Omega_2}}\right)
\right]\nonumber\\
&\times\frac{1}{8\pi^2\hbar^2}\exp\left[-\frac{1}{4\hbar\varrho V_{\rm m}}\left({\bf P}\cdot\Omega^{-1}{\bf P}+\varrho^2 V_{\rm m}^2{\bf Q}\cdot\Omega{\bf Q}\right)\right].\end{aligned}$$
The geometric factor $U_<(\sigma)$
----------------------------------
Assuming a continuous mass density, valid if $\hbar/\sigma_q\gg 5$Å, the geometric factor can be evaluated for the longitudinal mode as $$\begin{aligned}
\label{eq:U}
U_<(\sigma)\simeq&U(\sigma)=\frac{2\varrho^2\hbar^7}{m_e^2\sigma_q^7L_z^3}\left(1-e^{-L_x^2\sigma_q^2/2\hbar^2}+\frac{\sqrt{\pi}L_x\sigma_q}{\sqrt{2}\hbar}{\rm erf}\left[\frac{L_x\sigma_q}{\sqrt{2}\hbar}\right]\right)^2\nonumber\\
&\times\left[\sqrt{2\pi}\left(h\left[\frac{L_z\sigma_q}{\hbar},0\right]+e^{-L_z^2\sigma_q^2/2\hbar^2}{\rm Re}\left\{h\left[\frac{L_z\sigma_q}{\hbar},\frac{L_z\sigma_q}{\hbar}\right]\right\}\right)-\left(1+e^{-L_z^2\sigma_q^2/2\hbar^2}\right)\frac{L_z\sigma_q}{\hbar}\left(\pi^2-2\frac{L_z^2\sigma_q^2}{\hbar^2}\right)\right],\end{aligned}$$ with $$\begin{aligned}
h[a,b]=\sqrt{\frac{\pi}{2}}\left(i3a^2+\pi b^2-i\pi^2\right)\exp\left[\frac{(\pi/a-ib)^2}{2}\right]{\rm erf}\left[\frac{i \pi/a+b}{\sqrt{2}}\right].\end{aligned}$$
If $\hbar/\sigma_q$ is on the order of the lattice constant, the approximation of a continuous mass density fails. For even smaller $\hbar/\sigma_q$ the Gaussian in suppress all contributions involving more than a single atom, so that the modification acts on each of the $N$ atoms individually. The geometric factor then reads as $$\begin{aligned}
U_<(\sigma)=N\frac{m_{\rm Si}^2}{4\hbar^2m_e^2}{\rm erf}\left(\frac{q_c}{\sqrt{2}\sigma_q}\right)^2\left[\sigma_q^2{\rm erf}\left(\frac{q_c}{\sqrt{2}\sigma_q}\right)^2-\sqrt{\frac{2}{\pi}}\sigma_q q_c e^{-q_c^2/2\sigma_q^2}\right].\end{aligned}$$ Here, we averaged the mode function over the whole crystal, $\sum_n{\mathbf w}^2({\mathbf r}_n)\approx N/2$. As a result, the diffusion increases quadratically with $\sigma_q$ until the momentum displacements are strong enough to remove the particles from the crystal. In the limit that $\sigma_q\ll q_c$ one obtains $U_<(\sigma) \simeq Nm_{\rm Si}^2\sigma_q^2/4\hbar^2m_e^2$.
[61]{} 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 [****, ()](http://stacks.iop.org/0953-8984/14/i=15/a=201) [****, ()](\doibase
10.1103/RevModPhys.85.471) @noop [****, ()]{} [****, ()](\doibase 10.1126/science.290.5492.773) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1038/ncomms1263) @noop [****, ()]{} [****, ()](\doibase
10.1038/ncomms3077) [****, ()](\doibase
10.1038/nature16155) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevX.5.011003) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.121.220404) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.75.042106) [****, ()](\doibase 10.1103/PhysRevA.78.012109) [****, ()](http://stacks.iop.org/1367-2630/14/i=9/a=093039) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevX.8.041038) [****, ()](http://stacks.iop.org/1464-4266/6/i=11/a=001) [****, ()](\doibase 10.1103/PhysRevLett.106.220401) [****, ()](\doibase 10.1103/PhysRevLett.110.160403) @noop [**]{} (, ) @noop [****, ()]{} @noop [“,” ]{} () [****, ()](\doibase 10.1103/PhysRevA.90.052119) [****, ()](\doibase 10.1103/PhysRevLett.113.020405) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.119.110401) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.6.2211) [****, ()](\doibase 10.1103/PhysRevA.47.5138) [****, ()](\doibase
10.1016/j.physrep.2011.08.003) [****, ()](\doibase 10.1103/PhysRevA.85.022322) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.78.4675) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.86.4431) [****, ()](\doibase
10.1103/PhysRevLett.112.155304) @noop [****, ()]{} @noop [****, ()]{} [**** (), 10.1038/ncomms5009](\doibase 10.1038/ncomms5009) @noop [****, ()]{} @noop [**]{}, Vol. (, ) @noop [**]{} (, ) @noop [ ]{} [****, ()](\doibase 10.1103/PhysRevB.43.14248) [****, ()](\doibase 10.1103/PhysRevB.60.284)
[^1]: Starting with the lower hyperfine state one obtains the mirrored version of the distribution .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'By defining information entropy in terms of probabilities densities $|\Psi|^2$ ($\Psi$ is a wave function in the coordinate representation) it is explicitly shown how a loss of quantum information occurs in a transition from a quantum to a quasi-classical regime.'
author:
- 'Alex Granik[^1]'
title: 'On a loss of information in a transition from quantum to a quasi-classical regime'
---
The Boltzmann-Gibbs-Shannon entropy $$\label{eq:1} I=-\sum_{i=1}^Np_iLogp_i$$ ($p_i$ is the classical probability) has an obvious relation to von Neumann entropy $$S=-Tr\rho Log\rho$$ where the diagonal terms of the density matrix $\rho$ are the probabilties $p_i$. Usually in quantum mechanics the probability $p_i$ is considered as a function of energy $E_i$ (e.g. [@LL]). Since we study a rather narrow problem of information dynamics when the probability is described in terms of the wave function $\Psi$, we will not discuss a more sophisticated (and complete) approach based on the density matrix and information processing involving quantum bits (qubits) ( e.g., [@CA]).
To achieve our goal ,we proceed along the lines used in derivation of continuity equation for probability density. To this end consider the probability in its coordinate representation. Thus $p_i$ can be viewed as a probability of finding a particle in a spatial interval $\Delta q_i=q_{i+1}-q_i$: $$\label{eq:2}
p_i=\int_{q_i}^{q_{i+1}} \Psi(q)^*\Psi(q) dq$$ Using the mean value theorem, we rewrite Eq.(\[eq:2\]) as follows: $$\label{eq:3} p_i=|\Psi_i(\tilde{q}_i)|^2\Delta q_i,~~q_i\leq
\tilde{q_i}\leq q_{i+1}$$ where the sum of $p_i$ over all spatial intervals is $$\label{eq:4} \sum_{i=1}^{\infty} p_i
=\sum_{i=1}^{\infty}|\Psi_i(\tilde{q}_i)|^2\Delta q_i=1$$\
Using Eq.(\[eq:3\]) we get $$\label{eq:5} p_iLog p_i=|\Psi_i|^2\Delta q_i[Log(|\Psi_i|^2)+Log
(\Delta q_i)]$$ Now we take the following limit of (\[eq:5\]) $$\Delta q_i
\rightarrow 0; ~~~\Delta q_i \rightarrow dq$$ As a result, Eq.(\[eq:5\]) yields $$\label{eq:6} p_iLog
p_i\rightarrow|\Psi(q)|^2dq(Log|\Psi(q)|^2-1)+O(|\Delta q|^2)$$
and equation (\[eq:1\]) becomes respectively: $$\label{eq:7} I=-\int|\Psi(q)|^2(Log|\Psi(q)|^2-1)dq$$ Therefore we narrow our discussion even more by restricting it with continuous spectra.
In a 3-D case the integration Eq.(\[eq:7\]) is carried over a spatial volume $d^3q.$ This means that the expression $$\label{eq:8} \rho_I= -|\Psi(q)|^2(Log|\Psi(q)|^2-1)$$ can be interpreted as the volume information density $$\label{eq:A1}
\rho_I=dI/dV$$
We find its time variation: $$\label{eq:9}
\frac{\partial \rho_I}{\partial t}=
-\frac{\partial|\Psi(q)|^2(Log|\Psi(q)|^2-1)}{\partial
t}=-\frac{\partial |\Psi|^2}{\partial t}Log|\Psi|^2$$ If we use the continuity equation for probability density (e.g.,[@LL1]) $$\frac{\partial|\Psi|^2}{\partial t}=-
\frac{\hbar}{2im}div(|\Psi|^2\nabla Log\frac{\Psi}{\Psi^*})$$ then we obtain from Eq.(\[eq:9\]) $$\begin{aligned}
\label{eq:10}
\frac{\partial \rho_I}{\partial t}=
\frac{\hbar}{2im}div(|\Psi|^2\nabla Log\frac{\Psi}{\Psi^*})Log|\Psi|^2 \equiv\nonumber \\
div(|\Psi|^2Log|\Psi|^2\frac{\hbar}{2im}\nabla
Log\frac{\Psi}{\Psi^*})-\frac{\hbar}{2im}\nabla
Log\frac{\Psi}{\Psi^*}\bullet \nabla |\Psi|^2\end{aligned}$$ Inserting the definition of the information density $\rho_I$, Eq.(\[eq:8\]), and denoting the probability density $|\Psi|^2=\rho$ into the first term on the right hand side of Eq.(\[eq:10\]) , we obtain $$\label{eq:11} \frac{\partial\rho_I}{\partial t}
=-div[\frac{\hbar}{2im}(\rho_I-\rho)\nabla
Log\frac{\Psi}{\Psi^*}]-\frac{\hbar}{2im}\nabla
Log\frac{\Psi}{\Psi^*}\bullet \nabla |\Psi|^2$$\
Let us consider the transition to the classical case by representing the wave function as follows $$\label{eq:12} \Psi=\sqrt\rho e^{iS/\hbar}$$ Using (\[eq:12\]) in (\[eq:11\]), we get $$\label{eq:13} \frac{\partial\rho_I}{\partial t} =-div[\frac{\nabla
S}{m}(\rho_I-\rho)]-\frac{\nabla S}{m}\bullet \nabla \rho$$ In the classical limit $\hbar t_c/m L_c^2 \rightarrow 0$ ( where $L_c$ is a characteristic length and $t_c$ is the characteristic time) phase $S$ becomes a classical action. As a result $\nabla
S/m =\vec{v}$, that is a classical particle velocity. This means that in this limit Eq.(\[eq:13\]) yields $$\label{eq:14} \frac{\partial\rho_I}{\partial t}
=-div[\vec{v}(\rho_I-\rho)]-\vec{v}\bullet \nabla\rho$$\
If we integrate (\[eq:14\]) over a volume $V$, use definition of the information entropy (\[eq:A1\]), and Gauss’s theorem, we obtain $$\label{eq:15} \frac{\partial}{\partial t}\int\rho_i
dV=\frac{\partial I}{\partial t}= -\oint
(\rho_I-\rho)\vec{v}\bullet \vec{dA}- \int \vec{v}\bullet
\nabla\rho dV$$
By taking the boundary of the volume $V$ to infinity and assuming that $\rho(q\rightarrow\infty)\rightarrow 0$ we obtain from (\[eq:15\]) $$\label{eq:16} \frac{\partial I}{\partial t}= -\int \vec{v}\bullet
\nabla\rho dV$$
This means that in a transition from a quantum to classical regime the information entropy $I$ \[defined in the above narrow sense, Eq.(\[eq:7\])\] is not conserved. Instead, if the classical velocity is in the direction of an increase of probability density, the information entropy decreases. Inversely,if the classical velocity is in the direction of a decreasing probability density, the information density increases.\
These results are in agreement with the meaning of the information entropy as a number of states accessible to a system. In a transition to a classical regime this number drastically decreases, thus signaling a decrease in information entropy. Quite in agreement with that, and from another point of view, the less probable states carry more information than the more probable states.
One can interpret these results as a statement that a classical regime has less degree of freedom than its quantum counterpart. This is definitely true when there is such a counterpart. As we have already stated earlier, the above definition of information entropy is not fully appropriate for quantum mechanics because it does not take into account the information carried by a quantum phase. This serves as a strong indication that in the quantum region one needs to use another, more general, definition of information entropy , not necessarily associated with the qubits, which will account for the information associated with the quantum phase [@AG].\
[**[Acknowledgments]{}**]{}
The author thanks V.Panico and C.Wulfman for the illuminating discussions of the results.\
[99]{} L.Landau and E.Lifshitz, Statistical Physics,Pergamon Press, 1980 N.Cerf and C.Adams, Negative Entropy in quantum Information Theory, quant-ph/9610005. L.Landau and E.Lifshitz, Quantum Mechanics,Pergamon Press, 1974 A.Granik, G.Chapline, in preparation
[^1]: Department of Physics,UOP,Stockton,CA.95211;E-mail:[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Laser spectroscopy studies are being prepared to measure the $1s$ ground state hyperfine splitting in trapped cold highly charged ions. The purpose of such experiments is to test quantum electrodynamics in the strong electric field regime. These experiments form part of the HITRAP project at GSI. A brief review of the planned experiments is presented.'
address: 'Blackett Laboratory, Imperial College, Prince Consort Road, London SW7 2BW, United Kingdom'
author:
- 'D.F.A. Winters'
- 'A.M. Abdulla, J.R. Castrejón Pita, A. de Lange'
- 'D.M. Segal, R.C. Thompson'
title: 'Plans for laser spectroscopy of trapped cold hydrogen-like HCI'
---
,
,
Highly charged ion ,hyperfine splitting ,laser spectroscopy 32.30.Jc ,12.20.Fv ,21.10.Ky
Introduction
============
An accurate measurement of the hyperfine splitting (HFS) of the $1s$ ground state of hydrogen-like highly charged ions (HCI) is a good test of quantum electrodynamics (QED) in the limit of strong electric fields ($10^{15}$ V/cm [@bei00]). Such strong fields cannot be produced using conventional laboratory techniques, but naturally exist close to the stripped nuclei of heavy elements like Pb, Bi or U. By detecting the fluorescence from the laser-excited upper hyperfine state of such a trapped and cold ion, a high-precision measurement of the hyperfine splitting can be made.
HCI with extremely high charge states ([*e.g.*]{} Pb$^{81+}$ or U$^{92+}$) and relativistic energies (400 MeV/u) are created in the heavy ion facility (SIS) at GSI in Darmstadt. These HCI are injected into the Experimental Storage Ring (ESR) and can be decelerated (down to 4 MeV/u) forming 1 $\mu$s long bunches containing about $10^5$ ions, arriving every 10 seconds. In the HITRAP project [@qui01; @tdr03] these bunches will be extracted from the ESR, decelerated by linear (IH-LINAC) and radiofrequency (RFQ) stages, trapped and cooled in a Penning trap (cooler trap), and made available for experiments.
In the ESR at GSI, previous measurements of the HFS were made on bunches of relativistic HCI such as $^{209}$Bi$^{82+}$ [@kla94] and $^{207}$Pb$^{81+}$ [@see98]. HFS measurements were also made at the SuperEBIT on $^{165}$Ho$^{66+}$ [@cre96], $^{185,187}$Re$^{74+}$ [@cre98] and $^{203,205}$Tl$^{80+}$ [@bei01]. The resolution obtained in the above experiments is mainly limited by the Doppler effect. A measurement of the $1s$ ground state HFS of trapped cold HCI using laser spectroscopy should be even more accurate due to a cryogenic UHV environment, high ion cloud density, the absence of a large Doppler shift and virtually unlimited measurement time.
Laser spectroscopy offers the possibility of high-accuracy measurements of transition wavelengths in the visible region [@tho85]. In HCI, electronic transitions are generally in the far UV or X-ray regions of the spectrum. However, since the $1s$ groundstate HFS scales with the atomic number $Z$ as $Z^3$, the ground state HFS of hydrogen-like HCI can move into the visible spectrum for $Z>70$ [@bei00; @sha94]. The lifetime of this transition falls as $Z^{-9}$ and is of the order of milliseconds for $Z>70$. A measurement of this transition wavelength gives information on the QED corrections to the HFS or on the spatial distribution of the nuclear magnetisation (Bohr-Weisskopf effect), which is affected by core polarisation and is not really well understood [@bei00]. Its measurement thus allows for critical tests of nuclear models. From a comparison of measurements of the HFS of hydrogen-like and lithium-like HCI the nuclear effects can be eliminated so that an accurate measurement of the QED effects can be made [@tom98; @sun98].
There are several candidate systems that can be studied at HITRAP, including radioactive isotopes. An interesting first challenge would be to measure the HFS of the $1s$ ground state $M1$ transition of $^{207}$Pb$^{81+}$ ions ($\lambda=1020$ nm [@see98]) by means of laser spectroscopy. In order to reach the necessary accuracy, the HCI will be trapped in a cryogenic UHV environment. Once trapped, the ions can be easily stored for long times, therefore the lifetime $\tau=50$ ms [@see98] of the upper hyperfine state is not a problem. Electron capture (neutralisation) by collisions is strongly reduced by operating the trap at cryogenic temperatures (4 K) under UHV conditions (below $10^{-14}$ mbar).
![Zeeman and hyperfine splittings of the $1s$ ground states of $I=1/2$ nuclei plotted versus magnetic field: a) hydrogen atom $^1$H$^0$, b) highly charged lead ion $^{207}$Pb$^{81+}$. (The plots are drawn with different scales.)[]{data-label="fig1"}](fig1.eps){width="7.5cm"}
Zeeman versus hyperfine splitting
=================================
The high precision measurements of the HFS will be performed on HCI in a Penning trap, but the strong magnetic field will shift and split the hyperfine levels and thus the natural HFS. This ‘Zeeman effect’ [@zee96] is schematically illustrated in figure 1a) for the $1s$ ground state HFS in hydrogen, which has nuclear spin $I=1/2$. For hydrogen the HFS can be easily calculated using the famous Breit-Rabi formula [@bre31]. The zero-field splitting $A$ is modified by the interaction between the magnetic moment $\mu$ of the ion and the magnetic field strength $B$. In the simple hydrogen case at a field of 1 T, the Zeeman splitting of about $2.8 \cdot 10^{10}$ Hz is more than one order of magnitude larger than the HFS in hydrogen ($1.4 \cdot 10^9$ Hz or 21 cm). However, for HCI the HFS dominates the Zeeman splitting by several orders of magnitude. This is indicated in figure 1b) for the $1s$ ground state HFS of $^{207}$Pb$^{81+}$, which also has nuclear spin $I=1/2$.
![Wavelengths of the $1s$ ground state hyperfine splittings in the visible spectrum for atomic number $Z$ ranging from 50 to 100.[]{data-label="fig2"}](fig2.eps){width="7cm"}
In figure 2 the calculated wavelengths of the $1s$ ground state hyperfine transition (neglecting QED and the Bohr-Weisskopf effect) are plotted for HCI with atomic number $Z$ ranging from 50 to 100. Most of these heavier elements have their HFS in the visible region of the spectrum and are thus easily excited by conventional tunable laser systems.
The spectroscopy trap
=====================
The HCI will be extracted from the HITRAP cooler trap in a long narrow bunch with an average kinetic energy of a few eV and containing about $10^5$ ions. The radial energy spread will be reduced so as to maintain a parallel beam [@tdr03]. The HCI will then be loaded into the spectroscopy trap, which is a cylindrical open-endcap Penning trap [@gab89] with compensation electrodes to create a nearly perfect quadrupole potential at the trap centre. Figure 3 schematically shows the expected layout of the electrodes, the tank circuit for resistive cooling [@win75; @ver04] and the trap loading scheme: the HCI enter from the right, are reflected by the left capture electrode, enclosed by the right one, localised near the trap centre and finally cooled and compressed.
![Schematic of the proposed spectroscopy trap and the trap loading scheme.[]{data-label="fig3"}](fig3.eps){width="7.5cm"}
The principle of resistive cooling is fairly straightforward and well-known: the axial motion of the HCI in the trap induces image charges in the open-endcap electrodes, which can readily be detected by a frequency resonant electronic tank circuit (LCR-filter) attached to the electrodes. If the tank circuit is at cryogenic temperatures (4 K), kinetic energy of the HCI is dissipated by the circuit, thus effectively cooling the ions. Since the cooling time varies with the charge state $q$ of the ions as $1/q^2$, this cooling scheme is very effective for HCI [@sta04].
After the cooling procedure, the HCI will be focussed into a smaller cloud using the rotating wall technique [@dub99; @hua98]. If the coupling of an ion plasma in a Penning trap is strong enough, the density of the ion cloud depends on its global rotation frequency. By applying RF voltages to a segmented ring electrode, the radial electric quadrupole field will rotate and drive the ion cloud. There are two limits for the rotation frequency $\nu$. The lower limit, which leads to the minimal density, is reached when $\nu$ equals the single-ion magnetron frequency $\nu_m$. At the upper limit, when $\nu$ is set to half the cyclotron frequency $\nu_c$, the density is maximum (Brillouin limit). (See [*e.g.*]{} [@win83; @tho93] for the equations of motion of an ion in a Penning trap.)
Realistic values for the Penning trap are: a magnetic field of 6 T, an applied potential of 500 V, an ion cloud temperature of 4 K, and a trap parameter $d=17$ mm (see Ref [@gab89]). These values lead to the trap frequencies $\nu_m=23$ kHz, $\nu_c=36$ MHz, and an axial frequency $\nu_z=1.3$ MHz. If we set the rotation frequency to $\nu = 1$ MHz, the number density of a cloud of $10^5$ $^{207}$Pb$^{81+}$ ions is $5 \cdot 10^7$ cm$^{-3}$ (the Brillouin limit is $4.6 \cdot 10^8$ cm$^{-3}$). Such an ion cloud has a length of 6.8 mm and a diameter of 0.75 mm, which leads to an aspect ratio of 9.0. The Debye length of this cloud is 250 nm. The dependence of the ion cloud parameters on $\nu$ is shown in figure 4.
![The number density, length, aspect ratio and diameter of an ion cloud containing $10^5$ $^{207}$Pb$^{81+}$ ions, plotted as a function of the rotation frequency $\nu$. ($d=37.5$ mm)[]{data-label="fig4"}](fig4.eps){width="7.5cm"}
A spectrum of the ground state hyperfine transition can be obtained by scanning the laser wavelength across the resonance while recording the fluorescence from the trapped decaying HCI. Repeated measurements on a single cloud of ions can be made, as detection is not destructive. To obtain the highest spectral resolution, the laser beam must be parallel to the magnetic field (trap axis), so as to avoid the large Doppler shift associated with the global rotation of the ion cloud. Laser excitation along the trap axis, on an ion cloud smaller than the laser beam, will also help to ensure that the transition is fully saturated.
Calculated expected signal rates
================================
An ion temperature of 4 K corresponds to a Doppler width of 30 MHz. For $^{207}$Pb$^{81+}$ ions, the wavelength of 1020 nm corresponds to an upper state lifetime of 50 ms and a linewidth of 3 Hz. The laser power to fully saturate the Doppler-broadened transition is roughly 3 mW and is easily reached by [*e.g.*]{} a Ti:Sapphire or Ar$^+$ pumped dye laser. If the experiment is run in a continuous mode, assuming an overall detection efficiency of $4 \cdot 10^{-3}$, we expect to detect about $4 \cdot 10^3$ photon emissions per second for a completely saturated ion cloud. It should be possible to reduce the background signal to less than $10^2$ counts per second, yielding a typical S/N ratio of 40. Alternatively, if the laser excitation is pulsed with a duty cycle of 200 ms ($=4 \tau$), the signal is about $10^3$ counts per second, without any background from scattered laser light. These values are high enough to allow easy detection and measurement of the transition wavelength. Once the signal is seen, this will allow a wavelength determination to an accuracy which far exceeds the theoretical uncertainties. The linewidth of [*e.g.*]{} a Ti:Sapphire laser is 1 MHz at a wavelength of 1 $\mu$m, which is much smaller than the Doppler broadening of 30 MHz. Therefore the accuracy of the HFS measurement is estimated to be about $10^{-7}$, which is 3 orders of magnitude better than the current accuracy [@see98].
Acknowledgements
================
This work is supported by the European Commission within the RTD programme FP5 (HPRI-CT-2001-50036 HITRAP). JRCP acknowledges the support by CONACyT, SEP and the ORS Awards Scheme.
[99]{}
T. Beier, Phys. Rep. [**339**]{}, 79 (2000).
W. Quint [*et al.*]{}, Hyp. Int. [**132**]{}, 457 (2001).
www.gsi.de/documents/DOC-2003-Dec-69-2.pdf
I. Klaft [*et al.*]{}, Phys. Rev. Lett. [**73**]{}, 2425 (1994).
P. Seelig [*et al.*]{}, Phys. Rev. Lett. [**81**]{}, 4824 (1998).
J.R. Crespo López-Urrutia, P. Beiersdorfer, D.W. Savin, K. Widmann, Phys. Rev. Lett. [**77**]{} 826 (1996).
J.R. Crespo López-Urrutia, P. Beiersdorfer, K. Widmann, B.B. Birkett, A.-M. Mårtensson-Pendrill, M.G.H. Gustavsson, Phys. Rev. A [**57**]{}, 879 (1998).
P. Beiersdorfer [*et al.*]{}, Phys. Rev. A [**64**]{}, 032506 (2001).
R.C. Thompson, Rep. Prog. Phys. [**48**]{}, 531 (1985).
V.M. Shabaev, J. Phys. B: At. Mol. Opt. Phys. [**27**]{}, 5825 (1994).
M. Tomaselli, T. Kühl, P. Seelig, C. Holbrow, E. Kankeleit, Phys. Rev. C [**58**]{}, 1524 (1998).
P. Sunnergren [*et al.*]{} Phys. Rev. A [**58**]{}, 1055 (1998).
P. Zeeman, Versl. Kon. Ak. Wet. [**5**]{}, 181 (1896).
G. Breit, I.I. Rabi, Phys. Rev. [**38**]{}, 2082 (1931).
G. Gabrielse, L. Haarsma, S.L. Rolston, Int. J. Mass Spectr. Ion Proc. [**88**]{}, 319 (1989).
D.J. Wineland, H.G. Dehmelt, J. Appl. Phys. [**46**]{}, 919 (1975).
J. Verdú, S. Djekić, S. Stahl, T. Valenzuela, M. Vogel, G.Werth, T. Beier, H.-J. Kluge, W. Quint, Phys. Rev. Lett. [**92**]{}, 093002 (2004).
S. Stahl (private communication).
D.H.E. Dubin, T.M. O’Neil, Rev. Mod. Phys. [**71**]{}, 87 (1999).
X.-P. Huang, J.J. Bollinger, T.B. Mitchell, W.M. Itano, Phys. Rev. Lett. [**80**]{}, 73 (1998).
D.J. Wineland, W.M. Itano, R.S. Van Dyck, Adv. At. Mol. Phys. [**19**]{}, 135 (1983).
R.C. Thompson, Adv. At. Mol. Opt. Phys. [**31**]{}, 63 (1993).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the behavior as $t\to 0^+$ of nonnegative functions $$\label{0.1}
u\in C^{2,1} (\mathbb{R}^n\times (0,1))
\cap L^\lambda (\mathbb{R}^n\times (0,1)),\quad n\ge 1,$$ satisfying the parabolic Choquard-Pekar type inequalities $$\label{0.2}
0\leq u_t-\Delta u\leq(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad
\text{ in }B_1 (0)\times (0,1)$$ where $\alpha\in(0,n+2)$, $\lambda>0$, and $\sigma\geq0$ are constants, $\Phi$ is the heat kernel, and $*$ is the convolution operation in $\mathbb{R}^n\times (0,1)$. We provide optimal conditions on $\alpha,\lambda$, and $\sigma$ such that nonnegative solutions $u$ of , satisfy pointwise bounds in compact subsets of $B_1(0)$ as $t\to 0^+$. We obtain similar results for nonnegative solutions of , when $\Phi^{\alpha/n}$ in is replaced with the fundamental solution $\Phi_\alpha$ of the fractional heat operator $(\frac{\partial}{\partial t}-\Delta)^{\alpha/2}$.
address: 'Mathematics Department, Texas A&M University, College Station, TX 77843-3368'
author:
- 'Steven D. Taliaferro'
title: 'Initial Pointwise Bounds and Blow-up for Parabolic Choquard-Pekar Inequalities'
---
Introduction {#sec1}
============
In this paper we study the behavior as $t\to 0^+$ of nonnegative functions $$\label{1.1}
u\in C^{2,1} (\mathbb{R}^n\times (0,T))
\cap L^\lambda (\mathbb{R}^n\times (0,T)),\quad n\ge 1,$$ satisfying the nonlocal parabolic Choquard-Pekar type inequalities $$\label{1.2}
0\leq Hu\leq(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad
\text{ in }\Omega\times (0,T)$$ where $\alpha\in(0,n+2)$, $\lambda>0$, $\sigma\geq0$, and $T>0$ are constants, $\Omega$ is an open subset of $\mathbb{R}^n$, $Hu=u_t-\Delta u$ is the heat operator, $$\label{1.3}
\Phi(x,t)=\begin{cases}
\frac{1}{(4\pi t)^{n/2}}
e^{-\frac{|x|^2}{4t}}&\text{for }(x,t)\in \mathbb{R}^n\times (0,\infty)\\
0&\text{for }(x,t)\in \mathbb{R}^n\times (-\infty,0]
\end{cases}$$ is the heat kernel, and $*$ is the convolution operation in $\mathbb{R}^n\times(0,T)$, that is, $$(\Phi^{\alpha/n}*u^\lambda)(x,t)
=\iint_{\R^n\times(0,T)}\Phi(x-y,t-s)^{\alpha/n}u(y,s)^\lambda dy\,ds.$$ The regularity condition $u\in
L^\lambda(\mathbb{R}^n\times (0,T))$ in and the upper bound of $n+2$ for $\alpha$ are natural because one does not want the nonlocal convolution operation on the right side of to be infinite at every point in $\mathbb{R}^n\times (0,T)$.
We also obtain results on the behavior as $t\to 0^+$ of nonnegative solutions of , when $\Phi^{\alpha/n}$ in is replaced with the fundamental solution $\Phi_\alpha$ of the fractional heat operator $(\frac{\partial}{\partial t}-\Delta)^{\alpha/2}$. (See Remark \[rem2\].)
A motivation for the study of , comes from the nonlocal elliptic equation $$\label{prototype}
-\Delta u=(\Gamma^{\alpha/(n-2)}* u^{\lambda}) |u|^{\lambda-2}u\quad
\text{ in }\mathbb{R}^n,$$ where $\alpha\in (0,n)$, $\lambda>1$ and $\Gamma(x)=C(n)/|x|^{n-2}$ is a fundamental solution of $-\Delta$. For $n=3$, $\alpha=1$, and $\lambda=2$, equation is known in the literature as the [*Choquard-Pekar equation*]{} and was introduced in [@P1954] as a model in quantum theory of a polaron at rest (see also [@DA2010]). Later, the equation appears as a model of an electron trapped in its own hole, in an approximation to Hartree-Fock theory of one-component plasma [@L1976]. More recently, the same equation was used in a model of self-gravitating matter (see, e.g., [@J1995; @MPT1998]) and it is known in this context as the [*Schrödinger-Newton equation*]{}.
The Choquard-Pekar equation has been investigated for a few decades by variational methods starting with the pioneering works of Lieb [@L1976] and Lions [@Lions1980; @Lions1984]. More recently, new and improved techniques have been devised to deal with various forms of (see, e.g., [@MZ2010; @MZ2012; @MV2013a; @MV2013b; @MV2015; @WW2009] and the references therein).
Using nonvariational methods, the authors in [@MV2013b] obtained sharp conditions for the nonexistence of nonnegative solutions to $$-\Delta u \geq (\Gamma^{\alpha/(n-2)}* u^{\lambda}) u^{\sigma}$$ in an exterior domain of $\mathbb{R}^n$, $n\geq 3$.
For some very recent results on positive solutions Choquard-Pekar equations and inequalities which have an isolated singularity at the origin see [@CZ] and [@GT2016-2].
Other examples of nonlocal equations which have been studied extensively in recent years are equations containing the fractional Laplacian and some of these equations are equivalent to equations containing convolutions with powers of the fundamental solution $\Gamma$ of $-\Delta u$. For example, see [@Zhuo2016] and [@MCL2011].
On the other hand, we know of no results for nonlocal equations or inequalities when the nonlocal feature of the problem is due to convolutions with powers of the fundamental solution of the heat equation. Our results for , are, in this regard, new.
In this paper we consider the following question.
\[ques1\] Suppose $\alpha\in(0,n+2)$ and $\lambda>0$ are constants and $\Omega$ is an open subset of $\R^n$, $n\ge 1$. For which nonnegative constants $\sigma$, if any, does there exist a continuous function $\varphi:(0,1)\to(0,\infty)$ such that for all compact subsets $K$ of $\Omega$ and all nonnegative solutions $u$ of , we have $$\label{1.4}
\max_{x\in K}u(x,t)=O(\varphi(t))\quad\text{ as }t\to0^+$$ and what is the optimal such $\varphi$ when it exists?
We call the function $\varphi$ in a pointwise bound for $u$ on compact subsets of $\Omega$ as $t\to0^+$.
\[rem1\] Suppose $0<\lambda<(n+2)/n$. Then, since $u=\Phi$, where $\Phi$ is the heat kernel given by , is a solution of , and $\Phi(0,t)=(4\pi t)^{-n/2}$, we see that any pointwise bound for nonnegative solutions $u$ of , on compact subsets of $\Omega$ as $t\to0^+$ must be at least as large as $t^{-n/2}$ and whenever $t^{-n/2}$ is such a bound it is necessarily optimal.
In order to state our results for Question \[ques1\], we define for each $\alpha\in(0,n+2)$ the continuous, piecewise linear functions $g_\alpha,
G_\alpha:(0,\infty)\to [0,\infty)$ by $$\label{1.5}
g_\alpha (\lambda)=
\begin{cases}
\frac{n+2}{n} & \text{if }0<\lambda<\frac{n+2-\alpha}{n}\\
\frac{2(n+2)-\alpha}{n}-\lambda & \text{if }\frac{n+2-\alpha}{n}
\leq\lambda<\frac{n+2}{n}\\
\max\{0,1-\frac{\alpha-2}{n+2}\lambda \} & \text{if }
\lambda\geq\frac{n+2}{n}
\end{cases}$$ and $$G_\alpha (\lambda)=
\begin{cases}
\frac{2(n+2)-\alpha}{n}-\lambda & \text{if }
0<\lambda<\frac{n+2}{n}\\
\max\{0,1-\frac{\alpha-2}{n+2}\lambda \} & \text{if }
\lambda\geq\frac{n+2}{n}.
\end{cases}$$ These functions are graphed in Figure \[fig1\] (resp. Figure \[fig2\]) when $\alpha\in (2,n+2)$ (resp. $\alpha\in (0,2]$). Note that $$g_\alpha(\lambda)=G_\alpha(\lambda)
\quad\text{for }\frac{n+2-\alpha}{n}\le\lambda<\infty$$ and $$g_\alpha(\lambda)<G_\alpha(\lambda)
\quad\text{for }0<\lambda<\frac{n+2-\alpha}{n}.$$
According to the following theorem, if the point $(\lambda,\sigma)$ lies below the graph of $\sigma=g_\alpha(\lambda)$ then there exists a pointwise bound for nonnegative solutions $u$ of , on compact subsets of $\Omega$ as $t\to 0^+$ .
(6,3.2)(-0.2,0) (0.75,0.5)[(0,1)[2.50]{}]{} (0.75,0.5)[(1,0)[4.50]{}]{} (0.75,2.19)[(1,0)[0.75]{}]{} (0.75,2.94)[(1,-1)[1.69]{}]{} (2.44,0.50)[(0,1)[0.75]{}]{} (4.69,0.50)[(-3,1)[2.25]{}]{}
(1.20,0.12)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (2.14,0.12)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{} (4.39,0.12)[(0.6,0.3)[$\frac{n+2}{\alpha-2}$]{}]{} (0.12,1.10)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (0.12,2.04)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{}
(1.50,0.50)(0,0.2)[3]{}[(0,1)[0.1]{}]{} (1.50,2.19)(0,-0.2)[4]{}[(0,-1)[0.1]{}]{}
(0.75,1.25)(0.2,0)[2]{}[(1,0)[0.1]{}]{} (2.44,1.25)(-0.2,0)[3]{}[(-1,0)[0.1]{}]{}
(2.54,0.6)[(0.9,0.3)[$o\Bigl(t^{-\frac{n+2}{2\lambda}}\Bigr)$]{}]{} (1.1,1.10)[(0.8,0.3)[$O(t^{-n/2})$]{}]{} (3.6,0.8)[(1.2,0.3)[$\sigma=1-\frac{\alpha-2}{n+2}\lambda$]{}]{} (0.89,2.29)[(0.2,0.2)[?]{}]{} (2.03,1.6)[(1.5,0.3)[$\sigma=\frac{2(n+2)-\alpha}{n}-\lambda$]{}]{} (1.95,2.05)[(2.3,0.3)[*Arbitrarily large solutions*]{}]{}
(5.25,0.40)[(0.2,0.2)[$\lambda$]{}]{} (0.65,3.01)[(0.2,0.2)[$\sigma$]{}]{}
-0.25in
(6,3.2)(-0.2,0) (0.75,0.5)[(0,1)[2.50]{}]{} (0.75,0.5)[(1,0)[4.50]{}]{} (0.75,2.19)[(1,0)[0.75]{}]{} (0.75,2.94)[(1,-1)[1.69]{}]{} (2.44,0.50)[(0,1)[0.75]{}]{} (2.44,1.25)[(3,1)[2.75]{}]{}
(1.20,0.12)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (2.14,0.12)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{} (0.12,1.10)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (0.12,2.04)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{}
(1.50,0.50)(0,0.2)[3]{}[(0,1)[0.1]{}]{} (1.50,2.19)(0,-0.2)[4]{}[(0,-1)[0.1]{}]{}
(0.75,1.25)(0.2,0)[2]{}[(1,0)[0.1]{}]{} (2.44,1.25)(-0.2,0)[3]{}[(-1,0)[0.1]{}]{}
(2.90,0.75)[(0.9,0.3)[$o\Bigl(t^{-\frac{n+2}{2\lambda}}\Bigr)$]{}]{} (1.1,1.10)[(0.8,0.3)[$O(t^{-n/2})$]{}]{} (3.6,1.4)[(1.2,0.3)[$\sigma=1+\frac{2-\alpha}{n+2}\lambda$]{}]{} (0.89,2.29)[(0.2,0.2)[?]{}]{} (1.73,1.9)[(1.5,0.3)[$\sigma=\frac{2(n+2)-\alpha}{n}-\lambda$]{}]{} (1.95,2.35)[(2.3,0.3)[*Arbitrarily large solutions*]{}]{}
(5.25,0.40)[(0.2,0.2)[$\lambda$]{}]{} (0.65,3.01)[(0.2,0.2)[$\sigma$]{}]{}
-0.25in
\[thm1.1\] Suppose $u$ is a nonnegative solution of , where $\alpha\in(0,n+2)$, $\lambda>0$, $T>0$, and $$0\leq\sigma<g_\alpha (\lambda)$$ are constants and $\Omega$ is an open subset of $\R^n$. Then for each compact subset $K$ of $\Omega$ we have as $t\to 0^+$ that
[\_[xK]{}u(x,t)=]{} O(t\^[-n/2]{}) & if $0<\lambda<\frac{n+2}{n}$ \[1.6\]\
o(t\^[-(n+2)/(2)]{}) & if $\lambda\ge\frac{n+2}{n}.$ \[1.7\]
The estimate is optimal by Remark \[rem1\]. The exponent $-(n+2)/(2\lambda)$ in is also optimal by the following result.
\[thm1.2\] Suppose $$\lambda\ge\frac{n+2}{n}\quad\text{and}\quad
\gamma=\frac{n+2-\varepsilon}{2\lambda}$$ for some $\varepsilon\in(0,1)$. Then there exists a $C^\infty$ positive solution $u$ of $$Hu=0\quad\text{in }\R^n\times (0,\infty)$$ such that $$u\in L^\lambda(\R^n\times(0,T))\quad\text{for all } T>0$$ and $$u(0,t)=t^{-\gamma}\quad\text{for all } t>0.$$
By the next theorem, if the point $(\lambda,\sigma)$ lies above the graph of $\sigma=G_\alpha(\lambda)$ then there does not exist a pointwise bound for nonnegative solutions $u$ of , on compact subsets of $\Omega$ as $t\to 0^+$.
\[thm1.3\] Suppose $\alpha,\lambda$, and $\sigma$ are constants satisfying $$\alpha\in(0,n+2), \quad \lambda>0, \quad \text{and}\quad\
\sigma>G_\alpha (\lambda).$$ Let $\varphi:(0,1)\to(0,\infty)$ be a continuous function satisfying $$\lim_{t\to0^+}\varphi(t)=\infty.$$ Then there exists a positive solution $u$ of , with $T=1$ and $\Omega=\R^n$ such that $$u(0,t)\neq O(\varphi(t))\quad\text{ as }t\to 0^+.$$
Theorems \[thm1.1\]–\[thm1.3\] completely answer Question \[ques1\] when the point $(\lambda,\sigma)$ lies below the graph of $g_\alpha$ or above the graph of $G_\alpha$. In particular, if $u$ is a nonnegative solution of , where $(\lambda,\sigma)$ lies in the first quadrant of the $\lambda\sigma$-plane and either $\sigma<
g_\alpha(\lambda)$ or $\sigma> G_\alpha(\lambda)$ then according to Theorems \[thm1.1\]–\[thm1.3\] either
1. $\varphi(t)=t^{-n/2}$ is an optimal a priori pointwise bound for $u$ on compact subsets of $\Omega$ as $t\to 0^+$; or
2. $\varphi(t)=t^{-(n+2)/(2\lambda)}$ is an optimal a priori pointwise bound for $u$ on compact subsets of $\Omega$ as $t\to
0^+$; or
3. no pointwise a priori bound exists for $u$ on compact subsets of $\Omega$ as $t\to 0^+$, that is solutions can be arbitrarily large as $t\to 0^+$.
The regions in which these three possibilities occur are shown in Figures \[fig1\] and \[fig2\]. Also included in Figures \[fig1\] and \[fig2\] is an open triangular region marked with a question mark. For $(\lambda,\sigma)$ in this region we have no results for Question \[ques1\].
(6,3.2)(-0.2,0) (0.75,0.5)[(0,1)[2.50]{}]{} (0.75,0.5)[(1,0)[4.50]{}]{} (0.75,2.19)[(1,0)[0.75]{}]{} (0.75,2.94)[(1,-1)[1.69]{}]{} (2.44,0.50)[(0,1)[0.75]{}]{} (4.69,0.50)[(-3,1)[2.25]{}]{}
(1.20,0.12)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (2.14,0.12)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{} (4.39,0.12)[(0.6,0.3)[$\frac{n+2}{\alpha-2}$]{}]{} (0.12,1.10)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (0.12,2.04)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{}
(1.50,0.50)(0,0.2)[3]{}[(0,1)[0.1]{}]{} (1.50,2.19)(0,-0.2)[4]{}[(0,-1)[0.1]{}]{}
(0.75,1.25)(0.2,0)[2]{}[(1,0)[0.1]{}]{} (2.44,1.25)(-0.2,0)[3]{}[(-1,0)[0.1]{}]{}
(2.54,0.6)[(0.9,0.3)[$o\Bigl(t^{-\frac{n+2}{2\lambda}}\Bigr)$]{}]{} (1.1,1.10)[(0.8,0.3)[$O(t^{-n/2})$]{}]{} (3.6,0.8)[(1.2,0.3)[$\sigma=1-\frac{\alpha-2}{n+2}\lambda$]{}]{} (0.89,2.29)[(0.2,0.2)[?]{}]{} (2.03,1.6)[(1.5,0.3)[$\sigma=\frac{2(n+2)-\alpha}{n}-\lambda$]{}]{} (1.95,2.05)[(2.3,0.3)[*Arbitrarily large solutions*]{}]{}
(5.25,0.40)[(0.2,0.2)[$\lambda$]{}]{} (0.65,3.01)[(0.2,0.2)[$\sigma$]{}]{}
-0.25in
(6,3.2)(-0.2,0) (0.75,0.5)[(0,1)[2.50]{}]{} (0.75,0.5)[(1,0)[4.50]{}]{} (0.75,2.19)[(1,0)[0.75]{}]{} (0.75,2.94)[(1,-1)[1.69]{}]{} (2.44,0.50)[(0,1)[0.75]{}]{} (4.69,0.50)[(-3,1)[2.25]{}]{}
(1.20,0.12)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (2.14,0.12)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{} (4.39,0.12)[(0.6,0.3)[$\frac{n+2}{\alpha-2}$]{}]{} (0.12,1.10)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (0.12,2.04)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{}
(1.50,0.50)(0,0.2)[3]{}[(0,1)[0.1]{}]{} (1.50,2.19)(0,-0.2)[4]{}[(0,-1)[0.1]{}]{}
(0.75,1.25)(0.2,0)[2]{}[(1,0)[0.1]{}]{} (2.44,1.25)(-0.2,0)[3]{}[(-1,0)[0.1]{}]{}
(2.54,0.6)[(0.9,0.3)[$o\Bigl(t^{-\frac{n+2}{2\lambda}}\Bigr)$]{}]{} (1.1,1.10)[(0.8,0.3)[$O(t^{-n/2})$]{}]{} (3.6,0.8)[(1.2,0.3)[$\sigma=1-\frac{\alpha-2}{n+2}\lambda$]{}]{} (0.89,2.29)[(0.2,0.2)[?]{}]{} (2.03,1.6)[(1.5,0.3)[$\sigma=\frac{2(n+2)-\alpha}{n}-\lambda$]{}]{} (1.95,2.05)[(2.3,0.3)[*Arbitrarily large solutions*]{}]{}
(5.25,0.40)[(0.2,0.2)[$\lambda$]{}]{} (0.65,3.01)[(0.2,0.2)[$\sigma$]{}]{}
-0.25in
(6,3.2)(0.3,0) (0.75,0.5)[(0,1)[2.50]{}]{} (0.75,0.5)[(1,0)[4.50]{}]{} (0.75,2.19)[(1,0)[0.75]{}]{} (0.75,2.94)[(1,-1)[1.69]{}]{} (2.44,0.50)[(0,1)[0.75]{}]{} (2.44,1.25)[(3,1)[2.75]{}]{}
(1.20,0.12)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (2.14,0.12)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{} (0.12,1.10)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (0.12,2.04)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{}
(1.50,0.50)(0,0.2)[3]{}[(0,1)[0.1]{}]{} (1.50,2.19)(0,-0.2)[4]{}[(0,-1)[0.1]{}]{}
(0.75,1.25)(0.2,0)[2]{}[(1,0)[0.1]{}]{} (2.44,1.25)(-0.2,0)[3]{}[(-1,0)[0.1]{}]{}
(2.90,0.75)[(0.9,0.3)[$o\Bigl(t^{-\frac{n+2}{2\lambda}}\Bigr)$]{}]{} (1.1,1.10)[(0.8,0.3)[$O(t^{-n/2})$]{}]{} (3.6,1.4)[(1.2,0.3)[$\sigma=1+\frac{2-\alpha}{n+2}\lambda$]{}]{} (0.89,2.29)[(0.2,0.2)[?]{}]{} (1.73,1.9)[(1.5,0.3)[$\sigma=\frac{2(n+2)-\alpha}{n}-\lambda$]{}]{} (1.95,2.35)[(2.3,0.3)[*Arbitrarily large solutions*]{}]{}
(5.25,0.40)[(0.2,0.2)[$\lambda$]{}]{} (0.65,3.01)[(0.2,0.2)[$\sigma$]{}]{}
-0.25in
(6,3.2)(0.3,0) (0.75,0.5)[(0,1)[2.50]{}]{} (0.75,0.5)[(1,0)[4.50]{}]{} (0.75,2.19)[(1,0)[0.75]{}]{} (0.75,2.94)[(1,-1)[1.69]{}]{} (2.44,0.50)[(0,1)[0.75]{}]{} (2.44,1.25)[(3,1)[2.75]{}]{}
(1.20,0.12)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (2.14,0.12)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{} (0.12,1.10)[(0.6,0.3)[$\frac{n+2-\alpha}{n}$]{}]{} (0.12,2.04)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{}
(1.50,0.50)(0,0.2)[3]{}[(0,1)[0.1]{}]{} (1.50,2.19)(0,-0.2)[4]{}[(0,-1)[0.1]{}]{}
(0.75,1.25)(0.2,0)[2]{}[(1,0)[0.1]{}]{} (2.44,1.25)(-0.2,0)[3]{}[(-1,0)[0.1]{}]{}
(2.90,0.75)[(0.9,0.3)[$o\Bigl(t^{-\frac{n+2}{2\lambda}}\Bigr)$]{}]{} (1.1,1.10)[(0.8,0.3)[$O(t^{-n/2})$]{}]{} (3.6,1.4)[(1.2,0.3)[$\sigma=1+\frac{2-\alpha}{n+2}\lambda$]{}]{} (0.89,2.29)[(0.2,0.2)[?]{}]{} (1.73,1.9)[(1.5,0.3)[$\sigma=\frac{2(n+2)-\alpha}{n}-\lambda$]{}]{} (1.95,2.35)[(2.3,0.3)[*Arbitrarily large solutions*]{}]{}
(5.25,0.40)[(0.2,0.2)[$\lambda$]{}]{} (0.65,3.01)[(0.2,0.2)[$\sigma$]{}]{}
-0.25in
Concerning the case that $(\lambda,\sigma)$ lies on the graph of $g_\alpha$ we have the following result.
\[thm1.4\] Suppose $\alpha\in(0,n+2)$.
1. If $0<\lambda<\frac{n+2-\alpha}{n}$ and $\sigma=g_\alpha
(\lambda)$ then $\varphi(t)=t^{-n/2}$ is a poinwise bound for nonnegative solutions $u$ of , on compact subsets of $\Omega$ as $t\to 0^+$.
2. If $\alpha\in(2,n+2)$, $\lambda>\frac{n+2}{\alpha-2}$, and $\sigma=g_\alpha(\lambda)$ then there does not exist an a priori pointwise bound for nonnegative solutions $u$ of , on compact subsets of $\Omega$ as $t\to
0^+$.
When a pointwise a priori bound as $t\to 0^+$ for nonnegative solutions $u$ of , on compact subsets of $\Omega$ does not exist, as in Theorems \[thm1.3\] and \[thm1.4\](ii), we prove this by constructing for any given continuous function $\varphi:(0,1)\to(0,\infty)$ a nonnegative solution $u$ of , consisting of a sequence of smoothly connected peaks centered at $(x_j,t_j)$ where $t_j\to 0^+$ such that $$u(x_j,t_j)\not= O(\varphi(t_j)) \quad\text{as }j\to\infty.$$ When such a pointwise a priori bound does exist, as in Theorems \[thm1.1\] and \[thm1.4\](i), we reduce the proof of this fact to ruling out the possibility of such peaked solutions.
If $\alpha\in (0,n+2)$ and $\lambda>0$ then one of the following three conditions holds:
1. $0<\lambda<\frac{n+2-\alpha}{n}$;
2. $\frac{n+2-\alpha}{n}\leq\lambda<\frac{n+2}{n}$;
3. $\frac{n+2}{n}\leq\lambda<\infty$.
The proofs of Theorems \[thm1.1\]–\[thm1.4\] in case (i) (resp. (ii), (iii)) are given in Section \[sec3\] (resp. \[sec4\], \[sec5\]). In Section \[sec2\] we provide some lemmas needed for these proofs. Our approach relies on an integral representation formula for nonnegative supertemperatures (see Appendix \[secA\]), some integral estimates for heat potentials (see Appendix \[secB\]), and Moser’s iteration (see Lemmas \[lem4.1\] and \[lem5.2\]).
In this paper, we denote by $\mathcal{P}_r(x,t)$ the open circular cylinder in $\R^n\times\R$ of radius $\sqrt{r}$, height $r$, and top center point $(x,t)$. Thus $$\mathcal{P}_r(x,t)=\{(y,s)\in \R^n\times\R:|y-x|<\sqrt{r}\text{ and }t-r<s<t\}.$$
\[rem2\] Note that $$\label{1.9}
\Phi(x,t)^{\alpha/n}=\frac{1}{(4\pi)^{\alpha/2}}t^{-\alpha/2}
e^{-\frac{\alpha}{4n}\frac{|x|^2}{t}}\chi_{(0,\infty)}(t)\quad\text{in
}\R^n\times\R.$$ However, by checking the proofs of our results, we find that Theorems \[thm1.1\], \[thm1.3\], and \[thm1.4\] remain correct if $\Phi(x,t)^{\alpha/n}$ in is replaced with any function of the form $$\label{1.10}
C_1(n,\alpha)t^{-\alpha/2}
e^{-C_2(n,\alpha)\frac{|x|^2}{t}}\chi_{(0,\infty)}(t)\quad\text{in
}\R^n\times\R,$$ where $C_1(n,\alpha)$ and $C_2(n,\alpha)$ are any given positive constants. In particular, since the fundamental solution $\Phi_\alpha$ of the fractional heat operator $(\frac{\partial}{\partial t}-\Delta)^{\alpha/2}$, $\alpha\in(0,n+2)$, is given by $$\Phi_\alpha(x,t):=\frac{t^{\alpha/2-1}}{\Gamma(\alpha/2)}\Phi(x,t),$$ where $\Phi$ is the heat kernel (see [@S2002 Chapter 9, Section 2]), we find for $0<\alpha<n+2$ that $$\Phi_{n+2-\alpha}(x,t)=\frac{1}{(4\pi)^{n/2}\Gamma((n+2-\alpha)/2)}t^{-\alpha/2}
e^{-\frac{1}{4}\frac{|x|^2}{t}}\chi_{(0,\infty)}(t)$$ is of the form . Thus Theorems \[thm1.1\], \[thm1.3\], and \[thm1.4\] remain correct if $\Phi^{\alpha/n}$ in is replaced with $\Phi_{n+2-\alpha}$.
In particular, since the fundamental solution $\Phi_\beta$ of the fractional heat operator $(\frac{\partial}{\partial t}-\Delta)^\beta$, $\beta>0$, is given by $$\begin{aligned}
\notag
\Phi_\beta(x,t)&:=\frac{t^{\beta-1}}{\Gamma(\beta)}\Phi(x,t)\\
&=\frac{1}{(4\pi)^{n/2}\Gamma(\beta)}t^{\beta-1-n/2}
e^{-\frac{1}{4}\frac{|x|^2}{t}}\chi_{(0,\infty)}(t)\quad\text{in
}\R^n\times\R \label{1.10}\end{aligned}$$ (cf. [@S2002 Chapter 9, Section 2]), we see that Theorems \[thm1.1\], \[thm1.3\], and \[thm1.4\] remain correct if in we replace $\Phi^{\alpha/n}$ with $\Phi_\beta$ provided the exponents on $t$ in and are the same. That is, provided $\alpha=n+2-2\beta$ in the statements of Theorems \[thm1.1\], \[thm1.3\], and \[thm1.4\].
For example, since $$n+2-2\beta=\alpha\in(0,n+2)\quad \text{if and only if}\quad
\beta\in (0,\frac{n+2}{2})$$ and $$h_\beta(\lambda):= g_{n+2-2\beta} (\lambda)=
\begin{cases}
\frac{n+2}{n} & \text{if }0<\lambda<\frac{2\beta}{n}\\
\frac{n+2+2\beta}{n}-\lambda & \text{if }\frac{2\beta}{n}
\leq\lambda<\frac{n+2}{n}\\
\max\{0,1-\frac{n-2\beta}{n+2}\lambda \} & \text{if }
\lambda\geq\frac{n+2}{n}
\end{cases}$$ and $$\mathcal{H}_\beta(\lambda):=G_{n+2-2\beta} (\lambda)=
\begin{cases}
\frac{n+2+2\beta}{n}-\lambda & \text{if }0\leq\lambda<\frac{n+2}{n}\\
\max\{0,1-\frac{n-2\beta}{n+2}\lambda \} & \text{if }
\lambda\geq\frac{n+2}{n}
\end{cases}$$ we obtain the following analogs of Theorems \[thm1.1\] and \[thm1.3\] for nonnegative functions $$\label{1.11}
u\in C^{2,1} (\mathbb{R}^n\times (0,T))
\cap L^\lambda (\mathbb{R}^n\times (0,T)),\quad n\ge 1,$$ satisfying the nonlocal parabolic Choquard-Pekar type inequalities $$\label{1.12}
0\leq Hu\leq(\Phi_\beta*u^\lambda )u^\sigma \quad
\text{ in }\Omega\times (0,T).$$
\[thm1.5\] Suppose $u$ is a nonnegative solution of , where $\beta\in(0,\frac{n+2}{2})$, $\lambda>0$, $T>0$, and $$0\leq\sigma<h_\beta (\lambda)$$ are constants and $\Omega$ is an open subset of $\R^n$. Then for each compact subset $K$ of $\Omega$ we have as $t\to 0^+$ that
[\_[xK]{}u(x,t)=]{} O(t\^[-n/2]{}) & if $0<\lambda<\frac{n+2}{n}$ \[1.13\]\
o(t\^[-(n+2)/(2)]{}) & if $\lambda\ge\frac{n+2}{n}.$ \[1.14\]
The estimate is optimal by Remark \[rem1\]. The exponent $-(n+2)/(2\lambda)$ in is also optimal by Theorem \[thm1.2\].
\[thm1.6\] Suppose $\beta,\lambda$, and $\sigma$ are constants satisfying $$\beta\in(0,\frac{n+2}{2}), \quad \lambda>0, \quad \text{and}\quad\
\sigma>\mathcal{H}_\beta (\lambda).$$ Let $\varphi:(0,1)\to(0,\infty)$ be a continuous function satisfying $$\lim_{t\to0^+}\varphi(t)=\infty.$$ Then there exists a positive solution $u$ of , with $T=1$ and $\Omega=\R^n$ such that $$u(0,t)\neq O(\varphi(t))\quad\text{ as }t\to 0^+.$$
Theorems \[thm1.5\] and \[thm1.6\] are depicted graphically in Figure \[fig3\] when $\beta\in(0,n/2)$.
(6,3.2)(-0.2,0) (0.75,0.5)[(0,1)[2.50]{}]{} (0.75,0.5)[(1,0)[4.50]{}]{} (0.75,2.19)[(1,0)[0.75]{}]{} (0.75,2.94)[(1,-1)[1.69]{}]{} (2.44,0.50)[(0,1)[0.75]{}]{} (4.69,0.50)[(-3,1)[2.25]{}]{}
(1.20,0.12)[(0.6,0.3)[$\frac{2\beta}{n}$]{}]{} (2.14,0.12)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{} (4.39,0.12)[(0.6,0.3)[$\frac{n+2}{n-2\beta}$]{}]{} (0.12,1.10)[(0.6,0.3)[$\frac{2\beta}{n}$]{}]{} (0.12,2.04)[(0.6,0.3)[$\frac{n+2}{n}$]{}]{}
(1.50,0.50)(0,0.2)[3]{}[(0,1)[0.1]{}]{} (1.50,2.19)(0,-0.2)[4]{}[(0,-1)[0.1]{}]{}
(0.75,1.25)(0.2,0)[2]{}[(1,0)[0.1]{}]{} (2.44,1.25)(-0.2,0)[3]{}[(-1,0)[0.1]{}]{}
(2.54,0.6)[(0.9,0.3)[$o\Bigl(t^{-\frac{n+2}{2\lambda}}\Bigr)$]{}]{} (1.1,1.10)[(0.8,0.3)[$O(t^{-n/2})$]{}]{} (3.6,0.8)[(1.2,0.3)[$\sigma=1-\frac{n-2\beta}{n+2}\lambda$]{}]{} (0.89,2.29)[(0.2,0.2)[?]{}]{} (2.03,1.6)[(1.5,0.3)[$\sigma=\frac{n+2+2\beta}{n}-\lambda$]{}]{} (1.95,2.05)[(2.3,0.3)[*Arbitrarily large solutions*]{}]{}
(5.25,0.40)[(0.2,0.2)[$\lambda$]{}]{} (0.65,3.01)[(0.2,0.2)[$\sigma$]{}]{}
-0.25in
Preliminary Lemmas {#sec2}
==================
\[lem2.1\] Suppose $\alpha\in(0,n+2),\,\lambda>0,\,\sigma\geq0,\,T>0$, and $\beta\geq0$ are constants, $\Omega$ is an open subset of $\mathbb{R}^n$, and $K$ is a compact subset of $\Omega$, such that there exists a nonnegative solution $u$ of ,, where the convolution operation in is in $\mathbb{R}^n \times(0,T)$, satisfying $$\label{2.1}
\max_{x\in K}u(x,t)\neq O(t^{-\beta}),\,\quad (\text{resp. }o(t^{-\beta}))\quad\text{as }t\to0^+ .$$ Then there exists a nonnegative function $v(\xi,\tau)$ such that $$\label{2.2}
v\in C^{2,1}(\mathbb{R}^n \times(0,16))
\cap L^\lambda (\mathbb{R}^n \times(0,16)),$$ $$\label{2.3}
0\leq Hv\leq(\Phi^{\alpha/n}*v^\lambda )v^\sigma \quad\text{in }B_4 (0)\times(0,16),$$ where $*$ is the convolution operation in $\mathbb{R}^n \times(0,16)$, and $$\label{2.4}
\max_{|\xi|\leq1}v(\xi,\tau)\neq O(\tau^{-\beta}),
\quad (\text{resp. }o(\tau^{-\beta}))\quad\text{as }\tau\to 0^+ .$$
It follows from and the compactness of $K$ that there exist a sequence $\{(x_j ,t_j )\}\subset K\times(0,T)$ and $x_0 \in K$ such that $$\label{2.5}
(x_j ,t_j )\to(x_0 ,0)\quad\text{as }j\to\infty$$ and $$\label{2.6}
u(x_j ,t_j )\neq O(t^{-\beta}_{j})\quad (\text{resp. }o(t^{-\beta}_{j}))\quad\text{as }j\to\infty .$$ Choose $r\in(0,1)$ and $b>0$ such that $$\label{2.7}
\overline{B_{4r}(x_0 )}\times(0,16r^2 )\subset\Omega\times (0,T)$$ and $$\label{2.8}
b^{\lambda+\sigma-1}<r^{-(n+4-\alpha)}.$$ Define $v(\xi,\tau)$ by $u(x,t)=bv(\xi,\tau)$ where $x=x_0 +r\xi$ and $t=r^2 \tau$ and define $(\xi_j ,\tau_j )$ by $x_j =x_0 +r\xi_j$ and $t_j =r^2 \tau_j$. Then by $$\label{2.9}
(\xi_j ,\tau_j )\to(0,0)\quad\text{as }j\to\infty.$$ Clearly $(x,t)\in\mathbb{R}^n \times(0,16r^2 )$ if and only if $(\xi,\tau)\in\mathbb{R}^n \times(0,16)$. Also $16r^2 \leq T$ by . It therefore follows from that holds.
For $(x,t)\in \mathcal{P}_{16r^2}(x_0 ,16r^2 )$ (i.e. $(\xi,\tau)\in
\mathcal{P}_{16}(0,16)$) we have under the change of variables $y=x_0 +r\eta,\,s=r^2 \zeta$ that $$\iint_{\mathbb{R}^n \times(0,T)}\Phi(x-y,t-s)^{\alpha/n}u(y,s)^\lambda\,dy\,ds=\frac{b^\lambda r^{n+2}}{r^\alpha}\iint_{\mathbb{R}^n \times(0,16)}\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}v(\eta,\zeta)^\lambda \,d\eta\,d\zeta$$ where in the last integral we were able to replace the region of integration $\mathbb{R}^n \times(0,T/r^2 )$ with $\mathbb{R}^n\times(0,16)$ because $\tau<16\leq T/r^2$ and $\Phi(x,t)=0$ for $t<0$. Thus by and we find that $v$ satisfies .
Finally by we have $$\tau^{\beta}_{j}v(\xi_j ,\tau_j
)=\left(\frac{t_j}{r^2}\right)^\beta \frac{1}{b}u(x_j ,t_j
)\neq O(1)\quad (\text{resp. } o(1))\quad\text{as }j\to\infty$$ which together with implies .
\[rem2.1\] Suppose $\alpha$, $\lambda$, $\sigma$, $T$, $\beta$, $\Omega$, and $K$ are as in Lemma \[lem2.1\]. Then in order to show that all nonnegative solutions $u$ of , satisfy $$\max_{x\in K}u(x,t)=O(t^{-\beta})\quad (\text{resp. }o(t^{-\beta}))\quad\text{as }t\to 0^+$$ it suffices by Lemma \[lem2.1\] to show that all nonnegative solutions $u(x,t)$ of $$\label{2.10}
u\in C^{2,1}(\mathbb{R}^n \times(0,16))\cap L^\lambda (\mathbb{R}^n \times(0,16))$$ and $$\label{2.11}
0\leq Hu\leq(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad\text{in }B_4 (0)\times(0,16),$$ where $*$ is the convolution operation in $\mathbb{R}^n \times(0,16)$, satisfy $$\max_{|x|\leq1}u(x,t)=O(t^{-\beta})\quad (\text{resp. }o(t^{-\beta}))\quad\text{as }t\to 0^+.$$
Throughout this paper we will repeatedly use the following simple lemma.
\[lem2.2\] If $\gamma>0$ and $x\in\mathbb{R}^n ,\,n\geq1$, then $$\int_{|x-y|<r}e^{-\gamma|x-y|^2}dy=\gamma^{-n/2}\int_{|z|<\sqrt{\gamma}r}e^{-|z|^2}dz$$ and $$\int_{|x-y|>r}e^{-\gamma|x-y|^2}dy=\gamma^{-n/2}\int_{|z|>\sqrt{\gamma}r}e^{-|z|^2}dz.$$ In particular $$\int_{\mathbb{R}^n}e^{-\gamma|x-y|^2}dy=C(n)\gamma^{-n/2}.$$
Make the change of variables $z=\sqrt{\gamma}(x-y)$.
\[lem2.3\] Suppose for some constants $\alpha\in(0,n+2),\,\lambda>0$, and $\sigma\geq0$, the function $u$ is a nonnegative solution of , where $*$ is the convolution operation in $\mathbb{R}^n \times(0,16)$. Set $v=u+1$. Then $$\label{2.12}
v\in C^{2,1}(\mathbb{R}^n \times(0,16))\cap L^\lambda (B_{\sqrt{8}}(0)\times(0,8))$$ and for some positive constant $C$, $v$ satisfies $$\label{2.13}
\begin{rcases}
0\leq Hv\leq C(\Phi^{\alpha/n}*v^\lambda )v^\sigma\\
v\geq1
\end{rcases} \quad\text{in }B_2 (0)\times(0,8)$$ where $*$ is the convolution operation in $B_{\sqrt{8}}(0)\times(0,8)$. Also $$\label{2.14}
Hv,\,v^\beta \in L^1 (B_{\sqrt{8}}(0)\times(0,8))\quad
\text{for all }\beta\in\left[1,\frac{n+2}{n}\right)$$ and there exists a positive finite Borel measure $\mu$ on $B_{\sqrt{8}}(0)$ and a bounded function\
$h\in C^{2,1}(B_2 (0)\times(-4,4))$ satisfying $$\begin{aligned}
H&h=0\quad\text{in }B_2 (0)\times(-4,4)\\
&h=0\quad\text{in }B_2 (0)\times(-4,0]
\end{aligned}$$ such that $$\label{2.15}
v(x,t)=h(x,t)+\int^{8}_{0}\int_{|y|<\sqrt{8}}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds+\int_{|y|<\sqrt{8}}\Phi(x-y,t)\,d\mu(y)$$ for $(x,t)\in B_2 (0)\times(0,4)$.
\[rem2.2\] Under the assumptions of Lemma \[lem2.3\] we have $$(4\pi t)^{n/2}\int_{|y|<\sqrt{8}}\Phi(x-y,t)\,d\mu(y)\leq\int_{|y|<\sqrt{8}}d\mu(y)<\infty\quad\text{for }(x,t)\in\mathbb{R}^n \times(0,\infty).$$ Thus by we see that $$v(x,t)\leq C\left(\left(\frac{1}{\sqrt{t}}\right)^n +\int^{8}_{0}\int_{|y|<\sqrt{8}}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds\right)\quad\text{for }(x,t)\in B_2 (0)\times(0,4).$$
Clearly implies . For $$(x,t)\in B_2 (0)\times(0,8), \quad (y,s)\in(\mathbb{R}^n \times(0,16))\backslash(B_{\sqrt{8}}(0)\times(0,8)),$$ and $s<t$ we have $|x-y|>\sqrt{8}-2>1/\sqrt{2}$ and thus $$\begin{aligned}
\Phi(x-y,t-s)&\leq\frac{1}{(4\pi(t-s))^{n/2}}e^{-\frac{1}{8(t-s)}}\\
&\leq\sup_{0<\tau<t}\frac{e^{-\frac{1}{8\tau}}}{(4\pi\tau)^{n/2}}\leq C\frac{e^{-\frac{1}{8t}}}{t^{n/2}}.\end{aligned}$$ Hence for $(x,t)\in B_2 (0)\times(0,8)$ we have $$\begin{aligned}
\label{2.16}
\notag \iint_{\mathbb{R}^n \times(0,16)\backslash
B_{\sqrt{8}}(0)\times(0,8)}&
\Phi(x-y,t-s)^{\alpha/n}u(y,s)^\lambda
\,dy\,ds\\
\leq C&\left(\frac{e^{-\frac{1}{8t}}}{t^{n/2}}\right)^{\alpha/n}\iint_{\mathbb{R}^n \times(0,16)}u(y,s)^\lambda \,dy\,ds
\leq C\left(\frac{e^{-\frac{1}{8t}}}{t^{n/2}}\right)^{\alpha/n}\end{aligned}$$ by .
On the other hand, for $(x,t)\in B_2 (0)\times(0,8)$ we have $B_R (x)\subset B_{\sqrt{8}}(0)$ where $R=\sqrt{8}-2$ and thus by Lemma \[lem2.2\] we find that $$\begin{aligned}
\notag
\iint_{B_{\sqrt{8}}(0)\times(0,8)}\Phi(x-y,t-s)^{\alpha/n}\,dy\,ds
&\geq\int^{t}_{0}\left(\int_{B_R (x)}\Phi(y-x,t-s)^{\alpha/n}dy\right)ds\\
\notag &=\int^{t}_{0}\int_{B_R
(x)}\Phi(y-x,\tau)^{\alpha/n}dy\,d\tau\\
\notag &=\int_0^t\frac{1}{(4\pi\tau)^{\alpha/2}}
\left(\int_{|y-x|<R}e^{-\frac{\alpha}{4n\tau}|y-x|^2}dy\right)d\tau\\
\notag &=C\int^{t}_{0}\tau^{\frac{n-\alpha}{2}}\left(\int_{|z|<R\sqrt{\frac{\alpha}{4n\tau}}}e^{-|z|^2}dz\right)d\tau\\
\label{2.17}&\geq C\int^{t}_{0}\tau^{\frac{n-\alpha}{2}}d\tau=Ct^{\frac{n+2-\alpha}{2}}
\geq C\left(\frac{e^{-\frac{1}{8t}}}{t^{n/2}}\right)^{\alpha/n}.\end{aligned}$$ Hence for $(x,t)\in B_2 (0)\times(0,8)$ we obtain from and that $$\begin{aligned}
&\iint_{\mathbb{R}^n \times(0,16)}\Phi(x-y,t-s)^{\alpha/n}u(y,s)^\lambda \,dy\,ds\\
&\leq\iint_{B_{\sqrt{8}}(0)\times(0,8)}\Phi(x-y,t-s)^{\alpha/n}u(y,s)^\lambda
\,dy\,ds
+C\iint_{B_{\sqrt{8}}(0)\times(0,8)}\Phi(x-y,t-s)^{\alpha/n}1^\lambda \,dy\,ds\\
&\leq C \iint_{B_{\sqrt{8}}(0)\times(0,8)}\Phi(x-y,t-s)^{\alpha/n}v(y,s)^\lambda \,dy\,ds.\end{aligned}$$ Thus, since $u$ satisfies we see that $v$ satisfies . Finally, by , $Hv\ge 0$ in $B_4 (0)\times(0,16)$. Hence Theorem \[thmA\] and Remark \[remA\] with $R_1=4$, $R_2=8$, and $R_3=16$ imply and .
The following lemma will be needed to estimate the last integral in .
\[lem2.4\] Suppose $$\label{2.18}
u\in L^p (\Omega \times(0,T))$$ for some open subset $\Omega$ of $\mathbb{R}^n ,\,n\geq1$, and some constants $p\in[1,\infty)$ and $T>0$. Assume also that $$u(x,t)=\int_{\mathbb{R}^n}\Phi(x-y,t)\, d\mu(y)$$ for some finite positve Borel measure $\mu$ on $\mathbb{R}^n$. Then for each compact subset $K$ of $\Omega$ we have $$\label{2.19}
\max_{x\in K}u(x,t)=o\left(t^{-\frac{n+2}{2p}}\right)\quad\text{as }t\to 0^+ .$$
The proof consists of two steps.\
**Step 1.** In this step we prove Lemma \[lem2.4\] in the special case that $$\label{2.20}
\Omega=B_{3r}(x_0 )\quad\text{and}\quad K=\overline{B_r (x_0 )}$$ for some $x_0 \in\mathbb{R}^n$ and some $r>0$. Clearly we can assume $x_0 =0$. Since $u=v+w$ where $$v(x,t)=\int_{|y|<2r}\Phi(x-y,t)\,d\mu(y)$$ and $$w(x,t)=\int_{|y|\geq2r}\Phi(x-y,t)\,d\mu(y),$$ to complete step 1, it suffices to prove $v$ and $w$ satisfy when $\Omega$ and $K$ are given by .
Since for $|x-y|\geq r$ and $t>0$ $$\Phi(x-y,t)\leq\frac{1}{r^n}\left(\frac{r^2}{4\pi t}\right)^{n/2}e^{-\frac{r^2}{4t}}\leq r^{-n}C(n)$$ we have $$\max_{|x|\leq r}w(x,t)\leq r^{-n}C(n)\int_{\mathbb{R}^n}d\mu(y)<\infty\quad\text{for }t>0.$$ Thus $w$ satisfies when $\Omega$ and $K$ are given by .
For $|y|\leq 2r$ and $\tau>0$ it follows from Lemma \[lem2.2\] that $$\begin{aligned}
\int_{|x|\geq3r}\Phi(x-y,\tau)^p dx&=\frac{1}{(4\pi\tau)^{np/2}}\int_{|x|\geq3r}e^{-\frac{p|x-y|^2}{4\tau}}dx\\
&\leq\frac{1}{(4\pi\tau)^{np/2}}\int_{|x-y|\geq r}e^{-\frac{p|x-y|^2}{4\tau}}dx\\
&=\frac{C(n,p)}{r^{n(p-1)}}\left[\left(\frac{r}{\sqrt{\tau}}\right)^{n(p-1)}\int_{|z|>\sqrt{\frac{p}{4}}\frac{r}{\sqrt{\tau}}}e^{-|z|^2}dz\right]\\
&\leq C(n,p)/r^{n(p-1)}.
\end{aligned}$$ We obtain therefore from Jensen’s inequality and Fubini’s theorem that $$\begin{aligned}
\label{2.21}
\notag \| v\|^{p}_{L^p ((\mathbb{R}^n \backslash
B_{3r}(0))\times(0,t))}
&=\int^{t}_{0}\int_{|x|\ge 3r}\left(\int_{|y|<2r}\Phi(x-y,\tau)\,d\mu(y)\right)^p dx\,d\tau\\
\notag &\le |\mu|^{p-1}\int_{|y|<2r}\int^{t}_{0}\left(\int_{|x|\geq3r}\Phi(x-y,\tau)^p dx\right)d\tau \,d\mu(y)\\
&\leq|\mu|^p C(n,p)t/r^{n(p-1)}\quad\text{for all }t>0.
\end{aligned}$$ We now use to show $v$ satisfies .
For $0<\tau<t$ and $x\in\mathbb{R}^n$ it follows from standard $L^p$-$L^q$ estimates with $q=\infty$ (see [@QS Prop. 48.4] that $$v(x,t)\leq(4\pi)^{\frac{-n}{2p}}(t-\tau)^{\frac{-n}{2p}}\| v(\cdot,\tau)\|_{L^p (\mathbb{R}^n )}.$$ Hence $$v(x,t)^p \int^{t}_{0}(t-\tau)^{n/2}d\tau\leq(4\pi)^{-n/2}\| v\|_{L^p (\mathbb{R}^n \times(0,t))}^p$$ which implies $$\begin{aligned}
\max_{x\in\mathbb{R}^n}v(x,t)t^{\frac{n+2}{2p}}&\leq C(n,p)\| v\|_{L^p (\mathbb{R}^n \times(0,t))}\\
&\leq C(n,p)\left[\| u\|_{L^p (B_{3r}(0)\times(0,t))}+\| v\|_{L^p (\mathbb{R}^n \backslash B_{3r}(0)\times(0,t))}\right]\\
&\to0\quad\text{as }t\to0^+
\end{aligned}$$ by and . Thus $v$ satisfies when $\Omega$ and $K$ are given by .
**Step 2.** We now use Step 1 to complete the proof. For each $x\in K$ choose $r_x >0$ such that $B_{3r_x}(x)\subset\Omega$. Since $K$ is compact there exists finitely many points $x_1 ,...,x_m$ in $K$ such that $$\label{2.22}
K\subset\bigcup^{m}_{j=1}B_{r_j}(x_i )\quad \text{where }r_j =r_{x_j}.$$ For $j=1,2,...,m$ we have by Step 1 that $$\max_{|x-x_j |\leq r_j}u(x,t)=o\left(t^{-\frac{n+2}{2p}}\right)\quad\text{as }t\to0^+ .$$ Hence follows from .
\[lem2.5\] Suppose $r>0$ and $\beta>n+2$ are constants and $(x_0 ,t_0 )\in\mathbb{R}^n \times\mathbb{R}$. Then for $(x,t)\in\overline{\mathcal{P}_r (x_0 ,t_0 )}$ we have $$\iint_{\mathbb{R}^n \times\mathbb{R}\backslash\mathcal{P}_{2r}(x_0 ,t_0 )}\Phi(x-y,t-s)^{\beta/n}\,dy\,ds\leq\frac{C}{\sqrt{r}^{\beta-(n+2)}}$$ where $C=C(n,\beta)>0$.
Throughout this proof $(x,t)\in\overline{\mathcal{P}_r (x_0 ,t_0 )}$ and $C=C(n,\beta)$ is a positive constant whose value may change from line to line. Let $$A=\mathbb{R}^n \times(-\infty,t_0 -2r]\quad\text{and}\quad B=(\mathbb{R}^n \backslash B_{\sqrt{2r}}(x_0 ))\times(t_0 -2r,\,t_0 ).$$ For $(y,s)\in B$ we have $$\frac{|y-x|}{|y-x_0 |}\geq\frac{|y-x_0 |-|x-x_0 |}{|y-x_0 |}=1-\frac{|x-x_0 |}{|y-x_0 |}\geq1-\frac{\sqrt{r}}{\sqrt{2r}}=1-\frac{1}{\sqrt{2}}>\frac{1}{4}.$$ It therefore follows from Lemma \[lem2.2\] that $$\begin{aligned}
\iint_B \Phi(x-y,t-s)^{\beta/n}\,dy\,ds&\leq\int^{t}_{t_0 -2r}\frac{1}{(4\pi(t-s))^{\beta/2}}\left(\int_{|y-x_0 |>\sqrt{2r}}e^{-\frac{\beta|y-x_0 |^2}{64n(t-s)}}dy\right)ds\\
&=Cr^{\frac{n-\beta}{2}}\int^{t}_{t_0 -2r}\left(\frac{r}{t-s}\right)^{\frac{\beta-n}{2}}\int_{|z|>\sqrt{\frac{2\beta}{64n}}\sqrt{\frac{r}{t-s}}}e^{-|z|^2}dz\,ds\\
&\leq Cr^{\frac{n+2-\beta}{2}}.
\end{aligned}$$ Also, by Lemma \[lem2.2\], we obtain $$\begin{aligned}
\iint_{A}\Phi(x-y,t-s)^{\beta/n}\,dy\,ds&=\int^{t_0 -2r}_{-\infty}\frac{1}{(4\pi(t-s))^{\beta/2}}\int_{\mathbb{R}^n}e^{-\frac{\beta|x-y|^2}{4n(t-s)}}\,dy\,ds\\
&=C\int^{t_0 -2r}_{-\infty}(t-s)^{\frac{n-\beta}{2}}ds\\
&=C(t-t_0 +2r)^{\frac{n+2-\beta}{2}}\leq Cr^{\frac{n+2-\beta}{2}}.
\end{aligned}$$ Thus Lemma \[lem2.5\] follows from the fact that $$\iint_{\mathbb{R}^n \times\mathbb{R}\backslash\mathcal{P}_{2r}(x_0
,t_0 )}\Phi(x-y,t-s)^{\beta/n}\,dy\,ds=\iint_{A\cup B} \Phi(x-y,t-s)^{\beta/n}\,dy\,ds.$$
\[lem2.6\] Suppose $r>0$ and $0<\beta<n+2$ are constants and $(x_0 ,t_0 )\in\mathbb{R}^n \times\mathbb{R}$. Then $$\iint_{(y,s)\in\mathcal{P}_r (x_0 ,t_0 )}\Phi(x-y,t-s)^{\beta/n}\,dy\,ds\leq C\sqrt{r}^{n+2-\beta}\quad\text{for }(x,t)\in\mathcal{P}_r (x_0 ,t_0 )$$ where $C=C(n,\beta)>0$.
By Lemma \[lem2.2\], we have for $(x,t)\in\mathcal{P}_r (x_0 ,t_0 )$ that $$\begin{aligned}
\iint_{\mathcal{P}_r (x_0 ,t_0 )}\Phi(x-y,t-s)^{\beta/n}\,dy\,ds&\leq\int^{t}_{t_0 -r}\frac{1}{(4\pi(t-s))^{\beta/2}}\left(\int_{\mathbb{R}^n}e^{-\frac{\beta|x-y|^2}{4n(t-s)}}dy\right)ds\\
&=C\int^{t}_{t_0 -r}(t-s)^{\frac{n-\beta}{2}}ds\\
&=C(t-t_0 +r)^{\frac{n+2-\beta}{2}}\leq C\sqrt{r}^{n+2-\beta}.
\end{aligned}$$
\[lem2.7\] Suppose $\alpha\in(0,n+2)$ and $\beta\in[0,n+2)$ are constants. Then $$\label{2.23}
\iint_{\mathbb{R}^n \times (0,t)}\Phi(x-y,t-s)^{\alpha/n}\Phi(y-z,s)^{\beta/n}\,dy\,ds\leq\frac{C}{\sqrt{t}^{\alpha+\beta-(n+2)}}$$ for all $x,z\in\mathbb{R}^n$ and $t>0$ where $C=C(n,\alpha,\beta)>0$.
When $\beta =0$, Lemma \[lem2.7\] follows directly from Lemma \[lem2.2\]. Hence we can assume $\beta\in (0,n+2)$. Under the change of variables $$x-z=\sqrt{t}\xi,\quad y-z=\sqrt{t}\eta,\quad s=t\zeta$$ we see that the left side of equals $$\begin{aligned}
&\iint_{\mathbb{R}^n \times(0,1)}\Phi(\sqrt{t}(\xi-\eta),t(1-\zeta))^{\alpha/n}\Phi(\sqrt{t}\eta,t\zeta)^{\beta/n}\sqrt{t}^{n+2}\,d\eta\,d\zeta\\
&=\iint_{\mathbb{R}^n \times(0,1)}\left(\frac{1}{(4\pi
t(1-\zeta))^{n/2}}\right)^{\alpha/n}\left(\frac{1}{(4\pi
t\zeta)^{n/2}}\right)^{\beta/n}
e^{-\frac{\alpha}{n}\frac{|\xi-\eta|^2}{4(1-\zeta)}-\frac{\beta}{n}\frac{|\eta|^2}{4\zeta}}
\sqrt{t}^{n+2}\,d\eta\,d\zeta\\
&=\frac{C(n,\alpha,\beta)}{\sqrt{t}^{\alpha+\beta-(n+2)}}\int^{1}_{0}\frac{1}{(1-\zeta)^{\alpha/2}\zeta^{\beta/2}}\left(\int_{\mathbb{R}^n}
e^{-\frac{\alpha}{n}\frac{|\xi-\eta|^2}{4(1-\zeta)}-\frac{\beta}{n}\frac{|\eta|^2}{4\zeta}}d\eta\right)d\zeta\\
&\leq\frac{C(n,\alpha,\beta)}{\sqrt{t}^{\alpha+\beta-(n+2)}}\left[\int^{1/2}_{0}\frac{d\zeta}{\zeta^{\beta/2-n/2}}+\int^{1}_{1/2}\frac{d\zeta}{(1-\zeta)^{\alpha/2-n/2}}\right]
\end{aligned}$$ by Lemma \[lem2.2\].
\[lem2.8\] Suppose $(x_0,t_0)\in\R^n\times\R$ and $r>0$. If $$(x,t)\in\overline{\mathcal{P}_r(x_0,t_0)} \quad \text{and}\quad
(y,s)\in(\R^n\times\R)\setminus\overline{\mathcal{P}_{2r}(x_0,t_0)}$$ then $$\Phi(x-y,t-s)\le\frac{C(n)}{r^{n/2}}.$$
We consider three cases.
[**Case I**]{}. Suppose $t_0-2r\le s<t$. Then $|x-y|\ge
(\sqrt{2}-1) \sqrt{r}$ and hence $$\begin{aligned}
\Phi(x-y,t-s)&\le\frac{e^{-\frac{(\sqrt{2}-1)^2r}{4(t-s)}}}{(4\pi
(t-s))^{n/2}}\le\sup_{\tau>0}\frac{e^{-\frac{(\sqrt{2}-1)^2r}{4\tau}}}{(4\pi
\tau)^{n/2}}\\
&=\sup_{\zeta>0}\frac{e^{-(\sqrt{2}-1)^2\zeta}}{(\pi r/\zeta)^{n/2}}
=\frac{C(n)}{r^{n/2}}.\end{aligned}$$ [**Case II**]{}. Suppose $s<t_0-2r$. Then $t-s\ge r$ and hence $$\Phi(x-y,t-s)\le\frac{1}{(4\pi r)^{n/2}}=\frac{C(n)}{r^{n/2}}.$$ [**Case III**]{}. Suppose $s\ge t$. Then $\Phi(x-y,t-s)=0$.
\[lem2.9\] Suppose $\alpha\in(0,n+2)$ and $f\in L^p (\mathbb{R}^n \times(0,T))$ for some $$\label{2.24}
p>\frac{n+2}{n+2-\alpha}.$$ Then for all $(x,t)\in\mathbb{R}^n \times(0,T)$ we have $$\label{2.25}
\iint_{\mathbb{R}^n \times(0,T)}\Phi(x-y,t-s)^{\alpha/n}|f(y,s)|\,dy\,ds\leq Ct^{\frac{n+2-\alpha-(n+2)/p}{2}}\| f\|_{L^p (\mathbb{R}^n \times(0,t))}$$ where $C=C(n,\alpha,p)$ is a positive constant.
Define $q$ by $\frac{1}{q}=1-\frac{1}{p}$. Then by $$\frac{1}{q}>1-\frac{n+2-\alpha}{n+2}=\frac{\alpha}{n+2}$$ and thus using Hölders inequality and making the change of variables $$x=\sqrt{t}\xi,\quad y=\sqrt{t}\eta,\quad\text{and}\quad s=t\zeta$$ we have for all $(x,t)\in\mathbb{R}^n \times(0,T)$ that $$\begin{aligned}
\text{LHS of \eqref{2.25}}&=\iint_{\mathbb{R}^n \times(0,t)}\Phi(x-y,t-s)^{\alpha/n}|f(y,s)|\,dy\,ds\\
&\leq\left(\iint_{\mathbb{R}^n \times(0,t)}
\Phi(x-y,t-s)^{\alpha q/n}\,dy\,ds\right)^{1/q}\| f\|_{L^p (\mathbb{R}^n \times(0,t))}\\
&=\left(\iint_{\mathbb{R}^n \times(0,1)}\frac{1}{t^{\alpha q/2}}
\Phi(\xi-\eta,1-\zeta)^{\alpha q/n}\sqrt{t}^{n+2}\,d\eta\,d\zeta\right)^{1/q}\| f\|_{L^p (\mathbb{R}^n \times(0,t))}\\
&=\sqrt{t}^{\frac{n+2-\alpha q}{q}}\left(\iint_{\mathbb{R}^n
\times(0,1)}\Phi(-\eta,1-\zeta)^{\alpha q/n}\,d\eta\,d\zeta\right)^{1/q}\| f\|_{L^p (\mathbb{R}^n \times(0,t))}\\
&=C(n,\alpha,p)\sqrt{t}^{n+2-\alpha-(n+2)/p}\| f\|_{L^p (\mathbb{R}^n \times(0,t))}.
\end{aligned}$$
\[cor2.1\] Suppose $\alpha\in(0,n+2),\,0<\lambda<\frac{n+2-\alpha}{n}$, and $u\in L^{\frac{n+2}{n}}(\mathbb{R}^n \times(0,T))$ is a nonnegative function. Then for all $(x,t)\in\mathbb{R}^n \times(0,T)$ we have $$\iint_{\mathbb{R}^n
\times(0,T)}\Phi(x-y,t-s)^{\alpha/n}|u(y,s)|^\lambda \,dy\,ds\leq
Ct^{\frac{n+2-\alpha-n\lambda}{2}}
\| u\|^{\lambda}_{L^{\frac{n+2}{n}}(\mathbb{R}^n \times(0,t))}.$$
Since $u^\lambda \in L^{p}(\mathbb{R}^n \times(0,T))$, where $$p:=\frac{n+2}{n\lambda}>\frac{n+2}{n}\frac{n}{n+2-\alpha}=\frac{n+2}{n+2-\alpha},$$ and $$\| u^\lambda \|_{L^p}=\| u\|^{\lambda}_{L^{\lambda p}}=\| u\|^{\lambda}_{L^{\frac{n+2}{n}}}$$ the corollary follows from Lemma \[2.9\] with $f=u^\lambda$.
\[lem2.10\] Suppose $\alpha>0$ and $T$ are constants. Then for $s<t\leq T$ and $|x|\leq\sqrt{T-t}$ we have $$\int_{|y|<\sqrt{T-s}}\Phi(x-y,t-s)^{\alpha/n}dy\geq\frac{C}{(t-s)^{(\alpha-n)/2}}$$ where $C=C(n,\alpha)$ is a positive constant.
Making the change of variables $z=\frac{x-y}{\sqrt{t-s}}$ and letting $e_1 =(1,0,...,0)$ we get $$\begin{aligned}
\notag \int_{|y|<\sqrt{T-s}}\Phi(x-y,t-s)^{\alpha/n}dy&=\frac{1}{(4\pi)^{\alpha/2}}\frac{1}{(t-s)^{\alpha/2}}\int_{|y|<\sqrt{T-s}}e^{-\frac{\alpha|x-y|^2}{4n(t-s)}}dy\notag\\
&=\frac{1}{(4\pi)^{\alpha/2}(t-s)^{(\alpha-n)/2}}\int_{|z-\frac{x}{\sqrt{t-s}}|<\frac{\sqrt{T-s}}{\sqrt{t-s}}}e^{-\frac{\alpha}{4n}|z|^2}dz \label{2.26}\\
&\geq\frac{1}{(4\pi)^{\alpha/2}}\frac{1}{(t-s)^{(\alpha-n)/2}}\int_{|z-\frac{\sqrt{T-s}}{\sqrt{t-s}}e_1|<\frac{\sqrt{T-s}}{\sqrt{t-s}}}e^{-\frac{\alpha}{4n}|z|^2}dz
\label{2.27}\\
&\geq\frac{1}{(4\pi)^{\alpha/2}}\frac{1}{(t-s)^{(\alpha-n)/2}}\int_{|z-e_1
|<1}e^{-\frac{\alpha}{4n}|z|^2}dz, \label{2.28}
\end{aligned}$$ where the last two inequalities need some explanation. Since $|x|\leq\sqrt{T-t}<\sqrt{T-s}$, the center of the ball of integration in is closer to the origin than the center of the ball of integration in . Thus, since the integrand is a decreasing function of $|z|$, we obtain . Since $\sqrt{T-s}\geq\sqrt{t-s}$, the ball of integration in contains the ball of integration in and hence holds.
The case $0<\lambda<\frac{n+2-\alpha}{n}$ {#sec3}
=========================================
In this section we prove Theorems \[thm1.1\], \[thm1.3\], and \[thm1.4\] when $0<\lambda<(n+2-\alpha)/n$. For these values of $\lambda$, Remark \[rem2.1\] and the following theorem imply Theorems \[thm1.1\] and \[thm1.4\].
\[thm3.1\] Suppose $u$ is a nonnegative solution of , for some constants $\alpha\in(0,n+2)$, $$\label{3.1}
0<\lambda<\frac{n+2-\alpha}{n}\quad\text{and}\quad 0\leq\sigma\leq\frac{n+2}{n}.$$ Then $$\label{3.2}
\max_{|x|\leq1}u(x,t)=O(t^{-n/2})\quad\text{as }t\to0^+ .$$
Let $v=u+1$. Then by Lemma \[lem2.3\] we have – hold. To prove , it clearly suffices to prove $$\label{3.3}
\max_{|x|\leq1}v(x,t)=O(t^{-n/2})\quad\text{as }t\to0^+ .$$ Choose $\varepsilon\in(0,1)$ such that $$\label{3.4}
\lambda<\frac{n+2-\alpha}{n+\varepsilon}.$$ By , $$v^\lambda \in L^{\frac{n+2}{(n+\varepsilon)\lambda}}(\mathcal{P}_8 (0,8)).$$ Thus, since implies $$\frac{\lambda(n+\varepsilon)}{n+2}<\frac{n+2-\alpha}{n+2}$$ we have by Theorem \[thmB.2\] (with $\alpha$ replaced with $n+2-\alpha$) that $$\Phi^{\alpha/n}*v^\lambda \in L^\infty (\mathcal{P}_8 (0,8))$$ where the convolution operation is in $\mathcal{P}_8 (0,8)$. Hence by and , $v$ is a $C^{2,1}$ positive solution of $$0\leq Hv\leq Cv^\sigma \quad\text{in }B_2 (0)\times(0,8).$$ Thus by $_2$ and [@T2011 Theorem 1.1], $v$ satisfies .
The following theorem implies Theorem \[thm1.3\] when $0<\lambda<\frac{n+2}{n}$.
\[thm3.2\] Suppose $\alpha,\lambda$, and $\sigma$ are constants satisfying $$\label{3.5}
\alpha\in(0,n+2),\quad 0<\lambda<\frac{n+2}{n},\quad\text{and}\quad \sigma>\frac{2(n+2)-\alpha}{n}-\lambda.$$ Let $\varphi:(0,1)\to(0,\infty)$ be a continuous function satisfying $$\lim_{t\to0^+}\varphi(t)=\infty.$$ Then there exists a positive function $$\label{3.6}
u\in C^\infty (\mathbb{R}^n \times(0,1))\cap L^\lambda (\mathbb{R}^n \times(0,1))$$ satisfying $$\label{3.7}
0\leq Hu\leq(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad\text{in }\mathbb{R}^n \times(0,1),$$ where $*$ is the convolution operation in $\mathbb{R}^n \times(0,1)$, such that $$\label{3.8}
u(0,t)\neq O(\varphi(t))\quad\text{as }t\to0^+ .$$
By scaling $u$ and noting by that $\sigma+\lambda\not= 1$ we see that it suffices to prove Theorem \[thm3.2\] with replaced with the weaker statement that there exists a positive constant $C=C(n,\lambda,\sigma,\alpha)$ such that $u$ satisfies $$\label{3.9}
0\leq Hu\leq C(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad\text{in }\mathbb{R}^n \times(0,1)$$ where $*$ is the convolution operation in $\mathbb{R}^n \times(0,1)$.
By there exists $\varepsilon=\varepsilon(n,\lambda,\sigma,\alpha)\in(0,1)$ such that $$\label{3.10}
\sigma>\frac{2(n+2)-\alpha}{n-\varepsilon}-\lambda.$$ Let $$\label{3.11}
p=\frac{n-\varepsilon}{2}$$ and let $\{T_j \}\subset(0,1)$ be a sequence that $T_j \to0$ as $j\to\infty$. Define $w_j :(-\infty,T_j )\to(0,\infty)$ by $$\label{3.12}
w_j (t)=(T_j -t)^{-p}$$ and define $t_j \in(0,T_j )$ by $$\label{3.13}
w_j (t_j )=t^{-n/2}_{j}.$$ Then $$\label{3.14}
\frac{T_j -t_j}{t_j}=\frac{w_j (t_j )^{-1/p}}{t_j}=t^{n/(2p)-1}_{j}\to0\quad\text{as }j\to\infty$$ by .
Choose $a_j \in((t_j +T_j )/2,T_j )$ such that $w_j (a_j )>j\varphi(a_j )$. Then $$\label{3.15}
\frac{w_j (a_j )}{\varphi(a_j )}\to\infty\quad\text{as }j\to\infty.$$ Let $h_j (s)=\sqrt{a_j -s}$ and $H_j (s)=\sqrt{a_j +\varepsilon_j -s}$ where $\varepsilon_j >0$ satisfies $$\label{3.16}
a_j +2\varepsilon_j <T_j,
\quad t_j -\varepsilon_j >t_j /2,
\quad \varepsilon_j <T^{2}_{j},
\quad\text{and}\quad w_j (t_j -\varepsilon_j )>\frac{w_j (t_j )}{2}.$$ Define $$\begin{aligned}
&\omega_j =\{(y,s)\in\mathbb{R}^n \times\mathbb{R}:|y|<h_j (s)\quad\text{and}\quad t_j <s<a_j \}\\
&\Omega_j =\{(y,s)\in\mathbb{R}^n \times\mathbb{R}:|y|<H_j(s)\quad\text{and}\quad t_j -\varepsilon_j <s<a_j +\varepsilon_j \}.
\end{aligned}$$ By taking a subsequence we can assume the sets $\Omega_j$ are pairwise disjoint.
Let $\chi_j :\mathbb{R}^n \times\mathbb{R}\to[0,1]$ be a $C^\infty$ function such that $\chi_j \equiv1$ in $w_j$ and $\chi_j \equiv0$ in $\mathbb{R}^n \times\mathbb{R}\backslash\Omega_j$. Define $f_j ,\,u_j :\mathbb{R}^n \times\mathbb{R}\to[0,\infty)$ by $$\label{3.17}
f_j (y,s)=\chi_j (y,s)w^{\prime}_{j}(s)$$ and $$\label{3.18}
u_j (x,t)=\iint_{\mathbb{R}^n \times\mathbb{R}}\Phi(x-y,t-s)f_j (y,s)\,dy\,ds.$$ Then $f_j$ and $u_j$ are $C^\infty$ and $$\label{3.19}
Hu_j =f_j \quad\text{in }\mathbb{R}^n \times\mathbb{R}.$$ By Theorem \[thmB.2\] with $p=n+2$ and $q=\infty$ we see that $$\begin{aligned}
\label{3.20}
\notag\|\iint_{\Omega_j \backslash \omega_j}&\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds\|_{L^\infty (\mathbb{R}^n \times(0,1))}\\
\notag&=\|\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)\chi_{\Omega_j \backslash \omega_j}(y,s)w^{\prime}_{j}(s)\,dy\,ds\|_{L^\infty (\mathbb{R}^n \times(0,1))}\\
\notag&\leq C_n \| w^{\prime}_{j}(s)\|_{L^{n+2}(\Omega_j \backslash \omega_j )}\\
&\leq w_j (t_j )
\end{aligned}$$ provided we decrease $\varepsilon_j$ if necessary because $|\Omega_j \backslash \omega_j |\to0$ as $\varepsilon_j \to0$.
Also, for $(x,t)\in\Omega_j$ we have $|x|<\sqrt{T_j -t_j}$ by $_1$, and thus using $_2$ we obtain $$\sup_{(x,t)\in\Omega_j}\frac{|x|^2}{t}\leq\frac{T_j -t_j}{t_j -\varepsilon_j}\leq\frac{2(T_j -t_j )}{t_j}\to0\quad\text{as }j\to\infty$$ by . Hence by , $_2$, and there exists a positive number $M$, independent of $j$, such that $$\label{3.21}
M\Phi(x,t)\geq 2/t^{n/2}_{j}=2w_j (t_j)\quad\text{for }(x,t)\in\Omega_j .$$
In order to obtain a lower bound for $u_j$ in $\Omega_j$, note first that for $s<t\leq a_j +\varepsilon_j$ and $|x|\leq H_j (t)$ we have by Lemma \[lem2.10\] that $$\label{3.22}
\int_{|y|<H_j (s)}\Phi(x-y,t-s)\, dy\geq\beta$$ for some constant $$\label{3.23}
\beta=\beta(n)\in(0,1).$$ Next using and , we find for $(x,t)\in\Omega_j$ that $$\begin{aligned}
\iint_{\Omega_j}\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds&=\int^{t}_{t_j -\varepsilon_j}w^{\prime}_{j}(s)\left(\int_{|y|<H_j (s)}\Phi(x-y,t-s)dy\right)ds\\
&\geq\beta(w_j (t)-w_j (t_j -\varepsilon_j ))\\
&\geq\beta w_j (t)-w_j (t_j ).
\end{aligned}$$ It therefore follows from , , and that for $(x,t)\in\Omega_j$ we have $$\begin{aligned}
\label{3.24}
\notag u_j (x,t)&\geq\iint_{\omega_j}\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds\\
\notag &=\iint_{\Omega_j}\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds
-\iint_{\Omega_j \backslash \omega_j}\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds\\
&\geq\beta w_j (t)-2w_j (t_j ).
\end{aligned}$$ Also by , , and we obtain $$\begin{aligned}
\label{3.25}
\notag \iint_{\mathbb{R}^n \times\mathbb{R}}f_j (y,s)\,dy\,ds&\leq\iint_{\Omega_j}w^{\prime}_{j}(s)\,dy\,ds\\
\notag &\leq p\int^{T_j}_{0}(T_j -s)^{-(p+1)}\left(\int_{|y|<\sqrt{T_j -s}}dy\right)ds\\
\notag &=p|B_1(0)|\int^{T_j}_{0}(T_j -s)^{n/2-p-1}ds\\
&=p|B_1 (0)|\int^{T_j}_{0}\tau^{n/2-p-1}d\tau\to0\quad\text{as }j\to\infty
\end{aligned}$$ by . Hence for $1\leq\lambda<(n+2)/n$ it follows from and Theorem \[thmB.2\] that $$\label{3.26}
\| u_j \|_{L^\lambda (\mathbb{R}^n \times(0,1))}\to0\quad\text{as }j\to\infty.$$
We next prove when $$\label{3.27}
0<\lambda<1.$$ (Theorem \[thmB.2\] cannot be directly used in this case.) Choose $z_0 >1$ such that the expression $z^{n/2}e^{-z/4}$ is decreasing on the interval $z_0 \leq z<\infty$. Let $r_0 =\sqrt{z_0}+1$. Then $r_0 >2$ and by and we have $$\begin{aligned}
\label{3.28}
\notag\iint_{\mathbb{R}^n \times(0,1)}u_j (x,t)^\lambda \,dx\,dt&=\iint_{\mathbb{R}^n \times(0,1)}\left(\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)f_j (y,s)\,dy\,ds\right)^\lambda \,dx\,dt\\
&=I_j +J_j
\end{aligned}$$ where $$I_j :=\iint_{B_{r_0}(0)\times(0,1)}\left(\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)f_j (y,s)\,dy\,ds\right)^\lambda \,dx\,dt$$ and $$J_j :=\iint_{(\mathbb{R}^n \backslash B_{r_0}(0))\times(0,1)}\left(\iint_{\Omega_j}\Phi(x-y,t-s)f_j (y,s)\,dy\,ds\right)^\lambda \,dx\,dt.$$ By and Hölder’s inequality $$\begin{aligned}
\label{3.29}
\notag I_j &\leq\left(\iint_{B_{r_0}(0)\times(0,1)}\,dx\,dt\right)^{1-\lambda}\left(\iint_{B_{r_0}(0)\times(0,1)}\left(\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)f_j (y,s)\,dy\,ds\right)\,dx\,dt\right)^\lambda \\
&\to0\quad\text{as }j\to\infty
\end{aligned}$$ by and Theorem \[thmB.2\] with $p=q=1$. Also $$\label{3.30}
J_j \leq\iint_{(\mathbb{R}^n \backslash B_{r_0}(0))\times(0,1)}A_j (x,t)^\lambda \,dx\,dt\| f_j \|^{\lambda}_{L^1 (\Omega_j )}$$ where $$A_j (x,t):=\max_{(y,s)\in\Omega_j ,\,s<t}\Phi(x-y,t-s).$$ For $s<t,\,(y,s)\in\Omega_j$ and $(x,t)\in(\mathbb{R}^n \backslash B_{r_0}(0))\times(0,1)$ we have $0<s<t<1$ and $$|x-y|>|x|-|y|>|x|-1.$$ Thus $$\begin{aligned}
\label{3.31}
\notag (4\pi)^{n/2}\Phi(x-y,t-s)&\leq\frac{1}{(t-s)^{n/2}}e^{-\frac{(|x|-1)^2}{4(t-s)}}\\
&=\frac{1}{(|x|-1)^n}\left(\frac{(|x|-1)^2}{t-s}\right)^{n/2}e^{-\frac{(|x|-1)^2}{4(t-s)}}.
\end{aligned}$$ Since $|x|\geq r_0$ and $0<s<t<1$ we have $$\frac{(|x|-1)^2}{t-s}>(|x|-1)^2 \geq z_0$$ and thus by the definition of $z_0$ we obtain from that $$\begin{aligned}
(4\pi)^{n/2}\Phi(x-y,t-s)&\leq\frac{1}{(|x|-1)^n}((|x|-1)^2 )^{n/2}e^{-(|x|-1)^2 /4}\\
&=e^{-(|x|-1)^2 /4}.
\end{aligned}$$ Hence $$A_j (x,t)^\lambda \leq e^{-\lambda(|x|-1)^2 /4}\quad\text{for }(x,t)\in(\mathbb{R}^n \backslash B_{r_0}(0))\times(0,1).$$ It therefore follows from and that $J_j \to0$ as $j\to\infty$ which together with and yields when $\lambda$ satisfies .
By we find that $$\iint_{\mathbb{R}^n \times\mathbb{R}}\sum^{\infty}_{j=1}f_j (y,s)\,dy\,ds<\infty$$ provided we take a subsequence if necessary. Hence, since the $C^\infty$ functions $f_j$ have disjoint supports, we see that the function $u:(\mathbb{R}^n \times\mathbb{R})\backslash \{(0,0)\}\to[0,\infty)$ defined by $$\label{3.32}
u(x,t)=(M+1)\Phi(x,t)+\sum^{\infty}_{j=1}u_j (x,t)$$ is $C^\infty$ and by we have $$\label{3.33}
Hu=\sum^{\infty}_{j=1}f_j \quad\text{in }(\mathbb{R}^n \times\mathbb{R})\backslash \{(0,0)\}$$ $$u\equiv0\quad\text{in }\mathbb{R}^n \times(-\infty,0).$$ From we have $$u\in L^\lambda (\mathbb{R}^n \times(0,1))$$ provided we take a subsequence of $u_j$ if necessary. Thus holds.
We now prove . By and we have $Hu\equiv0$ in $(\mathbb{R}^n \times(0,1))\backslash \bigcup^{\infty}_{j=1}\Omega_j .$ Hence to prove , it suffice to prove there exists a positve constant $C=C(n,\lambda,\sigma,\alpha)$ such that $$\label{3.34}
0\leq Hu\leq C(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad\text{in }\Omega_j$$ for $j=1,2,...$.
By , , and we have for $(x,t)\in\Omega_j$ that $$\begin{aligned}
\label{3.35}
\notag u(x,t)&\geq(M+1)\Phi(x,t)+\beta w_j (t)-2w_j (t_j )\\
&\geq\Phi(x,t)+\beta w_j (t).
\end{aligned}$$ Thus for $(x,t)\in\Omega_j$ we see by , , and that $$\begin{aligned}
Hu(x,t)&=f_j (x,t)\leq w^{\prime}_{j}(t)=pw_j (t)^{1+1/p}\\
&=pw_j (t)^{1+1/p-\sigma}w_j (t)^\sigma
\leq\frac{p}{\beta^\sigma}w_j (t)^{1+1/p-\sigma}u(x,t)^\sigma.
\end{aligned}$$ Hence to prove it suffices to show $$\label{3.36}
w_j (t)^{1+1/p-\sigma}<C\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)^{\alpha/n}u(y,s)^\lambda \,dy\,ds\quad\text{for }(x,t)\in\Omega_j .$$ Our proof of consists of two cases.\
**Case I.** Suppose $$\label{3.37}
(x,t)\in\Omega_j \quad\text{and}\quad t\leq\frac{T_j +t_j}{2}.$$ Then using , , and the fact that $w_j$ is an increasing function we have $$\frac{1}{2}\leq\frac{w_j (t)}{2w_j (t_j -\varepsilon_j )}\leq\frac{w_j (t)}{w_j (t_j )}\leq\left(\frac{T_j -\frac{T_j +t_j}{2}}{T_j -t_j}\right)^{-p}=2^p <2^{n/2}.$$ Also by and $$\frac{w_j (t_j )}{T^{-n/2}_{j}}=\left(\frac{T_j}{t_j}\right)^{n/2}\in(1,2)$$ provided we take a subsequence if necessary. Thus implies $$\label{3.38}
\frac{1}{2}<\frac{w_j (t)}{T^{-n/2}_{j}}<2^{(n+2)/2}.$$ Next, making the change of variables $$x=\sqrt{T_j}\xi,\quad t=T_j \tau,\quad\text{and}\quad
y=\sqrt{T_j}\eta,\quad s=T_j \zeta,$$ we get $$\begin{aligned}
\label{3.39}
\notag \iint_{\mathbb{R}^n \times(0,1)}&\Phi(x-y,t-s)^{\alpha/n}\Phi(y,s)^\lambda \,dy\,ds\\
\notag &=\iint_{\mathbb{R}^n \times(0,\tau)}\frac{1}{T^{\frac{n}{2}\frac{\alpha}{n}}_{j}}\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}\frac{1}{T^{n\lambda/2}_{j}}\Phi(\eta,\zeta)^\lambda \sqrt{T_j}^{n+2}\,d\eta\,d\zeta\\
&\geq\frac{G(\xi,\tau)}{\sqrt{T_j}^{\alpha+n\lambda-(n+2)}}
\end{aligned}$$ where $$G(\xi,\tau):=\iint_{B_1(0)\times(1/2,\tau)}\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}\Phi(\eta,\zeta)^\lambda \,d\eta\,d\zeta.$$ Since by $_1$, $_1$, , and $_3$, $$1>\tau=\frac{t}{T_j}\geq\frac{t_j -\varepsilon_j}{T_j}\to1\quad\text{as }j\to\infty$$ we have by $_1$ that $$|\xi|=\frac{|x|}{\sqrt{T_j}}<\frac{\sqrt{T_j -t}}{\sqrt{T_j}}=\sqrt{1-\frac{t}{T_j}}\to0\quad\text{as }j\to\infty.$$ Thus, since $G$ is clearly continuous at $(\xi,\tau)=(0,1)$ and $G(0,1)>0$ we have by that $$\label{3.40}
\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)^{\alpha/n}\Phi(y,s)^\lambda \,dy\,ds\geq\frac{C}{\sqrt{T_j}^{\alpha+n\lambda-(n+2)}}\quad\text{for }(x,t)\in\Omega_j,$$ where $C:=G(0,1)/2>0$, provided we take a subsequence if necessary.
Since by and , $$\begin{aligned}
\sigma&>\frac{n+2}{n-\varepsilon}+\frac{n+2-\alpha}{n-\varepsilon}-\lambda\\
&>\frac{n+2-\varepsilon}{n-\varepsilon}+\frac{n+2-\alpha}{n}-\lambda\\
&=\frac{1}{p}+1+\frac{n+2-\alpha}{n}-\lambda
\end{aligned}$$ we have $$\begin{aligned}
\left(\frac{1}{p}+1-\sigma\right)\frac{n}{2}&<\frac{n}{2}\left(\lambda-\frac{n+2-\alpha}{n}\right)\\
&=\frac{\alpha+n\lambda-(n+2)}{2}.
\end{aligned}$$ Thus follows from , , and .\
**Case II.** Suppose $$\label{3.41}
(x,t)\in\Omega_j \quad\text{and}\quad t\geq\frac{T_j +t_j}{2}.$$ Then for $s<t$ we have by Lemma \[lem2.10\] with $T=a_j+\varepsilon_j$ that $$\int_{|y|<H_j (s)}\Phi(x-y,t-s)^{\alpha/n}dy\geq\frac{C}{(t-s)^{(\alpha-n)/2}}$$ for some positive constant $C=C(n,\alpha)$. Thus for $(x,t)$ satisfying we get $$\begin{aligned}
\label{3.42}
\notag \iint_{\Omega_j}\Phi&(x-y,t-s)^{\alpha/n}w_j (s)^\lambda \,dy\,ds\geq\int^{t}_{t_j}w_j (s)^\lambda \left(\int_{|y|<H_j (s)}\Phi(x-y,t-s)^{\alpha/n}dy\right)ds\\
\notag &\geq C\int^{t}_{t_j}\frac{ds}{(t-s)^a (T_j -s)^b}\quad
\text{where }a=(\alpha-n)/2 \text{ and } b=\lambda p\\
\notag &=\frac{C}{(T_j -t)^{a+b-1}}\int^{\frac{T_j -t_j}{T_j -t}}_{1}\frac{dz}{(z-1)^a z^b}\text{ under the change of variables }T_j -s=(T_j -t)z\\
&\geq\frac{C}{(T_j -t)^{a+b-1}}\int^{2}_{1}\frac{dz}{(z-1)^a z^b}
=\frac{C}{(T_j -t)^{(\alpha-n)/2+\lambda p-1}}.
\end{aligned}$$
Since by and $$\begin{aligned}
p(\sigma+\lambda-1)&>\frac{n-\varepsilon}{2}\left(\frac{2(n+2)-\alpha}{n-\varepsilon}-1\right)\\
&=n+2-\frac{\alpha}{2}-\frac{n}{2}+\frac{\varepsilon}{2}\\
&>\frac{n-\alpha}{2}+2,
\end{aligned}$$ we have $$p(1+\frac{1}{p}-\sigma)<\frac{\alpha-n}{2}+\lambda p-1.$$ Thus follows from , , and . This completes the proof of in all cases. Hence and hold.
Finally follows from and with $(x,t)=(0,a_j)$.
The case $\frac{n+2-\alpha}{n}\leq\lambda<\frac{n+2}{n}$ {#sec4}
========================================================
In this section we prove Theorem \[thm1.1\] when $$\label{4.1}
\frac{n+2-\alpha}{n}\leq\lambda<\frac{n+2}{n}.$$ (For these values of $\lambda$, Theorem \[thm1.3\] follows from Theorem \[thm3.2\] in the last section and Theorems \[thm1.2\] and \[thm1.4\] are vacuously true.)
For $\lambda$ satisfying , Remark \[rem2.1\] and the following theorem imply Theorem \[thm1.1\].
\[thm4.1\] Suppose $u$ is a nonnegative solution of , for some constants $\alpha\in(0,n+2)$, $$\label{4.2}
\frac{n+2-\alpha}{n}\leq\lambda<\frac{n+2}{n}\quad\text{and}\quad 0\leq\sigma<\frac{2(n+2)-\alpha}{n}-\lambda.$$ Then $$\label{4.3}
\max_{|x|\leq1}u(x,t)=O(t^{-n/2})\quad\text{as }t\to0^+ .$$
Let $v=u+1$. Then by Lemma \[lem2.3\] we have that – hold. To prove , it clearly suffices to prove $$\label{4.4}
\max_{|x|\leq1}v(x,t)=O(t^{-n/2})\quad\text{as }t\to0^+ .$$
Since increasing $\lambda$ or $\sigma$ increases the right side of the second inequality in $_1$, we can assume instead of that $$\label{4.5}
\frac{n+2-\alpha}{n}<\lambda<\frac{n+2}{n},\quad \sigma>0,\quad\text{and}\quad 1<\lambda+\sigma<\frac{2(n+2)-\alpha}{n}.$$ Since the increased value of $\lambda$ is less than $\frac{n+2}{n}$, it follows from that still holds.
By there exists $\varepsilon=\varepsilon(n,\lambda,\sigma,\alpha)\in(0,1)$ such that $$\label{4.6}
\left(\frac{n+4-\alpha}{n+4-\alpha-\varepsilon}\right)\frac{n+2-\alpha}{n}<\lambda<\frac{n+2}{n+\varepsilon}\quad\text{and}\quad \lambda+\sigma<\frac{2(n+2)-\alpha}{n+\varepsilon}$$ which implies $$\label{4.7}
\sigma<\frac{2(n+2)-\alpha}{n+\varepsilon}-\lambda<\frac{2(n+2)-\alpha}{n+\varepsilon}-\frac{n+2-\alpha}{n}<\frac{n+2}{n+\varepsilon}.$$
Suppose for contradition that is false. Then there is a sequence $\{(x_j ,t_j )\}\subset\overline{B_1(0)}\times(0,1)$ and $x_0 \in\overline{B_1 (0)}$ such that $(x_j ,t_j )\to(x_0 ,0)$ as $j\to\infty$ and $$\label{4.8}
\lim_{j\to\infty}t^{n/2}_{j}v(x_j ,t_j )=\infty.$$ By Lemma \[lem2.8\] we have for $(x,t)\in\mathcal{P}_{t_j /4}(x_j ,t_j )$ that $$\iint_{\mathcal{P}_8 (0,8)\backslash\mathcal{P}_{t_j /2}(x_j ,t_j )}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds\leq\frac{C(n)}{t^{n/2}_{j}}\iint_{\mathcal{P}_8 (0,8)}Hv(y,s)\,dy\,ds.$$ It therefore follows from and Remark \[rem2.2\] that $$\label{4.9}
v(x,t)\leq C\left[\left(\frac{1}{\sqrt{t}}\right)^n +\iint_{\mathcal{P}_{t_j /2}(x_j ,t_j ))}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds\right]\quad\text{for }(x,t)\in\mathcal{P}_{t_j /4}(x_j ,t_j ).$$ Substituting $x=x_j$ and $t=t_j$ in and using we find that $$\label{4.10}
t^{n/2}_{j}\iint_{\mathcal{P}_{t_j /2}(x_j ,t_j )}\Phi(x_j -y,t_j -s)Hv(y,s)\,dy\,ds\to\infty\quad\text{as }j\to\infty.$$ Also, by we have $$\label{4.11}
\iint_{\mathcal{P}_{t_j /2}(x_j ,t_j )}Hv(y,s)\,dy\,ds\to0\quad\text{as }j\to\infty.$$ Defining $$\label{4.12}
f_j (\eta,\zeta)=r^{\frac{n+2}{2}}_{j}Hv(x_j +\sqrt{r_j}\eta,t_j
+r_j \zeta)\quad \text{where }r_j =t_j /8$$ and making the change of variables $$\label{4.13}
y=x_j +\sqrt{r_j}\eta,\quad s=t_j +r_j \zeta$$ in and we get
$$\label{4.14}
\iint_{\mathcal{P}_4 (0,0)}f_j (\eta,\zeta)\,d\eta\,d\zeta\to0\quad\text{as }j\to\infty$$
and $$\label{4.15}
\iint_{\mathcal{P}_4 (0,0)}\Phi(-\eta,-\zeta)f_j (\eta,\zeta)\,d\eta\,d\zeta\to\infty\quad\text{as }j\to\infty.$$ Let $$N(y,s)=\iint_{\mathcal{P}_8
(0,8)}\Phi(y-\bar y,s-\bar s)Hv(\bar y,\bar s)\,
d\bar y\, d\bar s.$$ By and Theorem \[thmB.2\] we find that $N\in L^{\frac{n+2}{n+\varepsilon}}(\mathcal{P}_8 (0,8))$ and thus $N^\lambda \in L^{\frac{n+2}{\lambda(n+\varepsilon)}}(\mathcal{P}_8 (0,8))$. Hence by Hölder’s inequality, , and Lemma \[lem2.5\] we have for $R\in(0,1]$ and $(x,t)\in\mathcal{P}_{Rt_j /8}(x_j ,t_j )$ that $$\begin{aligned}
\label{4.16}
\notag &\iint_{\mathcal{P}_8 (0,8)\backslash\mathcal{P}_{Rt_j /4}(x_j ,t_j )}
\Phi(x-y,t-s)^{\alpha/n}N(y,s)^\lambda \,dy\,ds\\
\notag &\leq\| N^\lambda \|_{L^{\frac{n+2}{\lambda(n+\varepsilon)}}(\mathcal{P}_8 (0,8))}\left(\iint_{\mathbb{R}^n \times\mathbb{R}\backslash\mathcal{P}_{Rt_j /4}(x_j ,t_j )}\Phi(x-y,t-s)^{\frac{\alpha q}{n}}\,dy\,ds\right)^{1/q}\text{ where }\frac{\lambda(n+\varepsilon)}{n+2}+\frac{1}{q}=1\\
&\leq C\left(\frac{1}{\sqrt{Rt_j}^{\alpha q-(n+2)}}\right)^{1/q}\\
\label{4.17} &=C\frac{1}{\sqrt{t_j}^{(n+\varepsilon)\lambda-(n+2-\alpha)}}\end{aligned}$$ where $C>0$ depends on $R$ but not on $j$.
Since by and Lemma \[lem2.8\] we have $$N(y,s)\leq
C\left[\frac{1}{\sqrt{t_j}^n}+\iint_{(\bar y,\bar s)\in\mathcal{P}_{Rt_j /2}(x_j ,t_j
)}\Phi(y-\bar y,s-\bar s)Hv(\bar y,\bar s)\,
d\bar y\, d\bar s\right]\quad\text{for }(y,s)\in P_{Rt_j /4}(x_j ,t_j )$$ it follows from Lemma \[lem2.6\] that for $(x,t)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )$ we have
$$\begin{aligned}
\iint_{(y,s)\in\mathcal{P}_{Rt_j /4}(x_j,t_j)}&
\Phi(x-y,t-s)^{\alpha/n}N(y,s)^\lambda \,dy\,ds\notag\\
&\leq C\left[\frac{1}{\sqrt{t_j}^{n\lambda-(n+2-\alpha)}}
+\iint_{(y,s)\in\mathcal{P}_{Rt_j /4}(x_j,t_j)}\Phi(x-y,t-s)^{\alpha/n}\right.\notag\\
&\quad\left.\times\left(\iint_{(\bar y,\bar s)\in\mathcal{P}_{Rt_j /2}(x_j,t_j )}\Phi(y-\bar y,s-\bar s)Hv(\bar y,\bar s)\,d\bar y\,d\bar
s\right)^\lambda \,dy\,ds\right].
\label{4.17.1}\end{aligned}$$
Also by Jensen’s inequality, and Lemma \[lem2.7\] we have for $x\in\mathbb{R}^n$, $t>0$, and $\lambda\ge 1$ that $$\begin{aligned}
\iint_{\mathbb{R}^n \times\mathbb{R}}&\Phi(x-y,t-s)^{\alpha/n}\left(\int_{|z|<\sqrt{8}}\Phi(y-z,s)\,d\mu(z)\right)^\lambda \,dy\,ds\notag\\
&\leq C\int_{|z|<\sqrt{8}}\left(\iint_{\mathbb{R}^n \times (0,t)}\Phi(x-y,t-s)^{\alpha/n}\Phi(y-z,s)^{\lambda n/n}\,dy\,ds\right)d\mu(z)\notag\\
&\leq\frac{C}{\sqrt{t}^{\alpha+\lambda n-(n+2)}}.\label{4.17.2}
\end{aligned}$$
We claim that also holds for $0<\lambda<1$. To see this, let $x\in\R^n$ and $t>0$ be fixed and define $$f(y,s)=\Phi(x-y,t-s)^{\alpha/n} \quad\text{and}\quad
g(y,s)=\int_{|z|<\sqrt{8}}\Phi(y-z,s)\,d\mu(z).$$ Then by Lemma \[lem2.7\] with $\beta=0$ and $\beta=n$ we have $$\|f\|_1:=\iint_{\R^n\times(0,t)}\Phi(x-y,t-s)^{\alpha/n}dy\,ds\le C\sqrt{t}^{n+2-\alpha}$$ and $$\iint_{\R^n\times(0,t)}fg\,dy\,ds
=\int_{|z|<\sqrt{8}}\iint_{\R^n\times(0,t)}\Phi(x-y,t-s)^{\alpha/n}\Phi(y-z,s)\,dy\,ds\,d\mu(z)\le
C\sqrt{t}^{2-\alpha},$$ respectively, where $C$ depends on neither $x$ nor $t$. Thus by Jensen’s inequality we find for $(x,t)\in\R^n\times(0,\infty)$ and $0<\lambda<1$ that $$\begin{aligned}
\iint_{\R^n\times\R}fg^\lambda dy\,ds
&=\iint_{\R^n\times(0,t)}\left(g\|f\|_1^{1/\lambda}\right)^\lambda
\frac{f}{\|f\|_1}\,dy\,ds\\
&\le \left(\iint_{\R^n\times(0,t)}g\|f\|_1^{1/\lambda}
\frac{f}{\|f\|_1}\,dy\,ds\right)^\lambda\\
&=\|f\|_1^{1-\lambda}\left(\iint_{\R^n\times(0,t)}fg\,dy\,ds\right)^\lambda
\le C\sqrt{t}^{n+2-\alpha-\lambda n}.\end{aligned}$$ That is also holds for $0<\lambda<1$.
It therefore follows from , , , and Lemma \[lem2.7\] that for $(x,t)\in\mathcal{P}_{Rt_j /8}(x_j ,t_j )$ we have $$\begin{aligned}
\int_{(y,s)\in\mathcal{P}_8 (0,8)}&\Phi(x-y,t-s)^{\alpha/n}v(y,s)^\lambda \,dy\,ds\\
\begin{split}
\leq& C\left[\iint_{(y,s)\in\mathcal{P}_8(0,8)}\Phi(x-y,t-s)^{\alpha/n}
\left(\int_{|z|<\sqrt{8}}\Phi(y-z,s)\,d\mu(z)\right)^\lambda \,dy\,ds\right.\\
&\left.+\iint_{(y,s)\in\mathcal{P}_8 (0,8)}\Phi(x-y,t-s)^{\alpha/n}(N(y,s)+1)^\lambda \,dy\,ds\right]\\
\leq& C\left[\frac{1}{\sqrt{t_j}^{(n+\varepsilon)\lambda-(n+2-\alpha)}}+\iint_{(y,s)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )}\Phi(x-y,t-s)^{\alpha/n}\right.\\
&\times\left.\left(\iint_{(\bar y,\bar s)\in\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(y-\bar y,s-\bar s)Hv(\bar y,\bar s)\,d\bar y\,d\bar s\right)^\lambda \,dy\,ds\right]
\end{split}
\end{aligned}$$ where $C>0$ depends on $R$ but not on $j$.
Also, similar to the way was derived, we obtain $$v(x,t)\leq C\left[\frac{1}{\sqrt{t_j}^n}+\iint_{\mathcal{P}_{Rt_j /4}(x_j ,t_j )}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds\right]\quad\text{for }(x,t)\in P_{Rt_j /8}(x_j ,t_j ).$$ We see therefore from that for $(x,t)\in\mathcal{P}_{Rt_j /8}(x_j ,t_j )$ and $R\in(0,1]$ we have $$\begin{aligned}
\begin{split}
Hv(x,t)&\leq C\left[\frac{1}{\sqrt{t_j}^{(n+\varepsilon)\lambda-(n+2-\alpha)}}+\iint_{(y,s)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )}\Phi(x-y,t-s)^{\alpha/n}\right.\\
&\quad\times\left.\left(\iint_{(\bar y,\bar s)\in\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(y-\bar y,s-\bar s)Hv(\bar y,\bar s)\,d\bar y\,d\bar s\right)^\lambda \,dy\,ds\right]\\
&\quad\times\left[\frac{1}{\sqrt{t_j}^{n\sigma}}+\left(\iint_{\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds\right)^\sigma \right].
\end{split}
\end{aligned}$$ Hence under the change of variables , $$x=x_j +\sqrt{r_j}\xi,\quad t=t_j +r_j \tau,$$ and $$\bar y=x_j +\sqrt{r_j}\bar \eta,\quad \bar s=t_j +r_j \bar \zeta,$$ we obtain from and that $$\begin{aligned}
\label{4.18}
\notag &f_j (\xi,\tau)=r^{\frac{n+2}{2}}_{j}Hv(x,t)\leq r^{\frac{(n+\varepsilon)(\lambda+\sigma)-(n+2-\alpha)}{2}}_{j}Hv(x,t)\\
\notag &\leq C\left[1+\iint_{(\eta,\zeta)\in\mathcal{P}_{4R}(0,0)}\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}\left(\iint_{(\bar \eta,\bar \zeta)\in\mathcal{P}_{4r}(0,0)}\Phi(\eta-\bar \eta,\zeta-\bar \zeta)
f_j (\bar \eta,\bar \zeta)\,d\bar \eta\,d\bar \zeta\right)^\lambda d\eta\,d\zeta\right]\\
&\quad\times\left[1+\left(\iint_{\mathcal{P}_{4R}(0,0)}\Phi(\xi-\eta,\tau-\zeta)f_j (\eta,\zeta)\,d\eta\,d\zeta\right)^\sigma \right]
\end{aligned}$$ for $(\xi,\tau)\in\mathcal{P}_R (0,0)$ and $R\in(0,1]$ where $C>0$ depends on $R$ but not on $j$.
To complete the proof of Theorem \[thm4.1\] we will need the following lemma.
\[lem4.1\] Suppose the sequence $$\label{4.19}
\{f_j \}\text{ is bounded in }L^p (\mathcal{P}_{4R}(0,0))$$ for some constants $p\in[1,\frac{n+2}{2}]$ and $R\in(0,1]$. Then there exists a positive constant $C_0 =C_0 (n,\lambda,\sigma,\alpha)$ such that the sequence $$\label{4.20}
\{f_j \}\text{ is bounded in }L^q (\mathcal{P}_R(0,0))$$ for some $q\in(p,\infty)$ satisfying $$\label{4.21}
\frac{1}{p}-\frac{1}{q}\geq C_0.$$
For $R\in(0,1]$ we formally define operators $N_R$ and $I_R$ by $$(N_R f)(\xi,\tau)=\iint_{\mathcal{P}_{4R}(0,0)}
\Phi(\xi-\eta,\tau-\zeta)f(\eta,\zeta)\,d\eta\,d\zeta$$ and $$(I_R f)(\xi,\tau)=\iint_{\mathcal{P}_{4R}(0,0)}
\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}f(\eta,\zeta)\,d\eta\,d\zeta.$$ Define $p_2$ by $$\label{4.22}
\frac{1}{p}-\frac{1}{p_2}=\frac{2-\varepsilon}{n+2}$$ where $\varepsilon$ is as in . Then $p_2 \in(p,\infty)$ and thus by Theorem \[thmB.2\] we have $$\label{4.23}
\|(N_R f_j )^\lambda \|_{p_2 /\lambda}=\| N_R f_j \|^{\lambda}_{p_2}\leq C\| f_j \|^{\lambda}_{p}$$ and $$\label{4.24}
\|(N_R f_j )^\sigma \|_{p_2 /\sigma}=\| N_R f_j \|^{\sigma}_{p_2}\leq C\| f_j \|^{\sigma}_{p}$$ where $\|\cdot\|_p :=\|\cdot\|_{L^p (\mathcal{P}_{4R}(0,0))}$. Since $$\frac{1}{p_2}=\frac{1}{p}-\frac{2-\varepsilon}{n+2}
\leq 1-\frac{2-\varepsilon}{n+2}=\frac{n+\varepsilon}{n+2}$$ we see by that $$\label{4.25}
\frac{p_2}{\lambda}>1.$$ Now there are two cases to consider.
**Case I.** Suppose $$\label{4.26}
\frac{p_2}{\lambda}<\frac{n+2}{n+2-\alpha}.$$ Define $p_3$ and $q$ by $$\label{4.27}
\frac{\lambda}{p_2}-\frac{1}{p_3}=\frac{n+2-\alpha}{n+2}$$ and $$\label{4.28}
\frac{1}{q}:=\frac{1}{p_3}+\frac{\sigma}{p_2}=\frac{\lambda+\sigma}{p_2}-\frac{n+2-\alpha}{n+2}.$$ It follows from –, , and that $$\label{4.29}
1<\frac{p_2}{\lambda}<p_3 <\infty,\quad q>0,$$ and $$\begin{aligned}
\frac{1}{p}-&\frac{1}{q}
=\frac{1}{p}-\left((\lambda+\sigma)\left(\frac{1}{p}-\frac{2-\varepsilon}{n+2}\right)-\frac{n+2-\alpha}{n+2}\right)\\
&=\frac{(2-\varepsilon)(\lambda+\sigma)+(n+2-\alpha)}{n+2}-\frac{\lambda+\sigma-1}{p}\\
&\geq\frac{(2-\varepsilon)(\lambda+\sigma)+(n+2-\alpha)-(n+2)(\lambda+\sigma-1)}{n+2}\\
&=\frac{2(n+2)-\alpha-(n+\varepsilon)(\lambda+\sigma)}{n+2}.
\end{aligned}$$ Thus holds by .
By , , , and Theorem \[B.1\] we find that $$\begin{aligned}
\|(I_R ((N_R &f_j )^\lambda ))^q \|_{p_3/q}=\| I_R ((N_R f_j )^\lambda )\|^{q}_{p_3}\\
&\leq C\|(N_R f_j )^\lambda \|^{q}_{p_2 /\lambda}\\
&\leq C\| f_j \|^{\lambda q}_{p}.
\end{aligned}$$ Also by we get $$\|(N_R f_j )^{\sigma q}\|_{\frac{p_2}{\sigma q}}=\|(N_R f_j )^\sigma \|^{q}_{p_2 /\sigma}\leq C\| f_j \|^{\sigma q}_{p}.$$ It therefore follows from , , Hölder’s inequality, and that holds.
**Case II.** Suppose $$\label{4.30}
\frac{p_2}{\lambda}\geq\frac{n+2}{n+2-\alpha}.$$ Then by Theorem \[thmB.2\], , and we find that the sequence $$\label{4.31}
\{I_R ((N_R f_j )^\lambda)\}\text{ is bounded in }L^{\gamma}(\mathcal{P}_{4R}(0,0))
\quad\text{for all } \gamma\in(1,\infty).$$ Let $\hat q=p_2/\sigma$. Then by , $$\frac{1}{p}-\frac{1}{\hat q}=\frac{1}{p}-\frac{\sigma}{p_2}
=\frac{2-\varepsilon}{n+2}+\frac{1-\sigma}{p_2}.$$ Thus for $\sigma\le 1$ we have $$\frac{1}{p}-\frac{1}{\hat q}\ge\frac{2-\varepsilon}{n+2}>0$$ and for $\sigma>1$ it follows from and that $$\begin{aligned}
\frac{1}{p}-\frac{1}{\hat q}
&\ge\frac{2-\varepsilon}{n+2}-\frac{\sigma-1}{\frac{(n+2)\lambda}{n+2-\alpha}}\\
&\ge \frac{2-\varepsilon}{n+2}
-\frac
{\frac{2(n+2)-\alpha}{n}-\lambda-1}
{\frac{(n+2)\lambda}{n+2-\alpha}}\\
&=\frac{n+4-\alpha-\varepsilon}{(n+2)\lambda}
\left(\lambda-\left(\frac{n+4-\alpha}{n+4-\alpha-\varepsilon}\right)
\frac{n+2-\alpha}{n}\right)>0. \end{aligned}$$ Thus defining $q\in(p,\hat q)$ by $$\frac{1}{q}=\frac{\frac{1}{p}+\frac{1}{\hat q}}{2}$$ we have for $\sigma>0$ that $$\frac{1}{p}-\frac{1}{q}=\frac{1}{2}\left(\frac{1}{p}-\frac{1}{\hat
q}\right)
\ge C_0(n,\lambda,\sigma,\alpha)>0.$$ That is holds.
Since $q\sigma/p_2<\hat q\sigma/p_2=1$ there exists $\gamma\in(q,\infty)$ such that $$\label{4.32}
\frac{q}{\gamma}+\frac{q\sigma}{p_2}=1.$$ Also $$\|(I_R ((N_R f_j )^\lambda ))^q \|_{\gamma/q}
=\| I_R ((N_R f_j )^\lambda )\|^{q}_{\gamma}$$ and by $$\|(N_R f_j )^{\sigma q}\|_{\frac{p_2}{\sigma q}}=\|(N_R f_j )^\sigma
\|^{q}_{p_2 /\sigma}\leq C\| f_j \|^{\sigma q}_{p}.$$ It therefore follows from , , Hölder’s inequality, , and that holds.
We return now to the proof of Theorem \[thm4.1\]. By the sequence $$\label{4.33}
\{f_j \}\text{ is bounded in }L^1(\mathcal{P}_4 (0,0)).$$ Starting with this fact and iterating Lemma \[lem4.1\] a finite number of times ($m$ times is enough if $m>1/C_0$) we see that there exists $R_0 \in(0,1)$ such that the sequence $$\{f_j \}\text{ is bounded in }L^p (\mathcal{P}_{4R_0}(0,0))$$ for some $p>(n+2)/2$. Hence by Theorem \[thmB.2\] the sequence $\{N_{R_0}f_j \}$ is bounded in $L^\infty (\mathcal{P}_{4R_0}(0,0))$. Thus implies the sequence $$\label{4.34}
\{f_j \}\text{ is bounded in }L^\infty (\mathcal{P}_{R_0}(0,0)).$$ Since by Lemma \[lem2.8\], $$\begin{aligned}
\iint_{\mathcal{P}_{4}(0,0)}&\Phi(-\eta,-\zeta)f_j(\eta,\zeta)\,d\eta\,d\zeta\\
&\le \iint_{\mathcal{P}_{R_0}(0,0)}\Phi(-\eta,-\zeta)f_j(\eta,\zeta)\,d\eta\,d\zeta
+\frac{C(n)}{R_0^{n/2}}\iint_{\mathcal{P}_{4}(0,0)\setminus\mathcal{P}_{R_0}(0,0)}
f_j(\eta,\zeta)\,d\eta\,d\zeta\end{aligned}$$ we see that and contradict . This contradiction completes the proof of Theorem \[thm4.1\].
The case $\lambda\geq\frac{n+2}{n}$ {#sec5}
===================================
In this section we prove Theorems \[thm1.1\]–\[thm1.4\] when $\lambda\geq\frac{n+2}{n}$. For these values of $\lambda$, Remark 2.1 and the following theorem imply Theorem \[thm1.1\].
\[thm5.1\] Suppose $u$ is a nonnegative solution of , for some constants $\alpha\in(0,n+2)$, $$\label{5.1}
\lambda\geq\frac{n+2}{n}\quad\text{and}\quad 0\leq\sigma<1-\frac{\alpha-2}{n+2}\lambda.$$ Then $$\label{5.2}
\max_{|x|\leq1}u(x,t)=o(t^{-\frac{n+2}{2\lambda}})\quad\text{as }t\to0^+ .$$
Let $v=u+1$. Then by Lemma \[lem2.3\] we have that – hold. To prove it clearly suffices to prove $$\label{5.3}
\max_{|x|\leq1}v(x,t)=o(t^{-\frac{n+2}{2\lambda}})\quad\text{as }t\to0^+ .$$
Since increasing $\sigma$ increases the right side of the second inequality in $_1$, we can assume instead of that $$\label{5.4}
\lambda\geq\frac{n+2}{n}\quad\text{and}\quad 0<\sigma<1-\frac{\alpha-2}{n+2}\lambda$$ which implies $$\label{5.5}
\frac{\sigma}{\lambda}<\frac{2-\alpha}{n+2}+\frac{1}{\lambda}\leq\frac{2-\alpha}{n+2}+\frac{n}{n+2}=\frac{n+2-\alpha}{n+2}.$$ By there exists $\varepsilon=\varepsilon(n,\lambda,\sigma,\alpha)\in(0,1)$ such that $$\label{5.6}
\alpha+\varepsilon<n+2\quad\text{and}\quad \sigma<1-\frac{\alpha+\varepsilon-2}{n+2}\lambda$$ which implies $$\label{5.7}
\frac{\sigma-1}{\lambda}<\frac{2-\alpha-\varepsilon}{n+2}.$$
Part of the proof of Theorem \[thm5.1\] will consist of two lemmas, the first of which is the following.
\[lem5.1\] Suppose $\Omega$ is a bounded open subset of $\mathbb{R}^n \times\mathbb{R}$ and $$\label{5.8}
p\in\left[\lambda,\frac{(n+2)\lambda}{n+2-\alpha-\varepsilon}\right).$$ Then for all $w\in L^p (\Omega)$ we have $$\label{5.9}
\left\|\left(\iint_\Omega \Phi(\cdot-y,\cdot-s)^{\alpha/n}w(y,s)^\lambda \,dy\,ds\right)w^\sigma \right\|_{L^{p_3}(\Omega)}\leq C\| w\|^{\lambda+\sigma}_{L^p (\Omega)}$$ where $$\label{5.10}
\frac{1}{p_3}=\frac{\lambda+\sigma}{p}-\frac{n+2-\alpha-\varepsilon}{n+2}$$ and $C=C(n,\lambda,\sigma,\alpha,\Omega,p)$ is a positive constant. Moreover, $$\label{5.11}
p_3 >1.$$
Define $p_2$ by $$\frac{\lambda}{p}-\frac{1}{p_2}=\frac{n+2-\alpha-\varepsilon}{n+2}.$$ Then by and $_1$, $1\leq p/\lambda <p_2 <\infty$ and thus by Theorem \[thmB.2\] we have, letting $$I(f)=\iint_\Omega \Phi(\cdot-y,\cdot-s)^{\alpha/n}f(y,s)\,dy\,ds,$$ that $$\label{5.12}
\| I(w^\lambda )\|_{L^{p_2}(\Omega)}\leq C\| w^\lambda \|_{L^{p/\lambda}(\Omega)}=C\| w\|^{\lambda}_{L^p (\Omega)}.$$ Since $\frac{1}{p_3}=\frac{1}{p_2}+\frac{\sigma}{p}$ we have by Hölder’s inequality that $$\begin{aligned}
\| I(w^\lambda )w^\sigma &\|^{p_3}_{L^{p_3}(\Omega)}
=\|(I(w^\lambda )w^\sigma )^{p_3}\|_{L^1 (\Omega)}\\
&\leq\| I(w^\lambda )^{p_3}\|_{L^{p_2 /p_3}(\Omega)}\| w^{\sigma p_3}\|_{L^{\frac{p}{\sigma p_3}}(\Omega)}\\
&=\| I(w^\lambda )\|^{p_3}_{L^{p_2}(\Omega)}\| w\|^{\sigma p_3}_{L^p (\Omega)}.
\end{aligned}$$ Thus follows from .
Also from and we find that $$\frac{1}{p_3}\leq\frac{\lambda+\sigma}{\lambda}-\frac{n+2-\alpha-\varepsilon}{n+2}=\frac{\sigma}{\lambda}+\frac{\alpha+\varepsilon}{n+2}<\frac{1}{\lambda}+\frac{2}{n+2}.$$ Thus follows from $_1$.
We now continue with the proof of Theorem \[thm5.1\]. Suppose for contradiction that is false. Then there exists a sequence $\{(x_j ,t_j )\}\subset\overline{B_1 (0)}\times(0,1/2)$ and $x_0 \in\overline{B_1 (0)}$ such that $$\label{5.13}
(x_j ,t_j )\to (x_0 ,0)\quad\text{as }j\to\infty$$ and $$\label{5.14}
\liminf_{j\to\infty}t^{\frac{n+2}{2\lambda}}_{j}v(x_j ,t_j )>0.$$ Define $p_3 >0$ by $\frac{1}{p_3}=\frac{\alpha+\varepsilon}{n+2}+\frac{\sigma}{\lambda}$. Then by , and Lemma \[lem5.1\] with $\Omega=\mathcal{P}_8 (0,8),\,p=\lambda$, and $w=v$ we have $p_3 >1$ and $Hv\in L^{p_3}(\mathcal{P}_4 (0,4)$. Hence defining $p_4$ by $\frac{1}{p_3}+\frac{1}{p_4}=1$, using Hölder’s inequality, and making the change of variables $$\begin{array}{cc}
x=x_j +\sqrt{Rt_j}\xi, & t=t_j +Rt_j \tau\\
y=x_j +\sqrt{Rt_j}\eta, & s=t_j +Rt_j \zeta
\end{array}$$ we see for $R\in(0,1]$ that $$\begin{aligned}
&\sup_{(x,t)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )}
\iint_{\mathcal{P}_4 (0,4)\backslash\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(x-y,t-s)Hv(y,s)\,ds\\
&\leq\sup_{(x,t)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )}\left(\iint_{\mathcal{P}_4 \backslash\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(x-y,t-s)^{p_4}\,dy\,ds\right)^{1/p_4}\| Hv\|_{L^{p_3}(\mathcal{P}_4 (0,4))}\\
&\leq C\sup_{(\xi,\tau)\in\mathcal{P}_{1/4}(0,0)}\left(\iint_{\mathbb{R}^n \times\mathbb{R}\backslash\mathcal{P}_{1/2}(0,0)}\left(\frac{1}{(Rt_j )^{n/2}}\right)^{p_4}\Phi(\xi-\eta,\tau-\zeta)^{p_4}(Rt_j )^{\frac{n+2}{2}}\,d\eta\,d\zeta\right)^{1/p_4}\\
&=C\left(\frac{1}{Rt_j}\right)^{(\frac{np_4}{2}-\frac{n+2}{2})\frac{1}{p_4}}\sup_{(\xi,\tau)\in\mathcal{P}_{1/4}(0,0)}\left(\iint_{\mathbb{R}^n \times\mathbb{R}\backslash\mathcal{P}_{1/2}(0,0)}\Phi(\xi-\eta,\tau-\zeta)^{p_4}\,d\eta \,d\zeta\right)^{1/p_4}\\
&=C\left(\frac{1}{Rt_j}\right)^{\frac{n+2}{2\lambda}(\sigma-\frac{2-\alpha-\varepsilon}{n+2}\lambda)}
\end{aligned}$$ where $C$ depends on neither $R$ nor $j$ and $$\sigma-\frac{2-\alpha-\varepsilon}{n+2}\lambda<1$$ by $_2$.
Also, using , Lemma \[lem2.8\], and the fact that $\mathcal{P}_{t_j /4}(x_j ,t_j )\subset\mathcal{P}_2 (0,2)$ we see for $R\in(0,1]$ that $$\begin{aligned}
\sup_{(x,t)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )}&\iint_{\mathcal{P}_8 (0,8)\backslash\mathcal{P}_4 (0,4)}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds\\
&\leq C(n)\iint_{\mathcal{P}_8 (0,8)}Hv(y,s)\,dy\,ds<\infty.
\end{aligned}$$ Thus by , and Lemma \[lem2.4\] with $p=\lambda,\,\Omega\times(0,T)=B_2 (0)\times(0,4)$, and $K=\overline{B_{3/2}(0)}$ we have for $(x,t)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )$ and $R\in(0,1]$ that $$\label{5.15}
v(x,t)\leq\iint_{\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(x-y,t-s)Hv(y,s)\,dy\,ds+\frac{\varepsilon_j}{(Rt_j )^{\frac{n+2}{2\lambda}}}$$ for some sequence $\{\varepsilon_j \}\subset(0,1)$ which tends to zero as $j\to\infty$ and which depends in neither $(x,t)$ nor $R$.
Also, for $(x,t)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )$ and $R\in(0,1]$ we have by and Lemma \[lem2.8\] that $$\begin{aligned}
\iint_{\mathcal{P}_8 (0,8)\backslash\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(x-y,t-s)^{\alpha/n}v(y,s)^\lambda \,dy\,ds&\leq\left(C(n)\left(\frac{4}{Rt_j}\right)^{n/2}\right)^{\alpha/n}\| v^\lambda \|_{L^1 (\mathcal{P}_8 (0,8))}\\
&=\frac{C}{(Rt_j )^{\alpha/2}}
\end{aligned}$$ where $C$ depends on neither $(x,t),\,R$, nor $j$. Thus for $(x,t)\in\mathcal{P}_{Rt_j /4}(x_j ,t_j )$ and $R\in(0,1]$ we have by that $$\label{5.16}
0\leq Hv(x,t)\leq C\left(\frac{1}{(Rt_j )^{\alpha/2}}+\iint_{\mathcal{P}_{Rt_j /2}(x_j ,t_j )}\Phi(x-y,t-s)^{\alpha/n}v(y,s)^\lambda \,dy\,ds\right)v(x,t)^\sigma$$ where $C$ depends on neither $(x,t),\,R$, nor $j$.
Next, making the change of variables $$v(y,s)=t^{-\frac{n+2}{2\lambda}}_{j}v_j (\eta,\zeta),$$ $$x=x_j +\sqrt{t_j}\xi, \quad t=t_j +t_j \tau;\qquad
y=x_j +\sqrt{t_j}\eta, \quad s=t_j +t_j \zeta,$$ we obtain $$\label{5.17}
\iint_{\mathcal{P}_{1/2}(0,0)}v_j (\eta,\zeta)^\lambda \,d\eta\,d\zeta=\iint_{\mathcal{P}_{t_j /2}(x_j ,t_j )}v(y,s)^\lambda \,dy\,ds$$ and from and we find for $(\xi,\tau)\in\mathcal{P}_{R/4}(0,0)$ and $R\in(0,1]$ that $$\begin{aligned}
\label{5.18}
\notag v_j(\xi,\tau)&\leq\iint_{\mathcal{P}_{{R/2}}(0,0)}\frac{1}{t^{n/2}_{j}}\Phi(\xi-\eta,\tau-\zeta)t^{-1}_{j}Hv_j (\eta,\zeta)t^{\frac{n+2}{2}}_{j}\,d\eta\,d\zeta +\frac{\varepsilon_j}{R^{\frac{n+2}{2\lambda}}}\\
&=\iint_{\mathcal{P}_{R/2}(0,0)}\Phi(\xi-\eta,\tau-\zeta)Hv_j (\eta,\zeta)\,d\eta\,d\zeta+\frac{\varepsilon_j}{R^{\frac{n+2}{2\lambda}}},
\end{aligned}$$ where $\varepsilon_j \to0$ as $j\to\infty$ and $\varepsilon_j$ depends on neither $(\xi,\tau)$ nor $R$, and $$\begin{aligned}
\label{5.19}
\notag 0\leq Hv_j (\xi,\tau)&\leq Ct^{\frac{n+2}{2\lambda}+1-\frac{\alpha}{2}}_{j}\left(\frac{1}{R^{\alpha/2}}+\iint_{\mathcal{P}_{R/2}(0,0)}\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}v_j (\eta,\zeta)^\lambda \,d\eta\,d\zeta\right)\\
\notag &\quad\times(t^{-\frac{(n+2)\sigma}{2\lambda}}_{j}v_j (\xi,\eta)^\sigma )\\
&=C\hat{\varepsilon}_j\left(\frac{1}{R^{\alpha/2}}+\iint_{\mathcal{P}_{R/2}(0,0)}\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}v_j (\eta,\zeta)^\lambda \,d\eta\,d\zeta\right)v_j (\xi,\eta)^\sigma
\end{aligned}$$ where $C$ depends on neither $(\xi,\tau),\,R$, nor $j$ and $$\hat{\varepsilon}_j :=t^{-\frac{n+2}{2\lambda}(\sigma-1+\frac{\alpha-2}{n+2}\lambda)}_{j}\to0\quad\text{as }j\to\infty$$ by and .
Also by we have $$\label{5.20}
\liminf_{j\to\infty}v_j (0,0)>0.$$
To complete the proof of Theorem \[thm5.1\] we will require the following lemma.
\[lem5.2\] Suppose the sequence $$\label{5.21}
\{v_j \}\text{ is bounded in }L^p (P_{R/2}(0,0))$$ for some constants $R\in(0,1]$ and $$\label{5.22}
p\in\left[\lambda,\frac{(n+2)\lambda}{n+2-\alpha-\varepsilon}\right).$$ Then either the sequence $$\label{5.23}
\{v_j \}\text{ tends to zero in }L^{\frac{(n+2)\lambda}{n+2-\alpha-\varepsilon}}(P_{R/8}(0,0))$$ or there exists a positive constant $C_0 =C_0 (n,\lambda,\sigma,\alpha)$ such that the sequence $$\label{5.24}
\{v_j \}\text{ tends to zero in }L^q (P_{R/8}(0,0))$$ for some $q\in(p,\infty)$ satisfying $$\label{5.25}
\frac{1}{p}-\frac{1}{q}\geq C_0 .$$
It follows from , , , and Lemma \[lem5.1\] that the sequence $$\label{5.26}
\{Hv_j \}\text{ tends to $0$ in }L^{p_3}(P_{R/4}(0,0))$$ where $p_3$, defined by , satisfies .\
**Case I.** Suppose $p_3 \geq\frac{n+2}{2}$. Then by , , and Theorem \[thmB.2\] we have the sequence $$\{v_j \}\text{ tends to zero in }L^q (P_{R/8}(0,0))\quad\text{for all }q>1$$ which implies .
**Case II.** Suppose $p_3 <\frac{n+2}{2}$. Define $q$ by $$\label{5.27}
\frac{1}{p_3}-\frac{1}{q}=\frac{2}{n+2}.$$ Then by $$1<p_3 <q<\infty.$$ Hence by , and Theorem \[B.1\] we have holds.
Also by , , , and we get $$\begin{aligned}
\frac{1}{p}-\frac{1}{q}
&=\frac{1}{p}+\frac{2}{n+2}-\frac{1}{p_3}
=\frac{1}{p}+\frac{2}{n+2}-\frac{\sigma}{p}-\frac{\lambda}{p}+1-\frac{\alpha+\varepsilon}{n+2}\\
&=-\frac{\lambda+\sigma-1}{p}+1-\frac{\alpha+\varepsilon-2}{n+2}\\
&\geq\frac{1-(\lambda+\sigma)}{\lambda}+1+\frac{2-\alpha-\varepsilon}{n+2}>0.
\end{aligned}$$ Thus holds.
We now return to the proof of Theorem \[thm5.1\]. By and , the sequence $\{v_j \}$ tends to zero in $L^\lambda (P_{1/2}(0,0))$. Starting with this fact on iterating Lemma \[lem5.2\] a finite number of times we see that the sequence $$\label{5.28}
\{v_j \}\text{ tends to zero in }L^p (\mathcal{P}_{R/2}(0,0))$$ for some $R\in(0,1)$ and for some $$\label{5.29}
p>\frac{(n+2)\lambda}{n+2-\alpha}.$$ Hence the sequence $\{v^{\lambda}_{j}\}$ tends to zero in $L^{p/\lambda}(\mathcal{P}_{R/2}(0,0))$ and $\frac{p}{\lambda}>\frac{n+2}{n+2-\alpha}$. Thus by Theorem \[thmB.2\], the sequence whose $j$th term is the integral on the right side of , tends to zero in $L^\infty (\mathcal{P}_{R/2}(0,0))$. So by $$\label{5.30}
0\leq Hv_j <Cv^{\sigma}_{j}\quad\text{in }\mathcal{P}_{R/4}(0,0)$$ where $C$ does not depend on $j$. Hence by the sequence $\{Hv_j \}$ tends to zero in $L^{p/\sigma}(\mathcal{P}_{R/4}(0,0))$ and by and $$\frac{p}{\sigma}>\frac{(n+2)\lambda}{(n+2-\alpha)\sigma}>\left(\frac{n+2}{n+2-\alpha}\right)^2 >1.$$ Thus by and Theorem \[thmB.2\] the sequence $$\label{5.31}
\{v_j \}\text{ tends to zero in }L^q (\mathcal{P}_{R/8}(0,0))\text{ where }q=
\begin{cases}
\infty, & \text{if }\frac{p}{\sigma}\geq\frac{n+2}{2-\varepsilon}\\
\frac{1}{\frac{\sigma}{p}-\frac{2-\varepsilon}{n+2}}, & \text{if }\frac{p}{\sigma}<\frac{n+2}{2-\varepsilon}.
\end{cases}$$ However the possibility that $q=\infty$ is ruled out by . Hence we can assume $\frac{p}{\sigma}<\frac{n+2}{2-\varepsilon}$. Then by , $$\frac{1}{p}-\frac{1}{q}=\frac{1-\sigma}{p}+\frac{2-\varepsilon}{n+2}.$$ Thus, if $\sigma\in(0,1]$ then $$\frac{1}{p}-\frac{1}{q}>\frac{1}{n+2}.$$ On the other hand, if $\sigma>1$ then by and $$\begin{aligned}
\frac{1}{p}-\frac{1}{q}&=\frac{2-\varepsilon}{n+2}-\frac{\sigma-1}{p}\\
&>\frac{2-\varepsilon}{n+2}-\frac{\sigma-1}{\lambda}\frac{n+2-\alpha}{n+2}\\
&>\frac{2-\varepsilon}{n+2}-\frac{2-\alpha-\varepsilon}{n+2}
=\frac{\alpha}{n+2}.
\end{aligned}$$ Thus for $\sigma>0$ we have $$\frac{1}{p}-\frac{1}{q}>C(n,\alpha)>0.$$ Hence, after a finite number of iterations of the procedure of going from to we see that the sequence $\{v_j\}$ tends to zero in $L^\infty (\mathcal{P}_{\hat R}(0,0))$ for some $\hat R\in(0,R)$ which again contrdicts . This completes the proof of Theorem \[thm5.1\].
The following theorem implies Theorem \[thm1.2\].
\[thm5.2\] Suppose $$\label{5.32}
\lambda\geq\frac{n+2}{n}\quad\text{and}\quad \gamma=\frac{n+2-\varepsilon}{2\lambda}$$ for some $\varepsilon\in(0,1)$. Then the function $$\label{5.33}
u(x,t)=\Psi(x,t):=\int_{\mathbb{R}^n}\Phi(x-y,t)|y|^{-2\gamma}dy$$ is a $C^\infty$ positive solution of $$\label{5.34}
Hu=0\quad\text{in }\mathbb{R}^n \times(0,\infty)$$ such that $$\label{5.35}
u\in L^\lambda (\mathbb{R}^n \times(0,T))\quad\text{for all }T>0,$$ $$\label{5.36}
t^\gamma u(0,t)=u(0,1)\quad\text{for }0<t<\infty,$$ and $$\label{5.37}
t^\gamma u(x,t)\text{ is bounded between positive constants}$$ on $\{(x,t)\in\mathbb{R}^n \times(0,t):|x|<\sqrt{t}\}$.
By we have $2\gamma<n$. Thus is a $C^\infty$ positive solution of .
For $a>0$ and $(x,t)\in\mathbb{R}^n \times(0,\infty)$ we find making the change of variables $y=az$ that $$\begin{aligned}
\label{5.38}
\notag u(ax,a^2 t)&=\int_{\mathbb{R}^n}\Phi(ax-y,a^2 t)|y|^{-2\gamma}dy\\
\notag &=\int_{\mathbb{R}^n}\Phi(ax-az,a^2 t)a^{-2\gamma}|z|^{-2\gamma}a^n dz\\
\notag &=a^{-2\gamma}\int_{\mathbb{R}^n}\Phi(x-z,t)|z|^{-2\gamma}dz\\
&=a^{-2\gamma}u(x,t).
\end{aligned}$$ Taking $x=0$ and $t=1$ in we get $$\label{5.39}
u(0,a^2 )=a^{-2\gamma}u(0,1)\quad\text{for all }a>0.$$ Thus holds.
Taking $x\neq0,\,t>0$, and $a=1/|x|$ in and using the fact that $u(x,t)$ is radially symmetric in $x$ about the origin we get $$\label{5.40}
u(x,t)=a^{2\gamma}u(ax,a^2 t)=|x|^{-2\gamma}u(e_1 ,\frac{t}{|x|^2})
=|x|^{-2\gamma}g\left(\frac{t}{|x|^2}\right)$$ where $g(\zeta)=u(e_1 ,\zeta)$ and $e_1 =(1,0,...,0)\in\mathbb{R}^n$. By , $$\label{5.41}
g(\zeta)\to1\quad\text{as }\zeta\to0^+$$ and using and we obtain for $t>0$ that $$1=\lim_{x\to 0}\frac{u(x,t)}{u(0,t)}
=\lim_{x\to 0}\frac{|x|^{-2\gamma}g\left(\frac{t}{|x|^2}\right)}{u(0,1)t^{-\gamma}}
=\lim_{x\to0}\frac{1}{u(0,1)}\frac{g\left(\frac{t}{|x|^2}\right)}
{\left(\frac{t}{|x|^2}\right)^{-\gamma}}.$$ Thus $$\label{5.42}
\frac{g(\zeta)}{\zeta^{-\gamma}}\to u(0,1)\quad\text{as }\zeta\to\infty.$$ For $t>0$, it follows from – and that $$\begin{aligned}
\int_{\mathbb{R}^n}u(x,t)^\lambda dx&=\int_{\mathbb{R}^n}|x|^{-2\lambda\gamma}g\left(\frac{t}{|x|^2}\right)^\lambda dx\\
&\leq
C\left[\int_{\sqrt{t}<|x|}|x|^{-2\lambda\gamma}dx
+\int_{|x|<\sqrt{t}}|x|^{-2\lambda\gamma}\left(\frac{t}{|x|^2}\right)^{-\gamma\lambda}dx\right]\\
&\leq Ct^{-1+\varepsilon/2}
\end{aligned}$$ which implies .
Making the change of variables $$x=\sqrt{t}\xi\quad\text{and}\quad y=\sqrt{t}\eta$$ in we get $$u(x,t)=\frac{1}{t^\gamma}U\left(\frac{x}{\sqrt{t}}\right)\quad\text{for }(x,t)\in\mathbb{R}^n \times(0,\infty)$$ where $$U(\xi)=\frac{1}{(4\pi)^{n/2}}\int_{\mathbb{R}^n}e^{-|\xi-\eta|^2 /4}|\eta|^{-2\gamma}d\eta.$$ Thus since $U(\xi)$ is bounded between positive constants for $|\xi|\leq1$, we find that holds.
The following theorem implies Theorems \[thm1.3\] and \[thm1.4\] when $\lambda\geq(n+2)/n$.
\[thm5.3\] Suppose $\alpha,\,\lambda$, and $\sigma$ are constants satisfying $$\label{Q1}
\alpha\in(0,n+2),\quad\lambda\geq\frac{n+2}{n},\quad\sigma\geq0,\quad\text{and}\quad \sigma>1+\frac{2-\alpha}{n+2}\lambda.$$ Let $\varphi:(0,1)\to(0,\infty)$ be a continuous function satisfying $$\lim_{t\to0^+}\varphi(t)=\infty.$$ Then there exists a positive function $$\label{Q2}
u\in C^\infty (\mathbb{R}^n \times(0,1))\cap L^\lambda (\mathbb{R}^n \times(0,1))$$ satisfying $$\label{Q3}
0\leq Hu\leq(\Phi^{\alpha/n}*u^\lambda )u^\sigma\quad\text{in }\mathbb{R}^n \times(0,1),$$ where $*$ is the convolution operation in $\mathbb{R}^n \times(0,1)$, such that $$\label{Q4}
u(0,t)\neq O(\varphi(t))\quad\text{as }t\to0^+ .$$
By scaling $u$ and noting by that $\sigma+\lambda\neq1$ we see that it suffices to prove Theorem \[thm5.3\] with replaced with the weaker statement that there exists a positive constant $C=C(n,\lambda,\sigma,\alpha)$ such that $u$ satisfies $$\label{Q5}
0\leq Hu\leq C(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad\text{in }\mathbb{R}^n \times(0,1),$$ where $(*)$ is the convolution operation in $\mathbb{R}^n \times(0,1)$.
By there exists $\varepsilon=\varepsilon(n,\lambda,\sigma,\alpha)\in(0,1)$ such that $$\label{Q6}
2\varepsilon<\alpha$$ and $$\label{Q7}
\sigma>1+\frac{2-\alpha+2\varepsilon}{n+2-2\varepsilon}\lambda.$$ Let $$\label{Q8}
\gamma=\frac{n+2-\varepsilon}{2\lambda}\quad\text{and}\quad p=\frac{2\lambda}{n+2-2\varepsilon}.$$ Then $$\label{Q9}
\gamma p>1.$$ Let $\{T_j\}\subset(0,1)$ be a sequence such that $T_j \to0$ as $j\to\infty$. Define $w_j :(-\infty,T_j )\to(0,\infty)$ by $$\label{Q10}
w_j (t)=(T_j -t)^{-1/p}$$ and define $t_j \in(0,T_j )$ by $$\label{Q11}
w_j (t_j )=t^{-\gamma}_{j}.$$ Then $$\label{Q12}
\frac{T_j -t_j}{t_j}=\frac{w_j (t_j )^{-p}}{t_j}=t^{\gamma p-1}_{j}\to0\quad\text{as }j\to\infty$$ by .
Choose $a_j \in((t_j+T_j)/2,T_j )$ such that $w_j (a_j )>j\varphi(a_j )$. Then $$\label{Q13}
\frac{w_j (a_j )}{\varphi(a_j )}\to\infty\quad\text{as }j\to\infty.$$ Let $h_j (s)=\sqrt{a_j-s}$ and $H_j (s)=\sqrt{a_j +\varepsilon_j -s}$ where $\varepsilon_j >0$ satisfies $$\label{Q14}
a_j +2\varepsilon_j <T_j, \quad t_j-\varepsilon_j >t_j/2, \quad
\varepsilon_j <T^{2}_{j}, \quad \text{and} \quad w_j (t_j -\varepsilon_j )>\frac{w_j (t_j )}{2}.$$ Define $$\begin{aligned}
&\omega_j =\{(y,s)\in\mathbb{R}^n \times\mathbb{R}:|y|<h_j (s)\quad\text{and}\quad t_j <s<a_j \},\\
&\Omega_j =\{(y,s)\in\mathbb{R}^n \times\mathbb{R}:|y|<H_j (s)\quad\text{and}\quad t_j -\varepsilon_j <s<a_j +\varepsilon_j \}.
\end{aligned}$$ By taking a subsequence we can assume the sets $\Omega_j$ are pairwise disjoint.
Let $\chi_j :\mathbb{R}^n \times\mathbb{R}\to[0,1]$ be a $C^\infty$ function such that $\chi_j \equiv1$ in $\omega_j$ and $\chi_j \equiv0$ in $\mathbb{R}^n \times\mathbb{R}\backslash\Omega_j$. Define $f_j ,\,u_j :\mathbb{R}^n \times\mathbb{R}\to[0,\infty)$ by $$\label{Q15}
f_j (y,s)=\chi_j (y,s)w^{\prime}_{j}(s)$$ and $$\label{Q16}
u_j (x,t)=\iint_{\mathbb{R}^n \times\mathbb{R}}\Phi(x-y,t-s)f_j (y,s)\,dy\,ds.$$ Then $f_j$ and $u_j$ are $C^\infty$ and $$\label{Q17}
Hu_j =f_j \quad\text{in }\mathbb{R}^n \times\mathbb{R}.$$ By Theorem \[thmB.2\] with $p=n+2$ and $q=\infty$ we see that $$\begin{aligned}
\label{Q18}
\notag\biggl\|\iint_{\Omega_j \backslash \omega_j}&\Phi(x-y,t-s)
w^{\prime}_{j}(s)\,dy\,ds\biggr\|_{L^\infty (\mathbb{R}^n \times(0,1))}\\
\notag &=\biggl\|\iint_{\mathbb{R}^n \times(0,1)}
\Phi(x-y,t-s)\chi_{\Omega_j \backslash \omega_j}(y,s)
w^{\prime}_{j}(s)\,dy\,ds\biggr\|_{L^\infty (\mathbb{R}^n \times(0,1))}\\
\notag &\leq C_n \| w^{\prime}_{j}(s)\|_{L^{n+2}(\Omega_j \backslash \omega_j )}\\
&\leq w_j (t_j )
\end{aligned}$$ provided we decrease $\varepsilon_j$ if necessary because $|\Omega_j \backslash \omega_j |\to0$ as $\varepsilon_j \to0$.
Also, it follows from $_2$, $_1$, , $_1$, , and that there exists a positive constant $M$, independent of $j$, such that for $(x,t)\in\Omega_j$ we have $$\label{Q19}
M\Psi(x,t)>\frac{2^{\gamma+1}}{t^\gamma}>\frac{2^{\gamma+1}}{T^{\gamma}_{j}}>\frac{2^{\gamma+1}}{(2t_j )^\gamma}=2w_j (t_j ),$$ provided we take a subsequence if necessary, where $\Psi$ is defined by .
In order to obtain a lower bound for $u_j$ in $\omega_j$, note first that for $s<t\leq a_j +\varepsilon_j$ and $|x|\leq H_j (t)$ we have by Lemma \[lem2.10\] that $$\label{Q20}
\int_{|y|<H_j (s)}\Phi(x-y,t-s)\,dy\geq b$$ for some constant $$\label{Q21}
b=b(n)\in(0,1).$$ Next using and , we find for $(x,t)\in\Omega_j$ that $$\begin{aligned}
\iint_{\Omega_j}\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds&=\int^{t}_{t_j -\varepsilon_j}w^{\prime}_{j}(s)\left(\int_{|y|<H_j (s)}\Phi(x-y,t-s)\,dy\right)ds\\
&\geq b(w_j (t)-w_j (t_j -\varepsilon_j ))\\
&\geq bw_j (t)-w_j (t_j ).
\end{aligned}$$ It therefore follows from , , and that for $(x,t)\in\Omega_j$ we have $$\begin{aligned}
\label{Q22}
\notag u_j (x,t)&\geq\iint_{\omega_j }\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds\\
\notag &=\iint_{\Omega_j}\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds-\iint_{\Omega_j \backslash \omega_j}\Phi(x-y,t-s)w^{\prime}_{j}(s)\,dy\,ds\\
&\geq bw_j (t)-2w_j (t_j ).
\end{aligned}$$
Define $\beta>0$ by $$\label{Q23}
\frac{1}{\beta}-\frac{1}{\lambda}=\frac{2}{n+2}.$$ Then by $$\label{Q24}
\frac{2}{n+2}<\frac{1}{\beta}=\frac{1}{\lambda}+\frac{2}{n+2}\leq\frac{n}{n+2}+\frac{2}{n+2}=1$$ and by $$p>\frac{2\lambda}{n+2}=\frac{2}{(n+2)/\lambda}=\frac{2}{\frac{n+2}{\beta}-2}=\frac{2\beta}{n+2-2\beta}.$$ Thus $$\label{Q25}
\frac{n}{2}-\frac{\beta(p+1)}{p}+1=\frac{(n+2-2\beta)p-2\beta}{2p}>0.$$ Next we slightly increase $\beta$ in such a way that and the first inequality in still hold. Then instead of and we get $$\label{Q26}
\frac{1}{\beta}-\frac{1}{\lambda}<\frac{2}{n+2}$$ and $$\label{Q27}
\frac{2}{n+2}<\frac{1}{\beta}<1$$ respectively.
From , , $_1$ and we find that $$\begin{aligned}
\notag p^\beta &\iint_{\mathbb{R}^n \times\mathbb{R}}f_j (y,s)^\beta \,dy\,ds\leq p^\beta \iint_{\Omega_j}w^{\prime}_{j}(s)^\beta \,dy\,ds\\
\notag &\leq p^\beta \int^{T_j}_{0}\int_{|y|<\sqrt{T_j -s}}w^{\prime}_{j}(s)^\beta \,dy\,ds\\
\notag &=|B_1(0)|\int^{T_j}_{0}(T_j -s)^{n/2-\beta(p+1)/p}ds\\
\label{Q28} &=|B_1(0)|\int^{T_j}_{0}\tau^{n/2-\beta(p+1)/p}\,d\tau\to0\quad\text{as }j\to\infty.
\end{aligned}$$ Hence by , , , and Theorem \[thmB.2\] we obtain $$\label{Q29}
\| u_j \|_{L^\lambda (\mathbb{R}^n \times(0,1))}\to0\quad\text{as }j\to\infty.$$ Repeating the derivation of with $\beta$ replaced with $1$, we find that $$\iint_{\mathbb{R}^n \times\mathbb{R}}f_j (y,s)\,dy\,ds\to0\quad\text{as }j\to\infty.$$ Thus $$\iint_{\mathbb{R}^n \times\mathbb{R}}\sum^{\infty}_{j=1}f_j (y,s)d\eta ds<\infty$$ provided we take a subsequence if necessary. Hence, since the $C^\infty$ functions $f_j$ have disjoint supports, it follows from Theorem \[thm5.2\] that the function $u:\mathbb{R}^n \times(0,\infty)\to(0,\infty)$ defined by $$\label{Q30}
u(x,t)=(M+1)\Psi(x,t)+\sum^{\infty}_{j=1}u_j (x,t)$$ is $C^\infty$ and from and Theorem \[thm5.2\] we have $$\label{Q31}
Hu=\sum^{\infty}_{j=1}f_j \quad\text{in }\mathbb{R}^n \times(0,\infty).$$ By and Theorem \[thm5.2\], $$u\in L^\lambda (\mathbb{R}^n \times(0,1))$$ provided we take a subsequence of $u_j$ if necessary. Thus holds.
We now prove . By and we have $$Hu\equiv0\quad\text{in }(\mathbb{R}^n \times(0,1))\backslash\bigcup^{\infty}_{j=1}\Omega_j.$$ Hence to prove , it suffice to prove there exists a positive constant $C=C(n,\lambda,\sigma,\alpha)$ such that $$\label{Q32}
0\leq Hu\leq C(\Phi^{\alpha/n}*u^\lambda )u^\sigma \quad\text{in }\Omega_j$$ for $j=1,2,...$.
By , , and we have for $(x,t)\in\Omega_j$ that $$\label{Q33}
u(x,t)\geq(M+1)\Psi(x,t)+bw_j (t)-2w_j (t_j )\geq\Psi(x,t)+bw_j (t).$$ Thus for $(x,t)\in\Omega_j$ we see by , , and that $$\begin{aligned}
Hu(x,t)&=f_j (x,t)\leq w^{\prime}_{j}(t)=\frac{1}{p}w_j (t)^{1+p}\\
&=\frac{1}{p}w_j(t)^{1+p-\sigma}w_j(t)^\sigma
\le \frac{1}{pb^\sigma}w_j(t)^{1+p-\sigma}u(x,t)^\sigma.
\end{aligned}$$ Hence to prove it suffices to show $$\label{Q34}
w_j (t)^{1+p-\sigma}< C\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)^{\alpha/n}u(y,s)^\lambda \,dy\,ds\quad\text{for }(x,t)\in\Omega_j .$$
Our proof of consists of two cases.
**Case I.** Suppose $$\label{Q35}
(x,t)\in\Omega_j \quad\text{and}\quad t\leq\frac{T_j +t_j}{2}.$$ Then using $_4$, $_2$, $_2$ and the fact that $w_j$ is an increasing function we have $$\begin{aligned}
\frac{1}{2}&\leq\frac{w_j (t)}{2w_j (t_j -\varepsilon_j )}<\frac{w_j (t)}{w_j (t_j )}\\
&\leq\left(\frac{T_j -\frac{T_j +t_j}{2}}{T_j -t_j}\right)^{-1/p}=2^{1/p}<2^{n/2}.
\end{aligned}$$ Also by and $$\frac{w_j (t_j )}{T^{-\gamma}_{j}}=\left(\frac{T_j}{t_j}\right)^\gamma \in(1,2)$$ provided we take a subsequence if necessary. Thus implies $$\label{Q36}
\frac{1}{2}<\frac{w_j (t)}{T^{-\gamma}_{j}}<2^{(n+2)/2}.$$ Next making the change of variables $$x=\sqrt{T_j}\xi,\quad t=T_j \tau;\qquad y=\sqrt{T_j}\eta,\quad s=T_j
\zeta;\qquad\text{and}\quad \hat{y}=\sqrt{T_j}\hat{\eta},$$ we get for $(y,s)\in\mathbb{R}^n \times(0,\infty)$ that $$\begin{aligned}
\Psi(y,s)&=\int_{\mathbb{R}^n}\Phi(y-\hat{y},s)|\hat{y}|^{-2\gamma}d\hat{y}\\
&=\int_{\mathbb{R}^n}\frac{1}{T^{n/2}_{j}}\Phi(\eta-\hat{\eta},\zeta)T^{-\gamma}_{j}|\hat{\eta}|^{-2\gamma}T^{n/2}_{j}d\hat{\eta}\\
&=T^{-\gamma}_{j}\Phi(\eta,\zeta)
\end{aligned}$$ and thus for $(x,t)\in\Omega_j$ we obtain from $_1$ that $$\begin{aligned}
\notag \iint_{\mathbb{R}^n
\times(0,1)}&\Phi(x-y,t-s)^{\alpha/n}\Psi(y,s)^\lambda \,dy\,ds\\
&=\iint_{\mathbb{R}^n \times(0,\tau)}\left(\frac{1}{T^{n/2}_{j}}\Phi(\xi-\eta,\tau-\zeta)\right)^{\alpha/n}(T^{-\gamma}_{j}\Psi(\eta,\zeta))^\lambda
T^{\frac{n+2}{2}}_{j}\,d\eta\,d\zeta \notag\\
&\geq\frac{G(\xi,\tau)}{\sqrt{T_j}^{\alpha+2\gamma\lambda-(n+2)}}
=\frac{G(\xi,\tau)}{\sqrt{T_j}^{\alpha-\varepsilon}}\label{Q37}
\end{aligned}$$ where $$G(\xi,\tau):=\iint_{B_1 (0)\times(1/2,\tau)}\Phi(\xi-\eta,\tau-\zeta)^{\alpha/n}\Psi(\eta,\zeta)^\lambda \,d\eta\,d\zeta.$$ Since by $_1$, $_1$, , and $_3$, $$1>\tau=\frac{t}{T_j}\geq\frac{t_j -\varepsilon_j}{T_j}\to1\quad\text{as }j\to\infty$$ we have by $_1$ that $$|\xi|=\frac{|x|}{\sqrt{T_j}}<\frac{\sqrt{T_j -t}}{\sqrt{T_j}}=\sqrt{1-\frac{t}{T_j}}\to0\quad\text{as }j\to\infty.$$ Thus, since $G$ is clearly continuous at $(\xi,\tau)=(0,1)$ and $G(0,1)>0$ we have by that $$\label{Q38}
\iint_{\mathbb{R}^n \times(0,1)}\Phi(x-y,t-s)^{\alpha/n}\Psi(y,s)^\lambda \,dy\,ds\geq\frac{C}{\sqrt{T_j}^{\alpha-\varepsilon}}\quad\text{for }(x,t)\in\Omega_j$$ provided we take a subsequence if necessary.
Since by and $_2$ $$\sigma-1>\left(\frac{2-\alpha+2\varepsilon}{n+2-2\varepsilon}\right)\lambda =p-\frac{\alpha-2\varepsilon}{n+2-2\varepsilon}\lambda>p-\frac{\alpha-\varepsilon}{n+2-\varepsilon}\lambda$$ we have by $_1$ that $$\gamma(1+p-\sigma)<\gamma\left(\frac{(\alpha-\varepsilon)\lambda}{n+2-\varepsilon}\right)=\frac{\alpha-\varepsilon}{2}.$$ Thus follows from , , and .
**Case II.** Suppose $$\label{Q39}
(x,t)\in\Omega_j \quad\text{and}\quad t\geq\frac{T_j +t_j}{2}.$$ Then for $s<t$ we have by Lemma \[lem2.10\] that $$\int_{|y|<H_j (s)}\Phi(x-y,t-s)^{\alpha/n}dy\geq\frac{C}{(t-s)^{(\alpha-n)/2}}$$ for some positive constant $C=C(n,\alpha)$. Thus for $(x,t)$ satisfying we get $$\begin{aligned}
\notag \iint_{\Omega_j}\Phi&(x-y,t-s)^{\alpha/n}w_j (s)^\lambda \,dy\,ds\geq\int^{t}_{t_j}w_j (s)^\lambda \left(\int_{|y|<H_j (s)}\Phi(x-y,t-s)^{\alpha/n}dy\right)ds\\
\notag &\geq C\int^{t}_{t_j}\frac{ds}{(t-s)^{(\alpha-n)/2}(T_j -s)^{\lambda/p}}\\
\notag &=\frac{C}{(T_j -t )^{(\alpha-n)/2+\lambda/p-1}}\int^{\frac{T_j -t_j}{T_j -t}}_{1}\frac{dz}{(z-1)^{(\alpha-n)/2}z^{\lambda/p}}\text{ where }T_j -s=(T_j -t)z\\
\notag &\geq\frac{C}{(T_j -t)^{(\alpha-n)/2+\lambda/p-1}}\int^{2}_{1}\frac{dz}{(z-1)^{(\alpha-n)/2}z^{\lambda/p}}\\
\label{Q40} &=\frac{C}{(T_j -t)^{(\alpha-n)/2+\lambda/p-1}}=\frac{C}{(T_j -t)^{(\alpha-2\varepsilon)/2}}
\end{aligned}$$ by $_2$.
Since by and $_2$ $$\sigma-1>\frac{2-\alpha+2\varepsilon}{n+2-2\varepsilon}\lambda=p\frac{2-\alpha+2\varepsilon}{2}$$ we see that $$\frac{1}{p}(1+p-\sigma)=1+\frac{1-\sigma}{p}<1+\frac{\alpha-2-2\varepsilon}{2}=\frac{\alpha-2\varepsilon}{2}.$$ Thus follows from , , and .
Finally from and we get $$\frac{u(0,a_j )}{\varphi(a_j )}\geq\frac{bw_j (a_j )}{\varphi(a_j )}\to\infty\quad\text{as }j\to\infty,$$ which gives .
Representation formula {#secA}
======================
In this appendix we provide the following representation formula for nonnegative supertemperatures.
\[thmA\] Suppose $0<R_1<R_2<R_3$ are constants and $u$ is a $C^{2,1}$ nonnegative solution of $$\label{A.1}
Hu\ge 0\quad \text{in } B_{\sqrt{R_3}}(0)\times (0,R_3)
\subset \R^n \times \R,\ n\ge 1,$$ where $Hu = u_t-\Delta u$ is the heat operator. Then $$\label{A.2}
Hu \in L^1(B_{\sqrt{R_2}}(0) \times (0,R_2)),$$ $$\label{A.2.5}
u^\beta \in L^1(B_{\sqrt{R_1}}(0) \times (0,R_1))\quad\text{for }
1\le\beta <\frac{n+2}{n}$$ and there exist a finite positive Borel measure $\mu$ on $B_{\sqrt{R_2}}(0)$ and a bounded function\
$h\in C^{2,1}(B_{\sqrt{R_1}}(0) \times (-R_1,R_1))$ satisfying $$\begin{aligned}
{2}
\label{A.3}
Hh &= 0 \quad \text{in } &B_{\sqrt{R_1}}(0)\times (-R_1,R_1)\\
\label{A.4}
h &= 0 \quad \text{in } &B_{\sqrt{R_1}}(0)\times (-R_1,0] \end{aligned}$$ such that $$\label{A.5}
u = N +v+h\quad \text{in } B_{\sqrt{R_1}}(0)\times (0,R_1)$$ where $$\begin{aligned}
\label{A.6}
N(x,t) &:= \int^{R_2}_0 \int_{|y|<\sqrt{R_2}} \Phi(x-y,t-s) Hu(y,s)\,dy\,ds,\\
\label{A.7}
v(x,t) &:= \int_{|y|<\sqrt{R_2}} \Phi(x-y,t)\,d\mu(y),\end{aligned}$$ and $\Phi$ is the heat kernel .
When $\beta=1$, $R_1=1$, $R_2=4$, and $R_3=16$, Theorem \[thmA\] was proved in [@T2011]. The proof of Theorem \[thmA\] when $\beta=1$ is obtained by making straighforward changes to the proof in [@T2011]. It remains only to prove for $1<\beta<(n+2)/n$. To do this, it suffices by to show $$\label{A.8}
N^\beta\in L^1(\R^n\times(0,R_2))\quad\text{for }1<\beta<(n+2)/n$$ and $$\label{A.9}
v^\beta\in L^1(\R^n\times(0,R_2))\quad\text{for }1<\beta<(n+2)/n.$$ Theorem \[thmB.2\] and imply .
Finally, for $t>0$, $\beta>1$, and $\beta'$ the conjugate Hölder exponent of $\beta$ we have $$\begin{aligned}
\int_{\R^n}v(x,t)^\beta\,dx
&=\int_{\R^n}\left(\int_{|y|<\sqrt{R_2}}\Phi(x-y,t)\,d\mu(y)\right)^\beta dx\\
&\le \int_{\R^n}\left(\int_{|y|<\sqrt{R_2}}1^{\beta'}\,d\mu(y)\right)^{\beta/\beta'}
\int_{|y|<\sqrt{R_2}}\Phi(x-y,t)^\beta\,d\mu(y)\,dx\\
&=C\int_{|y|<\sqrt{R_2}}\left(\int_{\R^n}\Phi(x-y,t)^\beta\,dx\right)\,d\mu(y)\\
&=C\int_{|y|<\sqrt{R_2}}t^{-n\beta/2}\int_{\R^n}e^{-\frac{\beta|x-y|^2}{4t}}dx\,d\mu(y)\\
&=Ct^{n(1-\beta)/2}\quad\text{by Lemma \ref{lem2.2}}\end{aligned}$$ which implies .
\[remA\] If $u$ is a $C^{2,1}$ nonnegative solution of where $R_3>0$ then by Theorem \[thmA\], $$u^\beta \in L^1(B_{\sqrt{R}}(0) \times (0,R))\quad\text{for }
1\le\beta <\frac{n+2}{n}\text{ and }0<R<R_3.$$ Thus the conclusion in Theorem \[thmA\] can be replaced with $$u^\beta \in L^1(B_{\sqrt{R_2}}(0) \times (0,R_2))\quad\text{for }
1\le\beta <\frac{n+2}{n}.$$
\[thmA\] Suppose $u$ is a $C^{2,1}$ nonnegative solution of $$\label{A.1}
Hu\ge 0\quad \text{in } B_{\sqrt{4R}}(0)\times (0,4R)
\subset \R^n \times \R,\ n\ge 1,$$ where $Hu = u_t-\Delta u$ is the heat operator and $R$ is a positive constant. Then $$\label{A.2}
u^\beta,Hu \in L^1(B_{\sqrt{2R}}(0) \times (0,2R))\quad\text{for }
1\le\beta <\frac{n+2}{n}$$ and there exist a finite positive Borel measure $\mu$ on $B_{\sqrt{2R}}(0)$ and a bounded function\
$h\in C^{2,1}(B_{\sqrt{R}}(0) \times (-R,R))$ satisfying $$\begin{aligned}
{2}
\label{A.3}
Hh &= 0 \quad \text{in } &B_{\sqrt{R}}(0)\times (-R,R)\\
\label{A.4}
h &= 0 \quad \text{in } &B_{\sqrt{R}}(0)\times (-R,0] \end{aligned}$$ such that $$\label{A.5}
u = N +v+h\quad \text{in } B_{\sqrt{R}}(0)\times (0,R)$$ where $$\begin{aligned}
\label{A.6}
N(x,t) &:= \int^{2R}_0 \int_{|y|<\sqrt{2R}} \Phi(x-y,t-s) Hu(y,s)\,dy\,ds,\\
\label{A.7}
v(x,t) &:= \int_{|y|<\sqrt{2R}} \Phi(x-y,t)\,d\mu(y),\end{aligned}$$ and $\Phi$ is the heat kernel .
When $R=1$ and $\beta=1$, Theorem \[thmA\] was proved in [@T2011]. The proof of Theorem \[thmA\] for $R>0$ and $\beta=1$ is obtained by scaling the $R=1$ case. It remains only to prove for $1<\beta<(n+2)/n$. To do this, it suffices by to show $$\label{A.8}
N^\beta\in L^1(\R^n\times(0,2R))\quad\text{for }1<\beta<(n+2)/n$$ and $$\label{A.9}
v^\beta\in L^1(\R^n\times(0,2R))\quad\text{for }1<\beta<(n+2)/n.$$ Since $Hu\in L^1(B_{\sqrt{2R}}(0)\times(0,2R))$ we see that follows from heat potential estimates. (See [@GT2016 Theorem B.8].)
Finally, for $t>0$, $\beta>1$, and $\beta'$ the conjugate Hölder exponent of $\beta$ we have $$\begin{aligned}
\int_{\R^n}v(x,t)^\beta\,dx
&=\int_{\R^n}\left(\int_{|y|<\sqrt{2R}}\Phi(x-y,t)\,d\mu(y)\right)^\beta dx\\
&\le \int_{\R^n}\left(\int_{|y|<\sqrt{2R}}1^{\beta'}\,d\mu(y)\right)^{\beta/\beta'}
\int_{|y|<\sqrt{2R}}\Phi(x-y,t)^\beta\,d\mu(y)\,dx\\
&=C\int_{|y|<\sqrt{2R}}\left(\int_{\R^n}\Phi(x-y,t)^\beta\,dx\right)\,d\mu(y)\\
&=C\int_{|y|<\sqrt{2R}}t^{-n\beta/2}\int_{\R^n}e^{-\frac{\beta|x-y|^2}{4t}}dx\,d\mu(y)\\
&=Ct^{n(1-\beta)/2}\quad\text{by Lemma \ref{lem2.2}}\end{aligned}$$ which implies .
Heat potential estimates {#secB}
========================
In this appendix we provide estimates for the heat potentials $$(J_\alpha f)(x,t)=\iint_{\R^n\times\R}\Phi(x-y,t-s)^{\frac{n+2-\alpha}{n}}f(y,s)\,dy\,ds$$ and $$(V_\alpha f)(x,t)=\iint_{\Omega}
\Phi(x-y,t-s)^{\frac{n+2-\alpha}{n}}f(y,s)\,dy\,ds,$$ where $\Phi$ is given by , $\Omega=\R^n\times (a,b)$, and $\alpha\in(0,n+2)$. The proofs of these estimates are given in [@GT2016 Appendix B].
\[thmB.1\] Suppose $0<\alpha<n+2$ and $1<p<\frac{n+2}{\alpha}$ are constants and $f:\R^n \times\R\to\R$ is a nonnegative measurable function. Let $$q=\frac{(n+2)p}{n+2-\alpha p}.$$ Then $$\| J_\alpha f\|_{L^q (\R^n \times\R)}\leq C\| f\|_{L^p (\R^n
\times\R)}$$ where $C=C(n,p,\alpha)$ is a positive constant.
\[thmB.2\] Let $p,q\in[1,\infty]$, $\alpha$, and $\delta$ satisfy $$\label{B.1}
0\le \delta =\frac{1}{p}-\frac{1}{q}<\frac{\alpha}{n+2}<1.$$ Then $V_\alpha$ maps $L^p(\Omega)$ continuously into $L^q(\Omega)$ and for $f\in L^p(\Omega)$ we have $$\|V_\alpha f\|_{L^q(\Omega)} \le M\|f\|_{L^p(\Omega)},$$ where $$M=C(b-a)^{(\alpha-(n+2)\delta)/2}\quad\text{for some constant }
C=C(n,\alpha,\delta)>0.$$
Theorem \[thmB.2\] is weaker than Theorem \[thmB.1\] in that the second inequality in cannot be replaced with equality. However it is stronger in that the cases $p=1$ and $q=\infty$ are allowed.
[10]{} H. Chen and F. Zhou, Classification of isolated singularities of positive solutions for Choquard equations, arXiv:1512.03181 \[math.AP\].
J. T. Devreese and A. S. Alexandrov, Advances in polaron physics, Springer Series in Solid-State Sciences, vol. 159, Springer, 2010.
M. Ghergu and S. D. Taliaferro, *Isolated Singularities in Partial Differential Inequalities*, Cambridge University Press, 2016.
M. Ghergu and S. D. Taliaferro, Pointwise bounds and blow-up for Choquard-Pekar inequalities at an isolated singularity, *J. Differential Equations* **261** (2016), 189–217.
K.R.W. Jones, Newtonian quantum gravity, *Australian Journal of Physics* **48** (1995), 1055–1081.
E. H. Lieb, Existence and uniqueness of the minimizing solution of Choquard’s nonlinear equation, Studies in Appl. Math. 57 (1976/77), 93–105.
P.-L. Lions, The Choquard equation and related questions, *Nonlinear Anal.* **4** (1980), 1063–1072.
P.-L. Lions, The concentration-compactness principle in the calculus of variations. The locally compact case. I, *Ann. Inst. H. Poincaré Anal. Non Linéaire* **1** (1984), 109–145.
C. Ma, W. Chen, and C. Li, Regularity of solutions for an integral system of Wolff type, *Adv. Math.* **226** (2011), 2676–2699.
L. Ma and L. Zhao, Classification of positive solitary solutions of the nonlinear Choquard equation, *Arch. Ration. Mech. Anal.* **195** (2010), 455–467.
M. Melgaard and F. Zongo, Multiple solutions of the quasirelativistic Choquard equation, *J. Math. Phys.* **53** (2012), 033709.
I. M. Moroz, R. Penrose and P. Tod, Spherically-symmetric solutions of the Schrödinger-Newton equations, *Classical Quantum Gravity* **15** (1998), 2733–2742.
V. Moroz and J. Van Schaftingen, Groundstates of nonlinear Choquard equations: Existence, qualitative properties and decay asymptotics, *J. Funct. Anal.* **265** (2013), 153–184.
V. Moroz and J. Van Schaftingen, Nonexistence and optimal decay of supersolutions to Choquard equations in exterior domains, *J. Differential Equations* **254** (2013), 3089–3145.
V. Moroz and J. Van Schaftingen, Semi-classical states for the Choquard equation, *Calc. Var. Partial Differential Equations* **52** (2015), 199–235.
S. Pekar, Untersuchung über die Elektronentheorie der Kristalle, Akademie Verlag, Berlin, 1954.
P. Quittner and P. Souplet, Superlinear parabolic problems, blow-up, global existence and steady states, Birkhauser, Basel, 2007.
S. G. Samko, Hypersingular Integrals and Their Applications, Taylor and Francis, London, 2002.
S. D. Taliaferro, Initial blow-up of solutions of semilinear parabolic inequalities, *J. Differential Equations* **250** (2011), 892–928.
J. Wei and M. Winter, Strongly interacting bumps for the Schrödinger-Newton equations, *J. Math. Phys.* **50** (2009), 012905.
R. Zhuo, W. Chen, X. Cui, and Z. Yuan, Symmetry and non-existence of solutions for a nonlinear system involving the fractional Laplacian, *Discrete Contin. Dyn. Syst.* **36** (2016), 1125–1141.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The relation between nonanticipative Rate Distortion Function (RDF) and filtering theory is discussed on abstract spaces. The relation is established by imposing a realizability constraint on the reconstruction conditional distribution of the classical RDF. Existence of the extremum solution of the nonanticipative RDF is shown using weak$^*$-convergence on appropriate topology. The extremum reconstruction conditional distribution is derived in closed form, for the case of stationary processes. The realization of the reconstruction conditional distribution which achieves the infimum of the nonanticipative RDF is described. Finally, an example is presented to illustrate the concepts.'
author:
- 'Charalambos D. Charalambous, Photios A. Stavrou, and Nasir U. Ahmed[^1] [^2] [^3]'
bibliography:
- 'photis\_references\_filtering\_weakstar.bib'
title: Nonanticipative Rate Distortion Function and Relations to Filtering Theory
---
Nonanticipative Rate Distortion Function (RDF), filtering, realization, weak$^*$-convergence, optimal reconstruction conditional distribution
Introduction
============
Shannon’s information theory [@shannon1948] for reliable communication evolved over the years without much emphasis on nonanticipation imposed on the communication sub-systems. In particular, the classical rate distortion function (RDF) for source data compression or quantization deals with the characterization of the optimal reconstruction conditional distribution subject to a fidelity criterion [@berger; @cover-thomas], without regard to nonanticipation.
On the other hand, filtering theory is developed by imposing real-time realizability on the estimators with respect to measurement data. Although, both reliable communication and filtering (state estimation for control) are concerned with reconstruction of processes, the main underlying assumptions characterizing them are different.
In this paper, the intersection of rate distortion function (RDF) and real-time realizable filtering theory is established by invoking a nonanticipative constraint on the reconstruction conditional distribution to be realizable via real-time operations, while the optimal nonanticipative reconstruction conditional distribution is derived. Consequently, the connection between nonanticipative RDF, its characterization via the optimal reconstruction conditional distribution, and real-time realizable filtering theory is established under very general conditions on the source (including Markov sources).
The fundamental advantage of the new filtering approach based on nonanticipative RDF, is the ability to ensure average or probabilistic estimation error constraints, which is non-trivial task if Bayesian filtering techniques are employed to formulate such constraints. The motivations includes nonanticipative data compression over noisy channels, such as control over networks, where the controlled system and controller may be connected via a noisy channel [@yuksel2011; @gupta-dana-hespanha-murray-hassibi2009; @freudenberg-middleton2008; @borkar-mitter-tatikonda2001; @tatikonda-mitter2004; @yuksel-meyn2012]. In such applications, filtering via nonanticipative RDF approximates sensor measurements by the reconstruction process taking values in a set of smaller cardinality, while the approximation is quantified by the distortion function. Given the recent interest in developing controller and estimator architectures processing quantized information and, in general, communication schemes for control applications, nonanticipative RDF is necessary for developing zero-delay or limited delay quantization schemes. Moreover, nonanticipative RDF is necessary for the realization of the compression channel by communication systems processing information causally.
The first relation between information theory and filtering via distortion rate function is discussed by R. S. Bucy in [@bucy], by carrying out the computation of a realizable (nonanticipative) distortion rate function with square criteria for two samples of the Ornstein-Uhlenbeck Gaussian process. Related work on nonanticipative rate distortion theory is pursued by A. K. Gorbunov and M. S. Pinsker in [@gorbunov-pinsker; @gorbunov91]. Specifically, [@gorbunov-pinsker] discussed nonanticipative RDF for general stationary processes and establishes existence of the infinite horizon limit, while [@gorbunov91] computes a closed form expression for nonanticipative RDF (called $\epsilon$-entropy) for stationary Gaussian processes using power spectral methods. Further elaborations on the similarities and differences between [@bucy; @gorbunov-pinsker; @gorbunov91] and this paper will be discussed in subsequent parts of the paper. Moreover, over the years several papers appeared in the literature in which controller or estimator are designed based on information theoretic measures [@galdos-gustafson1977; @feng-loparo-fang1997; @guo-yin-wang-chai2009]. An earlier work designing filters via information theoretic measures is [@teneketzis-thesis], while [@caines1988] analyzes mutual information for Gaussian processes.
In this paper, the connection between nonanticipative rate distortion theory and filtering theory is further examined, under a nonanticipative condition defined by the family of conditional distributions (reconstructions), for general distortion functions and random processes on abstract Polish spaces. The connection is established via optimization on the space of conditional distributions with average distortion constraint and almost sure ($a.s.$) constraints to account for the nonanticipative condition on the reconstruction conditional distribution. The main results are the following.
(1)
: Existence of the nonanticipative RDF using the topology of weak$^*$-convergence;
(2)
: Closed form expression for reconstruction conditional distribution minimizing the nonanticipative RDF for stationary processes;
(3)
: Realization procedure of the filter based on the nonanticipative RDF;
(4)
: Example to demonstrate the realization of the filter.
It is important to point out that items (1)-(4) above are not addressed in the related papers [@bucy; @gorbunov-pinsker; @gorbunov91]. Moreover, (2) together with (3) are important in reliable communication for filtering and control applications, because one can develop communication architectures which operate with zero-delay or limited delay, as opposed to the classical RDF which is anticipative.
Next,we give a high level discussion on Bayesian filtering theory and nonanticipative RDF, and we present some aspects of the problem pursued in this paper. Consider a discrete-time process $X^n\triangleq\{X_0,X_1,\ldots,X_n\}\in{\cal X}_{0,n} \triangleq \times_{i=0}^n{\cal X}_i$, and its reconstruction $Y^n\triangleq\{Y_0,Y_1,\ldots,Y_n\}\in{\cal Y}_{0,n} \triangleq \times_{i=0}^n{\cal Y}_i$, where ${\cal X}_i$ and ${\cal Y}_i$ are Polish spaces (complete separable metric spaces). The objective is to reconstruct $X^n$ by $Y^n$ via nonanticipative operations subject to a distortion or fidelity criterion. That is, for each $i=0,1,\ldots$, the reconstruction $Y_i$ of $X_i$ should depend on past and present information $\{X_0,Y_0,X_1,Y_1,\ldots,X_{i-1},Y_{i-1},X_i\}$. Once this mapping is found a procedure is introduced to realize the filter of $Y_i$ from auxiliary[^4] measurements.
Bayesian Estimation Theory
--------------------------
In classical filtering [@elliott-aggoun-moore1995], one is given a mathematical model that generates the process $X^n$, via its conditional distribution $\{P_{X_i|X^{i-1}}$ $(dx_i|x^{i-1}):i=0,1,\ldots,n\}$ or via discrete-time recursive dynamics, a mathematical model that generates observed data obtained from sensors, say, $Z^n$, $\{P_{Z_i|Z^{i-1},X^i}$ $(dz_i|z^{i-1},x^i):i=0,1,\ldots,n\}$, while $Y^n$ are the causal estimates of some function of the process $X^n$ based on the observed data $Z^n$. Note that for a memoryless channel that generates the observation sequence $\{Z_i:~i=0,1,\ldots,n\}$ then $P_{Z_i|Z^{i-1},X^i}(dz_i|z^{i-1},x^i)=P_{Z_i|X_{i}}(dz_i|x_i)-a.s.,~i=0,1,\ldots,n$.\
In Bayesian estimation one is interested in causal estimators of some function $\Phi:{\cal X}_n\longmapsto\mathbb{R}$, $Y_n\triangleq\Phi(X_n)$ based on the observed data $Z^{n-1}\triangleq\{Z_0,Z_1,\ldots,Z_{n-1}\}$. With respect to minimizing the least-squares error pay-off, the best estimate of ${\Phi}(X_i)$ given $Z^{i-1}$, denoted by $\widehat{\Phi}(X_i)$, is given by the conditional mean $$\begin{aligned}
\widehat{\Phi}(X_i)\triangleq\mathbb{E}\Big\{\Phi(X_i)|Z^{i-1}\Big\}=\int_{{\cal X}_i}\Phi(x)P_{X_i|Z^{i-1}}(dx|z^{i-1}),~i=0,1,\ldots,n.\end{aligned}$$ For non-linear problems, Bayesian filtering is often addressed via the conditional distribution $\{P_{X_i|Z^{i-1}}(dx_i|z^{i-1}):i=0,1,\ldots,n\}$ or its unnormalized versions which satisfy discrete-recursions [@elliott-aggoun-moore1995], and forms a sufficient statistic for the filtering problem.\
Consider the simplified example of the multi-dimensional Gaussian-Markov processes modeled by $$\begin{aligned}
\left\{ \begin{array}{ll} X_{k+1}=A_kX_k+B_kW_k,~X_0{\sim}N(0;~\Sigma_{x_0}),~k=0,1,\ldots,n-1\\
Z_k=C_kX_k+D_kV_k,~k=0,1,\ldots,n \end{array} \right.\label{equation50}\end{aligned}$$ where $\{A_k, B_k, C_k, D_k\}$ are time-varying matrices having appropriate dimensions, $W_k{\sim}N(0;\Sigma_{W_k})$ (Gaussian with mean zero and covariance $\Sigma_{W_k}$), $V_k{\sim}N(0;\Sigma_{V_k})$, $k=0,1,\ldots,n$, while the processes $\{W_k:~k=0,1,\ldots,n-1\}, \{V_k:~k=0,1,\ldots,n\}$ are mutually independent, and independent of $X_0$. The classical Kalman Filter [@elliott-aggoun-moore1995] is a well-known example for which the optimal reconstruction $\widehat{X_i} =\mathbb{E}[X_i | Z^{i-1}],~i=0,1,\ldots,n$, is the conditional mean which minimizes the average least-squares estimation error. Thus, in classical filtering theory both models which generate the unobserved and observed processes, $X^n$ and $Z^n$, respectively, are given á priori, and the estimator $\widehat{X}_i$ is a nonanticipative function of the past information $Z^{i-1},~i=0,1,\ldots,n$. Fig. \[filtering\] illustrates the cascade block diagram of the Bayesian filtering problem.
![Block Diagram of Bayesian Filtering Problem[]{data-label="filtering"}](filtering_problem.jpg)
Nonanticipative Rate Distortion Theory and Estimation {#1.2}
-----------------------------------------------------
In nonanticipative rate distortion theory one is given the process $X^n$, which induces the conditional distributions $\{P_{X_i|X^{i-1}}(dx_i|x^{i-1}):~i=0,1,\ldots,n\}$ and determines the nonanticipative reconstruction conditional distribution $\{P_{Y_i|Y^{i-1},X^i}(dy_i|y^{i-1},x^i):~i=0,1,\ldots,n\}$ which minimizes the mutual information between $X^n$ and $Y^n$ subject to a distortion or fidelity constraint, via a nonanticipative or realizability constraint. The filter or estimate $\{Y_i:~i=0,1,\ldots,n\}$ of $\{X_i:~i=0,1,\ldots,n\}$ is found by realizing the reconstruction distribution $\{P_{Y_i|Y^{i-1},X^i}(dy_i|y^{i-1}$, $x^i):~i=0,1,\ldots,n\}$ via a cascade of sub-systems as shown in Fig. \[filtering\_and\_causal\].
![Block Diagram of Filtering via Nonanticipative Rate Distortion Function[]{data-label="filtering_and_causal"}](CRDF_and_filtering.jpg)
The point to be made here is that the auxiliary random sequence $\{Z_0,Z_1,\ldots\}$ which is the analogue of sensor measurements (in the above discussion of Bayesian estimation) is identified during the realization of the optimal reconstruction distribution $\{P_{Y_i|Y^{i-1},X^i}(dy_i|y^{i-1},x^i):~i=0,1,\ldots,n\}$. Thus, in Bayesian estimation, the sensor map is given á priori, while in nonanticipative rate distortion theory, this map is identified during the realization of the optimal reconstruction conditional distribution $\{P_{Y_i|Y^{i-1},X^i}(dy_i|y^{i-1},x^i):~i=0,1,\ldots,n\}$, so that the end-to-end nonanticipative RDF from $X^n$ to $Y^n$ is achieved.
The precise problem formulation of nonanticipative RDF is defined by first introducing the distortion or fidelity constraint and mutual information. The distortion function [@berger] or fidelity constraint between $x^n$ and its reconstruction $y^n$, is a measurable function defined by $$\begin{aligned}
d_{0,n} : {\cal X}_{0,n} \times {\cal Y}_{0,n} \mapsto [0, \infty], \: \: d_{0,n}(x^n,y^n)\triangleq\sum^n_{i=0}\rho_{0,i}(x^i,y^i).\nonumber\end{aligned}$$ For single letter distortion $d_{0,n}(x^n,y^n)\equiv\sum_{i=0}^n\rho(x_i,y_i)$, and for single letter square error distortion $d_{0,n}(x^n,y^n)\equiv\sum_{i=0}^n{||x_i-y_i||}^2$ [@berger]. Moreover, for finite alphabet spaces ${\cal X}_i$ and ${\cal Y}_i$, the distortion function can be defined in terms of the Hamming distance [@cover-thomas].\
The mutual information between $X^n$ and $Y^n$, for a given distribution ${P}_{X^n}(dx^n)$, and conditional distribution $P_{Y^n|X^n}(dy^n|x^n)$, is defined by[^5] [@berger] $${I}(X^n;Y^n)\triangleq \int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}}\log\Big(\frac{P_{Y^n|X^n}(dy^n|x^n)}{{P}_{Y^n}(dy^n)}\Big)
P_{Y^n|X^n}(dy^n|x^n)\otimes{P}_{X^n}(dx^n). \label{1}$$ Next, introduce the nonanticipative constraint on the reconstruction distribution. To this end, define the $(n+1)-$fold nonanticipative convolution measure $${\overrightarrow P}_{Y^n|X^n}(dy^n|x^n) \triangleq \otimes^n_{i=0}P_{Y_i|Y^{i-1},X^i}(dy_i|y^{i-1},x^i)-a.s. \label{9}$$ The set of nonanticipative reconstruction distributions is defined by $$\begin{aligned}
{\overrightarrow Q}_{ad} \triangleq&\Big\{ P_{Y^n|X^n}(dy^n|x^n) :~P_{Y^n|X^n}(dy^n|x^n) ={\overrightarrow P}_{Y^n|X^n}(dy^n|x^n)-a.s. \Big\}. \label{eq18}\end{aligned}$$ Note that without the nonanticipative constraint specified by ${\overrightarrow Q}_{ad}$, the connection between filtering theory and rate distortion theory cannot be established, since in general by Bayes’ rule $P_{Y^n|X^n}(dy^n|x^n)=\otimes_{i=0}^n{P}_{Y_i|Y^{i-1},X^n}(dy_i|y^{i-1},x^n)-a.s.$, and hence, for each $i=0,1, \ldots,n$, the conditional distribution ${P}_{Y_i|Y^{i-1},X^n}(\cdot|\cdot,\cdot)$ of $Y_i$ will depend on future symbols $\{X_{i+1},X_{i+2},\ldots,X_n\}$, in addition to the past and present symbols $\{Y^{i-1},X^i\}$. However, by imposing the nonanticipative constraint (\[eq18\]), then at each time instant $i=0,1,\ldots,$ the reconstruction $Y_i$ of $X_i$ will depend on the past reconstructions $\{Y_0,\ldots,Y_{i-1}\}$ and past and present symbols $\{X_0,\ldots,X_i\}$. For filtering and control applications, the nonanticipative constraint is necessary to avoid anticipative processing of information, while for quantization or compression applications it offers the possibility to realize the compression channel (optimal reconstruction distribution) via causal operations and achieve an end-to-end compression with zero-delay.
[*Nonanticipative Distortion Rate Function.*]{} The nonanticipative distortion rate function is defined by the minimization over ${ P}_{Y^n|X^n}(dy^n|x^n)$ of the average distortion function subject to a constraint on the mutual information rate $I(X^n;Y^n)\leq R$ and the nonanticipative constraint (\[eq18\]) as follows. $$\begin{aligned}
{D}^{na}_{0,n}(R)\triangleq \inf_{{ P}_{Y^n|X^n}(dy^n|x^n)\in \overrightarrow{Q}_{ad}: I(X^n ;{Y}^n)\leq R}\mathbb{E}\Big\{d_{0,n}(X^n,Y^n)\Big\}.\label{6}\end{aligned}$$ The classical distortion rate function does not imposes the nonanticipative constraint $P_{Y^n|X^n}(dy^n|$\
$x^n)={\overrightarrow P}_{Y^n|X^n}(dy^n|x^n)-a.s.$, hence the resulting optimal reconstruction distribution of symbol $y_i$ will depend on $(y^{i-1},x^i)$ and on future symbols $(x_{i+1},\ldots,x_n)$. Thus, by solving (\[6\]) and then realizing the conditional distribution the optimal causal filter will be defined.\
At this stage it is important to point out that the nonanticipative condition (\[eq18\]) is different from the realizability condition in [@bucy], in which is assumed that $Y_i$ is independent of $X_{j|i}^*\triangleq X_j -\mathbb{E}\Big(X_j|X^i\Big), j=i+1, i+2, \ldots,$. Moreover, the nonanticipative condition (\[eq18\]) is implied by the nonanticipative condition found in [@gorbunov-pinsker; @gorbunov91], defined by $X_{n+1}^\infty \leftrightarrow X^n \leftrightarrow Y^n$ forms a Markov chain for any $n=0,1,\ldots$ (e.g., $P_{Y^n|X^n,X_{n+1}^{\infty}}(dy^n|x^n,x_{n+1}^{\infty})=P_{Y^n|X^n}(dy^n|x^n),~n=0,1,\ldots$). The claim here is that the nonanticipative condition (\[eq18\]) is more natural and applies to processes which are not necessarily Gaussian with square error distortion function, and subject to slight modification to controlled sources in which the control is a function of the reconstruction process (we shall discuss this point further).
[*Nonanticipative Rate Distortion Function.*]{} An equivalent problem to (\[6\]) is the nonanticipative RDF defined by $${R}^{na}_{0,n}(D)\triangleq \inf_{{ P}_{Y^n|X^n}(dy^n|x^n)\in \overrightarrow{Q}_{ad}:~{E}\big\{d_{0,n}(X^n,Y^n)\leq{D}\big\}}I(X^n; Y^n).\label{7}$$ The two problems defined by (\[6\]) and (\[7\]) are equivalent in the sense that the solution of (\[6\]) gives that of (\[7\]) and vice-versa [@cover-thomas]. Moreover, it can be shown that $$\begin{aligned}
{P}_{Y^n|X^n}(dy^n|x^n)&=&{\overrightarrow P}_{Y^n|X^n}(dy^n|x^n)-a.s.
{\Longleftrightarrow} \nonumber\\
I(X^n;Y^n)&=&\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}} \log\Big(\frac{{\overrightarrow P}_{Y^n|X^n}(dy^n|x^n)}{{P}_{Y^n}(dy^n)}\Big){\overrightarrow P}_{Y^n|X^n}(dy^n|x^n)\otimes{P}_{X^n}(dx^n)\nonumber\\
&\equiv&{\mathbb I}(P_{X^n},{\overrightarrow P}_{Y^n|X^n})\label{eq6}\end{aligned}$$ where the notation ${\mathbb I}(P_{X^n},{\overrightarrow P}_{Y^n|X^n})$ is used to point out the functional dependence of $I(X^n;{Y^n})$ on $\{P_{X^n},{\overrightarrow P}_{Y^n|X^n}\}$. The nonanticipative distortion rate function and RDF can be generalized to controlled sources.
The paper is organized as follows. Section \[problem\_formulation\] discusses the problem formulation on abstract spaces. Section \[existence\] establishes existence of optimal minimizing reconstruction distribution, and Section \[necessary\] derives the stationary solution. Section \[realization\] describes the real-time realization of nonanticipative RDF. Finally, Section \[example\] demonstrates the filter realization via an example.
Formulation of Nonanticipative RDF on Abstract Spaces {#problem_formulation}
=====================================================
Throughout the paper we use the notation defined on Table \[notations\].
------------------------------------------------------------------------------------------------------------------------------------------------ --- ----------------------------------------------------------------------------------------------------
$\mathbb{N}\triangleq\{0,1,\ldots\}$ Set of nonnegative integers
$\mathbb{N}^n\triangleq\{0,1,\ldots,n\}$ Set of first $(n+1)$ nonnegative integers
${\cal X}_t$, ${\cal Y}_t$, $t\in\mathbb{N}$ Source and reconstruction alphabets
${\cal B}({\cal X}_t)$, ${\cal B}({\cal Y}_t)$ $\sigma$-algebra of Borel sets generated by ${\cal X}_t$, ${\cal Y}_t$
${\cal X}_{0,n}\triangleq {\times}_{k=0}^{n}{\cal X}_k$, ${\cal Y}_{0,n}\triangleq {\times}_{k=0}^{n}{\cal Y}_k$ Cartesian product of source and reconstruction alphabets
${\cal B}({\cal X}_{0,n})\triangleq\times_{k=0}^n{\cal B}({\cal X}_k)$, ${\cal B}({\cal Y}_{0,n})\triangleq\times_{k=0}^n{\cal B}({\cal Y}_k)$ $\sigma$-algebra of Borel sets generated by ${\cal X}_{0,n}$, ${\cal Y}_{0,n}$
$x^n \triangleq \{x_0,\ldots,x_n\}$, $y^n \triangleq \{y_0,\ldots,y_n\}$ Sequence of source and reconstruction symbols
${\cal M}_1({\cal Z})$ Set of probability measures on a measurable space $({\cal Z}, {\cal B}({\cal Z}))$
${\cal Q}({\cal Y};{\cal X})$ Set of stochastic kernels on $({\cal Y},{\cal B}({\cal Y}))$ given $({\cal X},{\cal B}(\cal X))$
$X\leftrightarrow Y \leftrightarrow Z\Leftrightarrow{P}_{Z|X,Y}(dz|x,y)=P_{Z|Y}(dz|y)-a.s.$ Markov Chain (MC) or conditional independence
$BC({\cal X})$ Vector space of bounded continuous real-valued functions defined on a Polish space ${\cal X}$
$L_1\big{(}\mu,BC({\cal X})\big{)}$ Set of all $\mu$-integrable functions defined on ${\cal X}$ with values in $BC({\cal X})$
$||\cdot||_{\mu}$ Norm with respect to $L_1\big{(}\mu,BC({\cal X})\big{)}$
${\cal X}^*$ Topological dual of a Banach space ${\cal X}$
${M}_{rba}({\cal X})$ Space of finitely additive regular bounded signed measures on $({\cal X},{\cal B}({\cal X}))$
${\Pi}_{rba}({\cal X})\subset{M}_{rba}({\cal X})$ Space of finitely additive regular bounded probability measures on $({\cal X},{\cal B}({\cal X}))$
------------------------------------------------------------------------------------------------------------------------------------------------ --- ----------------------------------------------------------------------------------------------------
: Table of Notation
\[notations\]
The source and reconstruction alphabets, respectively, are sequences of Polish spaces $\{ {\cal X}_t: t\in\mathbb{N}\}$ and $\{ {\cal Y}_t: t\in\mathbb{N}\}$, associated with their corresponding measurable spaces $({\cal X}_t,{\cal B}({\cal X}_t))$ and $({\cal Y}_t, {\cal B}({\cal Y}_t))$, $t\in\mathbb{N}$. Sequences of alphabets are identified with the product spaces $({\cal X}_{0,n},{\cal B}({\cal X}_{0,n})) \triangleq {\times}_{k=0}^{n}({\cal X}_k,{\cal B}({\cal X}_k))$, and $({\cal Y}_{0,n},{\cal B}({\cal Y}_{0,n}))\triangleq \times_{k=0}^{n}({\cal Y}_k,{\cal B}({\cal Y}_k))$. The source and reconstruction are random processes denoted by $X^n \triangleq \{X_t: t\in\mathbb{N}^n\}$, $X:~\{t\}\times\Omega\mapsto {\cal X}_t$, and by $Y^n\triangleq \{Y_t: t\in\mathbb{N}^n\}$, $Y:~\{t\}\times\Omega\mapsto {\cal Y}_t$, respectively.\
The reconstruction conditional distribution will be defined via stochastic kernels. Note that the random variable (RV) ${Z}$ is called conditional independent of RV $X$ given the RV $Y$ if and only if $X\leftrightarrow Y \leftrightarrow Z$ forms a MC in both directions, equivalently $P_{X,Z|Y}(dx,dz|y)=P_{X|Y}(dx|y)P_{Z|Y}(dz|y)-a.s.$, or $P_{Z|Y,X}(dz|y,x)=P_{Z|Y}(dz|y)-a.s.$
\[stochastic kernel\][@dupuis-ellis97] Let $({\cal X}, {\cal B}({\cal X})), ({\cal Y}, {\cal B}({\cal Y}))$ be measurable spaces in which $\cal Y$ is a Polish Space.\
A stochastic kernel on $\cal Y$ given $\cal X$ is a mapping $q: {\cal B}({\cal Y}) \times {\cal X} \rightarrow [0,1]$ satisfying the following two properties:
1. For every $x \in {\cal X}$, the set function $q(\cdot;x)$ is a probability measure (possibly finitely additive) on ${\cal B}({\cal Y})$;
2. For every $F \in {\cal B}({\cal Y})$, the function $q(F;\cdot)$ is ${\cal B}({\cal X})$-measurable.
Stochastic kernels can be used to define anticipative and nonanticipative convolution of reconstruction kernels and associated classical and nonanticipative RDF.
\[comprchan\] Given measurable spaces $({\cal X}_{0,n},{\cal B}({\cal X}_{0,n}))$, $({\cal Y}_{0,n},{\cal B}({\cal Y}_{0,n}))$, and their product spaces, data compression channels are classified as follows.
1. [*An Anticipative Data Compression Channel*]{} is a stochastic kernel $ q_{0,n} (dy^n; x^n) \in {\cal Q}({\cal Y}_{0,n} ;{\cal X}_{0,n})$. Such a kernel admits a factorization into a convolution of a sequence of anticipative stochastic kernels as follows $$q_{0,n}(dy^n; x^n)=\otimes_{i=0}^n q_i(dy_i;y^{i-1},x^n)-a.s.\label{eq.1}$$ where $q_i(dy_i;y^{i-1},x^n) \in {\cal Q}({\cal Y}_i;{\cal Y}_{0,i-1}\times{\cal X}_{0,n}), i=0,\ldots,n,~n \in \mathbb{N}$.
2. [*A Nonanticipative Convolution Data Compression Channel*]{} is a convolution of a sequence of nonanticipative stochastic kernels defined by $${\overrightarrow q}_{0,n}(dy^n;x^n) \triangleq \otimes_{i=0}^n q_i(dy_i;y^{i-1},x^i)-a.s.\label{eq.2}$$ where $q_i(dy_i;y^{i-1},x^i) \in {\cal Q}({\cal Y}_i;{\cal Y}_{0,i-1}\times{\cal X}_{0,i}), i=0,\ldots,n,~n \in \mathbb{N}$.
3. [*A Restricted Nonanticipative Data Compression Channel*]{} is a stochastic kernel $q_{0,n} (dy^n; x^n)$ $\in {\cal Q}({\cal Y}_{0,n} ;{\cal X}_{0,n})$ which is a convolution of a sequence of nonanticipative stochastic kernels obtained by imposing the almost sure (a.s.) constraint defined by $$q_{0,n}(dy^n; x^n)=\otimes_{i=0}^n q_i(dy_i;y^{i-1},x^i)-a.s.\label{eq.3}$$ where $q_i \in {\cal Q}({\cal Y}_i;{\cal Y}_{0,i-1}\times{\cal X}_{0,i}), i=0,\ldots,n,~n \in \mathbb{N}$.
As stated earlier, the classical RDF is concerned with optimizing (\[1\]) with respect to anticipative stochastic kernels (\[eq.1\]). In this paper we address problem (\[6\]) or (\[7\]), i.e., when the conditional distribution (stochastic kernel) is restricted nonanticipative, and we discuss generalizations based on (\[eq.2\]). That is, for a given distribution $P_{X^n}(dx^n)$, nonanticipative RDF imposes the a.s.-constraint (\[eq.3\]) on the reconstruction conditional distribution, and hence on the joint distribution $P_{X^n,Y^n}(dx^n,dy^n)$ generated by them, unlike the classical RDF which does not impose such a constraint. However, when the source is independently distributed, i.e., $P_{X^n}(dx^n)=\otimes_{i=0}^n{P}_{X_i}(dx_i)-a.s.$, it is well known that the optimal reconstruction conditional distribution of the classical RDF has the property $P^*_{Y^n|X^n}(dy^n|x^n)=\otimes_{i=0}^n{P}^*_{Y^i|X^i}(dy_i|x_i)-a.s.$ It is also well known that for sources with memory (i.e., Markov sources) the optimal reconstruction conditional distribution of the classical RDF is anticipative, i.e., $P^*_{Y^n|X^n}(dy^n|x^n)=\otimes_{i=0}^n{P}^*_{Y_i|Y^{i-1},X^n}(dy_i|y^{i-1},x^n)-a.s.$ Therefore, to ensure a nonanticipative reconstruction conditional distribution one has to impose the constraint (\[eq.3\]) to the classical RDF. On the other hand, a nonanticipative convolution data compression channel (\[eq.2\]) does not impose any constraint on the joint distribution $P_{X^n,Y^n}(dx^n,dy^n)$. This point is further explained below by discussing generalizations of distortion rate function (\[6\]) and RDF (\[7\]).
\[generalizations\] The nonanticipative distortion rate function and the nonanticipative RDF can be generalized as follows. Given a sequence of conditional distributions $\{{P}_{X_i|X^{i-1},Y^{i-1}}(dx_i|x^{i-1},y^{i-1})$ $:~i=0,1,\ldots,n\}$ then (\[6\]) and (\[7\]) can be generalized to $$\begin{aligned}
{\overrightarrow D}_{0,n}^{na}(R)&\triangleq \inf_{{\overrightarrow P}_{Y^n|X^n}(dy^n|x^n):I(X^n\rightarrow{Y^n})\leq R}E\Big\{d_{0,n}(X^n,Y^n)\Big\}\label{eq21}\\
{\overrightarrow R}_{0,n}^{na}(D)&\triangleq \inf_{{\overrightarrow P}_{Y^n|X^n}(dy^n|x^n):E\{d_{0,n}(X^n,Y^n)\leq{D}\}}I(X^n\rightarrow{Y^n})\label{eq22}\end{aligned}$$ where $I(X^n\rightarrow{Y^n})$ is the directed information measure from $X^n$ to $Y^n$ defined by $$\begin{aligned}
{I}(X^n\rightarrow{Y}^n)&\triangleq& \int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}}\log\Big(\frac{\overrightarrow{P}_{Y^n|X^n}(dy^n|x^n)}{{P}_{Y^n}(dy^n)}\Big)
\overrightarrow{P}_{Y^n|X^n}(dy^n|x^n)\otimes\overleftarrow{P}_{X^n|Y^{n-1}}(dx^n|y^{n-1})\nonumber\\
&\equiv&\mathbb{I}_{X^n\rightarrow{Y^n}}(\overleftarrow{P}_{X^n|Y^{n-1}},\overrightarrow{P}_{Y^n|X^n})\nonumber\end{aligned}$$ where $\overleftarrow{P}_{X^n|Y^{n-1}}$ is defined by $$\begin{aligned}
\overleftarrow{P}_{X^n|Y^{n-1}}(dx^n|y^{n-1})\triangleq\otimes_{i=0}^n{P}_{X_i|X^{i-1},Y^{i-1}}(dx_i|x^{i-1},y^{i-1})-a.s.\nonumber\end{aligned}$$ Clearly, (\[eq21\]) and (\[eq22\]) do not assume $P_{X_i|X^{i-1},Y^{i-1}}(dx_i|x^{i-1},y^{i-1})=P_{X_i|X^{i-1}}(dx_i|x^{i-1})-a.s.$, and hence the process $X^n$ can be affected by $Y^n$ causally. It is easy to show that if (\[eq.3\]) holds then $P_{X_i|X^{i-1},Y^{i-1}}(dx_i|x^{i-1},y^{i-1})=P_{X_i|X^{i-1}}(dx_i|x^{i-1})-a.s.,~i=0,1,\ldots,n$, also holds, and hence (\[eq21\]), (\[eq22\]) reduce to (\[6\]), (\[7\]). The generalizations (\[eq21\]), (\[eq22\]), covers conditionally Gaussian sources as a special case [@liptser-shiryaev1978]. It also covers the case when the source is a controlled process, controlled over a finite rate channel based on the quantized or reconstruction process. These generalizations will be investigated elsewhere, since they will require new topological spaces on which existence of optimal solution to (\[eq21\]) and (\[eq22\]) can be shown.
Nonanticipative RDF
-------------------
In this subsection the nonanticipative RDF is rigorously defined on abstract spaces. Given a source probability measure ${\cal \mu}_{0,n} \in {\cal M}_1({\cal X}_{0, n})$ (possibly finitely additive) and a reconstruction kernel $q_{0,n} \in {\cal Q}({\cal Y}_{0, n};{\cal X}_{0, n})$, one can define three probability measures as follows.
(P1): The joint measure $P_{0,n} \in {\cal M}_1({\cal Y}_{0,n}\times {\cal X}_{0, n})$: $$\begin{aligned}
P_{0,n}(G_{0,n})&\triangleq&(\mu_{0,n} \otimes q_{0,n})(G_{0,n}),\:G_{0,n} \in {\cal B}({\cal X}_{0,n})\times{\cal B}({\cal Y}_{0,n})\nonumber\\
&=&\int_{{\cal X}_{0,n}} q_{0,n}(G_{0,n,x^n};x^n) \mu_{0,n}(d{x^n})\nonumber\end{aligned}$$ where $G_{0,n,x^n}$ is the $x^n-$section of $G_{0,n}$ at point ${x^n}$ defined by $G_{0,n,x^n}\triangleq \{y^n \in {\cal Y}_{0,n}: (x^n, y^n) \in G_{0,n}\}$ and $\otimes$ denotes the convolution.
(P2): The marginal measure $\nu_{0,n} \in {\cal M}_1({\cal Y}_{0,n})$:$$\begin{aligned}
\nu_{0,n}(F_{0,n})&\triangleq& P_{0,n}({\cal X}_{0, n} \times F_{0,n}),~F_{0,n} \in {\cal B}({\cal Y}_{0,n})\nonumber\\
&=&\int_{{\cal X}_{0, n}} q_{0,n}(({\cal X}_{0, n}\times F_{0,n})_{{x}^{n}};{x}^{n}) \mu_{0,n}(d{x^n})=\int_{{\cal X}_{0, n}} q_{0,n}(F_{0,n};x^n) \mu_{0,n}(dx^n).\nonumber\end{aligned}$$
(P3): The product measure $\pi_{0,n}:{\cal B}({\cal X}_{0,n}) \times
{\cal B}({\cal Y}_{0,n}) \mapsto [0,1] $ of $\mu_{0,n}\in{\cal M}_1({\cal X}_{0, n})$ and $\nu_{0,n}\in{\cal M}_1({\cal Y}_{0, n})$ for $G_{0,n} \in {\cal B}({\cal X}_{0,n}) \times {\cal B}({\cal Y}_{0,n})$: $$\begin{aligned}
\pi_{0,n}(G_{0,n})\triangleq(\mu_{0,n} \times \nu_{0,n})(G_{0,n})=\int_{{\cal X}_{0, n}} \nu_{0,n}(G_{0,n,x^n}) \mu_{0,n}(dx^n).\nonumber\end{aligned}$$ The precise definition of mutual information between two sequences of Random Variables $X^n$ and $Y^n$, denoted $I(X^n; Y^n)$ is defined via the Kullback-Leibler distance (or relative entropy) between the joint probability distribution of $(X^n, Y^n)$ and the product of its marginal probability distributions of $X^n$ and $Y^n$, using the Radon-Nikodym derivative as follows.
\[relative\_entropy\][@ihara1993] Given a measurable space $({\cal X}, {\cal B}({\cal X}))$, the relative entropy between two probability measures $P, Q\in {\cal M}_1({\cal X})$ is defined by $$\begin{aligned}
\mathbb{D}(P || Q) \triangleq \left\{ \begin{array}{ll} \int_{{\cal X}} \log \Big( \frac{dP}{dQ}\Big) dP = \int_{{\cal X}} \log \Big(\frac{dP}{dQ}\Big) \frac{dP}{dQ} dQ & \mbox{if} \quad P << Q \\
+ \infty & \mbox{otherwise} \end{array} \right.
\nonumber\end{aligned}$$ where $ \frac{dP}{dQ}$ denotes the Radon-Nikodym derivative (density) of $P$ with respect to $Q$, and $P << Q$ denotes absolute continuity of $Q$ with respect to $P$.
Hence, by the construction of probability measures (P1)-(P3), and the chain rule of relative entropy [@dupuis-ellis97], the following equivalent definitions of mutual information are obtained. $$\begin{aligned}
I(X^n;Y^n) &\triangleq& \mathbb{D}(P_{0,n}|| \pi_{0,n})\label{re4}\\
&=&\int_{{\cal X}_{0,n} \times {\cal Y}_{0,n}}\log \Big( \frac{d (\mu_{0,n} \otimes q_{0,n}) }{d ( \mu_{0,n} \times \nu_{0,n} ) }\Big) d(\mu_{0,n} \otimes q_{0,n}) \label{eq.4}\\
& =& \int_{{\cal X}_{0,n} \times {\cal Y}_{0,n}} \log \Big( \frac{q_{0,n}(d y^n; x^n)}{ \nu_{0,n} (dy^n) } \Big)q_{0,n}(dy^n;dx^n)\otimes\mu_{0,n}(dx^n) \label{eq.5}\\
&=&\int_{{\cal X}_{0,n}} \mathbb{D}(q_{0,n}(\cdot;x^n)|| \nu_{0,n}(\cdot)) \mu_{0,n}(dx^n)\nonumber\\
&\equiv& \mathbb{I}(\mu_{0,n}, q_{0,n}). \label{re3}\end{aligned}$$ Note that (\[re3\]) states that mutual information is expressed as a functional of $\{\mu_{0,n}, q_{0,n}\}$ and it is denoted by $\mathbb{I}(\mu_{0,n},q_{0,n})$. Note also that $\mu_{0,n}\otimes{q}_{0,n}\ll{\mu}_{0,n}\times\nu_{0,n}$ if and only if $q(\cdot;x^n)\ll{\nu}_{0,n}(\cdot)$, $\mu_{0,n}-$a.s., which is used to established that (\[eq.4\]) is equivalent to (\[eq.5\]). Necessary and sufficient conditions for existence of a Radon-Nikodym derivative for finitely additive measures can be found in [@maynard79].\
Next, the classical RDF [@berger] is introduced, since the definition of nonanticipative RDF will be based on the classical definition by imposing the nonanticipative constraint (\[eq18\]) (or Definition \[comprchan\]-3).
\[classical\_rdf\] [@berger]$($[**Classical Rate Distortion Function**]{}$)$ Let $d_{0,n}: {\cal X}_{0,n} \times {\cal Y}_{0,n} \rightarrow [0, \infty]$, be an ${\cal B}({\cal X}_{0,n}) \times {\cal B }( {\cal Y}_{0,n})$-measurable distortion function, and let $Q_{0,n}(D) \subset {\cal Q}({\cal Y}_{0,n}; {\cal X}_{0,n})$ (assuming is nonempty) denotes the average distortion or fidelity constraint defined by $$\begin{aligned}
Q_{0,n}(D){{\stackrel}{\triangle}{=}}\Big\{ q_{0,n} \in {\cal Q}({\cal Y}_{0,n}; {\cal X}_{0,n}):\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}} d_{0,n}(x^n,y^n) q_{0,n}(dy^n;x^n)\otimes\mu_{0,n}(dx^n) \leq D \Big\}\label{dc1}\end{aligned}$$ for $D\geq 0$. The classical RDF associated with the anticipative kernel $q_{0,n} \in {\cal Q}({\cal Y}_{0,n}; {\cal X}_{0,n})$ is defined by $$R_{0,n}(D) \triangleq \inf_{q_{0,n} \in Q_{0,n}(D)}\mathbb{I}(\mu_{0,n},q_{0,n}). \label{f3s}$$
Existence in (\[f3s\]) is shown by assuming $d_{0,n}(x^n,\cdot)$ is bounded continuous on ${\cal Y}_{0,n}$ while ${\cal Y}_{0,n}$ is compact, using weak-convergence of probability measures in [@csiszar74], and for more general conditions $d_{0,n}(x^n,\cdot)$ which is only continuous on ${\cal Y}_{0,n}$ using weak$^*$-convergence of measures on Polish spaces [@farzad06].\
Unfortunately, for general sources and distortion function $d_{0,n}(x^n,y^n)$, the optimal reconstruction $q^*_{0,n}(dy^n;x^n)=\otimes^n_{i=0}q^*_i(dy_i;y^{i-1},x^n)$ is anticipative, and hence the link to filtering theory cannot be established due to dependence of $y_i$ on $(y^{i-1},x^i)$ and on future symbols $(x_{i+1},\ldots,x_n)$. This raises the question whether the classical RDF can be reformulated so that the optimal reconstruction kernel is nonanticipative. Before the definition of nonanticipative RDF we introduced a Lemma which gives insight into how classical and nonanticipative RDF are related.
The next lemma relates nonanticipative convolution reconstruction kernels and conditional independence.
\[lem1\] The following are equivalent for each $n\in\mathbb{N}$.
1. $q_{0,n} (dy^n; x^n)={\overrightarrow q}_{0,n}(dy^n;x^n)$-a.s., (see Definition \[comprchan\]-3).
2. For each $i=0,1,\ldots, n-1$, $Y_i \leftrightarrow (X^i, Y^{i-1}) \leftrightarrow (X_{i+1}, X_{i+2}, \ldots, X_n)$, forms a MC.
3. For each $i=0,1,\ldots, n-1$, $Y^i \leftrightarrow X^i \leftrightarrow X_{i+1}$ forms a MC.
Moreover, $X_{i+1}^n\leftrightarrow{X^i}\leftrightarrow{Y^i}$, forms a MC for each $i=0,1,\ldots,n-1$, implies any of the statements 1), 2), 3).
This is straight forward hence the derivation is omitted.
According to Lemma \[lem1\]$-1)$, for a restricted nonanticipative stochastic kernel the mutual information becomes $$\begin{aligned}
I(X^n;Y^n)&=&\int_{{\cal X}_{0,n} \times {\cal Y}_{0,n}} \log \Big( \frac{ \overrightarrow{q}_{0,n}(d y^n; x^n)}{\nu_{0,n}(dy^n)} \Big){\overrightarrow q}_{0,n}(dy^n;dx^n)\otimes\mu_{0,n}(dx^n) \nonumber \\
&\equiv& {\mathbb I}(\mu_{0,n},\overrightarrow{q}_{0,n} ) \label{ex11}\end{aligned}$$ where (\[ex11\]) states that $I(X^n;Y^n)$ is a functional of $\{\mu_{0,n},{\overrightarrow q}_{0,n}\}$. Hence, nonanticipative RDF is defined by optimizing ${\mathbb I}(\mu_{0,n},{q}_{0,n})$ over ${q}_{0,n}{\in}Q_{0,n}(D)$ subject to the realizability constraint $q_{0,n}(dy^n;x^n)={\overrightarrow q}_{0,n}(dy^n;x^n)-a.s.,$ which satisfies a distortion constraint.
\[def1\] $($[**Nonanticipative Rate Distortion Function**]{}$)$ Suppose $d_{0,n}(x^n,y^n)\triangleq\sum^n_{i=0}\rho_{0,i}$ $(x^i,y^i)$, where $\rho_{0,i}: {\cal X}_{0,i} \times {\cal Y}_{0,i}\rightarrow [0, \infty]$, is a sequence of ${\cal B}({\cal X}_{0,i}) \times {\cal B }( {\cal Y}_{0,i})$-measurable distortion functions, for $i=0,1,\ldots,n$, and let $\overrightarrow{Q}_{0,n}(D)$ (assuming is nonempty) denotes the average distortion or fidelity constraint defined by $$\begin{aligned}
\overrightarrow{Q}_{0,n}(D){{\stackrel}{\triangle}{=}}Q_{0,n}(D)\bigcap\Big\{q_{0,n}{\in}{\cal Q}({\cal Y}_{0,n};{\cal X}_{0,n}):&q_{0,n}(dy^n;x^n)={\overrightarrow q}_{0,n}(dy^n;x^n)-a.s.\Big\}\label{eq2}\end{aligned}$$ The nonanticipative RDF associated with the restricted nonanticipative stochastic kernel is defined by $$\begin{aligned}
{R}^{na}_{0,n}(D) {{\stackrel}{\triangle}{=}}\inf_{{{q}_{0,n}\in \overrightarrow{Q}_{0,n}(D)}}{\mathbb I}(\mu_{0,n},{q}_{0,n}).\label{ex12}\end{aligned}$$
Thus, ${R}^{na}_{0,n}(D)$ is characterized by minimizing mutual information or equivalently $\mathbb{I}(\mu_{0,n},{q}_{0,n})$ over the ${Q}_{0,n}(D)$ and the nonanticipative constraint (\[eq18\]). In the work of [@gorbunov-pinsker], nonanticipative RDF is called $\epsilon$-entropy and nonanticipation is defined via $X_{i+1}^n\leftrightarrow{X^i}\leftrightarrow{Y^i}$, forms a MC for each $i=0,1,\ldots,n-1$, which implies $q_{0,n} (dy^n; x^n)={\overrightarrow q}_{0,n}(dy^n;x^n)$. Clearly, Gorbunov and Pinsker [@gorbunov-pinsker] nonanticipative RDF which imposes the constraint $X_{i+1}^n\leftrightarrow{X}^i\leftrightarrow{Y^i}$ forms a MC for each $i=0,1,\ldots,n-1$, implies $P_{X_{i+1}|X^{i},Y^i}(dx_{i+1}|x^i,y^i)=P_{X_{i+1}|X^i}(dx_{i+1}|x^i)-a.s.,~i=0,1,\ldots,n-1$, and hence, it does not allow the generalizations discussed in Remark \[generalizations\].
Existence of Optimal Reconstruction Kernel {#existence}
==========================================
In this section, appropriate topologies and function spaces are introduced and existence of the minimizing nonanticipative product kernel in (\[ex12\]) is proved. The construction of spaces is based on [@farzad06].
Abstract Spaces
---------------
Let $BC({\cal Y}_{0,n})$ denote the vector space of bounded continuous real valued functions defined on the Polish space ${\cal Y}_{0,n}$. Furnished with the sup norm topology, this is a Banach space. Denote by $L_1(\mu_{0,n}, BC({\cal Y}_{0,n}))$ the space of all $\mu_{0,n}$-integrable functions defined on ${\cal X}_{0,n}$ with values in $BC({\cal Y}_{0,n}),$ so that for each $\phi \in L_1(\mu_{0,n}, BC({\cal Y}_{0,n}))$ its norm is defined by $$\begin{aligned}
\parallel \phi \parallel_{\mu_{0,n}} \triangleq \int_{{\cal X}_{0,n}} ||\phi(x^n,\cdot)||_{BC({\cal Y}_{0,n})} \mu_{0,n}(dx^n) <\infty\nonumber.\end{aligned}$$ The norm topology $\parallel{\phi}\parallel_{\mu_{0,n}}$, makes $L_1(\mu_{0,n}, BC({\cal Y}_{0,n}))$ a Banach space. The topological dual of $BC({\cal Y}_{0,n})$ denoted by $ \Big( BC({\cal Y}_{0,n})\Big)^*$ is isometrically isomorphic to the Banach space of finitely additive regular bounded signed measures on ${\cal Y}_{0,n}$ [@dunford1988], denoted by $M_{rba}({\cal Y}_{0,n})$. Let $\Pi_{rba}({\cal Y}_{0,n})\subset M_{rba}({\cal Y}_{0,n})$ denote the set of regular bounded finitely additive probability measures on ${\cal Y}_{0,n}$. Clearly if ${\cal Y}_{0,n}$ is compact, then $\Big(BC({\cal Y}_{0,n})\Big)^*$ will be isometrically isomorphic to the space of countably additive signed measures, as in [@csiszar74]. It follows from the theory of “lifting" [@tulcea1969] that the dual of the space $L_1(\mu_{0,n},BC({\cal Y}_{0,n}))$ is $L_{\infty}^w(\mu_{0,n}, M_{rba}({\cal Y}_{0,n}))$, denoting the space of all $M_{rba}({\cal Y}_{0,n})$ valued functions $\{q\}$ which are weak$^*$-measurable in the sense that for each $\phi \in
BC({\cal Y}_{0,n}),$ $x^n \rightarrow q_{x^n}(\phi) \triangleq \int_{{\cal Y}_{0,n}}\phi(y^n) q(dy^n;x^n)$ is $\mu_{0,n}$-measurable and $\mu_{0,n}$-essentially bounded.
Weak$^*$-Compactness and Existence
----------------------------------
Next, we prepare to prove existence of solution to $R_{0,n}^{na}(D)$. Define an admissible set of stochastic kernels associated with classical rate distortion function by $$\begin{aligned}
Q_{ad}\triangleq L_{\infty}^w(\mu_{0,n}, \Pi_{rba}({\cal Y}_{0,n})) \subset L_{\infty}^w(\mu_{0,n}, M_{rba}({\cal Y}_{0,n})).\end{aligned}$$ Clearly, $Q_{ad}$ is a unit sphere in $L_{\infty}^w(\mu_{0,n}, M_{rba}({\cal Y}_{0,n}))$. For each $\phi{\in}L_1(\mu_{0,n}, BC({\cal Y}_{0,n}))$ we can define a linear functional on $L_{\infty}^w(\mu_{0,n}, M_{rba}({\cal Y}_{0,n}))$ by $$\begin{aligned}
\ell_{\phi}(q_{0,n})\triangleq\int_{{\cal X}_{0,n}}\Big( \int_{{\cal Y}_{0,n}} \phi(x^n,y^n)q_{0,n}(dy^n;x^n) \Big)\mu_{0,n}(dx^n).\end{aligned}$$ This is a bounded, linear and weak$^*$-continuous functional on $L_{\infty}^w(\mu_{0,n}, M_{rba}({\cal Y}_{0,n}))$ as it is shown below. $$\begin{aligned}
|\ell_{\phi}(q_{0,n})|&=&\bigg{|}\int_{{\cal X}_{0,n}}\Big( \int_{{\cal Y}_{0,n}} \phi(x^n,y^n)q_{0,n}(dy^n;x^n) \Big)\mu_{0,n}(dx^n)\bigg{|}\nonumber\\
&\leq&\int_{{\cal X}_{0,n}}\bigg{|}\Big( \int_{{\cal Y}_{0,n}} \phi(x^n,y^n)q_{0,n}(dy^n;x^n) \Big)\bigg{|}\mu_{0,n}(dx^n)\nonumber\\
&\leq&\int_{{\cal X}_{0,n}}||\phi(x^n,\cdot)||_{BC({\cal Y}_{0,n})}||q_{0,n}(\cdot;x^n)||_{TV}\mu_{0,n}(dx^n)\nonumber\\
&\leq&\int_{{\cal X}_{0,n}}||\phi(x^n,\cdot)||_{BC({\cal Y}_{0,n})}\mu_{0,n}(dx^n)\nonumber\\
&=&||\phi||_{L_1(\mu_{0,n}, BC({\cal Y}_{0,n}))}<\infty.\end{aligned}$$ So given $\phi\in L_1(\mu_{0,n},BC({\cal Y}_{0,n}))$, there exists a $c_{\phi}<\infty$ such that $||\ell_{\phi}||<c_{\phi}$. Therefore, $\ell_{\phi}$ is a bounded, linear functional on $L_{\infty}^w(\mu_{0,n}, \Pi_{rba}({\cal Y}_{0,n}))$ and hence on $L_{\infty}^w(\mu_{0,n}, M_{rba}({\cal Y}_{0,n}))$. Thus, it is continuous in the weak$^*$-sense.
For $d_{0,n}: {\cal X}_{0,n} \times {\cal Y}_{0,n}\rightarrow [0, \infty)$ measurable and $d_{0,n}{\in}L_1(\mu_{0,n},BC({\cal Y}_{0,n}))$ the distortion constraint set of the classical RDF is given by $$\begin{aligned}
Q_{0,n}(D)\triangleq\{q{\in}Q_{ad}:\ell_{d_{0,n}}(q_{0,n}){\leq}D\}.\end{aligned}$$ The next result is shown in [@farzad06]; it utilizes the Alaoglu’s theorem [@dunford1988], which states that a closed and bounded subset of a weak$^*$-compact set is weak$^*$-compact. These will be used to establish existence of minimizer in $\overrightarrow{Q}_{ad}$ for the nonanticipative RDF $R_{0,n}^{na}(D)$.
\[$Q_ad$ w$^*$-closed\][@farzad06] For ${d_{0,n}}{\in}L_1(\mu_{0,n},BC({\cal Y}_{0,n}))$, the set $Q_{0,n}(D)$ is bounded and weak$^*$-closed subset of $Q_{ad}$ (hence weak$^*$-compact).
Now we prepare to consider the problem stated in Definition \[def1\]. First, we show weak$^*$-compactness of $\overrightarrow{Q}_{ad}$ defined as a subset of $Q_{ad}$ as follows. $$\begin{aligned}
{\overrightarrow Q}_{ad}=\Big\{q_{0,n}\in {Q_{ad}}:q_{0,n}(dy^n;x^n)={\overrightarrow q}_{0,n}(dy^n;x^n)-a.s.\Big\}.\end{aligned}$$ The average distortion function for the nonanticipative RDF is defined by $$\begin{aligned}
\overrightarrow{Q}_{0,n}(D)&\triangleq& \Big\{{q}_{0,n} \in {Q}_{ad} :\ell_{d_{0,n}}({q}_{0,n})\triangleq \int_{{\cal X}_{0,n}} \biggr(\int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n){q}_{0,n}(dy^n;x^n) \biggr)\\
&&\qquad\qquad\qquad\otimes\mu_{0,n}(dx^n)\leq D\Big\}\bigcap{\overrightarrow Q}_{ad}\\
&=&\Big\{{q}_{0,n} \in {\overrightarrow Q}_{ad} :\ell_{d_{0,n}}({q}_{0,n})\triangleq \int_{{\cal X}_{0,n}} \biggr(\int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n){ q}_{0,n}(dy^n;x^n) \biggr)\\
&&\qquad\qquad\qquad\otimes\mu_{0,n}(dx^n)\leq{D}\Big\},~D\geq0.\end{aligned}$$ Since we are interested in proving existence of nonanticipative RDF of Definition \[def1\], we shall first show that $\overrightarrow{Q}_{ad}$ is weak$^*$-closed, and then utilize Lemma \[$Q_ad$ w$^*$-closed\] to establish weak$^*$-compactness for $\overrightarrow{Q}_{ad}$ and then weak$^*$-compactness of $\overrightarrow{Q}_{0,n}(D)$.
\[weakstar-closed\] Let ${\cal X}_{0,n}$ and ${\cal Y}_{0,n}$ be Polish spaces and introduce the net $\{q^{\alpha}_i(dy_i;y^{i-1},x^i)\}$, where $\alpha\in({\cal D},\succeq)$, and $q^{\alpha}_i\in{\cal Q}({\cal Y}_i;{\cal Y}_{0,i-1},{\cal X}_{0,i})$. Assume
(a)
: $q^{\alpha}_i(\cdot;y^{i-1},x^i)\buildrel w^* \over {\longrightarrow}q_i^{0}(\cdot;y^{i-1},x^i)$ for $i=1,\ldots,n$;
(b)
: for all $h_i(\cdot,\cdot){\in}L_1(\mu_{i},BC({\cal Y}_{i}))$ the function $$\begin{aligned}
(x^{i},y^{i-1})\in{\cal X}_{0,i}\times{\cal Y}_{0,i-1}\longmapsto\int_{{\cal X}_i}\int_{{\cal Y}_i}h_i(y)q_i(dy;y^{i-1},x^i)\mu_i(dx_i;x^{i-1})\nonumber\end{aligned}$$ is $L_1(\mu_{0,i-1},BC({\cal Y}_{0,i-1}))$ for $i=0,1,\ldots,n$;
(c)
: for all $h_i(\cdot,\cdot){\in}L_1(\mu_{i},BC({\cal Y}_{i}))$ and $\forall~\epsilon>0$ there exists $\alpha\succ\alpha_{\epsilon}$ such that $$\begin{aligned}
&&\int_{{\cal X}_i}\sup_{y^{i-1}\in{\cal Y}_{0,i-1}}\bigg{|}\int_{{\cal Y}_i}h_i(x_i,y_i)q_i^{\alpha}(dy_i;y^{i-1},x^i)\nonumber\\
&&-\int_{{\cal Y}_i}h_i(x_i,y_i)q_i^0(dy_i;y^{i-1},x^i)\bigg{|}\mu_i(dx_i;x^{i-1})<\epsilon,\quad\forall~x^{i-1}\in{\cal X}_{0,i-1}\nonumber.\end{aligned}$$
Then the convolution of stochastic kernels converges in weak$^*$-sense as follows. $${\overrightarrow{q}}_{0,n}^{\alpha}\buildrel w^* \over \longrightarrow{\overrightarrow{q}}_{0,n}^{0}\label{eq.8}$$ e.g, the set $\overrightarrow{Q}_{ad}$ is weak$^*$-closed.
See Appendix.
Next, we utilize the weak$^*$-compactness of $\overrightarrow{Q}_{ad}$ to show that $\overrightarrow{Q}_{0,n}(D)$ is also weak$^*$-compact.
\[remove\_boundness\] There are certain important cases in which $d_{0,n}$ may not be bounded. This is the case when $d_{0,n}$ is a metric of a linear metric space. The next theorem is crucial in showing the weak$^*$-closedness property of $\overrightarrow{Q}_{0,n}(D)$ to those distortion functions $d_{0,n}$ which are not necessarily bounded, since they are measurable functions from the class $d_{0,n}\in{L}_1(\mu_{0,n},BC({\cal Y}_{0,n}))$.
\[weakstar-compact\_2\] Let ${\cal X}_{0,n},{\cal Y}_{0,n}$ be two Polish spaces and $d_{0,n} :{\cal X}_{0,n}\times{\cal Y}_{0,n}\mapsto[0,\infty]$, a measurable, nonnegative, extended real valued function, such that for a fixed $x^n \in {\cal X}_{0,n}$, $y^n \rightarrow d(x^n,\cdot)$ is continuous on ${\cal Y}_{0,n}$, for $\mu_{0,n}$-almost all $x^n \in {\cal X}_{0,n}$ and suppose the conditions of Lemma \[weakstar-closed\] hold. For any $D \in [0,\infty)$, the set ${\overrightarrow Q}_{0,n}(D)$ is a weak$^*$-closed subset of ${\overrightarrow Q}_{ad}$ and hence weak$^*$-compact.
Let $\{{\overrightarrow q}_{0,n}^{\alpha}\}\in {\overrightarrow Q}_{0,n}(D) \subset {\overrightarrow Q}_{ad}$ be a net. Since ${\overrightarrow Q}_{ad}$ is weak$^*$-compact, there exists a subnet of the net $\{{\overrightarrow q}_{0,n}^{\alpha}\},$ relabelled as the original net, and an element ${\overrightarrow q}_{0,n}^0 \in {\overrightarrow Q}_{ad}$ such that ${\overrightarrow q}_{0,n}^{\alpha} \buildrel w^* \over \longrightarrow {\overrightarrow q}_{0,n}^0$[^6]. We must show that ${\overrightarrow q}_{0,n}^0 \in {\overrightarrow Q}_{0,n}(D).$ Considering the sequence $\{d_{0,n}^k \triangleq d_{0,n}\wedge k, k \in N\},$ which are bounded, measurable functions (continuous in the second argument), it follows from the weak$^*$-convergence of the sequence $\{{\overrightarrow q}_{0,n}^{\alpha}\}$ to ${\overrightarrow q}_{0,n}^0$ that $$\begin{aligned}
&&\left.\int_{{\cal X}_{0,n}}\bigg( \int_{{\cal Y}_{0,n}}d_{0,n}^k(x^n,y^n){\overrightarrow q}_{0,n}^{0}(dy^n;x^n)\bigg)\mu_{0,n}(dx^n) \right.\nonumber \\[-1.5ex]
\label{eq.i.5}\\[-1.5ex]
&&\quad\left.=\lim_{\alpha}\int_{{\cal X}_{0,n}} \bigg( \int_{{\cal Y}_{0,n}}d_{0,n}^k(x^n,y^n){\overrightarrow q}_{0,n}^{\alpha}(dy^n;x^n)\bigg)\mu_{0,n}(dx^n) \right.\nonumber\end{aligned}$$ for each $k\in N$. Since $d_{0,n}$ is non-negative and $d_{0,n}^k\uparrow d_{0,n}$ as $k \longrightarrow \infty$ and ${\overrightarrow q}_{0,n}^{\alpha}\in {\overrightarrow Q}_{0,n}(D)$, we have $$\begin{aligned}
\int_{{\cal X}_{0,n}}\bigg( \int_{{\cal Y}_{0,n}}
&d_{0,n}^k&(x^n,y^n){\overrightarrow q}_{0,n}^{0}(dy^n;x^n)\bigg)\mu_{0,n}(dx^n)\nonumber\\
&=&\lim_{\alpha}\int_{{\cal X}_{0,n}} \biggl( \int_{{\cal Y}_{0,n}}d_{0,n}^k(x^n,y^n){\overrightarrow q}_{0,n}^{\alpha}(dy^n;x^n)\biggr)\mu_{0,n}(dx^n)\nonumber\\
&\leq&\lim_{\alpha}\int_{{\cal X}_{0,n}} \bigg( \int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n){\overrightarrow q}_{0,n}^{\alpha}(dy^n;x^n)\bigg)\mu_{0,n}(dx^n)\leq D \label{eq.i.6}\end{aligned}$$ which is valid for all $k \in N$. Since $d_{0,n}^k
\uparrow d_{0,n}$ and they are non-negative, it follows from Lebesgue’s monotone convergence theorem and non-negativity of stochastic kernels that $$\begin{aligned}
\int_{{\cal X}_{0,n}}\bigg( \int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n){\overrightarrow q}_{0,n}^{0}(dy^n;x^n)\bigg)\mu_{0,n}(dx^n)\leq D.\end{aligned}$$
This shows that the weak$^*$-limit ${\overrightarrow q}_{0,n}^0 \in{\overrightarrow Q}_{0,n}(D)$ and hence we have proved that the set ${\overrightarrow Q}_{0,n}(D)$ is a weak$^*$-closed subset of ${\overrightarrow Q}_{ad}$. By Alaoglu’s theorem [@dunford1988] being a weak$^*$-closed subset of a weak$^*$-compact set, it is weak$^*$-compact.
Based on Theorem \[weakstar-compact\_2\] and lower semicontinuity of relative entropy, we show existence of the optimal reconstruction conditional distribution for nonanticipative RDF.
\[th3\]$($[**Existence**]{}$)$ Under the conditions of Theorem \[weakstar-compact\_2\], $R^{na}_{0,n}(D)$ has a minimum.
This follows from Theorem \[weakstar-compact\_2\] provided lower semicontinuity of ${\mathbb{I}}(\mu_{0,n},\cdot)$ on $\overrightarrow{Q}_{ad}$ is established. First we prove that $\overrightarrow{q}_{0,n}\rightarrow \mathbb{I}({\mu}_{0,n},\cdot)$ is weak$^*$-lower semicontinuous. Let $\{\overrightarrow{q}_{0,n}^{\alpha}\}$ be a net from $\overrightarrow{Q}_{ad}$ and suppose it is weak$^*$-convergent to ${\overrightarrow q}_{0,n}^0$. Define the net $P_{0,n}^{\alpha} \in \Pi_{rba}({\cal X}_{0,n}\times {\cal Y}_{0,n})$ given by the convolution product $P_{0,n}^{\alpha} \equiv \mu_{0,n}(dx^n) \otimes {\overrightarrow q}_{0,n}^{\alpha}(dy^n;x^n)$. Take any $\varphi(\cdot)\in BC({\cal X}_{0,n}\times {\cal Y}_{0,n})$ and consider the expression $$\begin{aligned}
\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}} \varphi_{0,n}(x^n,y^n) P_{0,n}^{\alpha}(dx^n,dy^n)\equiv \int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}} \varphi_{0,n}(x^n,y^n) {\overrightarrow q}_{0,n}^{\alpha}(dy^n;x^n)\otimes\mu_{0,n}(dx^n)\nonumber.\end{aligned}$$ Since ${\overrightarrow q}_{0,n}^{\alpha}\buildrel w^* \over \longrightarrow {\overrightarrow q}_{0,n}^0$ in $L_{\infty}^w(\mu_{0,n},\Pi_{rba}({\cal Y}_{0,n}))$, it is clear from the above expression that $$\begin{aligned}
P_{0,n}^{\alpha} \buildrel w^* \over \longrightarrow P_{0,n}^0 \equiv \mu_{0,n} \otimes {\overrightarrow q}_{0,n}^0~~\hbox{in}~~\Pi_{rba}({\cal X}_{0,n}\times {\cal Y}_{0,n}).\end{aligned}$$ Similarly one can easily verify that the net of the product measures $\{\pi_{0,n}^{\alpha}\}$ converges to the product measure $\pi_{0,n}^0$, $$\begin{aligned}
\pi_{0,n}^{\alpha} \equiv \nu_{0,n}^{\alpha}\times \mu_{0,n} \buildrel w^* \over
\longrightarrow \nu_{0,n}^0 \times \mu_{0,n}\equiv \pi_{0,n}^0\nonumber\end{aligned}$$ where $\{\nu_{0,n}^{\alpha}\}$ are the marginals of $\{P_{0,n}^{\alpha}\}$ on ${\cal Y}_{0,n}$ and $\nu_{0,n}^0$ is its weak$^*$-limit. Now we use the lower semicontinuity property of relative entropy [@dupuis-ellis97 Lemma 1.4.3, p. 36]. Following [@dupuis-ellis97] it is verified that the same procedure holds true not only for countably additive measures but also for finitely additive ones. Using this fact we conclude that $$\begin{aligned}
\mathbb{D}(P_{0,n}||\pi_{0,n})\leq\liminf_{\alpha\longrightarrow\infty}\mathbb{D}(P_{0,n}^{\alpha}||\pi_{0,n}^{\alpha})\nonumber.\end{aligned}$$ By (\[re4\]), this is equivalent to $$\begin{aligned}
{\mathbb{I}}(\mu_{0,n},q_{0,n})\leq\liminf_{\alpha\longrightarrow\infty}{\mathbb{I}}(\mu_{0,n},q_{0,n}^{\alpha}).\end{aligned}$$ This proves weak$^*$-lower semicontinuity of ${\mathbb{I}}(\mu_{0,n},\cdot)$ on $\overrightarrow{Q}_{ad}$. We have already observed in Theorem \[weakstar-compact\_2\] that the set $\overrightarrow{Q}_{0,n}(D)$ is weak$^*$-compact, and we have just seen that ${\mathbb{I}}(\mu_{0,n},\cdot)$ is weak$^*$-lower semicontinuous. Hence ${\mathbb{I}}(\mu_{0,n},\cdot)$ attains its infimum on $\overrightarrow{Q}_{0,n}(D)$. So there exists a $\overrightarrow{q}_{0,n}^* \in \overrightarrow{Q}_{0,n}(D)$ such that $R_{0,n}^{na}(D)={\mathbb{I}}(\mu_{0,n},\overrightarrow{q}_{0,n}^*)$.
Necessary Conditions of Optimality for Nonanticipative RDF {#necessary}
==========================================================
In this section the form of the optimal nonanticipative convolution reconstruction kernels is derived under a stationarity assumption. The method is based on calculus of variations on the space of measures [@dluenberger69].
\[stationarity\] The family of measures $\overrightarrow{q}_{0,n}(dy^n;x^n)$ defined in (\[eq.2\]), is the convolution of stationary conditional distributions.
Assumption \[stationarity\] holds for stationary process $\{(X_i,Y_i):i\in\mathbb{N}\}$ and single letter distortion $d_{0,n}(x^n,y^n)\equiv\sum_{i=0}^n\rho(x_i,y_i)$. It also holds for distortion defined by $\rho(T^i{x^n},T^i{y^n})$, for which $T^i{x^n}=\tilde{x}^{n}$ is the $i^{th}$ shift operator on the input sequence $x^n$, where $\tilde{x}_{k}=x_{k+i}$ (similarly for $T^i{y^n}$), and $\sum_{i=0}^n\rho(T^ix^n,T^iy^n)$ depends only on the components of $(x^n,y^n)$ [@gray2010]. Utilizing Assumption \[stationarity\], which holds for stationary processes and a single letter distortion function, the Gateaux differential of $\mathbb{I}(\mu_{0,n},\overrightarrow{q}_{0,n})$ is taken at $\overrightarrow{q}_{0,n}^*$ in the direction of $\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*$, via the definition $\overrightarrow{q}_{0,n}^{\epsilon}\triangleq\overrightarrow{q}_{0,n}+\epsilon\big{(}
\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*\big{)}$, $\epsilon\in[0,1]$, since under the stationarity assumption, the functionals $\{q_i(dy_i;y^{i-1},x^i)\in{\cal Q}({\cal Y}_i;{\cal Y}_{0,i-1}\times{\cal X}_{0,i}):~i=0,1,\ldots,n\}$ are identical.
\[th5\] Suppose ${\mathbb I}_{\mu_{0,n}}(\overrightarrow{q}_{0,n}) \triangleq {\mathbb I}(\mu_{0,n},\overrightarrow{q}_{0,n})$ is well defined for every $\overrightarrow{q}_{0,n}\in L_{\infty}^w(\mu_{0,n},$ $\Pi_{rba}({\cal Y}_{0,n}))$ possibly taking values from the set $[0,\infty].$ Then $\overrightarrow{q}_{0,n} \rightarrow {\mathbb I}_{\mu_{0,n}}(\overrightarrow{q}_{0,n})$ is Gateaux differentiable at every point in $L_{\infty}^w(\mu_{0,n},\Pi_{rba}({\cal Y}_{0,n})),$ and the Gateaux derivative at the point $\overrightarrow{q}_{0,n}^*$ in the direction $\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*$ is given by $$\begin{aligned}
\delta{\mathbb I}_{\mu_{0,n}}(\overrightarrow{q}_{0,n}^*,\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)=\int_{{\cal X}_{0,n}}\int_{{\cal Y}_{0,n}}\log \Bigg(\frac{\overrightarrow{q}_{0,n}^*(dy^n;x^n)}{\nu_{0,n}^*(dy^n)}\Bigg)(\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)(dy^n;x^n)\otimes \mu_{0,n}(dx^n)\nonumber\end{aligned}$$ where $\nu_{0,n}^*\in{\cal M}_1({\cal Y}_{0,n})$ is the marginal measure corresponding to $\overrightarrow{q}_{0,n}^*\otimes\mu_{0,n}\in{\cal M}_1({\cal Y}_{0,n}\times{\cal X}_{0,n})$.
The proof, although lengthy, it is similar to the one in [@farzad06], hence it is omitted.
The constrained problem defined by (\[ex12\]) can be reformulated using Lagrange multipliers. The equivalence of constrained and unconstrained problems is established in the following theorem.
\[lagrange\_duality\] Suppose $d_{0,n}(x^n,y^n)\triangleq\sum_{i=0}^n\rho(T^i{x^n},T^i{y^n})$, where $d_{0,n}: {\cal X}_{0,n}\times{\cal Y}_{0,n} \rightarrow \overline R_0 \equiv [0,\infty]$ is continuous in the second argument and the set $\Gamma \equiv\{(x^n,y^n) \in {\cal X}_{0,n}\times{\cal Y}_{0,n}: d_{0,n}(x^n,y^n) < D \} $ is nonempty. Then the constrained problem as stated in Theorem \[th3\], is equivalent to an unconstrained problem stated below. $$\begin{aligned}
\inf_{\overrightarrow{q}_{0,n} \in \overrightarrow{Q}_{0,n}(D)} \mathbb{I}({\mu_{0,n}},\overrightarrow{q}_{0,n}) &= \max_{s \leq 0} \inf_{\overrightarrow{q}_{0,n}}\{ \mathbb{I}({\mu_{0,n}},\overrightarrow{q}_{0,n}) - sG(\overrightarrow{q}_{0,n})\},~G(\overrightarrow{q}_{0,n}){{\stackrel}{\triangle}{=}}\ell_{d_{0,n}}(\overrightarrow{q}_{0,n})-D \nonumber\\
&= \max_{s \leq 0}\inf_{\overrightarrow{q}_{0,n}} \Big\{ \mathbb{I}({\mu_{0,n}},\overrightarrow{q}_{0,n}) - s \Big(\int_{{\cal X}_{0,n}} \int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n)\nonumber\\
&\qquad{\overrightarrow q}_{0,n}(dy^n;x^n)\otimes\mu_{0,n}(dx^n) -D \Big)\Big\}\nonumber\end{aligned}$$ where $\overrightarrow{q}_{0,n}\equiv\overrightarrow{q}_{0,n}(dy^n;x^n)=\otimes_{i=0}^n{q}_i(dy_i;y^{i-1},x^i)$-a.s. Further the infimum occurs on the boundary of the set $\overrightarrow {Q}_{0,n}(D)$.
See Appendix.
Utilizing Theorem \[lagrange\_duality\], we can reformulate the constraint problem as an unconstrained problem, hence we have $${R}_{0,n}^{na}(D) = \sup_{s\leq{0}}\inf_{{\overrightarrow{q}_{0,n}}} \Big\{{{\mathbb I}}(\mu_{0,n},\overrightarrow{q}_{0,n})-s(\ell_{{d}_{0,n}}(\overrightarrow{q}_{0,n})-D)\Big\}. \label{ex13}$$ Note that ${\overrightarrow{q}_{0,n}} \in {\cal M}_1({\cal Y}_{0,n})$ are probability measures on ${\cal Y}_{0,n}$ therefore, one should introduce another set of Lagrange multipliers.\
Moreover, $\overrightarrow{q}_{0,n}(dy^n;x^n)=\otimes_{i=0}^n{q}_i(dy_i;y^{i-1},x^i)$ is a consistent probability measure on ${\cal Y}_{0,n}$, therefore for each $k=0,1,\ldots,n$, $\int_{{\cal Y}_{0,k}}\overrightarrow{q}_{0,k}(dy^k;x^k)=1$. This constraint is expressed via $$\begin{aligned}
&&\left.\sum_{i=0}^n\int_{{\cal X}_{0,i}\times{\cal Y}_{0,i}}\lambda_{i}(x^i,y^{i-1})\Big{(}\overrightarrow{q}_{0,i}(dy^i;x^i)-1\Big{)}\mu_{0,i}(dx^i)\right.\nonumber\\[-1.5ex]\label{eq.6}\\[-1.5ex]
&&\quad\left.=\sum_{i=0}^n\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}}\lambda_{i}(x^i,y^{i-1})\Big{(}\overrightarrow{q}_{0,n}(dy^n;x^n)-1\Big{)}\mu_{0,n}(dx^n)\right.\nonumber\end{aligned}$$ where $\{\lambda_{i}(\cdot,\cdot):~i=0,1,\ldots,n\}$ are Lagrange multipliers.\
Utilizing the additional constraint (\[eq.6\]) in (\[ex13\]), then we derive the optimal reconstruction kernel for the nonanticipative RDF, $R_{0,n}^{na}(D)$. This is given in the following theorem.
\[th6\] Suppose $d_{0,n}(x^n,y^n)=\sum_{i=0}^n\rho(T^i{x^n},T^i{y^n})$ and the conditions of Lemma \[weakstar-closed\] and Theorem \[weakstar-compact\_2\] hold. Then the infimum in (\[ex13\]) is attained at $\overrightarrow{q}^*_{0,n} \in{L}_{\infty}^w(\mu_{0,n},{\Pi}_{rba}({\cal Y}_{0,n}))$ given by[^7] $$\begin{aligned}
\overrightarrow{q}^*_{0,n}(dy^n;x^n)&=&\otimes_{i=0}^nq_i^*(dy_i;y^{i-1},x^i)-a.s\nonumber\\
&=&\otimes_{i=0}^n\frac{e^{s \rho(T^i{x^n},T^i{y^n})}\nu^*_i(dy_i;y^{i-1})}{\int_{{\cal Y}_i} e^{s \rho(T^i{x^n},T^i{y^n})} \nu^*_i(dy_i;y^{i-1})},~s\leq{0}\label{ex14}\end{aligned}$$ and $\nu^*_i(dy_i;y^{i-1})\in {\cal Q}({\cal Y}_i;{\cal Y}_{0,{i-1}})$. The nonanticipative RDF is given by $$\begin{aligned}
{R}_{0,n}^{na}(D)&=&sD -\sum_{i=0}^n\int_{{{\cal X}_{0,i}}\times{{\cal Y}_{0,i-1}}}\log \Big( \int_{{\cal Y}_i} e^{s\rho(T^i{x^n},T^i{y^n})} \nu^*_i(dy_i;y^{i-1})\Big)\nonumber\\
&&\quad\times{{\overrightarrow q}^*_{0,i-1}}(dy^{i-1};x^{i-1})\otimes\mu_{0,i}(dx^i)\label{ex15}\end{aligned}$$ where “$s$" is the optimal value of (\[ex13\]).\
If ${R}_{0,n}^{na}(D) > 0$ then $ s < 0$ and $$\begin{aligned}
\sum_{i=0}^n\int_{{\cal X}_{0,i}} \int_{{\cal Y}_{0,i}}
\rho(T^i{x^n},T^i{y^n}){\overrightarrow q}^*_{0,i}(dy^i;x^i)\otimes \mu_{0,i}(dx^i)=D\label{eq.7}\end{aligned}$$ and s is obtained from the equality condition (\[eq.7\]).
The fully unconstrained problem of (\[ex13\]) is obtained by introducing another set of Lagrange multipliers $\{\lambda_{i}(\cdot,\cdot):~i=0,1,\ldots,n\}$ as in (\[eq.6\]). Using the pair of Lagrange multipliers $\{s,\lambda\triangleq\{\lambda_{i}(\cdot,\cdot):~i=0,1,\ldots,n\}\}$ introduce the extended pay-off functional $$\begin{aligned}
\mathbb{I}^{s,\lambda}_D(\mu_{0,n},\overrightarrow{q}_{0,n})&\triangleq&\mathbb{I}(\mu_{0,n},\overrightarrow{q}_{0,n})-s\Big{(}\ell_{d_{0,n}}(\overrightarrow{q}_{0,n})-D\Big{)}\nonumber\\
&+&\sum_{i=0}^n\int_{{\cal X}_{0,n}} \int_{{\cal Y}_{0,n}}\lambda_i(x^i,y^{i-1})\Big{(}\overrightarrow{q}_{0,n}(dy^n;x^n)-1\Big{)}\mu_{0,n}(dx^n).\nonumber\end{aligned}$$ This is a fully unconstrained problem on the vector space ${L}_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))$. Utilizing Theorem \[th5\], the Gateaux derivative of $\mathbb{I}_D^{s,\lambda}$ on ${L}_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))$ at any point $\overrightarrow{q}_{0,n}^*$ in the direction $\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*$ is given by $$\begin{aligned}
\delta\mathbb{I}^{s,\lambda}_D(\overrightarrow{q}^*_{0,n};\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)&=\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}}\log\Bigg(\frac{\overrightarrow{q}_{0,n}^*(dy^n;x^n)}{\nu_{0,n}^*(dy^n)}\Bigg)(\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)(dy^n;x^n)\otimes\mu_{0,n}(dx^n)\nonumber\\
&\quad-s\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}}d_{0,n}(x^n,y^n)(\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)(dy^n;x^n)\otimes\mu_{0,n}(dx^n)\nonumber\\
&\quad+\sum_{i=0}^n\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}}\lambda_i(x^i,y^{i-1})(\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)(dy^n;x^n)\otimes\mu_{0,n}(dx^n)\nonumber\end{aligned}$$ $$\begin{aligned}
&=\int_{{\cal X}_{0,n}\times{\cal Y}_{0,n}}\log\Bigg(e^{\sum_{i=0}^n\big{(}-s\rho(T^i{x^n},T^i{y^n})+\lambda_i(x^i,y^{i-1})\big{)}}
\frac{\overrightarrow{q}_{0,n}^*(dy^n;x^n)}{\nu_{0,n}^*(dy^n)}\Bigg)\nonumber\\
&\qquad(\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)(dy^n;x^n)\otimes\mu_{0,n}(dx^n),~\forall\overrightarrow{q}_{0,n}\in{L}_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))\nonumber.\end{aligned}$$ Since $\mathbb{I}^{s,\lambda}_D(\mu_{0,n},\overrightarrow{q}_{0,n})$ is convex in $\overrightarrow{q}_{0,n}$, it follows from the calculus of variations principle that a necessary and sufficient condition for $\overrightarrow{q}^*_{0,n}$ to be a minimizer is $\delta\mathbb{I}^{s,\lambda}_D(\overrightarrow{q}^*_{0,n};\overrightarrow{q}_{0,n}-\overrightarrow{q}_{0,n}^*)=0$, $\forall\overrightarrow{q}_{0,n}\in{L}_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))$. Since the Gateaux derivative must be zero for all $\overrightarrow{q}_{0,n}\in{L}_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))$ then $$\begin{aligned}
\frac{\overrightarrow{q}_{0,n}^*(dy^n;x^n)}{\nu_{0,n}^*(dy^n)}=
e^{\sum_{i=0}^n\big{(}s\rho(T^i{x^n},T^i{y^n})-\lambda_i(x^i,y^{i-1})\big{)}}-a.s.\nonumber\end{aligned}$$ Equivalently, $$\begin{aligned}
\otimes_{i=0}^n\frac{q_i^*(dy_i;y^{i-1},x^i)}{\nu_{i}^*(dy_i;y^{i-1})}
=\otimes_{i=0}^n{e}^{\big{(}s\rho(T^i{x^n},T^i{y^n})-\lambda_i(x^i,y^{i-1})\big{)}}-a.s.\nonumber\end{aligned}$$ Since $\int_{{\cal Y}_i}q_i^*(dy_i;y^{i-1},x^i)=1$, then $$\begin{aligned}
\lambda_i(x^i,y^{i-1})=\log\int_{{\cal Y}_i}e^{s\rho(T^i{x^n},T^i{y^n})}\nu_i^*(dy_i;y^{i-1}),~i=0,1,\ldots,n\nonumber.\end{aligned}$$ Hence, $$\begin{aligned}
\overrightarrow{q}^*_{0,n}(dy^n;x^n)&=&\otimes_{i=0}^nq_i^*(dy_i;y^{i-1},x^i)-a.s\nonumber\\
&=&\otimes_{i=0}^n\frac{e^{s \rho(T^i{x^n},T^i{y^n})}\nu^*_i(dy_i;y^{i-1})}{\int_{{\cal Y}_i} e^{s \rho(T^i{x^n},T^i{y^n})} \nu^*_i(dy_i;y^{i-1})}.\nonumber\end{aligned}$$ Since $s\leq0$ and $\lambda_i\geq0$, $i=0,1,\ldots,n$ then $\overrightarrow{q}_{0,n}^*\in{L}_{\infty}^w(\mu_{0,n},{\Pi}_{rba}({\cal Y}_{0,n}))$. Substituting $\overrightarrow{q}_{0,n}^*$ into $\mathbb{I}_D^{s,\lambda}(\mu_{0,n},\overrightarrow{q}_{0,n})$ gives (\[ex15\]).\
Note that for $s=0$ then $R_{0,n}^{na}(D)=0$ and $\overrightarrow{q}_{0,n}^*(dy^n;x^n)=\nu_{0,n}^*(dy^n)$, $\mu_{0,n}-$almost all $x^n\in{\cal X}_{0,n}$. This is trivial so we must have $s<0$. From Theorem \[lagrange\_duality\] the solution occurs on the boundary of $\overrightarrow{Q}_{0,n}(D)$ giving (\[eq.7\]) for $s<0$.
Often it is interesting to identify conditions so that the optimal reconstruction is Markov with respect to $\{X_i:~i=0,1,\ldots,n\}$. The next remark discusses this case.
Note that if the distortion function satisfies $\rho(T^i{x^n},T^i{y^n})=\rho(x_i,T^i{y^n})$ then according to Theorem \[th6\] we have $${q}^*_{i}(dy_i;y^{i-1},x^i)=q_i^*(dy_i;y^{i-1},x_i)-a.s.,~i\in{\mathbb{N}^n}$$ that is, the reconstruction kernel is Markov in $X^n$. However, even if $\rho(T^ix^n,T^iy^n)=\rho(x_i,y_i)$ (single letter) one cannot claim that the optimal reconstruction distribution is also Markov with respect to $\{Y_i:~i=0,1,\ldots,n\}$ because the right hand side (RHS) of (\[ex14\]) does not satisfy $\nu_i(dy_i;y^{i-1})=\nu_i(dy_i;y_{i-1})$.
The relation between nonanticipative RDF and filtering theory is developed for fixed source distribution. In the next remark we discuss extensions of the nonanticipative RDF for a class of sources and relations to robust filtering.
Nonanticipative RDF can be generalized to a class of sources to address robustness of the filter. One such class is defined by a relative entropy constraint between the unknown or true distribution $P_{X^n}$ with respect to the nominal distribution $P^0_{X^n}$ via $$\begin{aligned}
{\cal M}_{P^0_{X^n}}(d)\triangleq\big\{P_{X^n}\in{\cal M}_1({\cal X}_{0,n}):\mathbb{D}(P_{X^n}||P^0_{X^n})\leq{d}\big\}\end{aligned}$$ where $d$ is the radius of uncertainty. Such a model of uncertainty or class of distributions is often employed in filtering and control applications because it is related to robust filtering and control using minimax methods [@xie-ugrinovskii-petersen2005; @charalambous-rezaei2007].\
Therefore, the nonanticipative RDF for the class of sources ${\cal M}_{P^0_{X^n}}(d)$ is now defined using minimax strategies by $$\begin{aligned}
R^{na,+}_{0,n}(D,d)=\inf_{\overrightarrow{P}_{Y^n|X^n}\in\overrightarrow{Q}_{0,n}(D)}\sup_{P_{X^n}\in{\cal M}_{P^0_{X^n}}(d)}\mathbb{I}(P_{X^n},\overrightarrow{P}_{Y^n|X^n})\label{equation111111}\end{aligned}$$ Through (\[equation111111\]) one can obtain relations to minimax filtering strategies via nonanticipative RDF. An example using this formulation for control of Gaussian state space systems over limited rate channels is found in [@farhadi-charalambous2010]. The investigation of the classical RDF for such a relative entropy class of soures is discussed in [@rezaei-charalambous-stavrou2010], where it is also shown that the Von-Neumann minimax theorem holds and hence one can interchange infimum and supremum operations. The validity of the Von-Neumann minimax theorem for (\[equation111111\]), will imply that the optimal reconstruction distribution for the minimax nonanticipative RDF is (\[ex14\]), and hence the remaining task is to perform the infimum operation over the relative entropy class of the solution to the nonanticipative RDF given by (\[ex15\]). This is the simplest approach to relate nonanticipative RDF for a class of sources and minimax filtering techniques. Unlike minimax filtering techniques, the filtering obtained from (\[equation111111\]) will always satisfy the fidelity criterion which can be defined with respect to probability of error or the average error.\
However, it is not clear how one can apply sensitivity minimization to nonanticipative RDF filter, because only the source distribution is given, while the observation map and filter are obtained from the realization of the optimal reconstruction distribution (see Fig. \[filtering\_and\_causal\]). This is contrary to sensitivity minimization approach, where the input-output maps are given and depend on design functions, such as, the controller or the filter [@hassibi-sayed-kailath1999; @zhou2010]. Nevertheless, when the source is a second order Gaussian process described by a Power Spectral Density (PSD) and the fidelity of reconstruction is the mean-square error, then it might be possible to apply robust filtering and control techniques to address uncertainty of the PSD similar to the computation of capacity of channels with memory [@denic-charalambous-djouadi2009].
Realization of Nonanticipative RDF {#realization}
==================================
The realization of the nonanticipative RDF (optimal reconstruction kernel and nonanticipative RDF) is equivalent to identifying the sensor mapping (see Fig. \[filtering\_and\_causal\]) which generates the auxiliary random process $\{Z_i:~i=0,1,\ldots,n\}$ so that the optimal reconstruction conditional distribution is matched from the output of the source to the output of the filter. This intermediate mapping consists of an encoder followed by a channel. Thus, the realization of the nonanticipative optimal reconstruction distribution consists of a communication channel, an encoder and a decoder such that the reconstruction from the sequence $X^n$ to the sequence $Y^n$ matches the nonanticipative rate distortion minimizing reconstruction kernel. Fig. \[realization2\] illustrates a cascade of subsystems that realizes the nonanticipative RDF. For the single letter expression of classical RDF this is related to the so-called source-channel matching of information theory [@gastpar2003]. It is also described in [@charalambous2008] and [@tatikonda2000] for control over finite capacity communication channels, since this technique allows one to design encoding/decoding schemes without encoding and decoding delays. The realization of the optimal reconstruction kernel is given below.
\[realization1\] Given a source $\{P_{X_i|X^{i-1}}(dx_i|x^{i-1}):i=0,\ldots,n\}$, a channel $\{P_{B_i|B^{i-1},A^{i}}(db_i|$ $b^{i-1},a^i):i=0,\ldots,n\}$ is a realization of the optimal nonanticipative reconstruction kernel $\{q_i^*(dy_i;y^{i-1},x^i):i=0,\ldots,n\}$ if there exists a pre-channel encoder $\{P_{A_i|A^{i-1},B^{i-1},X^i}(da_i|a^{i-1},$ $b^{i-1},x^i):i=0,\ldots,n\}$ and a post-channel decoder $\{P_{Y_i|Y^{i-1},B^i}(dy_i|y^{i-1},b^i):i=0,\ldots,n\}$ such that $$\begin{aligned}
{\overrightarrow q}_{0,n}^*(dy^n;x^n) &\triangleq&\otimes_{i=0}^n q_i^*(dy_i;y^{i-1},x^i)\nonumber\\
&=&\otimes_{i=0}^nP_{Y_i|Y^{i-1},X^i}(dy_i|y^{i-1},x^i)-a.s.\end{aligned}$$ where the joint distribution is $$\begin{aligned}
&P_{X^n,A^n, B^n, Y^n}(dx^n,da^n,db^n,dy^n)\\
&=\otimes_{i=0}^n{P}_{Y_i|Y^{i-1},B^i,A^i,X^i}(dy_i|y^{i-1},b^i,a^i,x^i)\nonumber\\
&\quad\otimes{P}_{B_i|B^{i-1},A^i,X^i,Y^{i-1}}(db_i|b^{i-1},a^i,x^i,y^{i-1})\otimes{P}_{A_i|A^{i-1},X^i,Y^{i-1},B^{i-1}}(da_i|a^{i-1},x^i,y^{i-1},b^{i-1})\nonumber\\
&\quad\otimes{P}_{X_i|X^{i-1},A^{i-1},B^{i-1},Y^{i-1}}(dx_i|x^{i-1},a^{i-1},b^{i-1},y^{i-1})-a.s.,\nonumber\\
&=\otimes_{i=0}^n P_{Y_i|Y^{i-1},B^i}(dy_i|y^{i-1},b^i)\otimes P_{B_i|B^{i-1},A^{i}}
(db_i|b^{i-1},a^i) \nonumber \\
&\quad\otimes P_{A_i|A^{i-1},B^{i-1},X^i}(da_i|a^{i-1},b^{i-1},x^i)\otimes P_{X_i|X^{i-1}}(dx_i|x^{i-1})-a.s. \nonumber\end{aligned}$$ The filter is given by $\{P_{X_i|B^{i-1}}(dx_i|b^{i-1}):i=0,\ldots,n\}$.
![Block Diagram of Realizable Nonanticipative Rate Distortion Function[]{data-label="realization2"}](realization_CRDF.jpg)
Thus, $\{B_i:~i=0,1,\ldots,n\}$ is the auxiliary random process which is obtained during the realization procedure in order to define the filter $\{P_{X_i|B^{i-1}}(dx_i|b^{i-1}):i=0,\ldots,n\}$. Note that unlike Bayesian filtering in which the auxiliary process represents the observations which are given á priori, in nonanticipative RDF this is identified during the realization procedure. In the Definition \[realization1\], the following MC assumptions are assumed.
1)
: $(X^i,A^i)\leftrightarrow(Y^{i-1},B^i)\leftrightarrow{Y_i}$;
2)
: $(X^i,Y^{i-1})\leftrightarrow(B^{i-1},A^i)\leftrightarrow{B_i}$;
3)
: $Y^{i-1}\leftrightarrow(A^{i-1},B^{i-1},X^i)\leftrightarrow{A_i}$;
4)
: $(A^{i-1},B^{i-1},Y^{i-1})\leftrightarrow{X}^{i-1}\leftrightarrow{X_i}$.
These conditional independent assumptions are natural since they correspond to data processing inequalities [@cover-thomas]. Thus, if $\{P_{B_i|B^{i-1},A^{i}}(db_i|b^{i-1},a^i):i=0,\ldots,n\}$ is a realization of the nonanticipative RDF minimizing kernel $\{q_i^*(dy_i;y^{i-1},x^i):i=0,\ldots,n\}$ then the channel connecting the source, encoder, channel, decoder achieves the nonanticipative RDF, and the filter is obtained via $\{P_{X_i|B^{i-1}}(dx_i|b^{i-1}):i=0,\ldots,n\}$. Moreover, the above MCs imply the following data processing inequality, $I (A^n \rightarrow B^n) {{\stackrel}{\triangle}{=}}\sum_{i=0}^n{I}(A^i;B_i|B^{i-1})\geq{I}(X^n;Y^n)$. The optimal realization (encoder-channel-decoder) is defined as the one for which the last inequality holds with equality.
Example
=======
In this section, we present the filter for Gaussian Markov partially-observable processes by utilizing the realization procedure of Section \[realization\].\
Consider the following discrete-time partially observed linear Gauss-Markov system described by $$\begin{aligned}
\left\{ \begin{array}{ll} X_{t+1}=AX_t+BW_t,~X_0=X,~t\in\mathbb{N}^n\\
Y_t=CX_t+NV_t,~t\in\mathbb{N}^n \end{array} \right.\label{equation51}\end{aligned}$$
![Communication System[]{data-label="communication_system"}](uncontrolled_communication_system2.jpg)
where $X_t\in\mathbb{R}^m$ is the state (unobserved) process of information source (plant), and $Y_t\in\mathbb{R}^p$ is the partially observed (measurement) process. The model in (\[equation51\]) consists of a process $\{X_t:~t\in\mathbb{N}^n\}$ which is not directly observed; instead what is directly observed is the process $\{Y_t:~t\in\mathbb{N}^n\}$ which is a noisy version of $\{X_t:~t\in\mathbb{N}^n\}$. This is a realistic model for any sensor which collects information for an underlying process, since the sensor is a measurement device which is often subject to additive Gaussian noise. Hence, in this application the objective is to compress the sensor data. Since we only treat the stationary case, we assume that ($C,A$) is detectable and ($A,\sqrt{BB^{tr}}$) is stabilizable, ($N\neq0$) [@caines1988]. The state and observation noise $\{(W_t,V_t):t\in\mathbb{N}^n\}$ are mutually independent, independent of the Gaussian RV $X_0$, with parameters $N(\bar{x}_0,\bar{\Sigma}_0)$, where $W_t\in\mathbb{R}^k$ and $V_t\in\mathbb{R}^d$, are Gaussian IID processes with zero mean and identity covariances.\
The realization will be done following Fig. \[communication\_system\]. The goal is to reconstruct $\{Y_t:~t\in\mathbb{N}^n\}$ by $\{\tilde{Y}_t:~t\in\mathbb{N}^n\}$ causally. The distortion is single letter defined by $$\begin{aligned}
d_{0,n}(y^n,\tilde{y}^n)\triangleq\frac{1}{n+1}\sum_{t=0}^n||y_t-\tilde{y}_t||^2.\end{aligned}$$ The objective is to compute $$\begin{aligned}
R_{0,n}^{na}(D)=\inf_{\overrightarrow{P}_{\tilde{Y}^n|Y^n}\in\overrightarrow{Q}_{0,n}(D)}\frac{1}{n+1}\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})\label{equation.1}\end{aligned}$$ where $\overrightarrow{Q}_{0,n}(D)\triangleq\big{\{}\overrightarrow{P}_{\tilde{Y}^n|Y^n}:~E\{d_{0,n}(Y^n,\tilde{Y}^n)\}\leq{}D\big{\}}$, and realize the reconstruction distribution. The reconstruction of $\{X_t:~t\in\mathbb{N}^n\}$ when it is fully observed, i.e., when $Y_t=X_t$, is realized over a scalar additive white Gaussian noise (AWGN) channel in [@tatikonda-mitter2004], while the partially observed scalar reconstruction of $\{Y_t:~t\in\mathbb{N}^n\}$ is realized over a scalar AWGN channel in [@charalambous2008] via indirect methods (utilizing upper bounds which are achievable).
Here, we consider the vector process $Y_t\in\mathbb{R}^p$ and realize it over a vector AWGN channel. The methodology is based on the explicit formulae of optimal reconstruction of Theorem \[th6\]. According to Theorem \[th6\], the optimal reconstruction is given by $$\begin{aligned}
\overrightarrow{P}^*_{\tilde{Y}^n|Y^n}(d\tilde{y}^n|y^n)=\otimes_{t=0}^n\frac{e^{s||\tilde{y}_t-y_t||^2}P_{\tilde{Y}_t|\tilde{Y}^{t-1}}(d\tilde{y}_t|\tilde{y}^{t-1})}{\int_{{\cal Y}_t}e^{s||\tilde{y}_t-y_t||^2}P_{\tilde{Y}_t|\tilde{Y}^{t-1}}(d\tilde{y}_t|\tilde{y}^{t-1})},~s\leq{0}\label{eq.9}\end{aligned}$$ where each term in the RHS is identical because our results are derived based on the stationarity assumption. Hence, from (\[eq.9\]) it follows that $P_{\tilde{Y}_t|\tilde{Y}^{t-1},Y^t}=P_{\tilde{Y}_t|\tilde{Y}^{t-1},Y_t}(d\tilde{y}_t|\tilde{y}^{t-1},y_t)-$a.s., that is the reconstruction is Markov with respect to the process $\{Y_t:~t\in\mathbb{N}^n\}$. Moreover, since the exponential term $||\tilde{y}_t-y_t||^2$ in the RHS of (\[eq.9\]) is quadratic in $(y_t,\tilde{y}_t)$, and $\{X_t:~t\in\mathbb{N}^n\}$ is Gaussian then $\{(X_t,{Y}_t):~t\in\mathbb{N}^n\}$ is jointly Gaussian, and it follows that a Gaussian distribution $P_{\tilde{Y}_t|\tilde{Y}^{t-1},Y_t}(\cdot|\tilde{y}^{t-1},y_t)$ (for a fixed realization of $(\tilde{y}^{t-1},y_t)$), and Gaussian distribution $P_{\tilde{Y}_t|\tilde{Y}^{t-1}}(\cdot|\tilde{y}^{t-1})$ can match the left and right side of (\[eq.9\]). Therefore, at time $t\in\mathbb{N}^n$, the output $\tilde{Y}_t$ of the optimal reconstruction channel depends on $Y_t$ and the previous channel outputs $\tilde{Y}^{t-1}$, and its conditional distribution is Gaussian. Hence, the channel connecting $\{Y_t:t\in\mathbb{N}^n\}$ to $\{\tilde{Y}_t:t\in\mathbb{N}^n\}$ has the general form $$\begin{aligned}
\tilde{Y}_t=\bar{A}_tY_t+\bar{B}_t\tilde{Y}^{t-1}+Z_t,~t\in\mathbb{N}^n\label{eq.10}\end{aligned}$$ where $\bar{A}_t\in\mathbb{R}^{p\times{p}}$, $\bar{B}_t\in\mathbb{R}^{p\times{t}p}$, and $\{Z_t:~t\in\mathbb{N}^n\}$ is an independent sequence of Gaussian vectors. Since we treat the stationary case, the finite horizon analysis below is only an intermediate state before we give the stationary solution.\
The communication channel (\[eq.10\]) can be realized via a memoryless additive Gaussian noise channel with feedback [@cover-thomas] defined by $$\begin{aligned}
B_t=A_t+Z_t,~t\in\mathbb{N}^n\label{eq.11}\end{aligned}$$ where the encoder, at time $t$, is a mapping $A_t=\Phi_t(Y_t,\tilde{Y}^{t-1})$ with power $P_t\triangleq{Trace}E\{A_tA_t^{tr}\}$, and the decoder at time $t\in\mathbb{N}^n$ receives $B^t$ and computes the reconstruction $\tilde{Y}_t=\Psi_t(B^t,\tilde{Y}^{t-1})$. By Section \[realization\], in view of the MCs we have the data processing inequality $\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})\leq I(A^n \rightarrow B^n) = \ {I}(A^n;B^n) $, where the last equality holds because the channel is memoryless [@cover-thomas].\
For the realization, the first step is the whitening of the source $\{Y_t:t\in\mathbb{N}^n\}$ by introducing the Gaussian innovation process $\{K_t:~t\in\mathbb{N}^n\}$, defined by $$\begin{aligned}
K_t\triangleq{Y}_t-E\Big{\{}Y_t|\sigma\{\tilde{Y}^{t-1}\}\Big{\}},~t\in\mathbb{N}^n\label{equation52}\end{aligned}$$ whose covariance is defined by $$\begin{aligned}
\Lambda_t\triangleq{E}\{K_tK_t^{tr}\},~t\in\mathbb{N}^n.\end{aligned}$$ The second step is the diagonalization of the covariance $\{\Lambda_t:t\in\mathbb{N}^n\}$ by introducing a unitary transformation $\{E_t:t\in\mathbb{N}^n\}$ such that $$\begin{aligned}
E_t\Lambda_t{E}_t^{tr}=diag\{\lambda_{t,1},\ldots\lambda_{t,p}\},~t\in\mathbb{N}^n.\label{equation53}\end{aligned}$$ Thus, $\Gamma_t\triangleq{E}_tK_t$, where $\{\Gamma_t:~t\in\mathbb{N}^{n}\}$ has independent components for each $t\in\mathbb{N}^n$. In practise, the encoder consists of a pre-encoder which preprocesses the observations $\{Y_t:t\in\mathbb{N}^n\}$ by generating $\{K_t:t\in\mathbb{N}^n\}$ and then applies $\{E_t:t\in\mathbb{N}^n\}$ to it. At the decoder end, there is a pre-decoder which generates $\{\tilde{K}_t:~t\in\mathbb{N}^n\}$ defined by $$\begin{aligned}
\tilde{K}_t\triangleq\tilde{Y}_t-E\Big{\{}Y_t|\sigma\{\tilde{Y}^{t-1}\}\Big{\}},~t\in\mathbb{N}^n\label{eq.12}\end{aligned}$$ on which the unitary transformation $\{E_t:t\in\mathbb{N}^n\}$ is applied to generate $\tilde{\Gamma}_t=E_t\tilde{K}_t$. Next, we calculate the RDF by taking advantage of the preprocessing at the encoder-decoder. Note that the fidelity criterion $d_{0,n}(\cdot,\cdot)$ is not affected by the preprocessing at the encoder-decoder since $d_{0,n}(Y^n,\tilde{Y}^n)=d_{0,n}(K^n,\tilde{K}^n)=\frac{1}{n+1}\sum_{t=0}^n||\tilde{K}_t-K_t||^2=\frac{1}{n+1}\sum_{t=0}^n||\tilde{\Gamma}_t-\Gamma_t||^2$. Now, we show that $$\begin{aligned}
\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})&=\sum_{t=0}^n\Big{(}H(\tilde{K}_t|\tilde{K}^{t-1})-H(\tilde{K}_t|\tilde{K}^{t-1},K_t)\Big{)}\\
&=\sum_{t=0}^n\Big{(}H(\tilde{\Gamma}_t|\tilde{\Gamma}^{t-1})-H(\tilde{\Gamma}_t|\tilde{\Gamma}^{t-1},\Gamma_t)\Big{)}.\end{aligned}$$ By (\[eq.9\]), $$\begin{aligned}
\overrightarrow{P}_{\tilde{Y}^n|Y^n}(d\tilde{y}^n|x^n)=\otimes_{t=0}^n{P}_{\tilde{Y}_t|\tilde{Y}^{t-1},Y_t}(d\tilde{y}_t|\tilde{y}^{t-1},y_t)-a.s.\end{aligned}$$ Hence, $$\begin{aligned}
\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})=\sum_{t=0}^n\Big{(}H(\tilde{Y}_t|\tilde{Y}^{t-1})-H(\tilde{Y}_t|\tilde{Y}^{t-1},Y_t)\Big{)}.\end{aligned}$$ Since conditional entropy is translation invariant, utilizing (\[eq.12\]) gives $$\begin{aligned}
H(\tilde{Y}_t|\tilde{Y}^{t-1})&=H(\tilde{K}_t|\tilde{Y}^{t-1})\nonumber\\
&=H(\tilde{K}_t|\tilde{Y}_{-1},\tilde{Y}_0,\ldots,\tilde{Y}_{t-1})\nonumber\\
&\stackrel{(a)}=H\big{(}\tilde{K}_t|\tilde{Y}_{-1},\tilde{Y}_0,\ldots,\tilde{Y}_{t-2},\tilde{K}_{t-1}+E(\tilde{Y}_{t-1}|\sigma\{\tilde{Y}^{t-2}\})\big{)}\nonumber\\
&=H(\tilde{K}_t|\tilde{Y}_{-1},\tilde{Y}_0,\ldots,\tilde{Y}_{t-2},\tilde{K}_{t-1})\nonumber\\
&=H(\tilde{K}_t|\tilde{K}^{t-1})\label{equation61}\end{aligned}$$ and repeated application of step $(a)$ gives (\[equation61\]). Similarly, $H(\tilde{Y}_t|\tilde{Y}^{t-1},Y_t)=H(\tilde{K}_t|\tilde{Y}^{t-1},K_t)=H(\tilde{K}_t|\tilde{K}^{t-1},K_t)$. Hence, $$\begin{aligned}
\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})=\sum_{t=0}^n\Big{(}H(\tilde{K}_t|\tilde{K}^{t-1})-H(\tilde{K}_t|\tilde{K}^{t-1},K_t)\Big{)}\equiv\sum_{t=0}^n{I}(K_t;\tilde{K}_t|\tilde{K}^{t-1}).\end{aligned}$$ Since the unitary transformation is non-singular then $\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})=\sum_{t=0}^n{I}(K_t;\tilde{K}_t|\tilde{K}^{t-1})=\sum_{t=0}^n{I}(\Gamma_t;\tilde{\Gamma}_t|\tilde{\Gamma}^{t-1}),~{t}\in\mathbb{N}^n$. Therefore, (\[equation.1\]) is equivalent to the following expression. $$\begin{aligned}
R^{na}_{0,n}(D)=R_{0,n}^{na,\Gamma^n,\tilde{\Gamma}^n}(D)\triangleq\inf_{\overrightarrow{P}_{\tilde{\Gamma}^n|\Gamma^n}:~E\big{\{}d_{0,n}(\Gamma^n,\tilde{\Gamma}^n)\leq{D}\big{\}}}\frac{1}{n+1}\mathbb{I}(P_{\Gamma^n},\overrightarrow{P}_{\tilde{\Gamma}^n|\Gamma^n}).\label{equation.2}\end{aligned}$$ By [@berger] (invoking an upper bound and Shannon’s lower bound if necessary) the stationary solution of (\[equation.2\]) is given by $$\begin{aligned}
\lim_{n \longrightarrow \infty} R_{0,n}^{na,\Gamma^n,\tilde{\Gamma}^n}(D)=\lim_{n \longrightarrow \infty}\frac{1}{2}\frac{1}{n+1}\sum_{t=0}^n\sum_{i=1}^p\log\Big{(}\frac{\lambda_{t,i}}{\delta_{t,i}}\Big{)}\end{aligned}$$ where $$\begin{aligned}
\delta_{t,i} \triangleq \left\{ \begin{array}{ll} \xi_t & \mbox{if} \quad \xi_t\leq\lambda_{t,i} \\
\lambda_{t,i} & \mbox{if}\quad\xi_t>\lambda_{t,i} \end{array} \right.,~t\in\mathbb{N}^n,~i=2,\ldots,p\nonumber\end{aligned}$$ and $\{\xi_t:~t\in\mathbb{N}^n\}$ is chosen such that $\sum_{i=1}^p\delta_{t,i}=D$. Define $\eta_{t,i}{{\stackrel}{\triangle}{=}}1-\frac{\delta_{t,i}}{\lambda_{t,i}}$, $i=1,\ldots,p$, $\Delta_t\triangleq{diag}\{\delta_{t,1},\ldots,\delta_{t,p}\}$, and $H_t\triangleq{diag}\{\eta_{t,1},\ldots,\eta_{t,p}\}\in\mathbb{R}^{p\times{p}}$.\
As a result, the reconstruction conditional distribution is given by $$\begin{aligned}
P^*_{\tilde{\Gamma}^n|\Gamma^n}(d\tilde{\gamma}^n|{\gamma}^n)=\otimes_{t=0}^n{P}^*_{\Gamma_t|\tilde{\Gamma}_t}(d\tilde{\gamma}_t|\gamma_t)-a.s.\nonumber\end{aligned}$$ where ${P}^*_{\tilde{\Gamma}_t|{\Gamma}_t}(\cdot|\cdot)\sim{N}(H_{t}\Gamma_t,H_{t}\Delta_{t})$.\
![Design of Discrete-Time Communication System[]{data-label="discrete_time_communication_system"}](discrete_time_communication_system_v2.jpg)
[*Realization of Nonanticipative RDF Over Vector AWGN Channel.*]{} Consider a vector channel $B_t=A_t+Z_t,~t\in\mathbb{N}^n$, where $Z_t$ is Gaussian zero mean, $Q\triangleq{C}ov(Z_t)=diag\{q_1,q_2,\ldots,q_p\}$, and $A_t\in\mathbb{R}^p$. By Section \[realization\], and the memoryless nature of the channel we know that $I(A^n \rightarrow B^n) ={I}(A^n;B^n)\geq\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})$. Hence, we compress the source and transmit it to the decoder over the vector channel, so that the RDF is equal to the capacity of the channel, i.e., $\lim_{n \longrightarrow \infty} R_{0,n}^{na}(D)= \lim_{n \longrightarrow \infty} \frac{1}{n+1} I(A^n;B^n)$. That is, we match the source to the channel. Therefore, we need to design the operators $\{({\cal A}_t,{\cal B}_t):~t\in\mathbb{N}^n\}$ so that the compressed signal $A_t={\cal A}_t\Gamma_t$, is sent through an AWGN channel with feedback (shown in Fig. \[discrete\_time\_communication\_system\]), after which the received signal is decompressed by $\tilde{\Gamma}_t={\cal B}_tB_t$ at the pre-decoder. By the knowledge of the channel output at the decoder, the mean square estimator $\hat{X}_t$ is generated at the decoder since $\hat{X}_t\triangleq{E}\big{\{}X_t|\sigma\{\tilde{Y}^{t-1}\}\big{\}}$ (one may also use $\sigma\{{B}^{t-1}\}$ to find the filter of $\{X_t:~t\in\mathbb{N}^n\}$).\
The compression operator $\{{\cal A}_t:t\in\mathbb{N}^n\}$ is chosen so that $\lim_{n \longrightarrow \infty} R^{na}_{0,n}(D)=\lim_{n \longrightarrow \infty} \frac{1}{n+1}I(A^n;B^n)$. Recall that $B_t=A_t+Z_t$, $A_t={\cal A}_tE_tK_t$, $Q\triangleq{C}ov(Z_t)$, $Trace{E}\{A_tA_t^{tr}\}={P}_t, t=0,1, \ldots n$. Hence, we find $\{{\cal A}_t:t\in\mathbb{N}^n\}$ so that the following holds. $$\begin{aligned}
\lim_{n \longrightarrow \infty} \frac{1}{n+1} C_{0,n}(P_0,\ldots{P}_n)&\triangleq& \lim_{n \longrightarrow \infty} \frac{1}{n+1}I(A^n;{B}^n)= \lim_{n \longrightarrow \infty}\frac{1}{2}\frac{1}{n+1}\sum_{t=0}^n\log|I+E\{A_tA_t^{tr}\}Q^{-1}|\nonumber\\
&=&\lim_{n \longrightarrow \infty}\frac{1}{2}\frac{1}{n+1}\sum_{t=0}^n\log\frac{|\Lambda_{t}|}{|\Delta_{t}|}=\lim_{n \longrightarrow \infty}R^{na}_{0,n}(D).\nonumber\end{aligned}$$ From the previous equality we obtain $${\cal A}_t\triangleq\sqrt{Q\Delta_t^{-1}H_t},~t\in\mathbb{N}^n.\nonumber$$ The decompression operator $\{{\cal B}_t:t\in\mathbb{N}^n\}$ is chosen so that the desired distortion is achieved by the above realization. The decompressed channel output $\tilde{\Gamma}_t={\cal B}_t{B}_t$ due to transmitting the compressed input $A_t={\cal A}_t{\Gamma}_t$ is $$\begin{aligned}
\tilde{\Gamma}_t&=&{\cal B}_tB_t={\cal B}_t(A_t+Z_t)={\cal B}_t({\cal A}_t{\Gamma}_t+Z_t),~\Gamma_t=E_tK_t\nonumber\\
&=&H_tE_tK_t+{\cal B}_tZ_t,~t\in\mathbb{N}^n.\label{equation56}\end{aligned}$$ By pre-multiplying $\tilde{\Gamma}_t$ by $E_t^{tr}$ we can construct $$\begin{aligned}
\tilde{K}_t&=&E_t^{tr}\tilde{\Gamma}_t=E_t^{tr}H_tE_tK_t+E_t^{tr}{\cal B}_tZ_t,~t\in\mathbb{N}^n.\end{aligned}$$ The reconstruction of $Y_t$ is given by the sum of $\tilde{K}_t$ and $C\hat{X}_t$ as follows. $$\begin{aligned}
\tilde{Y}_t&=&\Psi_t(B^t,\tilde{Y}^{t-1})\nonumber\\
&=&\tilde{K}_t+C\hat{X}_t,~\hat{X}_t=E\Big{\{}X_t|\sigma\{\tilde{Y}^{t-1}\}\Big{\}}\label{eq.13}\\
&=&E_t^{tr}H_tE_tK_t+E_t^{tr}{\cal B}_tZ_t+C\hat{X}_t,~t\in\mathbb{N}^n.\label{eq.14}\end{aligned}$$ Next, we determine $\{{\cal B}_t:t\in\mathbb{N}^n\}$.\
First, we notice that $$\begin{aligned}
E\Big{\{}(Y_t-\tilde{Y}_t)^{tr}(Y_t-\tilde{Y}_t)\Big{\}}=Trace\Big{(}E\Big{\{}(Y_t-\tilde{Y}_t)(Y_t-\tilde{Y}_t)^{tr}\Big{\}}\Big{)}.\nonumber\end{aligned}$$ Then we can compute $$\begin{aligned}
&&E\Big{\{}(Y_t-\tilde{Y}_t)^{tr}(Y_t-\tilde{Y}_t)\Big{\}}=Trace{E}\Big{\{}(K_t-\tilde{K}_t)(K_t-\tilde{K}_t)^{tr}\Big{\}}\nonumber\\
&&=Trace{E}\Big{\{}(K_t-E_t^{tr}\tilde{\Gamma}_t)(K_t-E_t^{tr}\tilde{\Gamma}_t)^{tr}\Big{\}}\nonumber\\
&&=Trace{E}\Big{\{}(K_t-E_t^{tr}H_tE_tK_t-E_t^{tr}{\cal B}_tZ_t)(K_t-E_t^{tr}H_tE_tK_t-E_t^{tr}{\cal B}_tZ_t)^{tr}\Big{\}}\nonumber\\
&&=Trace{E}\Big{\{}\big{(}(I-E_t^{tr}H_tE_t)K_t-E_t^{tr}{\cal B}_tZ_t\big{)}\big{(}(I-E_t^{tr}H_tE_t)K_t-E_t^{tr}{\cal B}_tZ_t\big{)}^{tr}\Big{\}}\nonumber\\
&&=Trace\Big{\{}(I-E_t^{tr}H_tE_t)\Lambda_t(I-E_t^{tr}H_tE_t)^{tr}+E_t^{tr}{\cal B}_tQ{\cal B}_t^{tr}E_t\Big{\}}\nonumber\\
&&=Trace\Big{\{}(I-E_t^{tr}H_tE_t)E_t^{tr}diag(\lambda_{t,1},\ldots,\lambda_{t,p})E_t(I-E_t^{tr}H_tE_t)^{tr}+E_t^{tr}{\cal B}_tQ{\cal B}_t^{tr}E_t\Big{\}}\nonumber\\
&&=Trace\Big{\{}E_t^{tr}\Big{(}(I-H_t)diag(\lambda_{t,1},\ldots,\lambda_{t,p})(I-H_t)^{tr}+({\cal B}_tQ{\cal B}_t^{tr})\Big{)}E_t\Big{\}}\nonumber\\
&&\stackrel{(b)}=Trace\Big{\{}diag(\delta_{t,1},\ldots,\delta_{t,p})\Big{\}}=D~~~~\nonumber$$ where $(b)$ holds if we set $${\cal B}_t\triangleq\sqrt{H_t\Delta_tQ^{-1}},~t\in\mathbb{N}^n.\nonumber$$ This shows that the realization of Fig. \[discrete\_time\_communication\_system\] achieves end-to-end average distortion equal to $D$.\
[*Decoder.*]{} The decoder is $\tilde{Y}_t=\tilde{K}_t+C\hat{X}_t$, where $\{\hat{X}_t:~t\in\mathbb{N}^n\}$ is obtained from the modified Kalman filter as follows. Recall that $$\begin{aligned}
\tilde{Y}_t&=&\tilde{K}_t+C\hat{X}_t\nonumber\\
&=&E_t^{tr}H_tE_t(Y_t-C\hat{X}_t)+E_t^{tr}{\cal B}_tZ_t+C\hat{X}_t\nonumber\\
&=&E_t^{tr}H_tE_t(CX_t+NV_t-C\hat{X}_t)+E_t^{tr}{\cal B}_tZ_t+C\hat{X}_t\nonumber\\
&=&E_t^{tr}H_tE_t(CX_t-C\hat{X}_t)+C\hat{X}_t+E_t^{tr}H_tE_tN V_t+E_t^{tr}{\cal B}_t Z_t\label{equation58}\end{aligned}$$ where $\{V_t:~t\in\mathbb{N}^n\}$ and $\{Z_t:~t\in\mathbb{N}^n\}$ are independent Gaussian vectors. Then $\hat{X}_t=E\big{\{}X_t|\sigma\{\tilde{Y}^{t-1}\}\big{\}}$ is given by the modified Kalman filter $$\begin{aligned}
\hat{X}_{t+1}&=&A\hat{X}_t+A\Sigma_t(E_t^{tr}H_tE_tC)^{tr}M_t^{-1}\big(\tilde{Y}_t-C\hat{X}_t\big),~\hat{X}_0=\bar{x}_0\label{10}\\
\Sigma_{t+1}&=&A\Sigma_tA^{tr}-A\Sigma_t(E_t^{tr}H_tE_tC)^{tr}M_t^{-1}(E_t^{tr}H_tE_tC)\Sigma_tA+BB_t^{tr},~\Sigma_0=\bar{\Sigma}_0\label{11}\end{aligned}$$ where $$\begin{aligned}
M_t=E_t^{tr}H_tE_tC\Sigma_t(E_t^{tr}H_tE_tC)^{tr}+E_t^{tr}H_tE_tNN^{tr}(E_t^{tr}H_tE_t)^{tr}+E_t^{tr}{\cal B}_t {\cal B}_t^{tr}E_t.\end{aligned}$$ [*Stationary Solution: Infinite Horizon.*]{} Now, we are ready to give the complete solution to the stationary nonanticipative RDF and its realization. As $t\longrightarrow\infty$, under the assumption that the linear Gauss-Markov system is stabilizable and detectable, we have $$\begin{aligned}
\Sigma_{\infty}&=&A\Sigma_\infty{A}^{tr}-A\Sigma_{\infty}(E_\infty^{tr}H_\infty{E}_{\infty}C)^{tr}M_{\infty}^{-1}(E_{\infty}^{tr}H_{\infty}E_{\infty}C)\Sigma_{\infty}A+BB_{\infty}^{tr}\nonumber\end{aligned}$$ where $$M_\infty=E_\infty^{tr}H_\infty{E}_{\infty}C\Sigma_{\infty}(E_{\infty}^{tr}H_{\infty}E_{\infty}C)^{tr}+E_{\infty}^{tr}H_{\infty}E_{\infty}NN^{tr}(E_{\infty}^{tr}H_{\infty}E_{\infty})^{tr}+E_{\infty}^{tr}{\cal B}_{\infty} {\cal B}_{\infty}^{tr}E_\infty\nonumber$$ and $E_{\infty}$ is the unitary matrix that diagonalizes $\Lambda_{\infty}$ given by $$\begin{aligned}
E_{\infty}\Lambda_{\infty}E_{\infty}^{tr}=diag(\lambda_{\infty,1},\ldots,\lambda_{\infty,p})\nonumber\end{aligned}$$ and $$\begin{aligned}
\delta_{\infty,i} \triangleq \left\{ \begin{array}{ll} \xi_\infty & \mbox{if} \quad \xi_\infty\leq\lambda_{\infty,i} \\
\lambda_{\infty,i} & \mbox{if}\quad\xi_\infty>\lambda_{\infty,i} \end{array} \right.,~i=1,\ldots,p\nonumber\end{aligned}$$ satisfying $\sum_{i=1}^p\delta_{\infty,i}=D$.\
Define $$\begin{aligned}
\Delta_{\infty}=diag(\delta_{\infty,1},\ldots,\delta_{\infty,p}),~H_{\infty}=diag(\eta_{\infty,1},\ldots,\eta_{\infty,p})\nonumber\end{aligned}$$ where $\eta_{\infty,i}=1-\frac{\delta_{\infty,i}}{\lambda_{\infty,i}}$. The nonanticipative RDF can be computed as follows. $$\begin{aligned}
R^{na}(D)&=&\lim_{n\longrightarrow\infty}\inf_{P_{\tilde{Y}^n|Y^n}(d\tilde{y}^n|{y}^n)\in\overrightarrow{Q}_{0,n}(D)}\frac{1}{n+1}\mathbb{I}(P_{Y^n},\overrightarrow{P}_{\tilde{Y}^n|Y^n})\nonumber\\
&=&\lim_{n\longrightarrow\infty}\Bigg{(}\frac{1}{2}\frac{1}{n+1}\sum_{t=0}^n\sum_{i=1}^p\log\Big(\frac{\lambda_{t,i}}{\delta_{t,i}}\Big)\Bigg{)}\nonumber\\
&=&\frac{1}{2}\sum_{i=1}^p\log\Big(\frac{\lambda_{\infty,i}}{\delta_{\infty,i}}\Big)=\frac{1}{2}\log\frac{|\Lambda_{\infty}|}{|\Delta_{\infty}|}=\frac{1}{2}\lim_{n\longrightarrow\infty}C_{0,n}(P_0,\ldots,P_n)\stackrel{(c)}\equiv{C}(P)\label{equation59}\end{aligned}$$ where $(c)$ comes from the fact that the power constraint satisfies $\lim_{t\longrightarrow\infty}Trace{E}\{A_tA_t^{tr}\}=\lim_{t\longrightarrow\infty}P_t=P$. Thus, for a given distortion level $D$, $C(P)=R^{na}(D)$ is the minimum capacity under which there exists a realizable filter for the data reconstruction of $\{Y_t:~t\in\mathbb{N}\}$ by $\{\tilde{Y}_t:~t\in\mathbb{N}\}$ ensuring an average distortion equal to $D$. Note that for $\frac{D}{p}<\min_{i}\lambda_{\infty,i}$ then $R^{na}(D)=\frac{1}{2}\log\frac{|\Lambda_{\infty}|}{(\frac{D}{p})^p}$, e.g., $\delta_{\infty,i}=\frac{D}{p}$. Hence, from (\[equation59\]) we have $D=p\Big(|\Lambda_{\infty}|e^{-2R^{na}}\Big)^{\frac{1}{p}}$. As a result, we have the direct relation between the reconstruction error $D$ and the rate $R^{na}$. Finally, the filter is the steady state version of (\[10\]), (\[11\]) with initial condition $\hat{X}_0 = E\{ X_0| Y^{-1}\}$ and $\Sigma_0$ the covariance of $X_0- \hat{X}$ which is Gaussian $N(0, \Sigma_\infty)$.
Conclusion
==========
This paper investigates nonanticipative RDF on abstract spaces. Existence of the optimal reconstruction conditional distribution is shown, while closed form expression is derived for the stationary case. The relation between filtering theory and nonanticipative rate distortion theory is discussed via a realization procedure. Finally, an example is presented which illustrates the realization of the nonanticipative RDF.
Proofs
========
Proof of Lemma \[weakstar-closed\]
----------------------------------
To show closedness of ${\overrightarrow Q}_{ad}$ as a subset of $Q_{ad}$ it suffices to show that $$\begin{aligned}
\otimes_{i=0}^n{q}_i^{\alpha}(\cdot;y^{i-1},x^{i})\buildrel w^* \over \longrightarrow\otimes_{i=0}^n{q}_i^{0}(\cdot;y^{i-1},x^{i}).\end{aligned}$$ This will be shown by induction. Consider $n=0$. For any $h_{0}(x_0,y_0)\in{L}_1(\mu_{0},{BC}({\cal Y}_{0}))$, by definition of weak$^*$-convergence it follows from (a) that $$\begin{aligned}
\lim_{\alpha\longrightarrow\infty}\int_{{\cal X}_0\times{\cal Y}_0}h_0(x_0,y_0)q_0^{\alpha}(dy_0;x_0)\mu_0(dx_0)=\int_{{\cal X}_0\times{\cal Y}_0}h_0(x_0,y_0)q_0^{0}(dy_0;x_0)\mu_0(dx_0)\nonumber.\end{aligned}$$ Consider $n=1$. For $\tilde{h}_0(\cdot,\cdot)\in{L}_1(\mu_0,BC({\cal Y}_0))$, $\tilde{h}_1(\cdot,\cdot)\in{L}_1(\mu_{1},BC({\cal Y}_{1}))$ We need to show that $$\begin{aligned}
&&\lim_{\alpha\longrightarrow\infty}\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{\alpha}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}q_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\nonumber\\
&&-\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{0}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}q_0^{0}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}=0\nonumber.\end{aligned}$$ The latter equation is written as follows. $$\begin{aligned}
&&\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{\alpha}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}q_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\nonumber\\
&&-\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{0}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}q_0^{0}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}\end{aligned}$$ $$\begin{aligned}
&&\leq\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\underbrace{\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{0}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}}_{\tilde{h}_1(x_0,y_0)}q_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\\
&&-\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{0}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}q_0^{0}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}\\
&&+\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{\alpha}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}q_0^{\alpha}dy_0;x_0)\bigg{)}\mu_0(dx_0)\nonumber\\
&&-\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)q_1^{0}(dy_1;y_0,x^1)\bigg{)}\mu_1(dx_1;x_0)\bigg{)}q_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}.\end{aligned}$$ We need to show that both RHS terms go to zero as $a\longrightarrow\infty$. Let $\epsilon>0$ be given. Then, there exists an $\alpha_\epsilon\in{\cal D}$ such that for all $\alpha\succeq{\alpha}_\epsilon$ the first RHS term can be written as $$\begin{aligned}
&&\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0){\tilde{h}_1(x_0,y_0)}\Big{(}q_0^{\alpha}(dy_0;x_0)-q_0^{0}(dy_0;x_0)\Big{)}\bigg{)}\mu_0(dx_0)\Bigg{|}\nonumber\\
&&=\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\tilde{h}_1(x_0,y_0)\Big{(}q_0^{\alpha}(dy_0;x_0)-q_0^{0}(dy_0;x_0)\Big{)}\bigg{)}\mu_0(dx_0)\Bigg{|}\nonumber\\
&&\leq\int_{{\cal X}_{0}}\Bigg{|}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\tilde{h}_1(x_0,y_0)\Big{(}q_0^{\alpha}(dy_0;x_0)-q_0^{0}(dy_0;x_0)\Big{)}\Bigg{|}\mu_0(dx_0)\nonumber\\
&&\leq\epsilon,~\forall~\epsilon>0~\mbox{and}~\forall\alpha\succ\alpha_\epsilon\nonumber\end{aligned}$$ where the last inequality follows from condition (b), e.g., $\tilde{h}_0(\cdot,\cdot)\in{L}_1(\mu_0,BC({\cal Y}_0))$.\
The second RHS term can be written as $$\begin{aligned}
&&\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\underbrace{\bigg{(}\int_{{\cal X}_{1}}\bigg{(}\int_{{\cal Y}_{1}}h_{1}(x_1,y_1)\Big{(}q_1^{\alpha}(dy_1;y_0,x^1)-q_1^0(dy_i;y_{0},x^1)\Big{)}\bigg{)}\mu_1(dx_1;x_0)\bigg{)}}_{\tilde{h}^\alpha_1(x_0,y_0)}
\nonumber\\
&&\otimes{q}_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}=\int_{{\cal X}_{0}}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\tilde{h}^\alpha_1(x_0,y_0)
q_0^{\alpha}(dy_0;x_0)\otimes\mu_0(dx_0).\label{eq.15}\end{aligned}$$ By condition (c) for $i=1$, and $\forall~\epsilon>0$ and $\alpha\succ\alpha_{\epsilon}$ we have $$\begin{aligned}
\sup_{y_0\in{\cal Y}_{0}}\int_{{\cal X}_1}\bigg{|}\int_{{\cal Y}_1}h_1(x_1,y_1)q_1^{\alpha}(dy_1;y_0,x^1)-\int_{{\cal Y}_1}h_1(x_1,y_1)q_1^0(dy_1;y^0,x^1)\bigg{|}\mu_1(dx_1;x_0)\leq\epsilon,~\forall~x_0\in{\cal X}_0.\nonumber\end{aligned}$$ Utilizing the last inequality into (\[eq.15\]) yields that in the limit as $\alpha\longrightarrow\infty$, then (\[eq.15\]) goes to zero.\
Next, suppose that for $n=k$ and for all $\epsilon>0$ there exists $\alpha_\epsilon\in{\cal D}$ such that for any $\alpha\succeq\alpha_\epsilon$ $$\begin{aligned}
&&\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\ldots\bigg{(}\int_{{\cal X}_{k}}\bigg{(}\int_{{\cal Y}_{k}}h_{k}(x_k,y_k){q}_k^{\alpha}(dy_k;y^{k-1},x^k)\bigg{)}\mu_k(dx_k;x^{k-1})\bigg{)}\nonumber\\
&&\ldots{q}_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\nonumber\\
&&-\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\ldots\bigg{(}\int_{{\cal X}_{k}}\bigg{(}\int_{{\cal Y}_{k}}h_{k}(x_k,y_k){q}_k^{0}(dy_k;y^{k-1},x^k)\bigg{)}\mu_k(dx_k;x^{k-1})\bigg{)}\nonumber\\
&&\ldots{q}_0^{0}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}\leq\epsilon.\nonumber\end{aligned}$$ To conclude the derivation we need to show that for $n=k+1$ $$\begin{aligned}
\otimes_{i=0}^{k+1}{q}_i^{\alpha}(\cdot;y^{i-1},x^{i})\buildrel w^* \over \longrightarrow\otimes_{i=0}^{k+1}{q}_i^{0}(\cdot;y^{i-1},x^{i}).\end{aligned}$$ Consider $n=k+1$. We need to show that for all $\epsilon>0$ there exists $\alpha_\epsilon\in{\cal D}$ such that for any $\alpha\succeq\alpha_\epsilon$ $$\begin{aligned}
&&\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\ldots\bigg{(}\int_{{\cal X}_{k+1}}\bigg{(}\int_{{\cal Y}_{k+1}}h_{k+1}(x_{k+1},y_{k+1}){q}_{k+1}^{\alpha}(dy_{k+1};y^{k},x^{k+1})\bigg{)}\mu_{k+1}(dx_{k+1};x^{k})\bigg{)}\\
&&\ldots{q}_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\nonumber\\
&&-\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\ldots\bigg{(}\int_{{\cal X}_{k+1}}\bigg{(}\int_{{\cal Y}_{k+1}}h_{k+1}(x_{k+1},y_{k+1}){q}_{k+1}^{0}(dy_{k+1};y^{k},x^{k+1})\bigg{)}\mu_{k+1}(dx_{k+1};x^{k})\bigg{)}\\
&&\ldots{q}_0^{0}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}\leq\epsilon.\nonumber\end{aligned}$$ Since, $$\begin{aligned}
&&\Bigg{|}\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\ldots\bigg{(}\int_{{\cal X}_{k+1}}\bigg{(}\int_{{\cal Y}_{k+1}}h_{k+1}(x_{k+1},y_{k+1}){q}_{k+1}^{\alpha}(dy_{k+1};y^{k},x^{k+1})\bigg{)}\mu_{k+1}(dx_{k+1};x^{k})\bigg{)}\\
&&\ldots{q}_0^{\alpha}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\nonumber\\
&&-\int_{{\cal X}_{0}}\bigg{(}\int_{{\cal Y}_0}h_{0}(x_0,y_0)\ldots\bigg{(}\int_{{\cal X}_{k+1}}\bigg{(}\int_{{\cal Y}_{k+1}}h_{k+1}(x_{k+1},y_{k+1}){q}_{k+1}^{0}(dy_{k+1};y^{k},x^{k+1})\bigg{)}\mu_{k+1}(dx_{k+1};x^{k})\bigg{)}\nonumber\\
&&\ldots{q}_0^{0}(dy_0;x_0)\bigg{)}\mu_0(dx_0)\Bigg{|}\end{aligned}$$ $$\begin{aligned}
&&\leq\Bigg{|}\int_{{\cal X}_{0,k}}\int_{{\cal Y}_{0,k}}\otimes_{i=0}^kh_{i}(x_i,y_i)\bigg{(}\int_{{\cal X}_{k+1}}\int_{{\cal Y}_{k+1}}h_{k+1}(x_{k+1},y_{k+1})\Big{(}{q}_{k+1}^{\alpha}(dy_{k+1};y^{k},x^{k+1})\nonumber\\
&&-q_{k+1}^0(dy_{k+1};y^{k},x^{k+1})\Big{)}\mu_{k+1}(dx_{k+1};x^{k})\bigg{)}
\otimes_{i=0}^k{q}_{i}^\alpha(dy_i;y^{i-1},x^i)\otimes\mu_{i}(dx_i;x^{i-1})\Bigg{|}\\
&&+\Bigg{|}\int_{{\cal X}_{0,k}}\int_{{\cal Y}_{0,k}}\otimes_{i=0}^kh_{i}(x_i,y_i)\underbrace{\bigg{(}\int_{{\cal X}_{k+1}}\int_{{\cal Y}_{k+1}}h_{k+1}(x_{k+1},y_{k+1}){q}_{k+1}^{0}(dy_{k+1};y^{k},x^{k+1})\mu_{k+1}(dx_{k+1};x^{k})\bigg{)}}_{\tilde{h}_{k+1}(x^k,y^k)}\\
&&\otimes_{i=0}^k\Big{(}{q}_{i}^\alpha(dy_i;y^{i-1},x^i)-{q}_{i}^0(dy_i;y^{i-1},x^i)\Big{)}\otimes\mu_{i}(dx_i;x^{i-1})\Bigg{|}.\nonumber\end{aligned}$$ By condition (c) the following inequality holds, $\forall x^k\in{\cal X}_{0,k}$, $$\begin{aligned}
&\sup_{y^k\in{\cal Y}_{0,k}}\int_{{\cal X}_{k+1}}\bigg{|}\int_{{\cal Y}_{k+1}}h_{k+1}(x_{k+1},y_{k+1})\Big{(}{q}_{k+1}^{\alpha}(dy_{k+1};y^{k},x^{k+1})-\\
&\qquad\qquad{q}_{k+1}^0(dy_{k+1};y^{k},x^{k+1})\Big{)}\bigg{|}\mu_{k+1}(dx_{k+1};x^{k})\nonumber\\
&\leq\epsilon,~\forall~\epsilon>0~\mbox{and}~\forall~\alpha\succ\alpha_{\epsilon}.\end{aligned}$$ Also, by condition (b), $\tilde{h}_{k+1}\in{L}_1(\mu_{0,k},BC({\cal Y}_{0,k}))$. Utilizing the previous observations and the induction hypothesis $\otimes_{i=0}^{k}{q}_i^{\alpha}(\cdot;y^{i-1},x^{i})\buildrel w^* \over \longrightarrow\otimes_{i=0}^{k}{q}_i^{0}(\cdot;y^{i-1},x^{i})$ in the two inequalities above, then in the limit as $\alpha\longrightarrow\infty$, the terms in the inequality go to zero.\
As a result, ${\overrightarrow Q}_{ad}$ is a weak$^*$-closed set. Being a weak$^*$-closed subset of the weak$^*$-compact set $Q_{ad}$, ${\overrightarrow{Q}}_{ad}$ is also weak$^*$-compact.
Proof of Theorem \[lagrange\_duality\]
--------------------------------------
The proof is based on Lagrange Duality theorem [@dluenberger69 Theorem 1, p. 224]. We choose $X \triangleq{L}_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))$ which is clearly a vector space. For the set $\Omega$ the natural choice is the set $ \Omega =\overrightarrow{Q}_{ad}\equiv{L}_{\infty}^w(\mu_{0,n},{\Pi}_{rba}({\cal Y}_{0,n}))\subseteq X$. Define $$\begin{aligned}
G(\overrightarrow{q}_{0,n})&\triangleq& \ell_{d_{0,n}}({\overrightarrow q}_{0,n})-D,\quad {\overrightarrow q}_{0,n} \in L_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))\nonumber\\
&\triangleq& \int_{{\cal X}_{0,n}}
\biggl( \int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n) {\overrightarrow q}(dy^n;x^n) \biggr) \mu_{0,n}(dx^n) - D.\nonumber\end{aligned}$$ It is clear that $G(\cdot)$ is a convex mapping from $L_{\infty}^{w}(\mu_{0,n},M_{rba}({\cal Y}_{0,n}))$ into the real line with the natural ordering $(\mathbb{R},\preceq) \triangleq Z$. Also recall that ${\overrightarrow q}_{0,n} \rightarrow {\mathbb{I}}(\mu_{0,n}; \overrightarrow{q}_{0,n})$ is convex and well defined on $\Omega$ and that, by Theorem \[th3\], $\inf_{\overrightarrow{q}_{0,n}\in \overrightarrow{Q}_{0,n}(D)}\mathbb{I}(\mu_{0,n}; \overrightarrow{q}_{0,n})$ exists and is finite. Thus, according to the Lagrange duality theorem referred to above, it suffices to show that there exists a ${\overrightarrow q}_{0,n}^{1} \in \Omega$ such that $$\begin{aligned}
G({\overrightarrow q}_{0,n}^{1})= \int_{{\cal X}_{0,n}} \big\{ \int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n){\overrightarrow q}_{0,n}^{1}(dy^n;x^n)\big\} \mu_{0,n}(dx^n) - D < 0.\nonumber\end{aligned}$$ Introduce the sets $A_1 \triangleq \{ x^n \in {\cal X}_{0,n}: \Gamma_{x^n} \ne \emptyset \} $ and $A_0 \triangleq {\cal X}_{0,n}\setminus A_1$, with $\Gamma_{x^n}$ denoting the $x^n$-section of $\Gamma.$ Define the measure valued function ${\overrightarrow q}_{0,n}^{1}$ as follows $$\begin{aligned}
{\overrightarrow q}_{0,n}^{1}(\Gamma_{x^n};x^n) = 0,~~ \forall ~~ x \in A_0;~~ {\overrightarrow q}_{0,n}^{1}({\cal Y}_{0,n};x^n)= 1,~~ \forall~~ x^n \in {\cal X}_{0,n}\nonumber\end{aligned}$$ $$\begin{aligned}
0 \leq {\overrightarrow q}_{0,n}^{1}(B;x^n)\leq 1, B\subset \Gamma_{x^n}, ~ {\overrightarrow q}_{0,n}^{1}(\Gamma_{x^n};x^n) =1,~~ \forall ~~ x^n \in A_1\nonumber\end{aligned}$$ where $B \in {\cal B}({\cal Y}_{0,n})$. Since by hypothesis $\Gamma \ne \emptyset$ we have $\mu_{0,n}(A_1)>0$ and thus the kernel ${\overrightarrow q}_{0,n}^{1}$ is well defined and it belongs to $L_{\infty}^w(\mu_{0,n},{\Pi}_{rba}({\cal Y}_{0,n}))$. Using this kernel in the expression for $\ell_{d_{0,n}}({\overrightarrow q}_{0,n})$, one can easily verify that $\ell_{d_{0,n}}({\overrightarrow q}_{0,n}^{1}) < D$ and hence $G( {\overrightarrow q}_{0,n}^{1}) < 0$. Then, by the Lagrange Duality theory, we arrive at the conclusion of the theorem as stated. Also it follows from the same duality theory that if the infimum is achieved by some ${\overrightarrow q}_{0,n}^* \in L_{\infty}^w(\mu_{0,n},{\Pi}_{rba}({\cal Y}_{0,n}))$, then $$\begin{aligned}
s \biggl(\int_{{\cal X}_{0,n}} \int_{{\cal Y}_{0,n}}d_{0,n}(x^n,y^n) {\overrightarrow q}_{0,n}^*(dy^n;x^n)\otimes\mu_{0,n}(dx^n) - D\biggr
)=0.\end{aligned}$$ In other words, for non-zero $s \in (-\infty,0]$, solution occurs on the boundary. This completes the proof.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors wish to thank the associate editor and anonymous reviewers for their valuable comments which improved significantly the presentation of this paper.
[^1]: This work was financially supported by a medium size University of Cyprus grant entitled “DIMITRIS" and by European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement no. INFSO-ICT-223844. Part of this work was presented in 20$^{th}$ International Symposium on Mathematical Theory of Networks and Systems (MTNS ’12)[@stavrou-charalambous2012b].
[^2]: Charalambos D. Charalambous and Photios A. Stavrou are with the Department of Electrical and Computer Engineering (ECE), University of Cyprus, 75 Kallipoleos Avenue, P.O. Box 20537, Nicosia, 1678, Cyprus, e-mail: [email protected], [email protected]
[^3]: Nasir U. Ahmed is with the School of Electrical Engineering and Computer Science, University of Ottawa, 161 Louis Pasteur, P.O. Box 450, Stn A, Ottawa, Ontario, Canada, K1N 6N5, e-mail: [email protected]
[^4]: This point is explained in Subsection \[1.2\].
[^5]: The precise definition of a convolution of measures denoted by $\otimes$ is given in Section \[problem\_formulation\].
[^6]: i.e.$\Big| \int_{{\cal X}_{0,n}} \int_{{\cal Y}_{0,n}} \phi(x^n,y^n)
\overrightarrow{q}_{0,n}^{\alpha}(dy^n;x^n)\otimes\mu_{0,n}(dx^n)-\int_{{\cal X}_{0,n}}\int_{{\cal Y}_{0,n}} \phi(x^n,y^n)\overrightarrow{q}_{0,n}^{0}(dy^n;x^n)\otimes\mu_{0,n}(dx^n)\Big| \longrightarrow 0$ for any $\phi \in L_1(\mu_{0,n};BC({\cal Y}_{0,n}))$.
[^7]: Due to stationarity assumption $\nu_i(\cdot;\cdot)=\nu(\cdot;\cdot)$ and $q^*_i(\cdot;\cdot,\cdot)=q^*(\cdot;\cdot,\cdot)$, $\forall~i=0,1,\ldots,n$.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'microstates.bib'
---
=10000
[**Microstate Geometries from Gauged Supergravity**]{}\
[**in Three Dimensions**]{}
$^1$Université Paris Saclay, CNRS, CEA,\
Institut de Physique Théorique,\
91191, Gif sur Yvette, France\
$^2$Department of Physics and Astronomy
and $^3$Department of Mathematics,
University of Southern California,
Los Angeles, CA 90089, USA
[daniel.mayerson @ ipht.fr, walkerra @ usc.edu, warner @ usc.edu]{}\
<span style="font-variant:small-caps;">Abstract</span>
[3mm]{}[3mm]{} The most detailed constructions of microstate geometries, and particularly of superstrata, are done using $\mathcal{N} = (1,0)$ supergravity coupled to two anti-self-dual tensor multiplets in six dimensions. We show that an important sub-sector of this theory has a consistent truncation to a particular gauged supergravity in three dimensions. Our consistent truncation is closely related to those recently laid out by Samtleben and Sar[i]{}oğlu [@Samtleben:2019zrh], which enables us to develop complete uplift formulae from the three-dimensional theory to six dimensions. We also find a new family of multi-mode superstrata, indexed by two arbitrary holomorphic functions of one complex variable, that live within our consistent truncation and use this family to provide extensive tests of our consistent truncation. We discuss some of the future applications of having an intrinsically three-dimensional formulation of a significant class of microstate geometries.
Introduction {#sec:Intro}
============
The construction of BPS/supersymmetric microstate geometries in five and six dimensions is now a well-developed art [@Bena:2007kg; @Bena:2013dka; @Bena:2015bea; @Bena:2016ypk; @Bena:2017xbt; @Heidmann:2019xrd; @Warner:2019jll]. In particular, superstrata represent one of the broadest families of such geometries and have the advantage of a highly developed holographic dictionary [@Kanitscheider:2006zf; @Kanitscheider:2007wq; @Giusto:2004id; @Ford:2006yb; @Lunin:2012gp; @Giusto:2013bda; @Bena:2015bea; @Bena:2016ypk; @Bena:2017xbt; @Shigemori:2020yuo]. Superstrata are based on the D1-D5 system, whose underlying CFT is created by open strings stretched between the branes, and so the field theory has a world-volume along the common directions of the branes. The most general known families of superstrata are supersymmetric and encode a variety of left-moving excitations of the CFT. Encoding such momentum waves in the dual geometries means that they necessarily depend non-trivially on five of the six dimensions. The construction of these geometries is only possible because of the dramatic simplification afforded by the linear structure of the BPS equations and the decomposition of the solution into its “linear pieces” [@Bena:2011dd]. Once these pieces are reassembled into the complete geometry, the metric appears to be remarkably complex, as it must be to encode all the physical data of the underlying CFT states.
One of the remarkable features that has become evident in recent constructions of asymptotically-AdS superstrata [@Bena:2016ypk; @Bena:2017upb; @Heidmann:2019xrd] is that most of the interesting physics of superstrata is encoded in a three-dimensional space-time, $\cK$. Indeed, the six-dimensional space-time of a superstratum naturally decomposes into the $S^3$ surface around the branes, the radial coordinate, $r$, and the common directions, $(t,y)$ along the branes.[^1] The manifold, $\cK$, is the geometry described by the coordinates $(t, y, r)$ that are complementary to the $S^3$.
Since we are working with the holographic dual of a $(1+1)$-dimensional CFT, the geometry is asymptotic to AdS$_3$ $\times S^3$, and the vacuum is simply global AdS$_3$ $\times S^3$. Superstrata involve turning on new fluxes and adding metric deformations, thereby creating a warped, fibered product, $\cK \times S^3$. The manifold, $\cK$, is a smooth, horizonless, three-dimensional space-time and the $S^3$ is usually deformed and fibered over $\cK$ by non-trivial Kaluza-Klein Maxwell fields.
The manifold, $\cK$, is best described as a “smoothly-capped BTZ geometry.” That is, like BTZ, it is asymptotic to AdS$_3$ at infinity and has a long AdS$_2$ $\times S^1$ throat, but unlike BTZ, this throat has a finite depth because it caps off smoothly without a horizon. These geometries thus look much like the horizon region of a black hole, except that there there is a finite redshift between the cap and any point in the asymptotic region. It is these three-dimensional geometries that have provided the basis of many of the recent studies and comparison between microstate geometries and black holes [@Tyukov:2017uig; @Raju:2018xue; @Bena:2018mpb; @Bena:2019azk; @Heidmann:2019zws].
The analysis of such microstate geometries was greatly facilitated by the fact that, for some superstrata, the massless scalar wave equation in six dimensions is separable [@Bena:2017upb; @Walker:2019ntz; @Heidmann:2019zws], reducing to a simple Laplacian on a “round” $S^3$ and a far more complicated wave equation on $\cK$. For these geometries, the physics of massless scalar waves could indeed be entirely reduced to a problem on $\cK$. It was also conjectured, based on indirect evidence, in [@Bena:2017upb] that some superstrata should be part of a consistent truncation to a gauged supergravity in three dimensions.
The purpose of this paper is to prove this conjecture by showing that the six-dimensional gauged supergravity that is the “work-horse” of superstrata construction, does indeed have a consistent truncation down to a three-dimensional gauged supergravity. We will also give some explicit superstrata solutions that are entirely captured by this truncation.
Consistent truncations have a long history in supergravity and we will not review this here. There are the relatively trivial consistent truncations that are based on reducing a higher-dimensional supergravity on a manifold that has isometries and restricting fields to singlets of those isometries. This includes all the standard torus compactifications. There are also highly non-trivial consistent truncations that involve sphere compactifications in which one keeps higher-dimensional fields that depend (at linear order) on particular sets of “lowest harmonics” on the sphere. These fields therefore, typically, transform non-trivially under the rotation group of the sphere. The isometries of the sphere also give rise to a non-abelian gauge symmetry in the lower dimension. The end result is a compactification that reduces a sector of the higher-dimensional supergravity to gauged supergravity in lower dimensions. Here we will be concerned with $S^3$ compactifications of six-dimensional supergravity coupled to some tensor multiplets, and the corresponding three-dimensional gauged supergravity theory. We will also show the consistent truncation encodes some rich families of superstrata, some of which have been constructed elsewhere [@Bena:2017upb; @Heidmann:2019xrd].
We also construct new families of superstrata that depend on two freely-choosable holomorphic functions and that live entirely within our consistent truncation.
The important point about consistent truncations is that they are not merely lower-dimensional effective field theories. If one solves the lower-dimensional equations of motion in a consistent truncation, the result is an [*exact*]{} solution of the higher-dimensional equations of motion. This fact can be immensely useful in simplifying the equations of motion. In particular, the sphere becomes an “auxiliary” space whose dynamics is entirely determined by the lower-dimensional theory and encoded in the details of the consistent truncation. In this way, one can reduce a higher dimensional problem to a much more tractable lower-dimensional problem.
Consistent truncations can prove to be a ‘Faustian Bargain.’ The price of the simplification is a huge restriction on the degrees of freedom: the higher dimensional theory has vastly more degrees of freedom than the lower-dimensional theory and these extra degrees of freedom may prove essential to capturing the correct physics. The study of holographic $(2+1)$- and $(3+1)$-dimensional field theories is littered with examples in which consistent truncations have captured the essential physics as well as examples in which the consistent truncation has lacked the necessary resolution to produce the correct physics. We will discuss this further in Section \[sec:Conclusions\].
We have several reason for constructing the consistent truncations that are relevant to superstrata.
First, motivated by the success of such a strategy for holographic field theories in $(2+1)$ and $(3+1)$ dimensions, we wish to mine everything that three-dimensional supergravities have to tell us about holographic field theories in $(1+1)$-dimensions, and the corresponding supergravity solutions in in six dimensions. Again, the lower dimensional BPS equations are much simpler than the higher-dimensional BPS equations, since solutions are functions of 2 rather than 5 variables, and thus may yield extremely interesting new holographic flows. The three-dimensional formulation may also lead to a deeper understanding of the moduli space of superstrata and the microstates they represent. For example, we know, from perturbation theory [@Ceplak:2018pws; @Tyukov:2018ypq], that there are supersymmetric metric perturbations of superstrata. As yet, we do not know how to “integrate” these perturbations up to finite moduli and thereby create new families of superstrata. It is possible that the three-dimensional formulation will simplify a class of these moduli and show us how to do this more generally.
Above all, is the possibility of getting a handle on non-supersymmetric, non-BPS superstrata.
Given the intrinsic complexity of even the supersymmetric superstrata in six-dimensions, it seems an overwhelming task to address the non-linear equations that necessarily underlie the construction of non-BPS superstrata. Indeed, such generic non-BPS superstrata are expected to depend non-trivially on all six dimensions. However, the consistent truncation we present in this paper reduces this problem, for some limited families of superstrata, to a three-dimensional problem. Solving the equations of motion for the three-dimensional supergravity will still be a formidable task, and we intend to explore this in future work. The importance of the results presented here is that they transform an impossible six-dimensional problem into a feasible three-dimensional problem.
Paper overview {#paper-overview .unnumbered}
--------------
In Section \[Sect:3Dsugr\], we describe the class of three-dimensional gauged supergravity theories that can encode superstrata; a summary of the supergravity theory, fields, and action is given in Section \[sec:3Dsummary\]. The details of how this theory uplifts to six-dimensional supergravity may be found in Section \[Sect:uplift\]. Specifically, we show how the consistent truncation works: how the three-dimensional fields are encoded in the six-dimensional supergravity and how the solutions of the three-dimensional equations yield a solution to the six-dimensional equations. In Section \[Sect:superstrata\] we describe a new class of six-dimensional BPS superstrata (whose computational details may be found in Appendix \[sect:6Dsuperstrata\]) that fit within the consistent truncation described in Section \[Sect:3Dsugr\]. We reduce these six-dimensional solutions to their three-dimensional data and use them to test the details of the consistent truncation. The BPS superstrata that we have construct are intrinsically new in that such a multi-function family, while in similar spirit to those in [@Heidmann:2019xrd], have not been constructed before.
In Section \[sec:Conclusions\] we make some final remarks and return to the discussion of the applications of our results.
The three-dimensional gauged supergravity {#Sect:3Dsugr}
=========================================
In this section, we will discuss a specific three-dimensional gauged supergravity theory which is relevant for the dimensional reduction of 6D superstrata. The summary of our resulting 3D theory is given in Section \[sec:3Dsummary\].
Some supergravity background {#sec:background}
----------------------------
If one reduces IIB supergravity on $\IT^4$ , one obtains the $\cN = (2,2)$ theory in six dimensions. Reducing on a $K3$, instead, halves the supersymmetry to those that are holonomy invariant, and the result is an $\cN = (2,0)$ supergravity theory coupled to 21 anti-self-dual tensor multiplets.
More generally, the “parent theories” of interest here are six-dimensional $\cN = (2,0)$ supergravity (with sixteen supersymmetries) coupled to $n$ tensor multiplets. In such theories, the graviton multiplet contains one graviton, two complex, left-handed gravitinos (or four symplectic-Majorana Weyl gravitinos) and five self-dual, rank-two tensors gauge fields. Each tensor multiplet contains one anti-self-dual, rank-two tensor gauge fields, two right-handed complex spinors (or four symplectic-Majorana Weyl spinors) and five real scalars. The $\cR$-symmetry is $SO(5) \cong USp(4)$, the tensor gauge fields transform in the fundamental of $SO(5,n)$ and the scalars are described in terms of a coset: $$\frac{SO(5,n)}{SO(5) \times SO(n)} \,.
\label{coset1}$$ In the fully non-linear theory, the scalar matrix plays an essential role in a twisted duality condition on the tensor gauge fields. We will discuss a reduced version of this below.
This six-dimensional supergravity can then be compactified on AdS$_3$ $\times S^3$ using a ‘Freund-Rubin’ Ansatz in which one of the self-dual field strengths is set equal to the volume form of AdS$_3$ and of $S^3$. This corresponds to the D1-D5 background in which the supergravity charges, $Q_1$ and $Q_5$, are set equal. If one wants unequal charges one must move some flux into an anti-self-dual tensor gauge field. The simple, self-dual flux breaks the $SO(5,n)$ symmetry to $SO(4,n)$.
There is now an extensive literature [@Cvetic:2000dm; @Cvetic:2000zu; @Nicolai:2001ac; @Nicolai:2003bp; @Nicolai:2003ux; @Deger:2014ofa; @Samtleben:2019zrh] on how this compactification leads to $\Neql8$ ($16$ supersymmetries) gauged supergravity in three dimensions. The gauge group is $SO(4) \cong SO(3)_+ \times SO(3)_-$ and comes from the isometries of $S^3$; the scalar coset becomes[^2]: $$\frac{SO(8,3+n)}{SO(8) \times SO(3+n)} \,.
\label{coset2}$$ The gauge group sits inside $SO(8,n+3)$ as the diagonal $SO(4)$ in the first and third factors of the decomposition [@Nicolai:2003ux]: $$SO(4) \times SO(4) \times SO(4) \times SO(n-1) \subset ~SO(8,3+n) \,.
\label{groups}$$ In particular, the precise relationship between $\cN = (2,0)$ supergravity coupled to one anti-self-dual tensor multiplet in six dimensions and the three-dimensional $\Neql8$, $SO(8,4)$ supergravity was recently laid out in [@Samtleben:2019zrh].
The construction of superstrata usually takes place in the less supersymmetric, $\cN = (1,0)$ theories in six dimensions. In such theories, the graviton multiplet contains one graviton, one complex, left-handed gravitino (or two symplectic-Majorana Weyl gravitinos) and one self-dual, rank-two tensors gauge field. The tensor multiplet contains one anti-self-dual, rank-two tensor gauge fields, one right-handed complex spinor (or two symplectic-Majorana Weyl spinors) and one real scalar.
The simplest version of the theory used in superstratum construction can be characterized [@Bena:2015bea] as taking the bosonic fields to be those obtained by making a $\IT^4$ compactification of IIB supergravity and then further restricting to only the fields that transform trivially under the $SO(4)$ global rotations on the tangent space of the $\IT^4$. This results in $\cN = (1,0)$ supergravity coupled to two anti-self dual tensor multiplets. To be precise, the ten-dimensional RR field, $C^{(2)}$, descends to the self-dual tensor in the gravity multiplet and one of the two anti-self-dual tensors. These two components are independent and account for the separate D1 and D5 pieces. The other anti-self-dual tensor descends from the ten-dimensional Kalb-Ramond field, $B^{(2)}$, and anti-self-duality is required by supersymmetry. Roughly, the only way that this can be compatible with $\cN = (1,0)$ supersymmetry in a D1-D5 system is if the F1 and NS5 fields are locked together via anti-self-duality.
Imposing invariance under global rotations on $\IT^4$ reduces the $SO(5,n)$ coset in six-dimensions to $SO(1,n)$. The compactification to three dimensions on AdS$_3$ $\times S^3$ then results in the scalar coset $$\frac{SO(4,3+n)}{SO(4) \times SO(3+n)} \,.
\label{coset3}$$ The relationship between $\cN = (1,0)$ supergravity coupled to one anti-self-dual tensor multiplet in six dimensions and the three-dimensional $\Neql4$, $SO(4,4)$ supergravity has been laid out in [@Cvetic:2000dm; @Cvetic:2000zu; @Cvetic:2000ah; @Deger:2014ofa].
As we noted above, superstrata require $\cN = (1,0)$ supergravity coupled to at least two tensor multiplets. We are therefore going to examine the corresponding three-dimensional, $(0,2)$ supergravity (with eight supersymmetries) for which the scalar coset is $$\frac{SO(4,5)}{SO(4) \times SO(5)} \,.
\label{coset4}$$ As will become evident, we will find the results in [@Samtleben:2019zrh] for $\cN = (2,0)$ supergravity immensely useful in extending the results of [@Deger:2014ofa] to obtain the three-dimensional supergravity corresponding to $\cN = (1,0)$ supergravity coupled to two tensor multiplets.
The relevant supergravity in three dimensions is fully defined by its amount of supersymmetry, the scalar coset, the gauge symmetry and the gauge couplings as defined by an embedding tensor. The number of bosonic degrees of freedom in the theory is equal to the dimension of the underlying coset. However, these degrees of freedom can be encoded in various ways in the action. In three dimensions, Yang-Mills gauge fields can be dualized into scalars and vice versa and this is how the Yang-Mills fields can be generated. In addition, one of the essential features of the three-dimensional theories is the appearance of massive Chern-Simons vector fields. These fields can be viewed as gauging non-semi-simple groups and can ultimately be integrated out. Thus, as explained in [@Nicolai:2003bp], the number of bosonic degrees of freedom, $d$, is given by: $$d~=~ {\rm dim}({\rm Coset})~=~ \# ({\rm Scalars}) ~+~ \# ({\rm YM \ vectors}) ~+~ \# ({\rm massive \ CS \ vectors})\,.
\label{bosedofs}$$ For the three-dimensional $(0,2)$ supergravity theories described above, with coset (\[coset3\]), the gauge group is actually a semi-direct product $SO(4) \ltimes \IT^6 \subset SO(4,3+n)$, where $SO(4)$ is the standard Yang-Mills gauge group coming from $S^3$ and $\IT^6$ is a translation that transforms in the adjoint of $SO(4)$. Thus the $4\times (3+n)$ degrees of freedom become $6$ YM vectors, $6$ CS vectors and $4n$ scalars.
The scalar degrees of freedom {#sec:scalars}
-----------------------------
To describe this theory we simply follow the discussion in [@Deger:2014ofa] but with an extra tensor multiplet. The $SO(4,5)$ group has an invariant metric: $$\eta ~\equiv~
\left( \begin{matrix}
0_{4 \times 4} & \oneone_{4 \times 4} & 0 \\
\oneone_{4 \times 4} & 0_{4 \times 4} & 0 \\
0 &0& \varepsilon
\end{matrix} \right) \,,
\label{invmat}$$ where $\varepsilon = \pm 1$. Note that we are using the anti-diagonal form of $\eta$ because it is far more convenient in describing the degrees of freedom and in expressing the gauging. For $\varepsilon =-1$ we have $G= SO(5,4)$ (in the conventions of [@Samtleben:2019zrh]) and for $\varepsilon =1$ we have $G= SO(4,5)$. While this might seem a trivial notational distinction, it is related to the self-duality or anti-self-duality of the additional tensor multiplet. The theory of interest to us has $\varepsilon =1$, whereas (a truncated version of) the theory in [@Samtleben:2019zrh] corresponds to $\varepsilon =-1$.
The generators of $G$ may be written as: $$\left( \begin{matrix}
A & B & \chi_A \\
C & -A^T & \lambda^A \\
-\varepsilon \lambda^A &-\varepsilon \chi_A & 0
\end{matrix} \right) \,,$$ where $A,B,C,D$ are $4 \times 4$ matrices with $B^T = -B$ and $C^T = -C$. The matrix $A$ generates a $GL(4,\IR)$ whose compact generators define the $SO(4)$ YM gauge group and whose non-compact generators are obtained by taking $A^T=A$. The remaining $10$ non-compact generators can be taken to be $B$ and $\chi_A$. That is, we will choose to parametrize the coset by setting $A^T=A$, $C=0$ and $\lambda^A =0$. The matrix $B$ describes the translation generators of $\IT^6$ transforming in the adjoint of $SO(4)$. In this formulation, the $20$ bosonic degrees of freedom are defined by $A^T=A$, $B^T = -B$ and $\chi_A$.
The simplest way to fix the $\IT^6$ gauge invariance is to set $B=0$, which we will now do. We could also fix the $SO(4)$ gauge invariance by reducing $A$ to a diagonal matrix. The $20$ degrees of freedom would then be the $4$ eigenvalues, the $4$ $\chi_A$’s and $6+6$ gauge fields. However, we will only go half-way: fixing the $\IT^6$ gauge and moving these degrees of freedom into the CS vectors. We will preserve the $SO(4)$ gauge invariance.
Thus our scalar matrix will be defined by: $$\begin{aligned}
{{\cal V}_{\bar M}}^{\bar K} & ~=~
\exp \left( \begin{matrix}
0 & 0& \chi_A \\
0& 0 & 0 \\
0 &-\varepsilon \chi_A & 0
\end{matrix} \right) \, \left( \begin{matrix}
{P_A}^B & 0& 0 \\
0& {(P^{-1})_B}^A & 1 \\
0 &0& 0
\end{matrix} \right) \\
& ~=~
\left( \begin{matrix}
{P_A}^B & - \frac{1}{2}\varepsilon\, \chi_A \, \big({(P^{-1})_B}^C \chi_C\big) & \chi_A \\
0& {(P^{-1})_B}^A & 0 \\
0 &-\varepsilon{(P^{-1})_B}^C \chi_C & 1
\end{matrix} \right) \,,
\end{aligned}
\label{scalmat}$$ where $P =P^T$ is a symmetric $GL(4,\IR)$ matrix.
Our index conventions will be as follows. A vector of $G$ will be denoted by $$\cX_{\bar M} ~\equiv~ ( \cX_A, \cX^A, \cX_0) \,, \qquad X^{\bar M} ~\equiv~ ( \cX^A, \cX _A, \varepsilon \cX_0) \,,$$ where the indices are raised and lowered using (\[invmat\]). The components, $X_A$ and $X^A$, transform, respectively, in the $4$ and $\overline 4$ of $GL(4,\IR)$. That is, they transform through multiplication by $P$ or $P^{-1}$, respectively.
Following [@Samtleben:2019zrh], we define: $$\begin{aligned}
M_{\bar A \bar B} ~=~ & ({\cal V} {\cal V}^T)_{\bar A \bar B} ~=~ {{\cal V}_{\bar A}}^{\bar C} {{\cal V}_{\bar B}}^{\bar C} \,, \qquad
M^{\bar A \bar B} ~=~ \big(({\cal V}^T)^{-1} \, ({\cal V}^{-1} )\big)^{\bar A \bar B} ~=~\big({ {\cal V}^{-1}\big)_{\bar C}}^{\bar A} \big({ {\cal V}^{-1}\big)_{\bar C}}^{\bar B} \,,\\
m_{A B} ~=~ & (P \, P^T)_{AB} ~=~{P_A}^C \, {P_B}^C \,, \qquad m^{A B} ~=~ \big( (P^T)^{-1} P^{-1}\big)^{AB} ~=~{(P^{-1})_C}^A\, {(P^{-1})_C}^B \,.
\end{aligned}
\label{scalarmats}$$
The gauge couplings {#sec:gaugecouplings}
-------------------
The embedding of the gauge group, $SO(4) \ltimes \IT^6$ in $G$ is defined through the embedding tensor, $\Theta$. Specifically, if $T^{\bar M \bar N}=-T^{\bar N \bar M}$ are the generators of $G$ then the covariant derivative is defined by: $$\widehat\cD_\mu \, \cX_{\bar P} ~\equiv~ \partial_\mu \, \cX_{\bar P} ~+~ {A_\mu}^{\bar K \bar L} \, \Theta_{\bar K \bar L, \bar M \bar N } \, (T^{\bar M \bar N})_{\bar P}{}^{\bar Q} (\cX_{\bar Q} ) \,,
\label{covderiv1}$$ where ${A_\mu}^{\bar K \bar L}$ are the gauge connections and $(T^{\bar M \bar N})_{\bar P}{}^{\bar Q} (\cX_{\bar Q} )$ represents the action of $T^{\bar M \bar N}$ on vectors. In the standard normalization, one has: $$(T^{\bar M \bar N})_{\bar P}{}^{\bar Q} (\cX_{\bar Q} ) ~=~ \delta^{\bar N}_{\bar P} \cX^{\bar M} ~-~ \delta^{\bar M}_{\bar P} \cX^{\bar N} \,,$$ along with $$\begin{aligned}
\big[ \, T^{\bar K \bar L}\,, T^{\bar M \bar N} \,\big] ~=~ & f^{\bar K\bar L,\bar M\bar N}{}_{\bar P\bar Q}\, T^{\bar P \bar Q} \\
~=~ & \eta^{\bar L \bar M} T^{\bar K \bar N}~+~ \eta^{\bar L \bar N} T^{\bar M \bar K } ~-~ \eta^{\bar K \bar M} T^{\bar L \bar N}~-~ \eta^{\bar K \bar N} T^{\bar M \bar L } \,.
\end{aligned}
\label{comms}$$ This defines the structure constants: $$\begin{aligned}
f^{\bar K\bar L,\bar M\bar N}{}_{\bar P\bar Q} ~=~ &
\coeff{1}{2}\,\eta^{\bar L \bar M} \, \big( \delta^{\bar K}_{\bar P} \,\delta^{\bar N}_{\bar Q} - \delta^{\bar K}_{\bar Q} \,\delta^{\bar N}_{\bar P}\big) ~+~ \coeff{1}{2}\, \eta^{\bar L \bar N} \, \big( \delta^{\bar M}_{\bar P} \,\delta^{\bar K}_{\bar Q} - \delta^{\bar M}_{\bar Q} \,\delta^{\bar K}_{\bar P}\big) \\
& ~-~ \coeff{1}{2}\, \eta^{\bar K \bar M} \, \big( \delta^{\bar L}_{\bar P} \,\delta^{\bar N}_{\bar Q} - \delta^{\bar L}_{\bar Q} \,\delta^{\bar N}_{\bar P}\big)-~ \coeff{1}{2}\,\eta^{\bar K \bar N} \, \big( \delta^{\bar M}_{\bar P} \,\delta^{\bar L}_{\bar Q} - \delta^{\bar M}_{\bar Q} \,\delta^{\bar L}_{\bar P}\big) \,,
\end{aligned}
\label{structureconsts}$$ in which indices are summed without any weight factors[^3].
The generic form of the embedding tensor is: $$\Theta_{\bar K \bar L, \bar M \bar N } ~=~ \theta_{\bar K \bar L \bar M \bar N } ~+~ \coeff{1}{2}\, \big( \eta_{{\bar M}[{\bar K}}\, \theta_{{\bar L}]{\bar N}} -\eta_{{\bar N}[{\bar K}}\, \theta_{{\bar L}]{\bar M}} \big) ~+~ \theta \, \eta_{{\bar M}[{\bar K}}\, \eta_{{\bar L}]{\bar N}}\,,$$ where $ \theta_{[\bar K \bar L \bar M \bar N ]} = \theta_{\bar K \bar L \bar M \bar N } $ and $ \theta_{\bar K \bar L} = \theta_{\bar L \bar K } $. However, for the gauged $SO(4) \times \IT^6$ theory of interest here, the only non-vanishing pieces are [@Nicolai:2003ux; @Deger:2014ofa; @Samtleben:2019zrh]: $$\theta_{A B CD } ~=~ -2\, \alpha \, \epsilon_{A B CD } \,, \qquad {\theta_{A B C }}^D ~=~ \gamma_0 \, \epsilon_{A B C E} \, \delta^{DE} \,,
\label{EmbTens}$$ for some coupling constants $\alpha$ and $\gamma_0$. It is in this expression that the $GL(4,\IR)$ formulation arising from the choice (\[invmat\]) leads to significant simplification.
In particular, this embedding tensor reduces ${A_\mu}^{\bar K \bar L}$ to the twelve independent gauge fields for $SO(4) \ltimes \IT^6$: $${A_\mu}^{AB} ~=~ - {A_\mu}^{BA} \,, \qquad A_{\mu}{}^{B}{}_A ~=~ -{A_{\mu \,B}}^A \,.$$ It is convenient to define: $${\widetilde A_\mu}{}^{AB} ~\equiv~ \coeff{1}{2} \,\epsilon_{ABCD}\,{A_\mu}^{CD} \,, \qquad \ {\widehat A_\mu}{}^{AB} ~\equiv~ \coeff{1}{2} \,\epsilon_{ABCD}\,{{A_\mu}{}^C}{}_D \,,
\label{dualGFs}$$ and introduce: $${B_\mu}{}^{AB} ~\equiv~ 8\, \big(\alpha\, {\widetilde A_\mu}{}^{AB} - \gamma_0 \, {\widehat A_\mu}{}^{AB} \big) \,.
\label{Bvecden}$$ One then finds that (\[covderiv1\]) can be written in terms of $GL(4,\IR)$ components as: $$\begin{aligned}
\widehat\cD_\mu \, \cX_{A} &~=~ \partial_\mu \, \cX_{A} ~+~ {B_\mu}{}^{AB} \, \cX^{B} ~-~ 4\, \gamma_0\,\widetilde A_\mu{}^{AB} \, \cX_{B} \,, \\
\widehat\cD_\mu \, \cX^{A} &~=~ \partial_\mu \, \cX^{A} ~-~ 4\, \gamma_0\,\widetilde A_\mu{}^{AB} \, \cX^{B} \,, \qquad \widehat\cD_\mu \, \cX_0 ~=~ \partial_\mu \, \cX_0 \,.
\end{aligned}
\label{covderiv2}$$ Note that, in terms of the matrices of $G$, the connection ${B_\mu}{}^{AB}$ has the form: $$\cB_\mu ~\equiv~
\left( \begin{matrix}
0 & {B_\mu}{}^{AB} & 0\\
0 & 0 & 0 \\
0 &0 & 0
\end{matrix} \right) \,.$$ These are therefore precisely the gauge fields of $\IT^6$. The vector fields $A_\mu{}^{AB}$ are those of $SO(4)$ but they act with their duals, and with a gauge coupling of $-4 \gamma_0$.
To make this more explicit, define the $SO(3)_+ \times SO(3)_-$ parts of the gauge connection, and its dual: $$A_\mu{}^{AB} ~=~ A^+_\mu{}^{AB} ~+~ A^-_\mu{}^{AB} \,, \qquad \widetilde A_\mu{}^{AB} ~=~ A^+_\mu{}^{AB} ~-~ A^-_\mu{}^{AB} \,,
\label{decomp}$$ and define the gauge couplings $$g_+ ~=~ -g_- ~=~ -4 \gamma_0 \,.
\label{gpmdefn}$$ Then one has $$\begin{aligned}
\widehat\cD_\mu \, \cX_{A} &~=~ \partial_\mu \, \cX_{A} ~+~ {B_\mu}{}^{AB} \, \cX^{B} ~+~ g_+ A^+_\mu{}^{AB} \, \cX_{B} ~+~ g_- A^-_\mu{}^{AB} \, \cX_{B} \,, \\
\widehat\cD_\mu \, \cX^{A} &~=~ \partial_\mu \, \cX^{A} ~+~ g_+ A^+_\mu{}^{AB} \, \cX^{B} ~+~ g_- A^-_\mu{}^{AB} \, \cX^{B}\,, \qquad \widehat \cD_\mu \, \cX_0 ~=~ \partial_\mu \, \cX_0 \,.
\end{aligned}
\label{covderiv3}$$ Finally, it is convenient to define the reduced, purely-$SO(4)$, covariant derivatives: $$\begin{aligned}
\cD_\mu \, \cX_{A} &~=~ \partial_\mu \, \cX_{A} ~-~ 4\, \gamma_0\,\widetilde A_\mu{}^{AB} \, \cX_{B} \,, \\
\cD_\mu \, \cX^{A} &~=~ \partial_\mu \, \cX^{A} ~-~ 4\, \gamma_0\,\widetilde A_\mu{}^{AB} \, \cX^{B} \,, \qquad \cD_\mu \, \cX_0 ~=~ \partial_\mu \, \cX_0 \,.
\end{aligned}
\label{covderiv4}$$
The scalar action {#sec:scalaraction}
-----------------
From [@Samtleben:2019zrh], the scalar action is $${\cal L}_{\rm scalar} ~=~ \frac{1}{32}\, \big( {\cal D_\mu} M_{\bar K \bar L} \big) \, \big( {\cal D^\mu} M^{\bar K \bar L}\big) ~-~ V \,,$$ where the potential, $V$, is given by: $$\begin{aligned}
V ~=~& \frac{1}{48} \, \theta_{\bar K\bar L\bar M\bar N}\, \theta_{\bar P\bar Q\bar R\bar S} \, \Big(M^{\bar K\bar P} M^{\bar L\bar Q} M^{\bar M\bar R} M^{\bar N\bar S}
~-~ 6\, M^{\bar K\bar P} M^{\bar L\bar Q} \eta^{\bar M\bar R} \eta^{\bar N\bar S} \\
& \qquad\qquad\qquad \qquad\qquad ~+~ 8 \, M^{\bar K\bar P} \eta^{\bar L\bar Q} \eta^{\bar M\bar R} \eta^{\bar N\bar S}
~-~ 3 \, \eta^{\bar K\bar P} \eta^{\bar L\bar Q} \eta^{\bar M\bar R} \eta^{\bar N\bar S} \Big) \\
&~+~ \frac{1}{32} \, \theta_{\bar K\bar L} \theta_{\bar P\bar Q} \Big(2 M^{\bar K\bar P} M^{\bar L\bar Q} - 2 \eta^{\bar K\bar P} \eta^{\bar L\bar Q}
~-~ M^{\bar K\bar L} M^{\bar P\bar Q} \Big) ~+~ \theta \, \theta_{\bar K\bar L} M^{\bar K\bar L} ~-~8\, \theta^2
\end{aligned}$$ Using the expressions above, we find the following result: $$\begin{aligned}
{\cal L}_{\rm scalar} ~=~ & -\coeff{1}{16}\, {\rm Tr}\big[ \big( \cD_\mu m \big) m^{-1}\, \big( \cD^\mu m \big) m^{-1}\, \big] -\coeff{1}{8}\, m^{AB} \,(\cD_\mu \chi_A)\,(\cD^\mu \chi_B)\\
&-\coeff{1}{16}\, m^{AC} \,m^{BD}\,\Big({B^\mu}{}^{AB} - \coeff{1}{2}\, \varepsilon \,Y^\mu _{ AB} \Big) \,\Big({B_\mu}{}^{CD} - \coeff{1}{2} \,\varepsilon\, Y_{\mu \, CD} \Big) ~-~ V \,.
\end{aligned}$$ where $\cD_\mu$ is the $SO(4)$ covariant derivative defined in (\[covderiv4\]), and where $$Y_{\mu \, AB} ~\equiv~ \chi_B \,\cD_\mu \chi_A ~-~ \chi_A \,\cD_\mu\chi_B \,,$$ and $$\label{eq:3Dpotential}
V ~=~ \det\big(m^{AB}\big)\, \Big[\, 2\,\big(\alpha ~+~ \coeff{1}{4} \,\varepsilon \, \gamma_0 (\chi_A \chi_A)\big)^2 ~+~ \gamma_0^2 \, \big( m_{AB}\,\big( m_{AB} ~+~ \coeff{1}{2} \, \chi_A \chi_B \big)~-~ \coeff{1}{2}m_{AA}\, m_{BB} \big) \Big] \,.$$
The Chern-Simons action {#sec:CSaction}
-----------------------
Again, from [@Samtleben:2019zrh], the general Chern-Simons term is $$\cL_{CS} ~=~ \frac{1}{4}\, \varepsilon^{\mu \nu \rho} \, A_\mu{}^{\bar K\bar L} \,\Theta_{\bar K\bar L,\bar M\bar N}\,
\Big( \partial_\nu A_\rho {}^{\bar M\bar N} ~+~ \frac{1}{4}\, f^{\bar M\bar N,\bar P\bar Q}{}_{\bar R\bar S}\,\Theta_{\bar P\bar Q,\bar U\bar V}\,A_\mu{}^{\bar R\bar S} A_\rho{}^{\bar U\bar V}\Big)\,.
\label{CSgeneric}$$ Using the embedding tensor (\[EmbTens\]) and the structure constants (\[structureconsts\]), we find $$\cL_{CS} ~=~ -\varepsilon^{\mu \nu \rho} \, \Big[ \alpha\,\big(A_\mu{}^{AB}\, \partial_\nu \widetilde A_\rho{}^{BA} ~-~\coeff{8}{3}\, \gamma_0 \, A_\mu{}^{AB} \, A_\nu{}^{BC}\, A_\rho{}^{CA} \,\big) ~-~ \coeff{1}{8}\, {B_\mu}{}^{BA} \, F_{\nu \rho}^{AB} \Big] \,,
\label{CSterm}$$ where $$F_{\nu \rho}^{AB} ~\equiv~ 2\,\big( \partial_{[\nu} A_{\rho]}{}^{AB} ~-~4 \, \gamma_0\, A_{[\nu} {}^{AC} \,\widetilde A_{\rho]} {}^{CB}\big) \,.
\label{Maxdefn}$$ When written in terms of $SO(3)_+$ and $SO(3)_-$, we note that this action takes the more familiar Chern-Simons form: $$\begin{aligned}
\cL_{CS} ~=~ - \alpha\,\varepsilon^{\mu \nu \rho} \, \Big[ &\big(A^+_\mu{}^{AB}\, \partial^{}_\nu A^+_\rho{}^{BA} ~+~\coeff{2}{3}\, g_+ \, A^+_\mu{}^{AB} \, A^+_\nu{}^{BC}\, A^+_\rho{}^{CA} \,\big) \\
&- \big(A^-_\mu{}^{AB}\, \partial^{}_\nu A^-_\rho{}^{BA} ~+~\coeff{2}{3}\, g_- \, A^-_\mu{}^{AB} \, \widetilde A^-_\nu{}^{BC}\,\widetilde A^-_\rho{}^{CA} \,\big) \Big] ~+~ \coeff{1}{8}\, \varepsilon^{\mu \nu \rho} \,{B_\mu}{}^{BA} \, F_{\nu \rho}^{AB} \,,
\end{aligned}
\label{CStermpieces}$$ with $$F_{\nu \rho}^{AB} ~=~ F^+{}_{\nu \rho}^{AB} ~+~ F^-{}_{\nu \rho}^{AB} \,, \qquad F^\pm{}_{\nu \rho}^{AB} ~\equiv~ 2\,\big( \partial^{}_{[\nu} A_{\rho]}^\pm{}^{AB} ~+~g_\pm\, A^\pm_{[\nu} {}^{AC } \, A^\pm_{\rho]} {}^{CB}\big) \,.
\label{SO3Fs}$$
Integrating out the Chern-Simons gauge fields {#sec:CSintout}
---------------------------------------------
The Chern-Simons gauge fields, ${B_\mu}{}^{AB}$, appear only quadratically in the action, and without derivatives. It is therefore trivial to integrate them out by completing the square. The complete bosonic action may be written $$\begin{aligned}
{\cal L} ~=~ &\coeff{1}{4}\,R ~-~\coeff{1}{16}\, {\rm Tr}\big[ \big( \cD_\mu m \big) m^{-1}\, \big( \cD^\mu m \big) m^{-1}\, \big] -\coeff{1}{8}\, m^{AB} \,(\cD_\mu \chi_A)\,(\cD^\mu \chi_B) ~-~V \\
&- \coeff{1}{8} m_{AC} \,m_{BD}\, F_{\mu \nu}^{AB}\, F^{\mu \nu}{}^{CD} ~-~ \alpha\,\varepsilon^{\mu \nu \rho} \, \big(A_\mu{}^{AB}\, \partial_\nu \widetilde A_\rho{}^{BA} - \coeff{8}{3}\, \gamma_0 \, A_\mu{}^{AB} \, A_\nu{}^{BC}\, A_\rho{}^{CA} \,\big) \\ & ~-~\coeff{1}{16}\, \varepsilon \, \varepsilon^{\mu \nu \rho} \, Y_\mu{}_{AB} \, F_{\nu \rho}^{AB} ~+~ {\cal L}_{\rm B} \,,
\end{aligned}
\label{action1}$$ where $$\begin{aligned}
{\cal L}_{\rm B} ~\equiv~ & -\coeff{1}{16}\, g_{\mu \nu} \, m^{AC} \,m^{BD}\,\Big({B^\mu}{}^{AB} + \varepsilon^{\mu \rho_1 \rho_2} \, m_{AE_1} m_{BE_2}\,F_{\rho_1 \rho_2}^{E_1E_2} - \coeff{1}{2}\, \varepsilon \,Y^\mu{}_{ \, AB} \Big) \\
&\qquad\qquad\qquad\qquad\qquad \times \Big({B^\nu}{}^{CD} + \varepsilon^{\nu \rho_3 \rho_4}\, m_{CE_3} \,m_{DE_4}\,F_{\rho_3 \rho_4}^{ E_3 E_4} - \coeff{1}{2} \,\varepsilon\, Y^\nu{}_{ CD} \Big) \,.
\end{aligned}
\label{Baction}$$ Thus the equations of motion for ${B_\mu}{}^{AB}$ are trivial, and yield $${B^\mu}{}^{AB} ~=~ \coeff{1}{2}\, \varepsilon \,Y^\mu{}_{ \, AB} ~-~ \varepsilon^{\mu \nu \rho} \, m_{AC} m_{BD}\,F_{\nu \rho}^{CD} \,,
\label{Beqn}$$ and these gauge fields drop out of the action entirely.
The three-dimensional supergravity: summary and comments {#sec:3Dsummary}
--------------------------------------------------------
The three-dimensional supergravity theory we consider is a $(0,2)$ gauged supergravity with eight supersymmetries. It has an $SO(4)$ gauge symmetry, with $6$ gauge fields, $A_\mu^{AB} = A_\mu^{[AB]}$, where the indices $A, B, \ldots =1,2,3,4$ transform in the vector of $SO(4)$. In addition to the graviton, and the gauge fields, $A_\mu^{AB}$, there are $14$ scalar fields in the bosonic sector. Four scalars are encoded in an $SO(4)$ vector, $\chi_A$, and the other ten are encoded as a general, symmetric $GL(4, \IR)$ matrix, $m_{AB}=m_{(AB)}$, with inverse, $m^{AB}$.
The bosonic action is $$\begin{aligned}
{\cal L} ~=~ &\coeff{1}{4}\,R ~-~\coeff{1}{16}\, {\rm Tr}\big[ \big( \cD_\mu m \big) m^{-1}\, \big( \cD^\mu m \big) m^{-1}\, \big] -\coeff{1}{8}\, m^{AB} \,(\cD_\mu \chi_A)\,(\cD^\mu \chi_B) ~-~V \\
&- \coeff{1}{8} m_{AC} \,m_{BD}\, F_{\mu \nu}^{AB}\, F^{\mu \nu}{}^{CD} ~-~ \alpha\,\varepsilon^{\mu \nu \rho} \, \big(A_\mu{}^{AB}\, \partial_\nu \widetilde A_\rho{}^{BA} - \coeff{8}{3}\, \gamma_0 \, A_\mu{}^{AB} \, A_\nu{}^{BC}\, A_\rho{}^{CA} \,\big) \\ & ~-~\coeff{1}{16}\, \varepsilon \, \varepsilon^{\mu \nu \rho} \, Y_\mu{}_{AB} \, F_{\nu \rho}^{AB} \,,
\end{aligned}
\label{action-final}$$ where $m$ denotes $m_{AB}$, and $$Y_{\mu \, AB} ~\equiv~ \chi_B \,\cD_\mu \chi_A ~-~ \chi_A \,\cD_\mu\chi_B \,.
\label{Ydefn}$$ The covariant derivative is defined on upper and lower $SO(4)$ indices as $$\cD_\mu \, \cX_{A} ~=~ \partial_\mu \, \cX_{A} ~-~ 4\, \gamma_0\,\widetilde A_\mu{}^{AB} \, \cX_{B} \,, \qquad
\cD_\mu \, \cX^{A} ~=~ \partial_\mu \, \cX^{A} ~-~ 4\, \gamma_0\,\widetilde A_\mu{}^{AB} \, \cX^{B} \,,
\label{covderiv-final}$$ where $${\widetilde A_\mu}{}^{AB} ~\equiv~ \coeff{1}{2} \,\epsilon_{ABCD}\,{A_\mu}^{CD} \,.
\label{dualGFsfinal}$$ The field strengths are given by $$F_{\nu \rho}^{AB} ~\equiv~ 2\,\big( \partial_{[\nu} A_{\rho]}{}^{AB} ~-~4 \, \gamma_0\, A_{[\nu} {}^{AC} \,\widetilde A_{\rho]} {}^{CB}\big) \,.
\label{Maxdefn-final}$$ The scalar potential is: $$V ~=~ \det\big(m^{AB}\big)\, \Big[\, 2\,\big(\alpha ~+~ \coeff{1}{4} \,\varepsilon \, \gamma_0 (\chi_A \chi_A)\big)^2 ~+~ \gamma_0^2 \, \big( m_{AB}\,\big( m_{AB} ~+~ \coeff{1}{2} \, \chi_A \chi_B \big)~-~ \coeff{1}{2}m_{AA}\, m_{BB} \big) \Big] \,.
\label{eq:3Dpotential-final}$$ The equations of motion following from (\[action-final\]) are given explicitly in Appendix \[App:3Deom\].
The bosonic theory has three parameters: the gauge coupling, $\gamma_0$, a scale parameter, $\alpha$, and a “signature,” $\varepsilon = \pm1$. In Appendix \[app:rescalings\], we discuss two scale invariances of the action that can be used to set $|\alpha| =1$ and $\gamma_0=1$. We will, however, retain these parameters.
The parameter $\varepsilon$, and the sign of $\alpha$ are extremely important to the supersymmetry. We will discuss this further in the six-dimensional context in Section \[sec:Establishing\]. Here we will simply note that sending $\varepsilon \to - \varepsilon, \alpha \to - \alpha$ leaves the potential invariant. In the rest of the action, the parameters $\varepsilon$ and $\alpha$ only appear as coefficients of the parity odd terms involving $\varepsilon^{\mu \nu \rho}$. Thus $\varepsilon \to - \varepsilon, \alpha \to - \alpha$, combined with an orientation reversal in the three-dimensional space-time is a symmetry of the action.
From three to six dimensions {#Sect:uplift}
============================
Here, we will describe how the 3D theory of Section \[Sect:3Dsugr\] can be obtained from a dimensional reduction from 6D. First, in Section \[sec:Establishing\], we discuss how we obtained our consistent truncation formulae. In Section \[sec:6Dtheorysub\], we describe the 6D theory at hand, which is the theory relevant for the description of superstrata. In Section \[sec:full6Duplift\], we give the full non-linear reduction ansatz from this 6D theory to the 3D theory of Section \[Sect:3Dsugr\]. Finally, in Section \[sec:U1sqtrunc\], we describe a useful $U(1)^2$ subsector of this truncation and its 3D counterpart.
Establishing the consistent truncation {#sec:Establishing}
--------------------------------------
The common core of superstrata and the consistent truncations of [@Cvetic:2000dm; @Cvetic:2000zu; @Cvetic:2000ah; @Deger:2014ofa; @Samtleben:2019zrh] is the basic $\mathcal{N}=(1,0)$ supergravity coupled to single anti-self-dual tensor multiplet. This has an $SO(1,1)$ scalar manifold. The $\mathcal{N}=(2,0)$ supergravity theory considered in [@Samtleben:2019zrh] involves adding four more self-dual tensors to complete the $(2,0)$ graviton multiplet, as well as adding scalars to extend the anti-self-dual tensor multiplet to $(2,0)$ supersymmetry. This extends the $SO(1,1)$ to $SO(4+1,1)$. Superstrata go in the opposite sense, in that one adds $(1,0)$ anti-self-dual tensor multiplets, which extend the $SO(1,1)$ to $SO(1,1+n)$, where $n$ is the number of added anti-self-dual tensor multiplets.
From the three-dimensional perspective, these two extensions of the six-dimensional theory involves extending the $SO(4,4)$ of the consistent truncation of the basic theory [@Cvetic:2000dm; @Cvetic:2000zu; @Cvetic:2000ah; @Deger:2014ofa] to $SO(4+4,4)$ or to $SO(4,4+n)$. The parameter, $\varepsilon$ thus directly encodes whether we are adding self-dual or anti-self-dual tensors to the basic theory.
As we noted earlier, flipping sign of $\varepsilon$ and $\alpha$ along with a change of orientation, leaves the three-dimensional bosonic equations of motion unchanged in three dimensions. The same is also true for the six-dimensional equations of motion: such an orientation flip on the three-dimensional base only changes the duality conditions (\[eq:6Dselfduality\]) in the six-dimensional bosonic equations of motion. Thus, even in the six-dimensional theory, an $SO(p,q)$ theory with $p$ self-dual and $q$ anti-self-dual multiplets, and the theory with $p$ and $q$ interchanged have exactly the same bosonic equations of motion (modulo orientations).
In six-dimensional supergravity theories, there is a correlation between the chirality of the supersymmetry and duality of the tensor gauge fields that belong to matter multiplets, or to the graviton multiplet. The convention that is used for superstrata, and is used in [@Samtleben:2019zrh], is that the self-dual tensors belong to the graviton multiplet and anti-self-dual tensors belong to matter multiplets. If one performed such an orientation flip on the $(2,0)$ theory, it would break the supersymmetry (to $(1,0)$) unless one also flipped the helicity of the spinors. However, this is irrelevant to our goals here. We are focussed on the consistent truncation of the equations of motion. We will return to the supersymmetry in future work.
Here, the important point is that the extensive work on consistent truncations to three dimensions that culminated in [@Samtleben:2019zrh], shows that solutions to the three-dimensional equations of motion for the $SO(4+4,4)$ theory, necessarily provide solutions to the equations of motion to the six-dimensional theory with five self-dual tensors multiplets and one anti-self-dual tensor multiplet. A trivial orientation flip, means that this result maps onto the $SO(4,4+4)$ theory in three dimensions and to a six-dimensional theory with one self-dual tensor multiplet and five anti-self-dual tensor multiplets. In particular, this theory can be truncated to the $SO(4,4+1)$ theory and to the six-dimensional theory with one self-dual tensor multiplet two anti-self dual tensor multiplets[^4].
Thus the existence of the consistent truncation we seek is already guaranteed by the results of [@Hohm:2017wtr; @Samtleben:2019zrh]. What remains is to adapt the uplift formulae of [@Samtleben:2019zrh] to the theory of interest to us. We will also subject our truncation and uplift formula to extensive and rigorous testing.
The six-dimensional theory {#sec:6Dtheorysub}
--------------------------
The general $\mathcal{N}=(1,0)$ six-dimensional supergravity theory coupled to an arbitrary number of tensor multiplets is discussed at length in [@Ferrara:1997gh; @Riccioni:2001bg]. We will consider the $\mathcal{N}=(1,0)$ supergravity multiplet coupled to two anti-self-dual tensor multiplets, as this is the relevant sector that captures the D1-D5-P solutions when reducing from ten-dimensional type IIB theory on a $\IT^4$ [@Giusto:2013rxa] (see also [@Duff:1998cr; @Lavrinenko:1998hf]; or Appendix B of [@deLange:2015gca] for a quick summary).
### The six-dimensional theory for superstrata {#sec:6Dtheory}
The six-dimensional bosonic field content we consider consists of the metric $g_{\hmu\hnu}$, two scalars $\varphi,X$, and three three-forms $G^\hI$, $\hI=1,2,4$.[^5] The theory has a $SO(1,2)$ global symmetry, where the two scalars parametrize a $SO(1,2)/SO(2)$ coset. The three-forms satisfy a self-duality relation: \[eq:6Dselfduality\] G\^I = G\^\_ dx\^dx\^dx\^= G\^,which serves as their equations of motion, together with their Bianchi identities: \[eq:6Dbianchi\] dG\^= 0.We have parametrized the scalar self-duality matrix $\mathcal{M}$ as: = e\^(
[ccc]{} X\^2 & 8 & -2 X\
18e\^[-2]{} (2 + e\^ X\^2)\^2 & X\^2 & -(2e\^[-]{}+X\^2)\
+(2e\^[-]{}+X\^2) & 4X & -2e\^[-]{} -2 X\^2
), \[Mdef\] Note that we use the following conventions for the $SO(1,2)$ metric which is used to raise or lower indices: \^ = (
[ccc]{} 0 & 1 & 0\
1 & 0 & 0\
0 & 0 & -2
). \[etaDef\] The other bosonic equations of motion can be obtained by varying the pseudo-Lagrangian [@Ferrara:1997gh; @Riccioni:2001bg]: \[eq:6Dlagr\] \_[6D]{} = R -12 (\_)\^2 -12 e\^(\_X)\^2 - 16 \_GG\^.Note that the scalar matrix $\mathcal{M}_{\hI\hJ}$ (with both indices down) is symmetric.
One should also note that the matrix $\tensor{\cM}{^\hI_\hJ}$ has one positive eigenvalue and two negative eigenvalues. It then follows from (\[eq:6Dselfduality\]) that for superstrata (with two anti-self dual tensors) one should take $\varepsilon=+1$, while for $\varepsilon=-1$, the theory has two self-dual tensors and hence the uplift formulae should reduce to a truncation of that given in [@Samtleben:2019zrh].[^6]
The full six-dimensional uplift {#sec:full6Duplift}
-------------------------------
Here, we give the full ansatz for the non-linear KK reduction of the six-dimensional theory (\[eq:6Dlagr\]) on an $S^3$, which gives the three-dimensional gauged supergravity discussed in Section \[Sect:3Dsugr\], with three-dimensional metric $g_{\mu\nu}$ (coordinates $x^\mu$), 14 scalars which consist of the four scalars $\chi_A$ and the 10 scalars parametrizing the symmetric matrix $m_{AB}$, and the six three-dimensional gauge fields parametrized by the antisymmetric $\tensor{\widetilde{A}}{_{\mu}^{AB}}$. This reduction ansatz follows from a simple adjustment of the ansatz considered in [@Samtleben:2019zrh] for six-dimensional $\mathcal{N}=(2,0)$ supergravity.
### The metric and scalars
The six-dimensional metric ansatz is: \[eq:6Dmetricansatz\] ds\_6\^2 = (m\_[AB]{})\^[-1/2]{} \^[1/2]{} ds\_3\^2 + g\_0\^[-2]{}(m\_[AB]{})\^[1/2]{} \^[-1/2]{} m\^[AB]{} \^A \^B, where we have made the convenient re-definition of the three-dimensional gauge coupling: \[eq:g0gamma0\] g\_0 = 2 \_0.We have also defined: = m\_[AB]{}\^A \^B. \[eq:DeltaDef\] The four Cartesian coordinates, $\mu^A$, on $\IR^4$ are required satisfy $\mu^A\mu^A=1$, so as to define a unit $S^3$. Their gauge-covariant derivatives are $\cD\mu^A = d\mu^A -2g_0 \tilde{A}^{AB}\mu^B$; see Appendix \[sec:app:S3\] for more details. Note that the metric ansatz only depends on the scalar matrix $m_{AB}$ (and its inverse $m^{AB}$) and not on the scalars $\chi_A$. The six-dimensional scalars are given by the simple expressions: \[eq:6Dscalaransatz\] e\^[-]{} = , X = \_A \^A .
### The tensor gauge fields
The expressions for the three-forms are quite unwieldy. It is easiest to give the two-form potentials $B^\hI$, related to the three-forms in the usual way: G\^= dB\^.The three-forms $G^\hI$ and its two-form potentials $B^\hI$ can be decomposed as:[^7] $$\begin{aligned}
G^\hI &= \frac{1}{3!} G^{\hI}_{ijk}\cD y^i\wedge \cD y^j\wedge \cD y^k + \frac12 G^\hI_{ij\mu}\cD y^i\wedge \cD y^j\wedge dx^\mu\\
\nonumber & + \frac12 G^\hI_{i\mu\nu}\cD y^i\wedge dx^\mu\wedge dx^\nu + \frac{1}{3!} G^\hI_{\mu\nu\rho}dx^\mu\wedge dx^\nu\wedge dx^\rho,\\
B^\hI &= \frac12 B^\hI_{ij} \cD y^i\wedge \cD y^j + B^\hI_{i\mu} \cD y^i\wedge dx^\mu + \frac12 B^\hI_{\mu\nu} dx^\mu\wedge dx^\nu .\end{aligned}$$
We will only give expressions for $B^\hI_{ij}$ and $B^\hI_{i\mu}$, which unambiguously determine the components $G^\hI_{ijk},G^\hI_{ij\mu}$ of the three-forms; the other components $G^\hI_{i\mu\nu},G^\hI_{\mu\nu\rho}$ (and thus also, by integration, $B^\hI_{\mu\nu}$) are then determined by the self-duality relation (\[eq:6Dselfduality\]). The ansatze for $B^\hI_{ij}$ is (using the round sphere quantities defined in Appendix \[sec:app:S3\]): $$\begin{aligned}
\label{eq:6DansatzB1ij} B^1_{ij} &= \left(-\frac{1}{g_0^2}\right)\left(-2\mathring{\omega}_{ijk}\mathring{\zeta}^k + \frac12\mathring{\omega}_{ijk}\mathring{g}^{kl}\Delta\partial_l \left[\Delta^{-1}\right] \right), \\
\label{eq:6DansatzB2ij} B^2_{ij} &= \left(-\frac{1}{4g_0^2}\right)\left(4\, \varepsilon\, g_0^{-1}\alpha\,\mathring{\omega}_{ijk}\mathring{\zeta}^k + \frac14\mathring{\omega}_{ijk}\mathring{g}^{kl}\Delta\partial_l \left[\Delta^{-1}X^2\right] \right),\\
\label{eq:6DansatzB4ij} B^4_{ij} &=\left(-\frac{\sqrt{2}}{g_0^2}\right) \frac12 \mathring{\omega}_{ijk}\mathring{g}^{kl}\Delta^{1/2}\partial_l \left(\Delta^{-1/2} X \right) .\end{aligned}$$ while the ansatze for the components $B^\hI_{i\mu}$ is: $$\begin{aligned}
\label{eq:6DansatzB1imu} B^1_{i\mu} &= \left(-\frac{1}{g_0^2}\right)\partial_i\mu^A \left[2g_0\right] A^{AB}_\mu\left(\mu^B - 2\mathring{\zeta}^k\partial_k\mu^B\right), \\
\label{eq:6DansatzB2imu} B^2_{i\mu} &= \left(-\frac{1}{4g_0^2}\right)\partial_i\mu^A\left[2g_0\right]\varepsilon \left( -\left[ \tensor{A}{_{\mu}^{A}_{B}} -2g_0^{-1}\alpha A_\mu^{AB}\right]\mu^B - 2g_0^{-1}\alpha A^{AB}_\mu\left[\mu^B- 2\mathring{\zeta}^k\partial_k\mu^B\right]\right),\\
\label{eq:6DansatzB4imu} B^4_{i\mu} &= 0 .\end{aligned}$$ In (\[eq:6DansatzB2imu\]), the auxiliary gauge field $A{\indices{^A_B}}$ features, but can be integrated out in favor of the fundamental fields in three-dimensions using (\[Beqn\]), (\[Bvecden\]), and (\[dualGFsfinal\]): g\_0 -2 A\^[AB]{}\_=- \_[ABCD]{}Y\^[CD]{}\_+ \_[ABCD]{} m\_[CC’]{}m\_[DD’]{} F\^[C’D’]{}\_.In Section \[sec:U1sqtrunc\], we will give explicit formulae for the entire three-forms $G^\hI$ in a specific sub-sector relevant for the $(1,0,n)$ superstrata.
The complete reduction ansatz is thus given by the metric ansatz (\[eq:6Dmetricansatz\]), the scalar ansatz (\[eq:6Dscalaransatz\]), and the two-form potential ansatze (\[eq:6DansatzB1ij\])-(\[eq:6DansatzB4imu\]). Note that there are two constant parameters $g_0$ and $\alpha$ in the uplift; these (or more precisely, their absolute value) can essentially be chosen at will, as there are two rescalings that one can perform on any six-dimensional solution which correspond to rescaling a three-dimensional theory; we discuss these in Appendix \[app:rescalings\]. For example, as we will do in Section \[subsect:(1,m,n) 3D data\], a natural choice would be to choose $g_0$ such that $m_{AB}=\oneone$ on an asymptotic $AdS_3$ boundary; then $g_0^{-1}$ is identified with the (asymptotic) $S^3$ radius in the six-dimensional solution. The sign of $\alpha$ can be changed by changing $\varepsilon$, as discussed in Section \[sec:3Dsummary\].
Our reduction presented here is a simple modification and extension of the $\mathcal{N}=(2,0)$ $S^3$ reduction ansatz in [@Samtleben:2019zrh]; to match their reduction ansatz, we need to take $\varepsilon = -1$ and the $SO(4)$ vector of three-forms $G^\alpha$ in [@Samtleben:2019zrh] truncates to our three-form $G^4$ as $G^\alpha \sim \delta_{\alpha,1}G^4$, and accordingly for the three-dimensional scalars. Appendix \[sec:app:Henning\] contains the explicit matching of our results to those of in [@Samtleben:2019zrh]. (This matching also involves some minor corrections to the uplift formulae presented [@Samtleben:2019zrh].)
### Testing the consistent truncation {#ss:testing}
As we noted earlier, the consistent truncation we are interested in is closely related to that of the $\mathcal{N}=(2,0)$ theory reduced on $S^3$. Indeed, our observations in Section \[sec:Establishing\] mean that our consistent truncation is essentially guaranteed. However, we still need to establish our uplift formulae and ensure that we have all the details correct. Our tests will also provide extensive and rigorous testing of the entire consistent truncation more broadly.
The first test will be to reduce the theory to a $U(1)^2$ truncation. Specifically, by imposing that the fields are invariant under a particular $U(1)$, we truncate the theory from an $SO(4)$ gauged theory to as $U(1)^2$ gauge theory. This is presented in Section \[sec:U1sqtrunc\]. For this reduced system we explicitly checked that the three-dimensional equation of motion (following from the truncated three-dimensional Lagrangian (\[eq:lagr3DU1sq\])), together with our uplift ansatz to six-dimensions, imply the six-dimensional three-form Bianchi identities (\[eq:6Dbianchi\]) and self-duality relations (\[eq:6Dselfduality\]), as well as the six-dimensional scalar equations of motion (i.e. the equations of motion for $X,\varphi$ coming from (\[eq:6Dlagr\])).
The second test involved constructing a new family of six-dimensional solutions: the $(1,m,n)$ superstrata, which depend on two independent, arbitrary holomorphic functions of one variable. We present these new six-dimensional solutions in Section \[Sect:superstrata\] and show that they precisely conform to our reduction ansatz. We then extract the three-dimensional data in Section \[subsect:(1,m,n) 3D data\] and use this as a detailed test of the three-dimensional equations of motion.
Finally, in Section \[subsect:(1,0,n) 3D data\] we examine the overlap of our two tests by looking at the six-dimensional $(1,0,n)$ superstrata considered as part of the $U(1)^2$ truncation of Section \[sec:U1sqtrunc\].
Needless to say, our uplift formulae, and the three-dimensional action, pass all of these tests. More to the point, these tests provide multiple, independent cross checks of all the functional forms and their coefficients in all of our uplift formulae. In particular, in the uplift we have thoroughly tested all the signs and numerical factors, as well as the appearances of $\alpha$, $g_0=2\gamma_0$ and $\varepsilon$, which correspond precisely to the parameters of the three-dimensional theory.
U(1)\^2 truncation {#sec:U1sqtrunc}
------------------
In this section, we focus on a consistent truncation of the general reduction given in Section \[sec:full6Duplift\]. This truncation is most simply defined by restricting to the fields that are invariant under the $O(2) = U(1)$ subgroup of the $SO(4)$ gauge group that rotates the gauge indices $A=3,4$ into each other. This truncation is the minimal one in which the $(1,0,n)$ superstrata sit, and provides an explicit, more approachable example of the complicated reduction formulae of Section \[sec:full6Duplift\]. There is an analogous truncation that restricts to the sector that is invariant under rotations that map the gauge indices $A=1,2$ into each other; this is the minimal truncation that contains the $(1,1,n)$ superstrata. Here we will focus on the first truncation.
We will use the explicit coordinates $(\theta,\varphi_1,\varphi_2)$ on the $S^3$, see (\[eq:appS3:explicitmu\]). The $O(2)$ invariant gauge fields are simply the $U(1)^2$ Cartan sub-sector: $$\begin{aligned}
A{\indices{_\mu^{AB}}} &= -\frac12 \left(\begin{array}{cc} \mathcal{A}^{\varphi_2}_{\mu} & 0_{2 \times 2}\\ 0_{2\times 2} & \mathcal{A}^{\varphi_1}_{\mu}\end{array}\right), & \mathcal{A}^{\varphi_i}_{\mu}&= \left(\begin{array}{cc} 0 & A^{\varphi_i}_{\mu}\\ -A^{\varphi_i}_{\mu} & 0 \end{array}\right),\end{aligned}$$ which have been parametrized such that the resulting gauge-covariant coordinates are given by (see (\[eq:appS3:Dmu\]) or (\[eq:appS3:Dy\])): = d, \_i = d\_i + g\_0A\^[\_i]{} .
Invariance under $O(2)$ means that we keep $\chi_1,\chi_2$ but set: \_3 = \_4 = 0. \[chiRestriction\]It also truncates the 10 scalars of $m_{AB}$ to 4 scalars $\xi_1,\xi_2,\xi_3,\xi_4$ defined as follows: m\_[AB]{} = (
[cc]{} e\^[-\_2]{} R & 0\_[2 2]{}\
0\_[2 2]{} & e\^[-\_1]{}\_[2 2]{}
) , R = (\_3 (
[cc]{} \_4 & \_4\
\_4 & -\_4
)).Note that the gauge covariant derivatives on the scalars are given by: $$\begin{aligned}
\cD_\mu\xi_{1,2,3} &= \partial_\mu \xi_{1,2,3}, & \cD_\mu\xi_4 &= \partial_\mu\xi_4 +2g_0 A_\mu^{\varphi_1} \\
\cD_\mu\chi_1 &= \partial_\mu \chi_1 +g_0 A^{\varphi_1}_\mu \chi_2, & \cD_\mu\chi_2 &= \partial_\mu \chi_2 -g_0 A^{\varphi_1}_\mu \chi_1. \end{aligned}$$
Thus, the three-dimensional fields in this truncation are the metric $g_{\mu\nu}$, the two (Abelian) gauge fields $A^{\varphi_i}_\mu$, and the six scalars $\xi_1,\xi_2,\xi_3,\xi_4,\chi_1,\chi_2$. The three-dimensional Lagrangian is given by the appropriate truncation of the full three-dimensional Lagrangian (\[action-final\]) and can be written explicitly as: $$\begin{aligned}
\label{eq:lagr3DU1sq} 4\mathcal{L}_{3D,U(1)^2} &= R - \frac12 (\partial_\mu \xi_1)^2 - \frac12 (\partial_\mu \xi_2)^2 - \frac12 (\partial_\mu \xi_3)^2 - \frac12 \sinh^2\xi_3 ( \cD_\mu\xi_4)^2\\
\nonumber & - \frac14 e^{-2\xi_1} F_{\mu\nu}^{\varphi_1} F^{\varphi_1,\mu\nu} - \frac14 e^{-2\xi_2} F_{\mu\nu}^{\varphi_2} F^{\varphi_2,\mu\nu}- \frac12e^{\xi_2} \left( \cosh \xi_3\left[(\cD_\mu\chi_1)^2+(\cD_\mu\chi_2)^2\right]\right.\\
\nonumber & \left.- \sinh \xi_3 \left[\sin \xi_4\left((\cD_\mu\chi_1)^2-(\cD_\mu\chi_2)^2\right) +2\cos \xi_4 \cD_\mu\chi_1 \cD^\mu\chi_2 \right] \right)\\
\nonumber& +e^{-1}\epsilon^{\mu\nu\rho}\left( 2\alpha A^{\varphi_1}_\mu F^{\varphi_2}_{\nu\rho} + \frac14 \,\varepsilon\, F_{\mu\nu}^{\varphi_2} (\chi_2\cD_\rho \chi_1 - \chi_1 \cD_\rho \chi_2) \right)- V ,\\
\nonumber V &= -2g_0^2 e^{\xi_1}\left(2 e^{\xi_2}\cosh\xi_3-e^{\xi_1}\sinh^2\xi_3\right) +\frac{g_0^2}{2} e^{2 \xi_1+\xi_2} \left[ e^{\xi_2}\left(\frac12\,\varepsilon\,\chi_1^2+\frac12\,\varepsilon\,\chi_2^2+4g_0^{-1}\alpha\right)^2 \right.\\
\nonumber & \left. +\cosh \xi_3 \left(\chi_1^2+\chi_2^2\right)+\sinh \xi_3\left( (\chi_1^2-\chi_2^2)\sin \xi_4 +2 \chi_1 \chi_2 \cos \xi_4\right) \right]\end{aligned}$$ Note that simply $F^{\varphi_i} = dA^{\varphi_i}$.
We have checked that the three-dimensional equations of motion following from (\[eq:lagr3DU1sq\]) and the reduction ansatz given in Section \[sec:full6Duplift\] imply the six-dimensional three-form Bianchi identities and self-duality relations, as well as the six-dimensional scalar equations of motion.
The $(1,0,n)$ superstrata solution sits in this truncation (see below in Section \[subsect:(1,0,n) 3D data\]). It is a solution of (\[eq:lagr3DU1sq\]) with $\xi_{3,4}=0$. Although one should note that setting these scalar fields to 0 does not give a consistent truncation of (\[eq:lagr3DU1sq\]), many of the reduction formulae of Section \[sec:full6Duplift\] simplify considerably when these scalars vanish. First of all, we have: .|\_[U(1)\^2, \_[3,4]{}=0]{} = m\_[AB]{}\^A \^B = e\^[-\_2]{}\^2+ e\^[-\_1]{}\^2, X = (\_1\_1+\_2\_1).The six-dimensional metric ansatz simplifies to: .ds\_6\^2|\_[U(1)\^2, \_[3,4]{}=0]{} = e\^[\_1+\_2]{}\^[1/2]{}ds\_3\^2 + g\_0\^[-2]{} .In fact, it was this simple metric structure that led to the original conjecture [@Bena:2017upb] that the $(1,0,n)$ superstrata should be part of a consistent truncation.
Finally, we can also explicitly calculate all components of the three-forms, including those determined by self-duality, which we give here (for $\xi_{3,4}=0$) in form notation:[^8] $$\begin{aligned}
\label{eq:U1sqG1} \left.G^1\right|_{U(1)^2,\xi_{3,4}=0} &= 2g_0^{-2}\Delta^{-2}e^{-\xi_1-\xi_2}\sin\theta\cos\theta d\theta\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber & + g_0^{-2}\Delta^{-2}e^{-\xi_1-\xi_2}\sin^2\theta\cos^2\theta(d\xi_1-d\xi_2)\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber &- g_0^{-1}\Delta^{-1}\left(e^{-\xi_2}\sin^2\theta F^{\varphi_2}\wedge \cD\varphi_1 + e^{-\xi_1}\cos^2\theta F^{\varphi_1}\wedge \cD\varphi_2\right)\\
\nonumber &- \frac12 e^{2\xi_1+2\xi_2}\,\varepsilon\,\left( 8\,\varepsilon\,\alpha + g_0(\chi_1^2 + \chi_2^2)\right) vol_3 ,\end{aligned}$$ $$\begin{aligned}
\label{eq:U1sqG2} \left.4G^2\right|_{U(1)^2,\xi_{3,4}=0} &= -g_0^{-2}\left(4\,\varepsilon\,\alpha g_0^{-1}+\frac12(\chi_1^2+\chi_2^2)\right)\sin\theta\cos\theta d\theta\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber & + g_0^{-2} e^{-\xi_1-\xi_2}X^2\left(\Delta^{-2}+e^{\xi_1}\Delta^{-1}\right)\sin\theta\cos\theta d\theta\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber & + g_0^{-2}\,\varepsilon\,\sin\theta\cos\theta(-e^{-2\xi_1}*F^{\varphi_1}\wedge d\theta\wedge \cD\varphi_1 + e^{-2\xi_2}*F^{\varphi_2}\wedge d\theta\wedge \cD\varphi_2)\\
\nonumber & +g_0^{-2}X \cos\theta (\cos\varphi_1 D\chi_1 -\sin\varphi_1 D\chi_2)\wedge d\theta\wedge \cD\varphi_2\\
\nonumber & +\frac12g_0^{-2}X^2\Delta^{-2}e^{-\xi_1-\xi_2} \sin^2\theta\cos^2\theta (d\xi_1-d\xi_2)\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber & -g_0^{-2} X e^{-\xi_1}\Delta^{-1}\sin\theta\cos^2\theta (\sin\varphi_1 D\chi_1+\cos\varphi_1 D\chi_2)\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber &-g_0^{-1}\,\varepsilon\,\cos\theta \left(\sin\theta(*d\xi_1-*d\xi_2)+e^{\xi_2} X (\sin\varphi_1*D\chi_1+\cos\varphi_1*D\chi_2)\right)\wedge d\theta\\
\nonumber & -\frac12 g_0^{-1}X^2\Delta^{-1} (e^{-\xi_1}\cos^2\theta F^{\varphi_1}\wedge \cD\varphi_2 + e^{-\xi_2}\sin^2\theta F^{\varphi_2}\wedge \cD\varphi_1)\\
\nonumber & +g_0^{-1}\,\varepsilon\, e^{\xi_2}X \sin\theta (\sin\varphi_1*D\chi_2-\cos\varphi_1 *D\chi_1)\wedge \cD\varphi_1\\
\nonumber & +\,\varepsilon\,\left( 2 e^{\xi_1+\xi_2}g_0 -e^{2\xi_1+\xi_2}X^2\left[g_0+ e^{\xi_2}\left(2\,\varepsilon\,\alpha + \frac14g_0(\chi_1^2+\chi_2^2)\right)\right] \right)vol_3,
\end{aligned}$$ $$\begin{aligned}
\label{eq:U1sqG4} \left.\sqrt{2}G^4\right|_{U(1)^2,\xi_{3,4}=0} &= g_0^{-2}\Delta^{-2}X e^{-\xi_1-\xi_2}(2+e^{\xi_1}\Delta)\sin\theta\cos\theta d\theta\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber & + g_0^{-2} \cos\theta (\cos\varphi_1 D\chi_1-\sin\varphi_1D\chi_2)\wedge d\theta\wedge \cD\varphi_2\\
\nonumber & + g_0^{-2}\Delta^{-2}\left[e^{-\xi_1-\xi_2}\sin^2\theta\cos^2\theta X(d\xi_1-d\xi_2) \right. \\
\nonumber & \left.- \sin\theta \cos^2\theta \Delta e^{-\xi_1}(\sin\varphi_1 D\chi_1 +\cos\varphi_1 D\chi_2) \right]\wedge \cD\varphi_1\wedge \cD\varphi_2\\
\nonumber & -g_0^{-1}\,\varepsilon\,e^{\xi_2}\cos\theta (\sin\varphi_1 *D\chi_1+\cos\varphi_1 * D\chi_2)\wedge d\theta \\
\nonumber & +g_0^{-1}\,\varepsilon\, e^{\xi_2} \sin\theta (\sin\varphi_1*D\chi_2-\cos\varphi_1 *D\chi_1)\wedge \cD\varphi_1\\
\nonumber & -g_0^{-1}\Delta^{-1}X\left( e^{-\xi_1}\cos^2\theta F^{\varphi_1}\wedge \cD\varphi_2 + e^{-\xi_2}\sin^2\theta F^{\varphi_2}\wedge \cD\varphi_1\right)\\
\nonumber & - \frac12\,\varepsilon\, e^{2\xi_1+\xi_2}X \left(2g_0 + e^{\xi_2}(8\,\varepsilon\,\alpha + g_0 (\chi_1^2+\chi_2^2))\right) vol_3 .\end{aligned}$$
Superstrata in three dimensions {#Sect:superstrata}
===============================
To further test our uplift formulae (as well as develop the theory of superstrata), we have constructed a novel family of multi-mode superstrata, which are solutions of the six-dimensional theory of \[sec:6Dtheory\]. We have also verified that they conform to the uplift formula of Section \[sec:full6Duplift\], and that they give a solution of the three-dimensional theory given by the action (\[action-final\]). In the language of [@Heidmann:2019xrd], this family is produced by superimposing the $(1,m,n)$ single-mode superstrata with $ m\in \{0,1 \} $ and $n\in \mathbb{Z}^{+}$. Since these are the maximal ranges allowed[^9] for $m$ and $n$, we refer to this family of solutions as the $(1,m,n)$ multi-mode family.
Appendix \[sect:6Dsuperstrata\] contains the full system of six-dimensional BPS equations used to construct these solutions, the solutions themselves, along with the regularity and asymptotic charge analysis. Here we give the truncated three-dimensional data, which solve the equations of motion for the action (\[action-final\]). In addition, in Section \[subsect:(1,0,n) 3D data\] we discuss the simpler $(1,0,n)$ sub-family, which fits in the simpler $U(1)^{2}$ sub-sector of the 6D reduction given in Section \[sec:U1sqtrunc\].
The holomorphic functions {#subsect:(1,m,n) Specifying Solutions}
-------------------------
We use the $S^{3}$ coordinates $(\theta,\varphi_{1},\varphi_{2})$, with metric (\[dsSphere\]). The $S^{3}$ is fibred over a “deformed" $AdS_{3}$, which we parametrize by $(u,v,r)$, where $$\begin{aligned}
u = \frac{1}{\sqrt{2}}(t-y) \,, \qquad v = \frac{1}{\sqrt{2}}(t+y) \,, \label{uvDef}\end{aligned}$$ are light cone coordinates, $t$ is the conventional time (in 3 dimensions) and $y$ parametrizes the common D1-D5 circle direction with radius $R_{y}$.
Following [@Heidmann:2019xrd], we introduce the complex coordinate $$\begin{aligned}
\xi ~\equiv~\frac{r}{\sqrt{r^2+ a^2}} \, e^{i \frac{\sqrt{2} v}{R_y} }\,. \label{xidef}\end{aligned}$$ A specific $(1,m,n)$ multi-mode superstrata is then fixed by specifying the two holomorphic functions: $$\begin{aligned}
F_{0}=\sum_{n=1}^{\infty}b_{n}\xi^{n} \qquad \text{and} \qquad F_{1}=\sum_{n=1}^{\infty}d_{n}\xi^{n}\,, \label{F0F1def}\end{aligned}$$ where $(b_{n},d_{n})$ are real numbers. Regularity of the solutions requires the introduction of the constant $$\begin{aligned}
c^{2} = \sum_{n=1}^{\infty} \left( b_{n}^{2} +d_{n}^{2} \right) \,, \label{cDef}\end{aligned}$$ with the constraint: $$\begin{aligned}
\frac{2Q_{1}Q_{5}}{R_{y}^{2}} = 2a^{2}+c^{2}\,.\end{aligned}$$ See Appendix \[SubSect:Regularity\] for details.
The $(1,0,n)$ multi-mode superstrata are recovered by setting $F_{1} =0 $, and the $(1,1,n)$ multi-mode superstrata are recovered by setting $F_{0} =0$. (These two multi-mode sub-families were first discussed in [@Heidmann:2019xrd].)
The three-dimensional description of (1,m,n) superstrata {#subsect:(1,m,n) 3D data}
--------------------------------------------------------
We use the freedom discussed in Appendix \[app:rescalings\] to rescale the six-dimensional uplift formulae of Section \[sec:full6Duplift\] (using $\Lambda_2 = 2\sqrt{Q_1/Q_5}$), and we choose: \[eq:3Dconsts1mn\] = - 12 g\_0, g\_0 = (Q\_1 Q\_5)\^[-1/4]{}.We have chosen these constant so that $g_0^{-1}=(Q_1 Q_5)^{1/4}$ corresponds to the radius of the $S^3$ in six-dimensions at the asymptotic $AdS_3\times S^3$ boundary, as appropriate for a D1-D5-P superstrata.
It is convenient to introduce the quantities: $$\begin{aligned}
S_{A} = - \frac{aR_{y}g_{0}^{2}}{2\sqrt{ 2(a^{2}+r^{2}) }} \left( iF_{0} ,F_{0},-ie^{i \frac{\sqrt{2}}{R_{y}} v} F_{1}, e^{i\frac{\sqrt{2}}{R_{y}} v }F_{1} \right) + c.c. \,. \label{Sdef}\end{aligned}$$ The four scalars, $\chi_A$, are then given by[^10]
$$\begin{aligned}
\chi_{A} = 2 S_{A} \,, \label{chiDef}\end{aligned}$$
and the ten scalars in $m_{AB}$ are: $$\begin{aligned}
m_{AB} &= \mathbb{I}-\begin{pmatrix}
S_{1}^{2}+S_{2}^{2} & 0 & S_{1}S_{3}-S_{2}S_{4} & S_{1}S_{4}+S_{2}S_{3} \\
0 & S_{1}^{2}+S_{2}^{2} & S_{1}S_{4}+S_{2}S_{3} & S_{2}S_{4}-S_{1}S_{3} \\
S_{1}S_{3}-S_{2}S_{4} & S_{1}S_{4}+S_{2}S_{3} & S_{3}^{2}+S_{4}^{4} & 0 \\
S_{1}S_{4}+S_{2}S_{3} & S_{2}S_{4}-S_{1}S_{3} & 0& S_{3}^{2}+S_{4}^{4} \,. \label{m1mnSol}
\end{pmatrix}\,,\end{aligned}$$ The three-dimensional metric takes the form of an $\mathbb{R}^{1}$ fiber over a conformally rescaled two-dimensional Kähler manifold: $$\begin{aligned}
ds_{3}^{2} = \frac{R_{y}^{2}g_{0}^{2}}{2} \left[ \Omega^{2}\, ds_{2}^{2} - a^{4}g_{0}^{4}\left(du + dv + \frac{\sqrt{2} }{a^{2}R_{y} g_{0}^{4}}\, \mathscr{A}\right)^{2} \right] \,, \label{ds3}\end{aligned}$$ where: $$ds_{2}^{2} = \frac{{\ensuremath{\left|d\xi\right|}}^{2}}{\left(1-{\ensuremath{\left|\xi\right|}}^{2} \right)^{2}} \,, \qquad \Omega^{2}= \frac{2 }{R_{y}^{2}g_{0}^{4}} \left( 1- S_{A}S_{A} \right) \,, \qquad
\mathscr{A} = \frac{i}{2} \left( \frac{\xi \, d\bar{\xi} - \bar{\xi} \, d\xi}{1-{\ensuremath{\left|\xi\right|}}^{2}} \right) \,.
\label{ds3Data}$$ This shows that the three-dimensional metric has the form of a non-trivial, warped time-fibration over a non-compact $\IC\IP^1$. This structure is almost certainly a consequence of supersymmetry and is extremely reminiscent of the structure used to find Gutowski-Reall black holes in AdS$_5$ [@Gutowski:2004ez; @Gutowski:2004yv].
Finally, the six vector fields $\tensor{\widetilde{A}}{_{\mu}^{AB}}$ read: $$\begin{aligned}
\tensor{\widetilde{A}}{_{\mu}^{AB}} dx^{\mu} &= \frac{1}{\sqrt{2}a^{2}R_{y}g_{0}} \left(C_{1}\tensor{\eta}{_{1}^{AB}} + C_{2}\tensor{\eta}{_{2}^{AB}} + C_{3}\tensor{\eta}{_{3}^{AB}} + \bar{C}_{3}\tensor{\bar{\eta}}{_{3}^{AB}}\right)\,, \label{Aform}\end{aligned}$$ where $$\begin{aligned}
C_{1} &= \left(S_{1}S_{3}-S_{2}S_{4} \right)\, \mathbf{d} \,, \\
C_{2} &= \left(S_{1}S_{4}+S_{2}S_{3} \right)\, \mathbf{d} \,, \\
C_{3} &= \left( \frac{a^{2}}{2} \right) \, dv -\frac{1}{2}\left( S_{1}^{2}+S_{2}^{2} -S_{3}^{2} - S_{4}^{2} \right)\, \mathbf{d} \,, \\
\bar{C}_{3} &= -\left( \frac{a^{2}+2r^{2}}{2} \right) \, dv + \left(1- \frac{S_{A}S_{A}}{2} \right)\, \mathbf{d} \,, \label{C3} \\
\mathbf{d} &= \frac{1}{\Omega^{2}}\left[ a^{4}( du+dv) +\frac{2r^{2}}{R_{y}^{2}g_{0}^{4}} \, dv \right] \,,\end{aligned}$$ and we have introduced the antisymmetric $4\times 4$ ’t Hooft matrices, which may be written: $$\begin{aligned}
\tensor{\eta}{_{1}^{AB}}&= \begin{pmatrix}
0 & \sigma_{x} \\
-\sigma_{x} & 0
\end{pmatrix} \,, & \tensor{\eta}{_{2}^{AB}} &= \begin{pmatrix}
0 & -\sigma_{z} \\
\sigma_{z} & 0
\end{pmatrix} \,, & \tensor{\eta}{_{3}^{AB}} &= \begin{pmatrix}
i\sigma_{y} & 0 \\ 0 & i\sigma_{y}
\end{pmatrix} \,, \\
\tensor{\bar{\eta}}{_{1}^{AB}}&= \begin{pmatrix}
0 & -i \sigma_{y} \\
i \sigma_{y} & 0
\end{pmatrix} \,, & \tensor{\bar{\eta}}{_{2}^{AB}} &= \begin{pmatrix}
0 & -\mathbb{I} \\
\mathbb{I} & 0
\end{pmatrix} \,, & \tensor{\bar{\eta}}{_{3}^{AB}} &= \begin{pmatrix}
i\sigma_{y} & 0 \\ 0 & -i\sigma_{y}
\end{pmatrix}\,.\end{aligned}$$
Note that the $\eta_j$ and the $\bar \eta_j$ generate the commuting $SU(2)$ factors of $SO(4) = SU(2) \times SU(2)$. In particular, this means that the gauge fields in (\[Aform\]) define an $SU(2) \times U(1)$ gauge connection.
We have explicitly checked that the three-dimensional fields $(ds_{3}^{2},\chi_{A},m_{AB},\tensor{\widetilde{A}}{_{\mu}^{AB}})$ given by (\[ds3\]), (\[chiDef\]), (\[m1mnSol\]), and (\[Aform\]) with the three-dimensional constants fixed by (\[eq:3Dconsts1mn\]) (and (\[eq:g0gamma0\])) satisfy the three-dimensional equations of motion coming from the three-dimensional Lagrangian (\[action-final\]) (these are given explicitly in Appendix \[App:3Deom\]). Note that the orientation of the three-dimensional manifold is tied to the sign of $\alpha$ as we must choose: $$e^{-1}\epsilon_{uvr} = -\varepsilon\,,
\label{eq:3Dorientation}$$ where $e \equiv \sqrt{|\det(g_{\mu\nu})|}$.
The solutions in U(1)\^2 truncations {#subsect:(1,0,n) 3D data}
------------------------------------
The $(1,m,n)$ multi-mode solution simplifies greatly when one sets either $F_{0}=0$, so that $S_{3}=S_{4}=0$, or $F_{1}=0$, so that $S_{1}=S_{2}=0$. These are the $(1,0,n)$ and $(1,1,n)$ multimode solutions respectively, both introduced and analyzed in [@Heidmann:2019xrd]. In each instance the expansion of $\tensor{\widetilde{A}}{_{\mu}^{AB}}$ in (\[Aform\]) simplifies with $C_{1}=C_{2}=0$, implying the $\tensor{\widetilde{A}}{_{\mu}^{AB}}$ define a $U(1)\times U(1)$ gauge connection.
The $(1,0,n)$ multi-mode family conforms exactly to the $U(1)^{2}$ truncation of Section (\[sec:U1sqtrunc\]). Using the notation of that section, the reduction data $(ds_{3}^{2},\chi_{1},\chi_{2},\xi_{1},\xi_{2},\xi_{3},\xi_4,A^{\varphi_{1}}_{\mu},A_{\mu}^{\varphi_{2}})$, are given by: $$\begin{aligned}
\chi_{1,2} = 2 S_{1,2}\,,\end{aligned}$$ with $$\begin{aligned}
S_{1} = - \frac{ia R_{y}g_{0}^{2}}{2\sqrt{2(a^{2}+r^{2})}} \left( F_{0}-\bar{F}_{0}\right) \qquad \text{and} \qquad S_{2}= - \frac{aR_{y}g_{0}^{2}}{2\sqrt{2(a^{2}+r^{2})}}\left( F_{0} + \bar{F}_{0}\right) \,.\end{aligned}$$ The three dimensional metric, $ds_{3}^{2}$, again takes the form (\[ds3\])-(\[ds3Data\]) but with the altered $$\begin{aligned}
\Omega^{2} = \frac{2}{R_{y}^{2}g_{0}^{4}} (1-S_{1}^{2}-S_{2}^{2})\,.\end{aligned}$$ The remainder of the scalars read: $$\begin{aligned}
\xi_{1} = \xi_{3}=\xi_{4}=0 \qquad \text{and} \qquad e^{-\xi_{2}} = \frac{1}{2}R_{y}^{2}g_{0}^{4} \Omega^{2} \,.\end{aligned}$$ While the vector fields reduce to: $$\begin{aligned}
A^{\varphi_{1}}_{\mu}\, dx^{\mu} &= -\frac{a^{2}R_{y}g_{0}^{3}}{\sqrt{2}}(du+dv) \,, \\
A^{\varphi_{2}}_{\mu}\, dx^{\mu} &= \frac{\sqrt{2}}{R_{y}g_{0} \Omega^{2}} \left[a^{2}(du+dv) + \frac{2}{a^{2}R_{y}^{2}g_{0}^{4}} \left( (a^{2}+r^{2})(S_{1}^{2}+S_{2}^{2}) -a^{2}\right)\, dv\right]\,.\end{aligned}$$
As noted earlier, the $(1,1,n)$ multi-mode family is part of another $O(2)$-invariant truncation. This has $\chi_{1}=\chi_{2}=0$ and non-trivial $(\chi_{3},\chi_{4})$. This truncated theory will involve a non-trivial gauge coupling in the $\varphi_{1}$ direction rather than the $\varphi_{2}$ direction. A priori, one might guess that the $(1,0,n)$ and $(1,1,n)$ families are related by a simple change of coordinates $\varphi_{1}\leftrightarrow \varphi_{2}$. However, we see from (\[C3\]) that the two solutions will have distinct gauge field expansions, even after re-labeling $\varphi_{1}\leftrightarrow \varphi_{2}$. This is in agreement with the work of [@Walker:2019ntz], where it was also shown that the $(1,0,n)$ and $(1,1,n)$ single mode solutions are only equivalent after a non-trivial spectral transformation and reduction to five dimensions.
Final comments {#sec:Conclusions}
==============
We have shown that the three-dimensional $(0,2)$ gauged $SO(4)$ supergravity described in Section \[Sect:3Dsugr\] is a consistent truncation of six-dimensional $(1,0)$ supergravity coupled to two tensor multiplets. We have also shown that this consistent truncation includes the newly-constructed family of $(1,m,n)$ superstrata, which involve momentum waves encoded in two freely-choosable holomorphic functions of one variable.
This raises the question as to whether there are other consistent truncations that might encode yet more classes of microstate geometries. The answer is almost certainly yes. First, the results of [@Nicolai:2003ux] suggest that there may well be consistent truncations that encode higher KK modes, and even entire towers of such modes. These KK towers include the modes of at least one tensor gauge field and so it seems likely that this work could be extended to the tensor gauge fields that one needs for superstrata.
There are also indications that the five-dimensional geometries that can be obtained from compactifications of the $(2,1,n)$ superstrata [@Bena:2017geu; @Bena:2017upb; @Walker:2019ntz], may also give rise to consistent truncations. These would be gauged supergravity theories in three dimensions obtained from AdS$_3$ $\times S^2$ compactifications of $\cN=2$ supergravity, coupled to vector multiplets, in five dimensions.
It therefore seems that the consistent truncations described here might be the tip of an iceberg: there are almost certainly extensive generalizations of our results.
As described in the introduction, our primary interest in examining these consistent truncations is to provide a new tool for the study of microstate geometries. In this paper, we have shown that the consistent truncation contains large and interesting families of BPS superstrata. We plan to see if the three-dimensional approach will enable us to find some new, broader families of BPS microstate geometries.
One of the remarkable things about the six-dimensional BPS equations is that, after specifying a hyper-Kähler base, the remaining equations reduce to a linear system [@Bena:2011dd]. [*A priori*]{}, it is not clear whether this simplification will be manifest in the three-dimensional BPS equations. Indeed, it seems likely that linearity in three dimensions will only emerge if one restricts the gauge fields, $\tensor{\widetilde{A}}{_{\mu}^{AB}}$, to an Abelian sub-sector. These gauge fields may also need to be locked onto the scalar fields in some manner. An important question then becomes, to what extent one can unlock all the non-abelian gauge fields, while still being able to solve the BPS system? The end result may well be intrinsically non-linear. If solutions can still be found, then their uplift to six dimensions might reveal new hyper-Kähler bases, which may give new and interesting microstate geometries. As mentioned in the introduction, there is evidence that such bases should exist, coming from perturbation theory, in both [@Ceplak:2018pws; @Tyukov:2018ypq].
Even more important is the possibility of constructing non-BPS microstate geometries. Through a simple parity flip, one can convert BPS superstata into anti-BPS superstata (ones that preserve a different, complementary set of supersymmetries). It is therefore possible to use the three-dimensional formulation to study non-BPS configurations that start from a combination of BPS and anti-BPS momentum waves. It should be relatively straightforward to set up a three-dimensional initial value problem that should produce such non-BPS microstate geometries as the result of ‘scattering’ BPS and anti-BPS waves. The extent to which this can be done analytically, or semi-analytically, is unclear, but it will certainly be possible to study this numerically.
In considering the outcome of such an approach to non-BPS solutions, it is important to remember the Faustian bargain of consistent truncations. It is quite possible that the combining of BPS and non-BPS solutions in three dimensions will evolve, at late times, into a singular solution. As we have seen in many examples of microstate geometries, the appearance of a singularity in supergravity does not invalidate the microstate geometry program, but usually indicates that one has suppressed degrees of freedom that are essential to resolving the singularity. Thus the appearance of a singularity at late times may simply be the result of limiting the degrees of freedom to a consistent truncation.
Even if singularities do arise in such non-BPS solutions, there will still be invaluable information to be gleaned from the three-dimensional analysis. One will see the early, time-dependent behavior and the radiation that comes from the scattering. By using the uplifts one may also be able to determine which degrees of freedom will be needed to resolve any singular behaviour.
It is also possible that the microstate geometries created in this way will be smooth and robust and provide families of non-BPS microstate geometries for which the holographic dictionary is precisely known.
Acknowledgments {#acknowledgments .unnumbered}
===============
NPW would like to thank Henning Samtleben for his patient and careful explanations of many aspects of his work on gauged supergravity in three dimensions. DRM is supported by the ERC Starting Grant 679278 Emergent-BH. The work of NPW and RAW was supported in part by ERC Grant number: 787320 - QBH Structure and by DOE grant DE-SC0011687. RAW is very grateful to the IPhT of CEA-Saclay for hospitality during this project, his research was also supported by a Chateaubriand Fellowship of the Office for Science Technology of the Embassy of France in the United States.
The Three-sphere {#sec:app:S3}
================
It is convenient to parametrize the unit radius, round $S^3$ with four restricted Cartesian coordinates $\mu^A$ of $\IR^4$ that satisfy $\mu^A\mu^A = 1$, or alternatively with three (unrestricted) coordinates $y^i$. The round, unit-sphere metric in coordinates $y^i$ is $\mathring{g}_{ij}$, with corresponding completely antisymmetric tensor $\mathring{\omega}_{ijk}$. Following [@Samtleben:2019zrh], we also use a vector $\mathring{\zeta}^i$ with unit divergence, \_i \^i = 1.
The gauge-covariant derivatives on the sphere are then: \[eq:appS3:Dmu\] \^A = d\^A -2 g\_0 \^[AB]{}\^B, where $g_0=2\gamma_0$ is the gauge coupling, and we have used the dual gauge fields given in (\[dualGFsfinal\]). We can rewrite this as: \^A = \_i\^A y\^i,with: \[eq:appS3:Dy\] y\^i = dy\^i -2 g\_0 \^i\_[AB]{} \^[AB]{} ,where we have used the Killing vectors on the sphere: \^i\_[AB]{} = \^[ij]{}\_j \^[\[A]{} \^[B\]]{}.There are many other identities involving the $\mu^A$ (which we will not explicitly need in this paper); see, for example, Appendix A of [@Samtleben:2019zrh].
An explicit coordinate basis that can be used is, for example, the standard coordinates $y^i=(\theta,\varphi_{1},\varphi_{2})$ with: \[eq:appS3:explicitmu\] \^1 = \_1, \^2 = \_1, \^3 = \_2, \^4 = \_2.The metric in these coordinates of the unit radius round $S^3$ is: \^2\_[S\^3]{} = \_[ij]{}dy\^idy\^j = d\^2 + \^2d\_1\^2 + \^2d\_2\^2, \[dsSphere\]so that $\mathring{\omega}_{ijk} = (\sin\theta\cos\theta) \epsilon_{ijk}$, with $\epsilon_{123}=+1$ and completely antisymmetric. In these coordinates, we can take: \^i = ( 12,0,0).
Six-dimensional and three-dimensional rescalings {#app:rescalings}
================================================
We wish to point out two rescalings of the six-dimensional fields which have a counterpart as a rescaling of three-dimensional fields, through the uplift formulae in Section \[sec:full6Duplift\].
The first rescaling is: $$\begin{aligned}
e^{\sqrt{2}\varphi} & \rightarrow \Lambda_1^{-2}\, e^{\sqrt{2}\varphi} , & X & \rightarrow \Lambda_1\, X, & g^{(6D)}_{\hmu\hnu} &\rightarrow \Lambda_1\, g^{(6D)}_{\hmu\hnu},\\
G^1 &\rightarrow G^1, & G^2 &\rightarrow \Lambda_1^2\, G^2, & G^4 &\rightarrow \Lambda_1\, G^4,\end{aligned}$$ which corresponds to the three-dimensional rescaling: $$\begin{aligned}
m_{AB} &\rightarrow \Lambda_1^2\, m_{AB}, & \chi_A & \rightarrow \Lambda_1\, \chi_A,\\
\alpha &\rightarrow \Lambda_1^2\, \alpha, & g^{(3D)}_{\mu\nu} &\rightarrow \Lambda_1^4\, g_{\mu\nu}^{(3D)}.\end{aligned}$$ Under this scaling, the six-dimensional Lagrangian (\[eq:6Dlagr\]), resp. three-dimensional action (\[action-final\]), scales as $\hat{e}\,\mathcal{L}_{6D}\rightarrow \Lambda_1^2\, \hat{e}\,\mathcal{L}_{6D}$, resp. $e\,\mathcal{L}_{3D}\rightarrow \Lambda_1^2\, e\, \mathcal{L}_{3D}$ (with $\hat{e} = \sqrt{-\det g_{\hmu\hnu}^{(6D)}}$ and $e=\sqrt{-\det g_{\mu\nu}^{(3D)}}$).
The second rescaling is: $$\begin{aligned}
e^{\sqrt{2}\varphi} & \rightarrow \Lambda_2^{-2}\, e^{\sqrt{2}\varphi} , & X & \rightarrow \Lambda_2\, X, & g^{(6D)}_{\hmu\hnu} &\rightarrow g^{(6D)}_{\hmu\hnu},\\
G^1 &\rightarrow \Lambda_2^{-1}\,G^1, & G^2 &\rightarrow \Lambda_2\, G^2, & G^4 &\rightarrow G^4,\end{aligned}$$ which has the three-dimensional counterpart: $$\begin{aligned}
m_{AB} &\rightarrow \Lambda_2^2\, m_{AB}, & \chi_A & \rightarrow \Lambda_2\, \chi_A,\\
\alpha &\rightarrow \Lambda_2^{5/2}\, \alpha, & g_{\mu\nu}^{(3D)} &\rightarrow \Lambda_2^3\, g_{\mu\nu}^{(3D)},\\
g_0 &\rightarrow \Lambda_2^{1/2}\,g_0, & A^{AB}_{\mu} &\rightarrow \Lambda_2^{-1/2}\, A^{AB}_\mu.\end{aligned}$$ Note that the rescaling of $g_0$ implies the same rescaling of $\gamma_0$ through (\[eq:g0gamma0\]). The six-dimensional action (\[eq:6Dlagr\]) is invariant under this scaling, $\hat{e}\,\mathcal{L}_{6D}\rightarrow \hat{e}\,\mathcal{L}_{6D}$, while the three-dimensional action (\[action-final\]) rescales as $e\,\mathcal{L}_{3D}\rightarrow \Lambda_2^{3/2}\, e\,\mathcal{L}_{3D}$.
A combination of both of these scalings can be used to rescale the two constants $|\alpha|$ and $g_0$ to any value in the reduced three-dimensional theory.
Matching with the conventions of [@Samtleben:2019zrh] {#sec:app:Henning}
=====================================================
Our goal here is to provide a map between our conventions and those of [@Samtleben:2019zrh]. We start by noting that our three-dimensional formulation matches that of [@Samtleben:2019zrh] if one sets $$\label{eq:3Dmatch} \gamma_0 ~=~ 1 \qquad \Leftrightarrow \qquad g_0 ~=~ 2 \ \,.$$ The parameter $\alpha$ is the same in both sources.
Turning to the six-dimensional theory, our scalar matrix is $$\cM^{\hI \hJ} ~=~ \frac{1}{2}\, e^{\sqrt{2}\varphi}\,\left( \begin{array}{ccc} 8 & X^2 &4 \sqrt{2} X\\
X^2 & \frac18e^{-2\sqrt{2}\varphi} (2 + e^{\sqrt{2}\varphi} X^2)^2 & \frac{1}{\sqrt{2}}(2e^{-\sqrt{2}\varphi}+X^2)X
\\ 4\sqrt{2}X &\frac{1}{\sqrt{2}}(2e^{-\sqrt{2}\varphi}+X^2) X&4 \,(e^{-\sqrt{2}\varphi} + X^2) \end{array}\right) \,, \label{eq:MUs}$$ where we have raised the second index using our $SO(1,2)$ metric: $$\eta^{\hI\hJ} ~=~ \left( \begin{array}{ccc} 0 & 1 & 0\\ 1 & 0 & 0\\ 0 & 0 & -2 \end{array}\right)\,.
\label{eq:etaUs}$$ One should also recall (\[eq:DeltaDef\]) and (\[eq:6Dscalaransatz\]): $$e^{-\sqrt{2}\varphi} ~=~ \Delta ~=~ m_{AB}\mu^A \mu^B \,.
$$ The corresponding objects in [@Samtleben:2019zrh] are $$\begin{aligned}
& \widetilde \cM^{a b} ~=~ \\
& \frac{1}{8} \,\left( \begin{array}{ccc} 4 \Delta+ 4 X^2+ \Delta^{-1}(2+X^2)^2 & 4 \Delta^{-1} - \Delta (2+ \Delta^{-1} X^2)^2 & -2 \sqrt{2} (2 + \Delta^{-1}(2+X^2)) X\\
4 \Delta^{-1} - \Delta (2+\Delta^{-1} X^2)^2 & 4 \Delta+ 4 X^2+ \Delta^{-1}(2-X^2)^2 & 2 \sqrt{2} (2 - \Delta^{-1}(2-X^2)) X
\\ -2 \sqrt{2} (2 + \Delta^{-1}(2+X^2)) X &2 \sqrt{2} (2 - \Delta^{-1}(2-X^2)) X &1+ \Delta^{-1} X^2 \end{array}\right) \,,
\end{aligned}
\label{eq:MUSSs}$$ and $$\tilde \eta^{ab} ~=~ \left( \begin{array}{ccc} 1 & 0 & 0\\ 0 & -1 & 0\\ 0 & 0 & -1 \end{array}\right)\,,
\label{eq:etaSS}$$ where the indices $a,b,\ldots$ take the values $0, \bar 0, \bar1$. To compare with the results of [@Samtleben:2019zrh], one should note that [@Samtleben:2019zrh] uses two expressions: $\Delta$ and $\tilde\Delta$. We re-label the $\Delta$ of [@Samtleben:2019zrh] as $\hat\Delta$ here; the relations between these quantities and our $\Delta$ are: $$\label{eq:phiDelta}
\hat \Delta ~=~ \big(\det(m_{AB})\big)^{\frac{1}{4}} \, \Delta^{-\frac{1}{4}} \, , \qquad \tilde \Delta ~=~ \Delta^{-\frac{1}{4}} \,.$$
Define the matrix $$P ~=~ \left( \begin{array}{ccc} -\frac{1}{2\sqrt{2}} & - \sqrt{2} &0 \\
-\frac{1}{2\sqrt{2}} & \sqrt{2} & 0
\\ 0 &0 & \frac{1}{\sqrt{2}} \end{array}\right) \,, \label{eq:Pdefn}$$ Then one can easily verify that $$\tilde \eta ~=~ P \, \eta \, P^t \,, \qquad \widetilde \cM ~=~ P \, \cM \, P^t \,, \label{eq:Props}$$ Thus $P$ provides a change of basis from our fields to those of [@Samtleben:2019zrh]. In particular, performing the change of basis on the gauge potentials (\[eq:6DansatzB1ij\])–(\[eq:6DansatzB4ij\]) yields $$\begin{aligned}
B^0_{ij} ~=~ & -\frac{1}{2\sqrt{2}} \, B^1_{ij} ~-~ \sqrt{2} \, B^2_{ij} \nonumber \\
~=~ &\frac{1}{\sqrt{2} \,g_0^2} \,\Big[ -\big(1-2\,\varepsilon \alpha g_0^{-1}\big) \,\mathring{\omega}_{ijk}\mathring{\zeta}^k - \coeff{1}{4}\,\mathring{\omega}_{ijk}\mathring{g}^{kl}\partial_l (\log \Delta) ~+~ \coeff{1}{8} \, \mathring{\omega}_{ijk}\mathring{g}^{kl}\Delta\partial_l ( \Delta^{-1}X^2) \Big] \,, \label{eq:6DB0ij} \\
B^{\bar 0}_{ij} ~=~ & -\frac{1}{2\sqrt{2}} \, B^1_{ij} ~+~ \sqrt{2} \, B^2_{ij} \nonumber \\
~=~ &\frac{1}{\sqrt{2} \,g_0^2} \,\Big[ -\big(1 + 2\,\varepsilon \alpha g_0^{-1}\big) \,\mathring{\omega}_{ijk}\mathring{\zeta}^k - \coeff{1}{4}\,\mathring{\omega}_{ijk}\mathring{g}^{kl}\partial_l (\log \Delta) ~-~ \coeff{1}{8} \, \mathring{\omega}_{ijk}\mathring{g}^{kl}\Delta\partial_l ( \Delta^{-1}X^2) \Big] \,, \label{eq:6DB0barij} \\
B^{\bar 1}_{ij} ~=~ & \frac{1}{\sqrt{2}} \, B^4_{ij} ~=~ - \frac{1}{2\,g_0^2} \,\mathring{\omega}_{ijk}\mathring{g}^{kl}\Delta^{1/2}\partial_l \left(\Delta^{-1/2} X \right) \,, \label{eq:6DB1barij} \end{aligned}$$ Similarly, transforming (\[eq:6DansatzB1imu\]) – (\[eq:6DansatzB2imu\]) yields $$\begin{aligned}
B^0_{\mu i} ~=~ & -\frac{1}{2\sqrt{2}} \, B^1_{i \mu} ~+~ \sqrt{2} \, B^2_{i \mu} \nonumber \\
~=~ & -\frac{1}{\sqrt{2} \,g_0} \,\big( \partial_i\mu^A\big)\, \Big[ \big(A^{AB}_\mu - \varepsilon \, {A_\mu}^A{}_B \big) \mu^B ~-~2\, \big(1-2\,\varepsilon \alpha g_0^{-1}\big)\,A^{AB}_\mu \big(\mathring{\zeta}^k\partial_k\mu^B\big) \Big] \,, \label{eq:6DB0imu} \\
B^{\bar 0}_{\mu i} ~=~ & \frac{1}{2\sqrt{2}} \, B^1_{i \mu} ~-~ \sqrt{2} \, B^2_{i \mu} \nonumber \\
~=~ & -\frac{1}{\sqrt{2} \,g_0} \,\big( \partial_i\mu^A\big)\, \Big[ \big(A^{AB}_\mu + \varepsilon \, {A_\mu}^A{}_B \big) \mu^B ~-~2\, \big(1+ 2\,\varepsilon \alpha g_0^{-1}\big)\,A^{AB}_\mu \big(\mathring{\zeta}^k\partial_k\mu^B\big) \Big]\,,\label{eq:6DB0barimu} \\
B^{\bar 1}_{ \mu i } ~=~ & - \frac{1}{2\sqrt{2}} \, B^4_{i \mu} ~=~ 0 \,. \label{eq:6DB1barimu} \end{aligned}$$ Note that we have reversed the indices, ${ \mu i }$, on the left-hand side to facilitate comparison with [@Samtleben:2019zrh]. We find a perfect match for the $B_{ij}$ components, up to an overall factor of $g_0^{-2}$, provided that one uses (\[eq:3Dmatch\]) and takes $$\varepsilon ~=~ -1 \,. \label{eq:translation1}$$ This choice was anticipated in Sections \[sec:scalars\] and \[sec:6Dtheory\].
Using (\[eq:3Dmatch\]) and (\[eq:translation1\]) we also find a nearly perfect match for the $B_{\mu i}$ components, up to an overall factor of $-g_0^{-1}$.
The overall factors of $g_0^{-2}$ in $B_{ij}$ and $-g_0^{-1}$ in $B_{\mu i}$ are easily fixed using the scalings in Appendix \[app:rescalings\]. Indeed, choosing $$\label{eq:scales} \Lambda_1 ~=~ \Lambda_2^{-1} ~=~ \Lambda$$ results in the six-dimensional rescaling $$g^{(6D)}_{\hmu\hnu} ~\rightarrow~ \Lambda \, g^{(6D)}_{\hmu\hnu} \,, \qquad G^\hI ~\rightarrow~ \Lambda \, G^\hI \,.$$ One can then take $\Lambda =g_0^2 =4$ to match the overall scale in $B_{ij}$.
This scaling then creates an overall factor of $-g_0$ in $B_{\mu i}$. This can then be compensated by coordinate changes: $$x^\mu ~=~ \coeff{1}{2} \, \tilde x^\mu \,, \qquad y^i ~=~ - \tilde{y}^i \,,$$ which then rescale $B_{\mu i}$ by $-\frac{1}{2}$, while leaving $B_{ij}$ unchanged. There are two discrepancies between our analysis and that of [@Samtleben:2019zrh] that may be transcription errors in [@Samtleben:2019zrh]. First, although our expressions for $B^{0}_{ij},B^{\bar 0}_{ij},B^{\bar 1}_{ij}$ match those in [@Samtleben:2019zrh] precisely, our expressions for $B^{0}_{\mu i}$ and $B^{\bar 0}_{\mu i}$ match the expressions in [@Samtleben:2019zrh] for ${B _{\mu i}}{}_0$ and ${B_{\mu i}}{}_{\bar 0}$. When $0$ and ${\bar 0}$ are lowered using $\tilde \eta_{ab}$ (see, \[eq:etaSS\]), they get a relative minus sign and so they cannot be reconciled simultaneously.
Second, we have, using (\[eq:3Dmatch\]): $$\label{eq:appS3:Dya} \cD y^i ~=~ dy^i -2 g_0 \mathcal{K}^i_{AB} \tilde{A}^{AB} ~=~ -d\tilde y^i - 4\, \mathcal{K}^i_{AB} \tilde{A}^{AB} \,,$$ whereas, [@Samtleben:2019zrh] defines $$\label{eq:appS3:Dyb} \cD y^i ~=~ dy^i + \mathcal{K}^i_{AB} \tilde{A}^{AB}\,,$$ so these expressions do not match. We also note that the combination $g_0 A^{AB}_\mu dx^\mu$ is both coordinate invariant and invariant under both the rescalings described in Appendix \[app:rescalings\], and so we cannot reconcile our expressions for $\cD y^i $ with those of [@Samtleben:2019zrh].
Therefore, up to two minor discrepancies, our results match the expressions in [@Samtleben:2019zrh]. As we indicated in Section \[ss:testing\], we have subjected our uplift formulae to rigorous testing in both three and six dimensions, and have every confidence in our expressions and normalizations.
(1,m,n) Superstrata in six dimensions {#sect:6Dsuperstrata}
=====================================
This appendix summarizes the six-dimensional BPS equations for the D1-D5-P system, and the novel construction of the $(1,m,n)$ multi-mode superstrata family of solutions.
Six-dimensional BPS equations {#app:BPS equations}
-----------------------------
All solutions of 6 dimensional, $\mathcal{N}=(1,0)$ supergravity, coupled to 2 tensor multiplets, with the same charges as the D1-D5-P system, satisfy a “layered" set of linear equations. These equations were first developed in [@Bena:2011dd] for a single tensor multiplet, and extended to include a second tensor multiplet in [@Giusto:2013rxa]. We follow [@Bena:2017geu], introducing the equations in an explicitly $SO(1,2)$ covariant form, using the $SO(1,2)$ indices: $\hat{I},\hat{J},\hat{K},\cdots \in \{1,2,4 \}$, with non-zero $SO(1,2)$ metric components (see (\[etaDef\])): $$\begin{aligned}
\eta^{12}=\eta^{21}=1 \qquad \text{and} \qquad \eta^{44}=-2~. \label{etadef2}\end{aligned}$$
The full six-dimensional geometry, constrained by supersymmetry, may be written as a $(1+1)$-dimensional Lorentzian fiber, parametrized by the light cone coordinates $(u,v)$, over a four dimensional hyper-Kähler base $ds_4^2(\mathcal{B})$ as: $$\begin{aligned}
ds_6^2 &= -\frac{2}{\sqrt{\mathcal{P}}} \, (dv+\beta) \big(du + \omega + \tfrac{1}{2}\, \mathcal{F} \, (dv+\beta)\big)
+ \sqrt{\mathcal{P}} \, ds_4^2(\mathcal{B})~. \label{ds6}\end{aligned}$$ The metric data consists of the base $\mathcal{B}$, the functions $(\mathcal{P},\mathcal{F})$ and one forms $(\beta,\omega)$. The one forms must have legs only on the base $\mathcal{B}$, while the complete data can have functional dependence on all coordinates except for $u$. This metric data is fixed by solving the BPS equations, which are written in terms of a set of three functions $Z_{\hI}$ and three two forms $\Theta^{\hI}$. In terms of this data, the three-form fields encoding the multiplets read: $$\begin{aligned}
G^{\hat{I}} = d \left[ -\frac{1}{2}\frac{\eta^{\hat{I}\hat{J}}Z_{\hat{J}}}{\mathcal{P}}(du+\omega)\wedge (dv+\beta) \right] + \frac{1}{2}\eta^{\hat{I}\hat{J}}*_{4}DZ_{\hat{J}}+\frac{1}{2}(dv+\beta)\wedge \Theta^{\hat{I}} \label{Gdef}\end{aligned}$$ where: $$\begin{aligned}
\mathcal{P} = \frac{1}{2}\eta^{\hat{I}\hat{J}}Z_{\hat{I}}Z_{\hat{J}}= Z_{1}Z_{2}-(Z_{4})^{2}~,\end{aligned}$$ $D$ is defined to act on forms $\Phi$ by: $$\begin{aligned}
D\Phi=d_{4}\Phi-\beta\wedge \dot{\Phi}\end{aligned}$$ where overhead dots denote $\partial_{v}$ derivatives, $(d_{4},*_{4})$ are the exterior derivative and Hodge star with respect to the four-dimensional hyper-Kähler base $\mathcal{B}$, and their non-subscript and hatted counterparts refer to the full six-dimensional geometry (\[ds6\]). These three forms satisfy the twisted self duality constraint:[^11] $$\begin{aligned}
\hat{*} G^{\hat{I}} = \tensor{M}{^{\hat{I}}_{\hat{J}}}G^{\hat{J}} \qquad \text{where} \qquad M_{\hat{I}\hat{J}} = \frac{Z_{\hat{I}}Z_{\hat{J}}}{P}-\eta_{\hat{I}\hat{J}}~. \label{M1mnDef}\end{aligned}$$
Solving the BPS equations takes the layered form:
- Fix a hyper-Kähler base $ds_4^2(\mathcal{B})$ and choose a $\beta$ satisfying: $$\begin{aligned}
d\beta = *_{4}d\beta ~.\label{BPSlayer0}\end{aligned}$$
- Find a set $(Z_{\hat{I}},\Theta^{\hat{I}})$ that solve the “first layer"[^12]: $$\begin{aligned}
*_{4}D\dot{Z}_{\hat{I}} = \eta_{\hat{I}\hat{J}}D\Theta^{\hat{J}} ~,\qquad D*_{4}DZ_{\hat{I}}=-\eta_{\hat{I}\hat{J}} \Theta^{\hat{J}}\wedge d\beta ~, \qquad \Theta^{\hat{I}} = *_{4} \Theta^{\hat{I}}~. \label{BPSlayer1}\end{aligned}$$
- Find $(\mathcal{F},\omega)$ that solve the “second layer:" $$\begin{aligned}
(1+*_{4})D\omega +\mathcal{F}\, d\beta &= Z_{\hat{I}}\Theta^{\hat{I}} ~, \label{BPSlayer2a}\\
*_{4}D*_{4} \left(\dot{\omega}- \frac{1}{2}D\mathcal{F} \right) &= \frac{1}{4}\eta_{\hat{I}\hat{J}}\left[ 4\ddot{Z}^{\hat{I}}Z^{\hat{J}}+2\dot{Z}^{\hat{I}}\dot{Z}^{\hat{J}} - *_{4}\left( \Theta^{\hat{I}}\wedge \Theta^{\hat{J}} \right)\right]~. \label{BPSlayer2b}\end{aligned}$$
The solution
------------
The standard hyper-Kähler base used in the construction of six-dimensional superstrata is flat $\mathbb{R}^{4}$, which is most conveniently written in spherical bipolar coordinates $(r,\theta,\varphi_{1},\varphi_{2})$, with metric: $$ds_4^2(\mathcal{B}) ~=~ \Sigma \, \left(\frac{d r^2}{r^2+a^2}+ d\theta^2\right)+(r^2+a^2)\sin^2\theta\,d\varphi_1^2+r^2 \cos^2\theta\,d\varphi_2^2\, .
\label{ds4flat}$$ where $a$ is a positive constant and $$\Sigma~\equiv~ r^2+a^2 \cos^2\theta \,.
\label{Sigdefn}$$
In terms of the complex coordinates: $$\chi ~\equiv~\frac{a}{\sqrt{r^2+ a^2}} \, \sin \theta \, e^{i \varphi_1} \,, \qquad \mu ~=~\cot \theta \, e^{i \big (\frac{\sqrt{2} v}{R_y} - \varphi_1- \varphi_2\big)} \,, \qquad \xi ~\equiv~\frac{r}{\sqrt{r^2+ a^2}} \, e^{i \frac{\sqrt{2} v}{R_y} }\, , \label{cplxcoorddefn}$$ the $(1,m,n)$ multi-mode solution can be written in terms of the basic function: $$\begin{aligned}
F(\chi,\mu,\xi)= \chi F_{0}(\xi)+\chi \mu F_{1}(\xi)~, \end{aligned}$$ where $F_{0,1}$ are holomorphic functions of $\xi$, with expansions in terms of the real coefficients $(c_{n},d_{n})$: $$\begin{aligned}
F_{0}=\sum_{n=1}^{\infty}b_{n}\xi^{n} \qquad \text{and} \qquad F_{1}=\sum_{n=1}^{\infty}d_{n}\xi^{n}~. \end{aligned}$$
We define the auxiliary data: $$\begin{aligned}
A= \chi\mu\left(1+\xi\partial_{\xi} \right)F_{1} \qquad \text{and} \qquad B=\chi\xi\partial_{\xi}F_{0} \, , \label{ABdef}\end{aligned}$$ and self dual forms $$\begin{aligned}
\Omega_{y} = \frac{1}{\sqrt{2}} \left(-\Omega^{(2)}+i r\sin\theta \, \Omega^{(1)} \right)\, , \qquad
\Omega_{z} = \frac{1}{\sqrt{2}}\left(\Omega^{(3)}+i \left(r\sin\theta - \frac{\Sigma}{r\sin\theta} \right) \Omega^{(1)} \right)\, , \label{SDyz}\end{aligned}$$ where $$\label{selfdualbasis}
\begin{aligned}
\Omega^{(1)} &~\equiv~ \frac{dr\wedge d\theta}{(r^2+a^2)\cos\theta} + \frac{r\sin\theta}{\Sigma} d\varphi_1\wedge d\varphi_2\,,\\
\Omega^{(2)} &~\equiv~ \frac{r}{r^2+a^2} dr\wedge d\varphi_2 + \tan\theta\, d\theta\wedge d\varphi_1\,,\\
\Omega^{(3)} &~\equiv~ \frac{dr\wedge d\varphi_1}{r} - \cot\theta\, d\theta\wedge d\varphi_2\,.
\end{aligned}$$
To solve the BPS equations, first, one fixes: $$\begin{aligned}
\beta = \frac{a^{2}R_{y}}{\sqrt{2}} \left(\sin^{2}\theta \, d\varphi_{1} -\cos^{2} \theta \, d\varphi_{2} \right)\,, \label{beta}\end{aligned}$$ then the solution to the first BPS layer is given by the data: $$\begin{aligned}
\begin{split}
Z_{1} &= \frac{Q_{1}}{\Sigma} + \frac{R_y^{2}}{4Q_{5}\Sigma} \left( F^{2}+\bar{F}^{2}\right) \,, \\
Z_{2}&= \frac{Q_{5}}{\Sigma} \,, \\
Z_{4} &= \frac{R_y}{2\Sigma} \left(F+\bar{F} \right) \,,
\end{split}
~~
\begin{split}
\Theta^1 &= 0 \,,\\
\Theta^2 &= \frac{R_y}{Q_{5}}\,F \left(A\, \Omega_{y} + B\, \Omega_{z} \right)\,+\,c.c. \,, \\
\Theta^4 &= -2\left( A\, \Omega_{y} + B\, \Omega_{z} \right)+\,c.c. \,.
\end{split}
\label{1stLayerGenHolo}\end{aligned}$$
The solution to the second BPS layer can then be written in the form: $$\begin{aligned}
\mathcal{F}&= \mathcal{F}^{(p)}+ c^{2}\mathcal{F}^{(c)}\\
\omega &=\frac{4}{\sin 2\theta}\omega_{\mu}^{(p)} \, d\theta +2\left(\omega_{\chi}^{(0)}+\omega_{\chi}^{(p)}+c^{2}\,\omega_{\chi}^{(c)} \right)\, d\varphi_{1} +2\left(\omega_{\delta}^{(0)}+\omega_{\delta}^{(p)} \right) \, d\varphi_{2}~, \label{FomegaAnsatz2}\end{aligned}$$ where $c$ is a constant.[^13] The “round supertube" part is given by: $$\begin{aligned}
\omega^{(0)}_{\chi}= \frac{\omega^{(0)}_{\delta}}{{\ensuremath{\left|\mu\right|}}^{2}} = \frac{R_{y}{\ensuremath{\left|\chi\right|}}^{2}}{2\sqrt{2}(1-{\ensuremath{\left|\chi\right|}}^{2})}~. \end{aligned}$$ The homogeneous part is given by: $$\begin{aligned}
\mathcal{F}^{(c)}&= - \frac{1}{a^{2}} \qquad \text{and} \qquad \omega^{(c)}_{\chi} = \frac{R_{y} {\ensuremath{\left|\chi\right|}}^{2}}{2\sqrt{2}a^{2}(1-{\ensuremath{\left|\chi\right|}}^{2})}~.\end{aligned}$$ Finally, the solution is completed by adding the particular part: $$\begin{aligned}
\mathcal{F}^{(p)}&= \frac{1}{a^{2}} \left({\ensuremath{\left|F_{0}\right|}}^{2}+{\ensuremath{\left|\xi\right|}}^{2} {\ensuremath{\left|F_{1}\right|}}^{2} \right) \,,
&\omega_{\chi}^{(p)}&= - \frac{R_{y} }{4\sqrt{2}a^{2}(1-{\ensuremath{\left|\chi\right|}}^{2})}\left( \bar{\chi}\bar{F}_{0}F+\chi F_{0}\bar{F}\right) \,,\\
\omega_{\mu}^{(p)} &= - \frac{iR_{y}{\ensuremath{\left|\chi\right|}}^{2}}{4\sqrt{2}a^{2}} \left( \mu \bar{F}_{0}F_{1}-\bar{\mu}F_{0}\bar{F}_{1} \right) \,,
&\omega_{\delta}^{(p)}&= \frac{R_{y} {\ensuremath{\left|\xi\right|}}^{2}}{4\sqrt{2}a^{2}(1-{\ensuremath{\left|\chi\right|}}^{2})} \left( \chi \mu F_{1}\bar{F}+\bar{\chi}\bar{\mu}\bar{F}_{1}F\right) \,.\end{aligned}$$
Tuning the asymptotic geometry {#SubSect:Asypmtotic Geom}
------------------------------
Setting $F_{0}=F_1=0$ in the $(1,m,n)$ solution of the previous section gives the “round supertube" solution, which is globally $AdS_{3}\times S^{3}$. This solution has $$\begin{aligned}
\cF = 0 \qquad \text{and} \qquad \omega = \omega_{0} = \frac{a^{2}R_{y}}{\sqrt{2}} \left(\sin^{2}\theta \, d\varphi_{1} +\cos^{2} \theta \, d\varphi_{2} \right) \,. \label{Supertube}\end{aligned}$$
To ensure the $(1,m,n)$ solutions have the same asymptotics, one must arrange for $(\cF,\omega)$ to have at most $\mathcal{O}(r^{-2})$ corrections to the round supertube solution (\[Supertube\]). We achieve this by first defining: $$\begin{aligned}
F_{0}^{(\infty)}(v) \equiv \lim_{{\ensuremath{\left|\xi\right|}}\to\infty}F_{0}(\xi)=\lim_{r\to\infty}F_{0}(\xi) \qquad \text{and} \qquad F_{1}^{(\infty)}(v) \equiv \lim_{{\ensuremath{\left|\xi\right|}}\to\infty}F_{1}(\xi)=\lim_{r\to\infty}F_{1}(\xi) ~, \label{AsymtoticFs}\end{aligned}$$ and then fixing $c^{2}$ by: $$\begin{aligned}
c^{2}\equiv \frac{1}{\sqrt{2}\pi R} \int_{0} dv \, \left( {\ensuremath{\left|F_{0}^{(\infty)}(v)\right|}}^{2}+{\ensuremath{\left|F_{1}^{(\infty)}(v)\right|}}^{2}\right)=b^{2}+d^{2}~, \label{cDef2}\end{aligned}$$ where we have implicitly defined $$\begin{aligned}
b^{2}=\sum_{n=1}^{\infty}b_{n}^{2} \qquad \text{and} \qquad d^{2}=\sum_{n=1}^{\infty}d_{n}^{2}~. \label{c01Def}\end{aligned}$$ Then, we can use a gauge transformation which leaves the six-dimensional BPS equations (\[BPSlayer0\])-(\[BPSlayer2b\]) and metric (\[ds6\]) invariant: $$\begin{aligned}
u \to u+f(v,r,\theta,\varphi_{1},\varphi_{2}) \qquad \iff \qquad \omega \to \omega - d_{4}f+ \dot{f}\beta \,, \qquad \cF\to \cF-2\dot{f}\,,\end{aligned}$$ with gauge parameter chosen as: $$\begin{aligned}
\label{eq:fdef}
f(v)\equiv \frac{1}{2a^{2}}\int_{0}^{v} dv' \, \left( {\ensuremath{\left|F_{0}^{(\infty)}(v')\right|}}^{2}+{\ensuremath{\left|F_{1}^{(\infty)}(v')\right|}}^{2}-c^{2}\right)\,,\end{aligned}$$ which brings the $(\cF,\omega)$ for the $(1,m,n)$ family to the form $$\begin{aligned}
\cF &= -\frac{1}{r^{2}} \left(d^{2}+ \sum_{n=1}n(b_{n}^{2}+d_{n}^{2}) + {\it oscillating \ terms} \right) + \mathcal{O}(r^{-4}) \,, \label{FAsym} \\
\omega &= \omega_{0} + \frac{R_{y}d^{2}}{\sqrt{2}r^{2}} \left( \sin^{2}\theta\, d\varphi_{1} +\cos^{2}\theta \, d\varphi_{2}+ {\it oscillating \ terms}\right) + \mathcal{O}(r^{-4})\,. \label{omegaAsym}\end{aligned}$$ This gauge-transformed geometry is now asymptotically the same as the round supertube, i.e. $AdS_{3}\times S^{3}$. Note that the relation (\[cDef2\]) is crucial to make the gauge parameter (\[eq:fdef\]) a well-defined, periodic function of $v$; without (\[cDef2\]) it is not possible to retrieve the correct asymptotics.
Regularity and CTC analysis {#SubSect:Regularity}
---------------------------
There are four distinct ways in which the six-dimensional metric (\[ds6\]), for the $(1,m,n)$ solution of the previous section, which also takes the form (\[eq:6Dmetricansatz\]), may fail to be regular:
- The metric is singular where the data $(\beta,\omega,\mathcal{F})$ are singular, at the locus: $$\begin{aligned}
\Sigma = 0~. \label{SigmaLocus}\end{aligned}$$
- The warp factors $(\Delta^{-1}\det m_{AB})^{\pm 1/2}$ are singular.
- The sphere deformations $\tensor{\widetilde{A}}{_{\mu}^{AB}}$ of (\[Aform\]) are singular.
- The $ds_{3}^{3}$ metric (\[ds3\])-(\[ds3Data\]) possesses a conical singularity at $r=0$, where the $y$-circle pinches off.
- $ds_{3}^{3}$ possesses closed time-like curves (CTCs).
Upon expanding and analyzing the metric along the locus (\[SigmaLocus\]), the only potentially singular part was found to be the $d\varphi_{1}^{2}$ coefficient. Setting $$\begin{aligned}
r=a \epsilon \qquad \text{and} \qquad \theta=\frac{\pi}{2}-\epsilon~,\end{aligned}$$ and expanding in powers of $\epsilon$, this term reads: $$\begin{aligned}
\frac{1}{\epsilon^{2}}\left(\frac{2 Q_{1}Q_{5}-(2a^{2}+c^{2})R^{2}}{2\sqrt{4Q_{1}Q_{5}-2{\ensuremath{\left|F_{0}\right|}}^{2}R^{2}}}\right)\, d\varphi_{1}^{2} + O(\epsilon^0)~.\end{aligned}$$ To remove this singularity one must tune: $$\begin{aligned}
\frac{Q_{1}Q_{5}}{R_{y}^{2}}=\frac{1}{g_{0}^{4}R_{y}^{2}}=a^{2}+\frac{c^{2}}{2}~. \label{RegCond}\end{aligned}$$
Now consider $$\begin{aligned}
\det m_{AB} &=\left( 1-S_{A}S_{A}\right)^{2} =\left(1- {\ensuremath{\left|z_{1}\right|}}^{2} -{\ensuremath{\left|z_{2}\right|}}^{2} \right)^{2} \,,\end{aligned}$$ where $$\begin{aligned}
z_{1}= S_{1}+iS_{2} \qquad \text{and} \qquad z_{2}=S_{3}+iS_{4}\,.\end{aligned}$$ Since $m^{AB}=(m_{AB})^{-1}$ appears in the kinetic term of the 3 dimensional Lagrangian (\[action-final\]), its solutions must necessarily bound $\det m_{AB}$ away from zero[^14]. Since $\lim_{r\to\infty}S_{A}=0$ we conclude that: $$\begin{aligned}
0<(\det m_{AB})^{1/2} = 1- {\ensuremath{\left|z_{1}\right|}}^{2} -{\ensuremath{\left|z_{2}\right|}}^{2} \,, \label{detmPositive}\end{aligned}$$ for all solutions of the action (\[action-final\]), which includes the $(1,m,n)$ solution. Now one can also calculate that $$\begin{aligned}
\Delta = 1 - {\ensuremath{\left|\tilde{z}_{1}+\tilde{z}_{2}\right|}}^{2}\,,\end{aligned}$$ where $$\begin{aligned}
\tilde{z}_{1} = (S_{1}-iS_{2})\sin\theta \, e^{-i\varphi_{1}} \qquad \text{and} \qquad \tilde{z}_{2} = -(S_{3}+iS_{4})\cos\theta \, e^{i\varphi_{2}}\,.\end{aligned}$$ The triangle inequality and (\[detmPositive\]) then imply that $$\begin{aligned}
0<\Delta\,. \end{aligned}$$ Hence the warp factors $(\Delta^{-1}\det m_{AB})^{\pm 1/2}$ are regular.
In passing, looking at the form of $\Omega^{2}$ in (\[ds3Data\]), it also follows that: $$\begin{aligned}
0<\Omega^{2}\,. \label{Omega2Positive}\end{aligned}$$ Hence the sphere deformations $\tensor{\widetilde{A}}{_{\mu}^{AB}}$ of (\[Aform\]), are clearly regular by inspection.
Setting $\rho=r/a$, the three dimensional metric can be written as: $$\begin{aligned}
ds_{3}^{2} &= \frac{1}{g_{0}^{2}}\left[\frac{d\rho^{2}}{1+\rho^{2}}-g_{0}^{8}a^{4}R_{y}^{2}(1+\rho^{2})dt^{2} +\frac{\rho^{2}}{R_{y}^{2}}\left(dy +(1-g_{0}^{4}a^{2}R_{y}^{2})dt \right)^{2} \right] \notag \\
& \qquad \qquad \qquad \qquad \qquad - \frac{g_{0}^{2}}{2}\left( {\ensuremath{\left|F_{0}\right|}}^{2}+{\ensuremath{\left|F_{1}\right|}}^{2} \right) \left[ \rho^{2}(dt+dy)^{2}+\frac{R_{y}^{2}\,d\rho^{2}}{(1+\rho^{2})^{2}} \right]\,. \label{ds3SupertubePert}\end{aligned}$$ This form of the metric makes it clear that there is no conical singularity when the $y$-circle pinches off at $\rho=0$.
Proving there are no CTCs in (\[ds3SupertubePert\]), when suitably regularized by (\[RegCond\]), and tuned to be asymptotically $AdS_{3}$ by (\[cDef2\]), is a delicate business. A proof for the $(1,0,n)$ and $(1,1,n)$ families appears in [@Heidmann:2019xrd]. It relies on properties of ${\ensuremath{\left|F_{0,1}\right|}}^{2}$ following from the analyticity of $F_{0,1}$, which do not easily generalize to the sum ${\ensuremath{\left|F_{0}\right|}}^{2}+{\ensuremath{\left|F_{1}\right|}}^{2}$, as is required for the $(1,m,n)$ family. Although we do not have a proof, we expect $ds_{3}^{2}$ to be free of CTCs when tuned with (\[cDef2\]) and (\[RegCond\]) for the $(1,m,n)$ family. This expectation is based on examining many examples with explicit expansions of $F_{0,1}$, as well as the fact that the holographic duals should be well defined CFT states. Intuitively, one can think of the second line of (\[ds3SupertubePert\]) as a “perturbation" of a regular $AdS_{3}$ seed. Upon fixing the CFT charges $Q_{1,5}$, the magnitude of ${\ensuremath{\left|F_{0,1}\right|}}^{2}$ are restricted by (\[cDef2\]) and (\[RegCond\]), and so the negative contribution coming from the “perturbation" is sufficiently controlled so as to avoid CTCs.
Conserved charges
-----------------
A detailed analysis of computing the conserved charges for the 6-dimensional superstrata appears in [@Heidmann:2019xrd], where the explicit calculations for the $(1,0,n)$ and $(1,1,n)$ multi-mode families are also given. Here we give a short summary of the procedure, which are closely analogous to the individual $(1,0,n)$ and $(1,1,n)$ family analysis, and the result. The analysis requires $c^{2}$ to be tuned as in (\[cDef2\]), so that the geometry is asymptotic to $AdS_{3}\times S^{3}$.
The D1-D5-P system possesses five conserved charges: the net charge of each type of brane $Q_{1,5}$, the momentum in the common D1-D5 direction $Q_{P}$, and the two angular momenta $J_{L,R}$. The brane charges $Q_{1,5}$ can be simply read off from the $r^{-2}$ coefficient of $Z_{1,2}$ in (\[1stLayerGenHolo\]), when expanded about $r\to\infty$.
Using the solution as presented in the gauge of (\[FAsym\])-(\[omegaAsym\]), with $c^{2}$ fixed by (\[cDef2\]), the remaining charges can be read off from the expansions: $$\beta_1+\beta_2 +\omega_1 + \omega_2~=~ \frac{ \sqrt{2}}{r^2}\, \big[ \, (J_R - J_L \cos 2\theta ) ~+~ {\it oscillating \ terms}\, \big] ~+~ \mathcal{O}(r^{-4})
\label{asmpmoms}$$ where $ \beta_1, \omega_1$ and $ \beta_2, \omega_2$ are the components of $\beta$ and $\omega$ along $d\varphi_1$ and $d\varphi_2$ respectively, and $$\mathcal{F} ~=~ - \frac{1}{r^2} \,\big(2\, Q_P ~+~{\it oscillating \ terms} \big) ~+~ \mathcal{O}(r^{-4}) \,.
\label{Fexp}$$ For the $(1,m,n)$ multi-mode solution this procedure gives (using (\[c01Def\])): $$\begin{aligned}
J_{L}=\frac{a^{2}R_{y}}{2} \,, \qquad \qquad J_{R} = \frac{R_{y}}{2}\left(a^{2}+d^{2} \right) \,, \qquad Q_{P} = \frac{1}{2}\left[ d^{2}+\sum_{n=1}^{\infty} n \left(b_{n}^{2} + d_{n}^{2} \right) \right]\,.\end{aligned}$$
Uplifting the six-dimensional scalars to ten dimensions
-------------------------------------------------------
The parametrization of the six-dimensional scalars $\varphi,X$ that we used in the uplift formulae in Section \[sec:full6Duplift\] is convenient from the point of view of the six-dimensional reduction, as $\varphi$, resp. $X$, only depend on $m_{AB}$, resp. $\chi_A$. However, for the D1-D5 system, they require some rearranging to consider the uplift to ten dimensions; in particular, the ten-dimensional dilaton is not given by $\varphi$. With the following rearrangement of the scalars, introducing an arbitrary constant $\tilde{Z}_2$: e\^[2\^]{} = , C\_[(0)]{}\^=-\_2,the six-dimensional scalar kinetic action becomes: \_[6D,]{} = -12 ()\^2 -12 e\^(X)\^2 = -(\^)\^2 - e\^[2\^]{}(C\_[(0)]{}\^)\^2.Then, for the superstrata uplift to ten dimensions (using the conventions of, for example, Appendix B of [@deLange:2015gca]), $\phi^{\text{(10D)}}$ can be identified with the ten-dimensional dilaton after uplifting on a $\IT^4$ while $C_{(0)}^{\text{(10D)}}$ is the ten-dimensional axion. Setting $\tilde{Z}_2 = Z_2(g_0^2\Sigma)$ then gives the traditional form for these scalars: e\^[2\^]{} = , C\_[(0)]{}\^ = ,whereas the six-dimensional fields used in Section \[sec:full6Duplift\] were simply: e\^[-]{} = (g\_0\^2)\^2, X = -(g\_0\^2) Z\_4.
Three-dimensional equations of motion {#App:3Deom}
=====================================
Here, we will give for reference the equations of motion following from the three-dimensional action (\[action-final\]).
The scalar equations of motion are: $$\begin{aligned}
0&= \cD_\mu \cD^\mu \chi_A
- \cD^\mu\chi_B \left(\varepsilon\, \epsilon{\indices{_\mu^{\nu\rho}}}F^{BC}_{\nu\rho}m_{CA}
+m^{BC} \cD_\mu m_{CA}\right)\\
\nonumber & -(\det m^{AB})\left( 8\varepsilon\, \alpha \gamma_0 m_{AB}\chi_B +4\gamma_0^2 m_{AB}m_{BC}\chi_C + 2\gamma_0^2m_{AB}\chi_B (\chi_C\chi_C) \right),\\
0 &= \cD_\mu \cD^\mu m_{AB}
+ \cD_\mu \chi_A \cD^\mu \chi_B
-m^{CD}\cD^\mu m_{AC} \cD_\mu m_{BD}
-2 F^{CD}_{\mu\nu} F^{EF\, \mu\nu} m_{AC}m_{BE}m_{DF}\\
\nonumber &+ (\det m^{AB})\left( \left[16\alpha^2+8\gamma_0^2m_{CD}m_{CD}-4\gamma_0^2m_{CC}m_{DD}\right]m_{AB} \right.\\
\nonumber & \left. +8\gamma_0^2m_{CC}m_{AD}m_{BD} -16\gamma_0^2m_{CD}m_{AC}m_{BD} \right.\\
\nonumber & \left.+ 8\varepsilon\, \gamma_0\alpha m_{AB} \chi_C\chi_C + 4\gamma_0^2(m_{CD}m_{AB}-m_{AC}m_{BD})\chi_C\chi_D +\gamma_0^2m_{AB}(\chi_C\chi_C)^2 \right),\end{aligned}$$ the Einstein equations are: $$\begin{aligned}
R_{\mu\nu} - \frac12 R g_{\mu\nu} &= 2 g_{\mu\nu}\left( -~\coeff{1}{16}\, {\rm Tr}\big[ \big( \cD_\rho m \big) m^{-1}\, \big( \cD^\rho m \big) m^{-1}\, \big] \right.\\
\nonumber & \left. -\coeff{1}{8}\, m^{AB} \,(\cD_\rho\chi_A)\,(\cD^\rho \chi_B) ~-~V - \coeff{1}{8} m_{AC} \,m_{BD}\, F_{\rho \sigma}^{AB}\, F^{\rho \sigma}{}^{CD} \right)\\
\nonumber & + \frac14 m^{AC}m^{BD}\cD_\mu m_{AB}\cD_\nu m_{CD} + \frac12 m^{AB}\cD_\mu \chi_A \cD_\nu \chi_B + m_{AC}m_{BD} F{\indices{^{AB}_{\mu}^{\rho}}}F^{CD}_{\nu\rho},\end{aligned}$$ and finally the gauge field equations of motion are: $$\begin{aligned}
0&= m_{AC}m_{BD}\cD_\nu F{\indices{^{CD}_\mu^\nu}} + \left( F^{CD}_{\mu\nu} m_{BD} \cD^\nu m_{AC} - F^{CD}_{\mu\nu} m_{AD} \cD^\nu m_{BC} \right)\\
\nonumber & +\gamma_0\epsilon_{ABCD}m^{DE}( \cD_\mu m_{CE} + \chi_C \cD_\mu \chi_E) - \alpha\epsilon_{ABCD} \epsilon{\indices{_\mu^{\nu\rho}}} F^{CD}_{\nu\rho}\\
\nonumber & +\frac12\varepsilon\left( -\epsilon_{\mu\nu\rho} \cD^\nu\chi_A \cD^\rho \chi_B +\frac12 \epsilon_{\mu\nu\rho} (\chi_B\cD^\nu \cD^\rho\chi_A-\chi_A\cD^\nu \cD^\rho\chi_B) - \gamma_0\epsilon{\indices{_\mu^{\nu\rho}}}\epsilon_{ABCD}F^{ED}_{\nu\rho}\chi_E\chi_C \right) . \end{aligned}$$
[-1mm]{}[-1mm]{}
[^1]: The remaining directions of the D5 are compactified on the $\IT^4$ that reduces IIB supergravity to six dimensions.
[^2]: Here we are going to consider the lowest KK towers in the compactification. Remarkably, it seems that one can consistently truncate in a manner that allows higher modes in the KK towers [@Nicolai:2003ux]. As we will discuss in Section \[sec:Conclusions\], this might prove immensely useful in using three-dimensional supergravity to construct much more general classes of superstrata.
[^3]: This “double counts” the generators because $T^{\bar P \bar Q} =-T^{\bar Q \bar P}$. This is, however, a completely standard convention that we use everywhere in this paper.
[^4]: This truncation is trivially achieved by by imposing an invariance under the $SO(3)$ that acts of three of the five anti-self-dual tensor multiplets.
[^5]: We conform to the idiosyncratic notation and conventions for the six-dimensional three-forms that is used in the superstrata literature. This slightly odd notation of omitting the index 3 is historical. In reduction of the six-dimensional system to five dimensions the $F$ and $d\beta$ fields are identified with $Z_{3}$ and $\Theta_{3}$ respectively.
[^6]: Our conventions for the 6D Hodge dual are given explicitly in (\[eq:6Dselfduality\]). While never explicitly mentioned in [@Samtleben:2019zrh], their convention for Hodge duals is such that their self-duality relation receives a relative minus sign compared to ours in (\[eq:6Dselfduality\]).
[^7]: Our six-dimensional Hodge dual conventions are given explicitly in (\[eq:6Dselfduality\]); for completeness, note that we take the six-dimensional Levi-Civita tensor to decompose as $\epsilon_{\mu\nu\rho ijk} = + \epsilon_{\mu\nu\rho} \epsilon_{ijk}$.
[^8]: Note that all Hodge stars in (\[eq:U1sqG1\])-(\[eq:U1sqG4\]) refer to three-dimensional Hodge stars with metric $ds_3^2$.
[^9]: These modes are restricted by regularity or equivalently CFT considerations, see [@Shigemori:2020yuo] for a discussion.
[^10]: An interesting perspective can also be gained from introducing the complex combinations: $$\begin{aligned}
z_{1}= S_{1}+iS_{2} \qquad \text{and} \qquad z_{2}=S_{3}+iS_{4}\,\end{aligned}$$ which simplifies some of the following expressions since: $$\begin{aligned}
S_{1}^{2}+S_{2}^{2}={\ensuremath{\left|z_{1}\right|}}^{2}\,, \qquad S_{3}^{2}+S_{4}^{2}={\ensuremath{\left|z_{2}\right|}}^{2}\,, \qquad S_{1}S_{3}-S_{2}S_{4} = \Re \left\lbrace z_{1}z_{2} \right\rbrace \,, \qquad S_{1}S_{4}+S_{2}S_{3} = \Im \left\lbrace z_{1}z_{2} \right\rbrace\,.\end{aligned}$$
[^11]: Note that this corresponds to choosing $\varepsilon=+1$ in (\[eq:6Dselfduality\]). Also note that our Hodge dual conventions (as given in (\[eq:6Dselfduality\])) imply that there should indeed be two anti-self-dual tensors and one self-dual tensor for the superstrata.
[^12]: Significant process was made in solving this layer in general in [@Tyukov:2018ypq]. For any harmonic functions $\Phi_{\hI}$ on $\mathcal{B}$, one can derive from them and a complex structure a self dual two forms $\Theta^{\hI}$. If it is known what modulus of $\mathcal{B}$ these two form control as a Kähler deformation, then the $Z_{\hI}$ which solve the fist BPS layer together with these $\Theta^{\hI}$ can be found directly from the $\Phi_{\hI}$.
[^13]: Note that we introduce this constant of integration as $c^2$, whereas $c$ (i.e. unsquared) was used in [@Heidmann:2019xrd]. As we will see in (\[cDef2\]), $c^2$ is naturally a positive number.
[^14]: This argument is plausible, rather than providing a strict proof. The action would become infinite wherever $\det m_{AB}=0$, so the minimization procedure should ensure solutions avoid this condition.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We use three-dimensional smoothed particle hydrodynamics simulations together with a dynamical ray-tracing scheme to investigate the build-up of the first H [ii]{} regions around massive Population III stars in minihaloes. We trace the highly anisotropic breakout of the ionizing radiation into the intergalactic medium, allowing us to predict the resulting recombination radiation with greatly increased realism. Our simulations, together with Press-Schechter type arguments, allow us to predict the Population III contribution to the radio background at $\sim 100~{\rm MHz}$ via bremsstrahlung and 21-cm emission. We find a global bremsstrahlung signal of around $1~{\rm mK}$, and a combined 21-cm signature which is an order of magnitude larger. Both might be within reach of the planned Square Kilometer Array experiment, although detection of the free-free emission is only marginal. The imprint of the first stars on the cosmic radio background might provide us with one of the few diagnostics to test the otherwise elusive minihalo star formation site.'
author:
- |
Thomas H. Greif$^{1,4}$[^1], Jarrett L. Johnson$^{2,3}$, Ralf S. Klessen$^{1}$ and Volker Bromm$^{2,3}$\
$^{1}$ Zentrum für Astronomie der Universität Heidelberg, Institut für Theoretische Astrophysik,\
Albert-Ueberle-Straße 2, 69120 Heidelberg, Germany\
$^{2}$ Department of Astronomy, University of Texas, Austin, TX 78712, USA\
$^{3}$ Texas Cosmology Center, University of Texas, Austin, TX 78712, USA\
$^{4}$ Fellow of the International Max Planck Research School for Astronomy and Cosmic Physics at the University of Heidelberg
title: 'The observational signature of the first H [ii]{} regions'
---
-1cm
cosmology: observations – cosmology: theory – early Universe – stars: formation
Introduction
============
One of the most important questions in modern cosmology is to understand how the first stars, the so-called Population III (Pop III), ended the cosmic dark ages at redshifts $z\la 30$ [@bl01; @bl04a; @cf05]. Their emergence led to a fundamental transformation in the early Universe, from its simple initial state to one of ever-increasing complexity. The emission from the hot, $T_{\rm eff}\sim 10^5~{\rm K}$, photospheres of Pop III stars began the reionization of primordial hydrogen and helium in the intergalactic medium (IGM), although this process was completed only later on, when more massive galaxies formed [@fck06]. In addition, the supernova explosions that ended the lives of massive Pop III stars distributed the first heavy elements into the IGM [@byh03; @greif07; @tfs07; @wa08b]. This latter process might have had a significant impact on the physics of early star formation, as metal-enriched gas can cool more efficiently than primordial gas [@bl03a; @omukai05; @jappsen07a; @jappsen09a].
Based on numerical simulations, a general consensus has emerged that the first stars formed in dark matter minihaloes at $z\sim 20 - 30$, in isolation or at most as a small stellar multiple, and with typical masses of $M_{*}\sim 100~{\rm M}_{\odot}$ [for a recent review, see @bromm09]. It is crucial to observationally test this key prediction. However, it has become evident that this will be very challenging. Even the exquisite near-IR ($\sim{\rm nJy}$) sensitivity of the upcoming [*James Webb Space Telescope (JWST)*]{} will not suffice to directly image such massive, single Pop III stars [@bkl01; @gardner06], unless they explode as energetic pair-instability supernovae [@hw02; @scannapieco05a]. The direct spectroscopic detection of recombination line emission from the H [ii]{} region surrounding the Pop III star, as well as from the relic H [ii]{} region left behind once the star had died, is beyond the capability of [*JWST*]{} as well, although such line emission might be detectable from primordial stellar populations inside more massive host haloes [@schaerer02; @schaerer03; @johnson09].
An alternative approach is to search for the global signature from many Pop III stars that formed in minihaloes over large cosmic volumes [@hl97]. One such probe is the optical depth to Thomson scattering of cosmic microwave background (CMB) photons off free electrons along the line of sight, determined by the five-year [*Wilkinson Microwave Anisotropy Probe (WMAP)*]{} measurement to be $\tau_{e}\simeq 0.09\pm 0.02$ [@komatsu09]. This signal, however, is dominated by ionizing sources that must have formed closer to the end of reionization, with only a small contribution from Pop III stars formed in minihaloes [@gb06; @sbk08]. A second empirical signature is the combined bremsstrahlung emission from the H [ii]{} regions in their active and relic states around those minihaloes that hosted Pop III stars. The resulting free-free radio emission leads to spectral distortions that might be detectable in the Rayleigh-Jeans part of the CMB spectrum. Recently, the ARCADE 2 experiment has attempted to measure such a free-free contribution from the epoch of the first stars [@kogut06]. The surprisingly strong signal found, however, cannot originate in early Pop III stars, and in any case would overwhelm the much weaker contribution from the first stars and galaxies [@seiffert09]. The most promising detection strategy might be to scrutinize the background from the redshifted 21-cm line of neutral hydrogen [@fob06]. Once the central Pop III star has died, the relic H [ii]{} region left behind would provide a bright source of 21-cm emission [@tokutani09]. Again, individual sources are much too weak to leave a detectable imprint, but the planned Square Kilometer Array (SKA) might be able to detect the cumulative signal [@furlanetto06; @lazio08].
We here carry out radiation hydrodynamics simulations of the evolution of H [ii]{} regions around massive Pop III stars in minihaloes, giving us a detailed understanding of the properties of individual sources. We combine this with an approximate, Press-Schechter type analysis of the cosmological number density of minihaloes as a function of redshift to derive the observational signature of the first H [ii]{} regions as well as relic H [ii]{} regions, where we specifically focus on the free-free and 21-cm probes. We note that we do not include the feedback effects exerted by black holes or supernovae, which are possible end products of massive Pop III stars [@hw02]. In this sense, we organize our work as follows. In Section 2, we describe the simulation setup and our implementation of the radiative transfer scheme in the smoothed particle hydrodynamics (SPH) code [GADGET]{}-2 [@springel05]. In Section 3, we discuss the properties of the first H [ii]{} regions in their active as well as relic states and their observational signature in terms of recombination radiation, bremsstrahlung and 21-cm emission. Finally, in Section 4 we summarize our results and assess their implications. For consistency, all quoted distances are physical, unless noted otherwise.
Numerical methodology
=====================
Our treatment of ionizing and photodissociating radiation emitted by massive Pop III stars is very similar to the methodology introduced in @jgb07 and @yoshida07, with the exception that we here take the hydrodynamical response into account, self-consistently coupled to the chemical and thermal evolution of the gas. This allows us to model dense (D-type) as well as rarefied (R-type) ionization fronts, which is crucial for a proper treatment of the breakout of ionizing radiation. In the following, we describe our simulation setup, as well as the numerical implementation of the ray-tracing algorithm.
Simulation set-up
-----------------
We perform our simulations in a cosmological box with linear size $200~{\rm kpc}$ (comoving), and $256^{3}$ particles per species, corresponding to a particle mass of $\simeq 17~{\rm M}_{\odot}$ for dark matter and $\simeq 3~{\rm M}_{\odot}$ for gas. The simulations are initialized at $z=99$ with a fluctuation power spectrum determined by a $\Lambda$ cold dark matter ($\Lambda$CDM) cosmology with matter density $\Omega_{m}=1-\Omega_{\Lambda}=0.27$, baryon density $\Omega_{b}=0.046$, Hubble parameter $h=H_{0}/100~{\rm km}~{\rm s}^{-1}~{\rm Mpc}^{-1}=0.7$, where $H_{0}$ is the Hubble expansion rate today, and spectral index $n_{s}=0.96$ [@komatsu09]. We use an artificially high fluctuation power of $\sigma_{8}=1.6$ to accelerate structure formation in our relatively small box, although the cosmological mean is given by $\sigma_{8}=0.81$. We take the chemical evolution of the gas into account by following the abundances of H, H$^{+}$, H$^{-}$, H$_{2}$, H$_{2}^{+}$, He, He$^{+}$, He$^{++}$, and e$^{-}$, as well as the three deuterium species D, D$^{+}$, and HD. We include all relevant cooling mechanisms, i.e. H and He collisional ionization, excitation and recombination cooling, bremsstrahlung, inverse Compton cooling, and collisional excitation cooling via H$_{2}$ and HD [@gj07]. We explicitly include H$_{2}$ cooling via collisions with protons and electrons, which is important for the chemical and thermal evolution of relic H [ii]{} region gas [@ga08].
We run the simulations until the first minihalo in the box has collapsed to a density of $n_{\rm H}=10^{4}~{\rm cm}^{-3}$, at which point the gas has cooled to $\simeq 200~{\rm K}$ and becomes Jeans-unstable [@abn02; @bcl02]. The first halo that fulfils this criterion collapses at $z_{*}\simeq 20$ and has a virial mass of $\simeq 9.4\times 10^{5}~{\rm M}_{\odot}$ and a virial radius of $\simeq 90~{\rm pc}$. Highly resolved simulations have shown that at later times, the gas condenses further under the influence of self-gravity to $n_{\rm H}\sim 10^{21}~{\rm cm}^{-3}$, where it becomes optically thick and forms a protostellar seed [@yoh08]. Due to its residual angular momentum, the central clump flattens and likely evolves into an accretion disk. In our case, we find a flattened structure already at a density of $n_{\rm H}=10^{4}~{\rm cm}^{-3}$ (see Fig. 1). Subsequently, the star grows to as massive as $\sim 100~{\rm M}_{\odot}$ within its lifetime of a few million years [@bl04b]. However, we note that under certain conditions the disk may fragment to form multiple objects of smaller masses [@cgk08]. Unfortunately, the details of the accretion phase and the concomitant radiative feedback are poorly understood, although some analytic investigations have been carried out [@tm04; @mt08]. Under these circumstances, it seems best to initialize the calculation of the H [ii]{} region at the onset of the initial Jeans-instability, when the density exceeds $n_{\rm H}=10^{4}~{\rm cm}^{-3}$.
Ray-tracing scheme
------------------
The procedure used to calculate the Strömgren sphere around the star for a given time-step $\Delta t$ is similar to the ray-tracing scheme used in @jgb07. We first designate an individual SPH particle as the source of ionizing radiation and create a spherical grid with typically $10^{5}$ rays and $500$ logarithmically spaced radial bins around the source particle. The minimum radius is set by the smoothing length of the central particle, while the maximum radius is chosen appropriately to encompass the entire H [ii]{} region. This approach may seem crude compared to existing methods that use adaptive grids [e.g. [HEALPIX]{}; @gorski05], but the increased angular and radial resolution towards the center tend to mirror the existing density profile. However, one must proceed with care if the ionization front encounters dense clumps far from the source, where the resolution may no longer be sufficient.
In a single, parallel loop, the Cartesian coordinates of all particles are converted to spherical coordinates, such that their density and chemical abundances are mapped to the bins corresponding to their radius, zenith angle and azimuth, denoted by $r$, $\theta$ and $\phi$, respectively. The volume of each particle is approximately given by $\Delta V\simeq h^{3}$, which transforms to $\Delta r=h$, $\Delta\theta=h/r$ and $\Delta\phi=h/(r\sin \theta)$. If the volume element of a particle intersects with the volume element of a bin, the particle contributes to the bin proportional to the density of the particle squared. This dependency ensures that overdense regions are not missed if the bin size is much larger than the smoothing length, which could occur far from the source where the grid resolution is poor. Accidental flash-ionization of minihaloes is thus avoided. Once the above steps are complete, it is straightforward to solve the ionization front equation along each ray: $$n_{n}r_{\rm I}^{2}\frac{{\rm d}r_{\rm I}}{{\rm d}t}=\frac{\dot{N}_{\rm ion}}{4\pi}-\alpha_{\rm B}\int_{0}^{r_{\rm I}}n_{e}n_{+}r^{2}{\rm d}r\mbox{\ ,}$$ where $r_{\rm I}$ denotes the position of the ionization front, $\dot{N}_{\rm ion}$ the number of ionizing photons emitted per second, $\alpha_{\rm B}$ the case B recombination coefficient, and $n_{n}$, $n_{e}$ and $n_{+}$ the number densities of neutral particles, electrons and positively charged ions, respectively. We assume that the recombination coefficient remains constant at its value for $10^{4}~{\rm K}$, which is roughly the temperature of the H [ii]{} and He [iii]{} region.
The numbers of H [i]{}/He [i]{} and He [ii]{} ionizing photons are given by $$\dot{N}_{\rm ion}=\frac{\pi L_{*}}{\sigma T_{\rm eff}^{4}}\int_{\nu_{\rm min}}^{\infty}\frac{B_{\nu}}{h\nu}{\rm d}\nu\mbox{\ ,}$$ where $h$ from now on denotes Planck’s constant, $\sigma$ denotes Boltzmann’s constant, and $\nu_{\rm min}$ is the minimum frequency corresponding to the ionization threshold of H [i]{} and He [ii]{}. We assume that massive Pop III stars emit a blackbody spectrum $B_{\nu}$ (in ${\rm erg}~{\rm s}^{-1}~{\rm cm}^{-2}~{\rm Hz}^{-1}~{\rm sr}^{-1}$) with an effective temperature $T_{\rm eff}={\rm dex}(4.922$, $4.975$, $4.999)~{\rm K}$ and luminosity $L_{*}={\rm dex}(5.568$, $6.095$, $6.574)~{\rm L}_{\odot}$ for a $50$, $100$ and $200~{\rm M}_{\odot}$ star, respectively [@schaerer02]. This yields $$\dot{N}_{\rm ion,HI/HeI}=[2.80,9.14,26.99]\times 10^{49}~{\rm s}^{-1}$$ and $$\dot{N}_{\rm ion,HeII}=[0.72,4.14,15.43]\times 10^{48}~{\rm s}^{-1}\mbox{\ .}$$ We do not distinguish between the H [ii]{} and He [ii]{} region, which is a good approximation for massive Pop III stars [@of06]. The lifetimes of the stars are given by $t_{*}=3.7$, $2.7$ and $2.2~{\rm Myr}$, respectively. We neglect the effects of stellar evolution, which might lead to a decrease of the number of ionizing photons emitted at the end of the main sequence [@marigo01; @schaerer02], although recent investigations have shown that rotating Pop III stars remain on bluer evolutionary tracks and this effect might not be so strong [@yl05; @wh06; @vazquez07].
To obtain a discretisation of the ionization front equation, we replace the integral on the right-hand side of equation (1) by a discrete sum: $$\int_{0}^{r_{\rm I}}n_{e}n_{+}r^{2}{\rm d}r\simeq\sum_{i}n_{e,i}n_{+,i}r_{i}^{2}\Delta r_{i}\mbox{\ ,}$$ where $\Delta r_{i}$ is the radial extent of bin $i$, and the sum extends from the origin to the position of the ionization front at the end of the current time-step $\Delta t$. The above equation describes the advancement of the ionization front due to an excess of ionizing photons compared to recombinations. Similarly, the left-hand side of equation (1), which models the propagation of the ionization front into neutral gas, is discretized by $$n_{n}r_{\rm I}^{2}\frac{{\rm d}r_{\rm I}}{{\rm d}t}\simeq\frac{1}{\Delta t}\sum_{i}n_{n,i}r_{i}^{2}\Delta r_{i} \mbox{\ ,}$$ where the sum now extends from the position of the ionization front at the previous time-step to its position at the end of the current time-step. We perform the above steps separately for the H [ii]{} and He [iii]{} region, since they require distinct heating and ionization rates. For the He [iii]{} region, we replace the quantities $n_{n}$ and $n_{+}$ in equation (1) with $n_{n}=f_{\rm HeII}\,n_{\rm H}$ and $n_{+}=f_{\rm HeIII}\,n_{\rm H}$, where $f_{X}$ is the number density of species $X$ relative to $n_{\rm H}$. We adopt a case B recombination rate of $\alpha_{\rm B}=1.3\times 10^{-12}~{\rm cm}^{3}~{\rm s}^{-1}$ for He [iii]{} recombinations to He [ii]{} [@of06]. Applying the same prescription to the H [ii]{} region, we find $n_{n}=(f_{\rm HI}+f_{\rm HeI})\,n_{\rm H}$ and $n_{+}=(f_{\rm HII}+f_{\rm HeII})\,n_{\rm H}$. Similarly, we adopt a case B recombination rate of $\alpha_{\rm B}=2.6\times 10^{-13}~{\rm cm}^{3}~{\rm s}^{-1}$ for hydrogen and helium recombinations from their first ionized states to the ground state [@of06]. We initialize the calculation of the H [ii]{} region at the boundary of the He [iii]{} region, since hydrogen and helium are maintained in their first ionization states by recombinations of He [iii]{} to He [ii]{} [@of06]. We note that the exact position of the ionization front is not restricted to integer multiples of our pre-defined radial bins, but may instead lie anywhere in between. For this purpose we adopt a simple linear scaling of the number of ionizations and recombinations as a function of the relative position of the ionization front. The most expensive step in terms of computing time is the assignment of the density and the chemical abundances to the grid, while the ray-tracing itself requires only a negligible amount of time.
Photoionization and photoheating
--------------------------------
Once the extent of the H [ii]{} and He [iii]{} region have been determined, the SPH particles within these regions are assigned an additional variable that stores their distance from the source. This information is then passed to the chemistry solver, which determines the ionization and heating rates, given by $$k_{\rm ion}=\int_{\nu_{\rm min}}^{\infty}\frac{F_{\nu}\sigma_{\nu}}{h\nu}{\rm d}\nu$$ and $$\Gamma=n_{n}\int_{\nu_{\rm min}}^{\infty}F_{\nu}\sigma_{\nu}\left(1-\frac{\nu_{\rm min}}{\nu}\right){\rm d}\nu\mbox{\ ,}$$ where $F_{\nu}$ and $\sigma_{\nu}$ denote the incoming specific flux and ionization cross section, respectively. For the case of a blackbody, $$F_{\nu}=\frac{L_{*}}{4\sigma T_{\rm eff}^{4}r^{2}}B_{\nu}\mbox{\ ,}$$ where $r$ is the distance from the source. The resulting rates are given by $$k_{\rm ion,HI}=\frac{[0.45,1.32,3.69]\times 10^{-6}}{\left(r/{\rm pc}\right)^{2}}~{\rm s}^{-1}\mbox{\ ,}$$ $$k_{\rm ion,HeI}=\frac{[0.42,1.43,4.29]\times 10^{-6}}{\left(r/{\rm pc}\right)^{2}}~{\rm s}^{-1}\mbox{\ ,}$$ $$k_{\rm ion,HeII}=\frac{[0.67,3.72,13.57]\times 10^{-8}}{\left(r/{\rm pc}\right)^{2}}~{\rm s}^{-1}\mbox{\ ,}$$ $$\Gamma_{\rm HI}=n_{\rm HI}\frac{[0.40,1.28,3.74]\times 10^{-17}}{\left(r/{\rm pc}\right)^{2}}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{-3}\mbox{\ ,}$$ $$\Gamma_{\rm HeI}=n_{\rm HeI}\frac{[0.41,1.57,4.94]\times 10^{-17}}{\left(r/{\rm pc}\right)^{2}}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{-3}\mbox{\ ,}$$ $$\Gamma_{\rm HeII}=n_{\rm HeII}\frac{[0.72,4.46,17.13]\times 10^{-19}}{\left(r/{\rm pc}\right)^{2}}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{-3}$$ for a $50$, $100$ and $200~{\rm M}_{\odot}$ Pop III star, respectively. These are taken into account every time-step, while the ray-tracing is performed only every fifth time-step. Since the hydrodynamic time-step is generally limited to one-twentieth of the sound-crossing time through the kernel, our treatment of the coupled evolution of the ionization front and the hydrodynamic shock is roughly correct. The computational cost of runs with and without ray-tracing are typically within a factor of a few.
Photodissociation and photodetachment
-------------------------------------
The final ingredient in our algorithm is the inclusion of molecule-dissociating radiation. This effect turns out to be of only minor importance in the present study, but will render our algorithm capable of addressing a general set of early Universe applications. Molecular hydrogen is the most important coolant in low-temperature, primordial gas, but is easily destroyed by radiation in the Lyman-Werner (LW) bands between $11.2$ and $13.6~{\rm eV}$. The small residual H$_{2}$ fraction in the IGM leads to a very small optical depth over cosmological distances, such that even a small background can have a significant effect [@har00; @gb01; @jgb07]. In our implementation, we do not take self-shielding into account, which becomes important for H$_{2}$ column densities $\ga 10^{14}~{\rm cm}^{-2}$ [@db96]. Such a high column density is difficult to achieve in minihaloes, and is more likely to occur within the virial radius of the first galaxies [@oh02]. However, the onset of turbulence in the first galaxies likely leads to a reduction of self-shielding via Doppler shifting [@wa07b; @greif08]. For this reason we treat the photodissociation of H$_{2}$ in the optically thin limit, such that the dissociation rate in a volume limited by causality to a radius $r=c\,t_{*}$ is given by $k_{{\rm H}_{2}}=1.1\times 10^{8}F_{\rm LW}~{\rm s}^{-1}$, where $F_{\rm LW}$ is the integral of the specific flux $F_{\nu}$ over the LW bands, resulting in $$k_{{\rm H}_{2}}=\frac{[1.27,3.38,9.07]\times 10^{-7}}{\left(r/{\rm pc}\right)^{2}}~{\rm s}^{-1}$$ for a $50$, $100$ and $200~{\rm M}_{\odot}$ Pop III star, respectively. We equate the photodissociation rate of hydrogen deuteride to that of molecular hydrogen. In the present work, we do not explicitly include photodetachment of H$^{-}$ and photodissociation of H$^{+}_{2}$, which might be problematic outside of the H [ii]{} region, where molecules survive collisional destruction. However, in the context discussed here, this caveat is not important [see @jgb07].
Observational signature
=======================
In the following, we discuss the direct observational signature of the first H [ii]{} regions and relic H [ii]{} regions in terms of recombination radiation, as well as their indirect signature in terms of a global radio background produced by bremsstrahlung and 21-cm emission.
Build-up of H [ii]{} and He [iii]{} region
------------------------------------------
The build-up of the first H [ii]{} regions by Pop III stars in minihaloes was treated in one dimension by @kitayama04 and @wan04, and in three dimensions by @abs06, @awb07 and @yoshida07. The latter also treated the build-up of a smaller He [iii]{} region, which is created by the very hard spectrum of massive Pop III stars. The consensus was that recombinations initially balanced ionizations within the virial radius of the host halo, leading to the formation of a D-type ionization front. Breakout occured after the density dropped sufficiently for the ionization front to race ahead of the hydrodynamic shock, becoming R-type. The hydrodynamic response of the gas is self-similar, since minihaloes approximately resemble singular isothermal spheres [@shu02; @abs06]. The relevant parameters are set by the temperature of the singular isothermal sphere and the H [ii]{} region, which in our case are $T \simeq 200$ and $\simeq 10^{4}~{\rm K}$, respectively.
In Fig. 2, we compare the density profile of the @shu02 solution to the simulation for the case of a $100~{\rm M}_{\odot}$ Pop III star. Interestingly, we find a clear deviation from the ideal, spherically symmetric solution already during the D-type phase, which is caused by the anisotropic collapse of the minihalo. Due to angular momentum conservation, the gas spins up and forms a flattened, disk-like structure at a density of $10^{4}~{\rm cm}^{-3}$, which can be seen in the right panel of Fig. 1, and in the left panel of Fig. 2, where it is evident that the density dispersion is almost an order of magnitude within the central $\simeq 10~{\rm pc}$. In response to this anisotropy, which is further amplified by the density-squared dependence of recombinations, the ionization front first breaks out perpendicular to the disk, where the column density is lowest. This is visible in the left and middle panels of Fig. 3, as well as in Fig. 2, where the @shu02 solution is approximately reproduced perpendicular to the disk, while the plane of the disk remains neutral and dense. Once the ionization front becomes R-type, spherical symmetry is asymptotically restored and the H [ii]{} region expands to $r_{\rm HII}\simeq 1.9$, $2.7$ and $3.7~{\rm kpc}$ for the $50$, $100$ and $200~{\rm M}_{\odot}$ Pop III star, respectively. We find that He [ii]{} ionizing photons within the He [iii]{} region increase the central temperature by a factor of $\simeq 1.5$, leaving only a small imprint on the dynamical evolution of the H [ii]{} region [see @yoshida07]. However, the He [ii]{} $\lambda 1640$ recombination line within the He [iii]{} region may be used as a distinct probe for the presence of massive Pop III stars [@bkl01; @oh01; @tgs01; @schaerer02]. In the following, we use the results obtained in this section to determine the recombination signature of the first H [ii]{} and He [iii]{} regions in their active as well as relic states.
Recombination radiation from individual H [ii]{} and He [iii]{} regions
-----------------------------------------------------------------------
The strongest direct signature of the first H [ii]{} and He [iii]{} regions is likely generated by recombination radiation, since ionizing photons are absorbed by dense gas in the host halo. We here concentrate on the H$\alpha$ and He [ii]{} $\lambda 1640$ lines, since more energetic photons photons are scattered out of resonance by the neutral IGM, creating extended haloes around high-redshift sources [e.g. @lr99]. The resulting fluxes may then be compared to the expected sensitivity of the Mid-Infrared Instrument (MIRI) on [*JWST*]{} at $\sim 10~\mu{\rm m}$ wavelengths [@gardner06]. The spatial resolution is limited by diffraction, such that a scale of $\simeq 1~{\rm kpc}$ at $z=20$ is marginally resolved, which allows us to approximate the region of emission as a point source. Using the simulation output, the total luminosities are given by $$L_{{\rm H}\alpha}=j_{{\rm H}\alpha}\sum_{i}\frac{m_{i}}{\rho_{i}}\left(\frac{X\rho_{i}}{m_{\rm H}}\right)^{2}f_{e,i}f_{{\rm HII},i}$$ and $$L_{1640}=j_{1640}\sum_{i}\frac{m_{i}}{\rho_{i}}\left(\frac{X\rho_{i}}{m_{\rm H}}\right)^{2}f_{e,i}f_{{\rm HeIII},i}\mbox{\ ,}$$ where $j_{{\rm H}\alpha}$ and $j_{1640}$ are the emissivity of the H$\alpha$ and He [ii]{} $\lambda 1640$ lines at $10^{4}~{\rm K}$ [@of06], $X=0.76$ is the primordial mass fraction of hydrogen, $m_{\rm H}$ is the mass of the hydrogen atom, $m_{i}$ and $\rho_{i}$ are the mass and density of particle $i$, respectively, and the sum is over all particles in the simulation box. From the total luminosity, we determine the observed flux with the inverse-square law $$F=\frac{L}{4\pi D_{\rm L}^{2}}\mbox{\ ,}$$ where $D_{\rm L}$ is the cosmological luminosity distance. In Fig. 4, we show the observed flux for a $50$, $100$ and $200~{\rm M}_{\odot}$ Pop III star as a function of time. The emission peaks before breakout, when the density in the host halo is still high, reaching a maximum flux of $\simeq 10^{-23}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{-2}$. Once the star turns off, the H$\alpha$ emission drops quite rapidly over the course of a few $10~{\rm Myr}$, while the He [ii]{} $\lambda 1640$ emission drops almost instantaneously in the $100$ and $200~{\rm M}_{\odot}$ cases, due to the high recombination coefficient of He [iii]{} to He [ii]{}. It is noteworthy that the emission in the He [ii]{} $\lambda 1640$ line is generally not much lower than that in the H$\alpha$ line, which may be used as an indicator for massive Pop III stars [@bkl01; @oh01; @tgs01; @schaerer02].
For a $10\,\sigma$ detection with an exposure time of $100$ hours, the spectrograph on MIRI exhibits a typical limiting sensitivity of $\simeq 10^{-18}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{-2}$ [@panagia05], implying that the first H [ii]{} regions are typically five orders of magnitude too faint for a direct detection. We must therefore resort to indirect methods that rely on their cumulative signal. One such signature is the cosmic infrared background (CIB), where the redshifted Ly$\alpha$ recombination photons from $z\sim 10$ – $20$ might contribute at a detectable level [@sbk02; @kashlinsky05]. Minihaloes, however, are not expected to be important sources for the CIB, as opposed to more massive dark matter haloes that host the first galaxies [@gb06]. This leads us to consider the radio background as a key diagnostic of the Pop III minihalo formation site.
Radio background produced by bremsstrahlung
-------------------------------------------
The first H [ii]{} regions in their active as well as relic states also emit bremsstrahlung via thermal motions of electrons in an ionized medium. In line with our conlusions of the previous section, the signature from an individual source is much too faint to be detected. However, the cumulative radio signal might be strong enough to be detected by the upcoming SKA. We will here further explore this possibility [for a review of earlier work, see @fob06].
Solving the cosmological radiative transfer equation, it is straightforward to derive a simple expression for the observed radio background $J_{\nu}$ (in ${\rm erg}~{\rm s}^{-1}~{\rm cm}^{-2}~{\rm Hz}^{-1}~{\rm sr}^{-1}$): $$J_{\nu}=\int_{0}^{t_{H,0}}\frac{j_{\nu}}{(1+z)^{3}}c\,{\rm d}t\mbox{\ ,}$$ where $t_{H,0}$ is the present Hubble time and $j_{\nu}$ is the specific emissivity of bremsstrahlung, given by $$j_{\nu}=\epsilon_{\rm ff}\left<n_{e}^{2}\right>\left(T/10^{3}~{\rm K}\right)^{-1/2}\mbox{\ ,}$$ where $\epsilon_{\rm ff}\simeq 10^{-39}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{3}~{\rm Hz}^{-1}~{\rm sr}^{-1}$, $\left<n_{e}^{2}\right>$ is the volume-averaged electron density, and $T$ is the temperature [@rl79]. We universally assume $T=10^{3}~{\rm K}$, since the relic H [ii]{} region cools quite rapidly to $\sim 10^{3}~{\rm K}$ via inverse Compton losses and adiabatic expansion once the star has died [e.g. @greif07; @yoshida07]. Furthermore, we assume $j_{\nu}=0$ at $z<6$, since photoheating during reionization evaporates minihaloes [@dijkstra04]. This leads to: $$J_{\nu}=c\,\epsilon_{\rm ff}\int_{\infty}^{6}\frac{\left<n_{e}^{2}\right>}{\left(1+z\right)^{3}}\left|\frac{{\rm d}t}{{\rm d}z}\right|{\rm d}z\mbox{\ ,}$$ where we relate $\left<n_{e}^{2}\right>$ to the number density of minihaloes according to: $$\left<n_{e}^{2}\right> \simeq t_{\rm rec}\,n_{\rm H,b}^{2}\,V_{\rm HII}\left|\frac{{\rm d}N_{\rm ps}}{{\rm d}z}\right|\left|\frac{{\rm d}z}{{\rm d}t}\right|\mbox{\ .}$$ Here, $t_{\rm rec}=\left(\alpha_{\rm B}\,n_{\rm H,b}\right)^{-1}$ denotes the recombination time for hydrogen atoms, $\alpha_{\rm B}$ the case B recombination rate for $T=10^{3}~{\rm K}$, $n_{\rm H,b}$ the background density, $N_{\rm ps}$ the number of minihaloes per comoving volume, $V_{\rm HII}=N_{\rm ion}/n_{{\rm H,b},0}$ the comoving volume of an individual H [ii]{} region in its active as well as relic state, which is independent of redshift, and $N_{\rm ion}=\dot{N}_{\rm ion}\,t_{*}$ the total number of ionizing photons emitted per Pop III star (see Section 2). In the above equation, we have implicitly assumed that (relic) H [ii]{} regions survive for a recombination time, and that all ionizing photons escape into the IGM, which is a good approximation for massive Pop III stars in minihaloes [@alvarez06]. We note that in the range of redshifts considered here, the recombination time is larger than the stellar lifetime and smaller than the age of the Universe. In principle, one must also account for the clustering of minihaloes (biasing), which reduces the net volume filling factor of H [ii]{} regions [@mw96; @iliev03; @gao05; @reed05; @gb06]. However, it is extremely difficult to determine the importance of this effect, since (i) the actual overlap depends on the relative separation of minihaloes, and (ii) previous ionization allows a nearby H [ii]{} region to become larger than usual. We therefore neglect biasing, but keep in mind that the actual signal may be somewhat lower.
![The observed recombination flux in H$\alpha$ and He [ii]{} $\lambda 1640$ for a $50$, $100$ and $200~{\rm M}_{\odot}$ Pop III star, shown as a function of time after reaching the zero-age main-sequence (the black dots denote the end of their lifetimes). The emission peaks before breakout, when the density in the host halo is still high, reaching a maximum flux of $\simeq 10^{-23}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{-2}$. Once the star turns off, the H$\alpha$ emission drops quite rapidly over the course of a few $10~{\rm Myr}$, while the He [ii]{} $\lambda 1640$ emission drops almost instantaneously in the $100$ and $200~{\rm M}_{\odot}$ cases, due to the high recombination coefficient of He [iii]{} to He [ii]{}. The emission in the He [ii]{} $\lambda 1640$ line is generally not much lower than that in the H$\alpha$ line, which is characteristic for a top-heavy IMF and may be used as an indicator for massive Pop III stars. For a $10\,\sigma$ detection and an exposure time of $100$ hours, the limiting sensitivity of the MIRI spectrograph on [*JWST*]{} is approximately $10^{-18}~{\rm erg}~{\rm s}^{-1}~{\rm cm}^{-2}$, indicating that the first H [ii]{} regions are typically five orders of magnitude too faint to be detected directly.](f4.eps){width="8.5cm"}
In equation (23), the number density of minihaloes is given by $$N_{\rm ps}\left(z\right)=\int_{M_{\rm min}}^{M_{\rm max}}n_{\rm ps}\left(z,M\right){\rm d}M\mbox{\ ,}$$ where $n_{\rm ps}$ is the well-known Press-Schechter mass function [@ps74]. The minimum mass required for efficient cooling within a Hubble time may be found in @yoshida03a and @ts09: $$M_{\rm min}\simeq 10^{6}~{\rm M}_{\odot}\left(\frac{1+z}{10}\right)^{-2}\mbox{\ ,}$$ while the maximum mass is set by the requirement that cooling must be dominated by molecular hydrogen, i.e. the virial temperature must not exceed $T \simeq 10^{4}~{\rm K}$ for atomic hydrogen cooling, resulting in [e.g. @bl01] $$M_{\rm max}\simeq 2.5\times 10^{7}~{\rm M}_{\odot}\left(\frac{1+z}{10}\right)^{-3/2}\mbox{\ .}$$ We have found that our results are only marginally affected by the upper mass limit, but depend sensitively on the lower mass limit, since most minihaloes reside at the lower end of the halo distribution function.
After combining the above equations, we obtain $$J_{\nu}\simeq \frac{c\,\epsilon_{\rm ff}\,N_{\rm ion}}{\alpha_{\rm B}}N_{\rm ps}\left(z=6\right)\mbox{\ ,}$$ which, for an IMF consisting solely of $100~{\rm M}_{\odot}$ Pop III stars, yields $$J_{\nu}\simeq 300~{\rm mJy}~{\rm sr}^{-1}\mbox{\ .}$$ The brightness temperature, $T_{b}=c^{2}J_{\nu}/2k_{\rm B}\nu^{2}$, is given by $$T_{b}\simeq 1~{\rm mK}\left(\frac{\nu}{100~{\rm MHz}}\right)^{-2}\mbox{\ .}$$ In the following, we investigate whether a signal of this magnitude is observable by the upcoming SKA.
The sensitivity of radio instruments is generally defined by the ratio of the effective collecting area $A_{e}$ to the system temperature $T_{\rm sys}$. For the SKA with its proposed aperture array configuration at low frequencies, $A_{e}/T_{\rm sys}\simeq 5\times 10^{3}~{\rm m}^{2}~{\rm K}^{-1}$ at $100~{\rm MHz}$ [^2]. In this range, the system temperature is dominated by Galactic synchrotron emission, for which a useful approximation is given by $T_{\rm sky}\simeq 180~{\rm K}\left(\nu/180~{\rm MHz}\right)^{-2.6}$ [@fob06], resulting in $T_{\rm sys}\simeq 800~{\rm K}$ and $A_{e}\simeq 4\times 10^{6}~{\rm m}^{2}$. The minimum angular resolution for an array filling factor of unity at $100~{\rm MHz}$ is approximately $15~{\rm arcmin}$. At higher resolutions, the sensitivity decreases much too rapidly for effective imaging. In Fig. 5, we compare the sensitivity of the SKA for a $10\,\sigma$ detection, a bandwidth of $\Delta\nu_{\rm obs}=1~{\rm MHz}$, and an integration time of $1000~{\rm h}$ to the brightness temperature and specific flux expected for free-free emission. Although the figure implies that the free-free signal is detectable by the SKA, we have neglected biasing as well as radiative feedback in the form of a global LW background, which attenuates star formation in minihaloes [@jgb07; @jgb08]. Another complicating issue is the overlap with 21-cm emission, which makes it nearly impossible to isolate the contribution from bremsstrahlung. In consequence, we do not believe that this signal will be observable in the near future.
![The brightness temperature and specific flux of the radio background produced by bremsstrahlung, shown as a function of observed frequency. We have chosen a beam size of $15~{\rm arcmin}$ to achieve the highest possible resolution and sensitivity at $100~{\rm MHz}$ for the currently planned configuration of the SKA. The dot-dashed, dotted and solid lines correspond to an initial mass function consisting solely of $50$, $100$ and $200~{\rm M}_{\odot}$ Pop III stars, respectively. The dashed line shows the sensitivity of the SKA for a $10\,\sigma$ detection, a bandwidth of $1~{\rm MHz}$, and an integration time of $1000~{\rm h}$. Although the free-free signal is in principle detectable by the SKA, we have here neglected biasing and radiative feedback, which act to reduce the signal. For this reason we do not believe that the free-free signal of the first H [ii]{} regions in their active or relic states will be observable in the near future.](f5.eps){width="8.5cm"}
Radio background produced by 21-cm emission
-------------------------------------------
Perhaps the most promising observational signature comes from 21-cm emission of the relic H [ii]{} region gas once the star has died, a prospect that was already investigated by @tokutani09. An emission signal requires the spin temperature $T_{S}$ of neutral hydrogen to be greater than the temperature of the CMB, with its relative brightness determined by $T_{S}$ and the size of the relic H [ii]{} region. The spin temperature is set by collisional coupling with neutral hydrogen atoms, protons and electrons, as well as radiative coupling to the CMB. Furthermore, it may be modified by the so-called Wouthuysen-Field effect, which describes the mixing of spin states due to the absorption and re-emission of Ly$\alpha$ photons [@wouthuysen52; @field59]. The color temperature of the Ly$\alpha$ background is determined by the ratio of excitations to de-excitations, which approaches the kinetic gas temperature at high redshifts, where the optical depth to Ly$\alpha$ scattering is very large [@fob06]. In this case, adopting the Rayleigh-Jeans approximation and assuming $T_{S}\gg T_{*}$, where $T_{*}=h\nu_{21}/k_{\rm B}=68~{\rm mK}$ is the temperature associated with the 21-cm transition, the spin temperature may be written as [@mmr97] $$T_{S}=\frac{T_{\gamma}+\left(y_{c}+y_{\alpha}\right)\,T}{1+y_{c}+y_{\alpha}}\mbox{\ ,}$$ where $T_{\gamma}$ is the temperature of the CMB. The collisional coupling coefficient $y_{c}$ is approximately given by $$y_{c}=\frac{T_{*}}{A_{21}T}\left(n_{\rm HI}\kappa_{\rm HI}+n_{e}\kappa_{e}\right)\mbox{\ ,}$$ where $A_{21}=2.85\times 10^{-15}~{\rm s}^{-1}$ is the Einstein A-coefficient for the 21-cm transition, and $\kappa_{\rm HI}$ and $\kappa_{e}$ are the effective single-atom rate coefficients for collisions with neutral hydrogen atoms and electrons, respectively. Good functional fits in the temperature range $100~{\rm K}\la T\la 10^{4}~{\rm K}$ are given by $$\kappa_{\rm HI}=10^{-11}\,T^{1/2}~{\rm cm}^{3}~{\rm s}^{-1}$$ and $$\kappa_{e}=2\times 10^{-10}\,T^{1/2}~{\rm cm}^{3}~{\rm s}^{-1}\mbox{\ ,}$$ which we have obtained from the rates quoted in @kmm06. At $z\la 20$, the electron fraction in the IGM remains above $f_{e}=0.1$ for most of the lifetime of the relic H [ii]{} region. In this case, the collisional coupling coefficient is given by $$y_{c}\simeq 0.015\,\left(\frac{f_{e}}{0.5}\right)\,\left(\frac{T}{10^{3}~{\rm K}}\right)^{-1/2}\,\left(\frac{1+z}{10}\right)^{3}\mbox{\ .}$$ A derivation of the Ly$\alpha$ coupling coefficient $y_{\alpha}$ requires radiative transfer of local as well as global Ly$\alpha$ radiation, which is beyond the scope of this work. We therefore consider two limiting cases: one in which we only consider collisional coupling, and the other in which a strong Ly$\alpha$ background drives the spin temperature towards the gas temperature (i.e. $y_{\alpha}\gg 1$ or $T_{S}=T$).
The differential brightness temperature with respect to the CMB may then be derived as follows. In the Rayleigh-Jeans limit and for $T_{S}\gg T_{*}$, the monochromatic radiative transfer equation for a ray passing through a cloud, evaluated in its comoving frame, may be written in terms of the brightness temperature $T_{b}$: $$T_{b}=T_{\gamma}\,e^{-\tau}+\int_{0}^{\tau}T_{S}\,e^{-\tau'}{\rm d}\tau'\mbox{\ ,}$$ where the optical depth at the 21-cm line is given by $${\rm d}\tau=\frac{3c^{2}A_{21}n_{\rm HI}}{32\pi\nu_{21}^{2}}\,\phi(\nu_{21})\,\frac{T_{*}}{T_{S}}\,{\rm d}s\mbox{\ .}$$ Here, $\phi(\nu_{21})$ is the normalized line profile at the resonance frequency $\nu_{21}$ and ${\rm d}s$ is the distance traveled by the ray. In our case, the line profile is dominated by thermal broadening, with a Doppler width given by $$\Delta\nu_{\rm D}=\nu_{21}\sqrt{\frac{2k_{\rm B}T}{\mu m_{\rm H}c^{2}}}\mbox{\ .}$$ The amplitude of the line profile at the resonance frequency may be replaced by the Doppler width, i.e. $\phi(\nu_{21})=\Delta\nu_{\rm D}^{-1}$. With this definition, equation (35) yields the differential brightness temperature $\delta T_{b}=T_{b}-T_{\gamma}$, which becomes particularly simple for a constant spin temperature and the fact that the relic H [ii]{} regions considered here are optically thin: $$\delta T_{b}=(T_{S}-T_{\gamma})\,\tau\mbox{\ .}$$ The observed differential brightness temperature is then simply given by $\delta T_{{\rm b,obs}}=\delta T_{b}/\left(1+z\right)$.
In Fig. 6, we show the observed differential brightness temperature for a $100~{\rm M}_{\odot}$ star and the two limiting cases discussed above. Note that we have only taken into account ionized gas along the line of sight. For collisional coupling, the observed differential brightness temperature is of order a few $10~{\rm mK}$ for $\simeq 100~{\rm Myr}$, while for perfect coupling the signal is elevated by an order of magnitude to a few $100~{\rm mK}$ for well over $\simeq 100~{\rm Myr}$. In reality, the expected signal lies between these extremes and is a function of redshift, since collisional coupling becomes weaker as the background density drops, while Ly$\alpha$ coupling becomes stronger as the Ly$\alpha$ background rises. At $z\la 20$, where the observationally accessible signal is produced, the Ly$\alpha$ background is likely strong enough for the latter to be more important [@furlanetto06; @pf07].
Next, we determine the radio background produced by the integrated 21-cm emission of relic H [ii]{} regions. The differential specific flux observed at the redshifted 21-cm line from a single relic H [ii]{} region with differential brightness temperature $\delta T_{b}$ is given by $$\delta F_{\nu}=\frac{2k_{\rm B}\nu_{21}^{2}}{c^{2}}\,\left(1+z\right)^{-3}\Delta\Omega\,\delta T_{b}\mbox{\ ,}$$ where $\Delta\Omega=A/D_{\rm A}^{2}$ denotes the solid angle subtended by the relic H [ii]{} region, $A=\pi r_{\rm HII}^{2}$ its area, $r_{\rm HII}=\left(3N_{\rm ion}/4\pi n_{\rm H,b}\right)^{1/3}$ its radius, and $D_{\rm A}$ the angular diameter distance. The average differential specific flux $\left<\delta F_{\nu}\right>$ within a beam size $\Delta\Omega_{\rm beam}$ and bandwidth $\Delta\nu_{\rm obs}$ is then given by $$\left<\delta F_{\nu}\right>=\delta F\,N_{\rm ps}(z)\,\frac{{\rm d}^{2}V(z)}{{\rm d}z\,{\rm d}\Omega}\,\frac{\Delta z\,\Delta\Omega_{\rm beam}}{\Delta\nu_{\rm obs}}\mbox{\ ,}$$ where $\delta F=\delta F_{\nu}\,\Delta\nu_{\rm D}/\left(1+z\right)$, $N_{\rm ps}(z)$ is the Press-Schechter mass function defined in equation (24), $\Delta z=\Delta\nu_{\rm obs}\left(1+z\right)^{2}/\nu_{21}$, and ${\rm d}^{2}V(z)/{\rm d}z\,{\rm d}\Omega$ is the comoving volume per unit redshift and solid angle: $$\frac{{\rm d}^{2}V(z)}{{\rm d}z\,{\rm d}\Omega}=\frac{c\,D_{\rm A}^{2}\left(1+z\right)^{2}}{H(z)}\mbox{\ ,}$$ where $H(z)$ is the Hubble expansion rate. With the definition of the brightness temperature, the average differential antenna temperature $\left<\delta T_{b}\right>$ is given by $$\left<\delta T_{b}\right>=\frac{\pi c}{\nu_{21}}\,\frac{\left(1+z\right)^{2}N_{\rm ps}(z)}{H(z)}\,\Delta\nu_{\rm D}\,r_{\rm HII}^{2}(z)\,\delta T_{b}(z)\mbox{\ .}$$ Based on our argument above, we assume that the Ly$\alpha$ background is strong enough for perfect coupling at all redshifts. In this case, and for $T\gg T_{\gamma}$, the average differential antenna temperature becomes independent of electron fraction and temperature: $$\left<\delta T_{b}\right>=\frac{9c^{3}A_{21}T_{*}N_{\rm ion}}{128\pi \nu_{21}^{3}H_{0}\sqrt{\Omega_{m}}}\,\left(1+z\right)^{1/2}N_{\rm ps}(z)\mbox{\ ,}$$ where we have set $n_{\rm HI}=n_{\rm H,b}$ in equation (36). We note that the observed frequency is related to the redshift via $\nu_{\rm obs}=\nu_{21}/\left(1+z\right)$. We have further assumed that the relic H [ii]{} region produced by each star-forming minihalo persists until the Universe is reionized (i.e. $z\simeq 6$), which is a good approximation for perfect coupling and $T\gg T_{\gamma}$. Equation (43) thus provides a robust upper limit for the collective 21-cm emission from the first relic H [ii]{} regions.
In Fig. 7, we compare the average differential antenna temperature and specific flux for a beam size of $\Delta\theta_{\rm beam}=15'$ to the sensitivity of the SKA, assuming a $10\,\sigma$ detection, a bandwidth of $\Delta\nu_{\rm obs}=1~{\rm MHz}$, and an integration time of $1000~{\rm h}$. At all frequencies, the maximum 21-cm signal from the first relic H [ii]{} regions is of order $10~{\rm mK}$, which is well detectable by the SKA. The effects of biasing and radiative feedback will reduce this signal, but probably not enough to fall below the sensitivity of the SKA. Compared to free-free emission, the 21-cm signal is typically an order of magnitude stronger, and offers the best prospect for indirectly probing the first stars. Furthermore, the 21-cm signal is explicitly frequency-dependent, while this is not the case for bremsstrahlung, where a flat spectrum is produced (see equation 28). This dependency might allow for a better distinction from other sources of radio emission at these wavelengths.
{width="8.5cm"}
Summary and conclusions
=======================
We have introduced a general-purpose radiative transfer scheme for cosmological SPH simulations that treats ionizing and photodissociating radiation from massive Pop III stars in the early Universe. Based on this methodology, we have investigated the build-up of the first H [ii]{} regions and relic H [ii]{} regions around Pop III stars formed in minihaloes, and predicted their contribution to the extragalactic radio background via bremsstrahlung and 21-cm emission. Although recombination radiation from individual H [ii]{} regions in their active as well as relic states is too faint to be directly detectable even with [*JWST*]{}, their collective radio emission might be strong enough to be within reach of the planned SKA. In particular, we have found that the integrated free-free emission results in a maximum differential antenna temperature of $\simeq 1~{\rm mK}$, while the 21-cm emission is an order of magnitude stronger. Considering the effects of biasing and negative radiative feedback, which would act to reduce the predicted signal, the free-free signal is likely beyond the capability of the SKA, while the 21-cm signal will most likely be observable, providing an excellent opportunity for indirectly probing the first stars.
We note that an analysis of the angular fluctuation power spectrum will be essential to isolate the 21-cm signal from other backgrounds [@fo06], although the frequency-dependence of the 21-cm signal might already prove useful. Among these are neutral minihaloes, which appear in emission due to their enhanced density and temperature [@iliev02], or IGM gas heated by X-rays from supernovae [@oh01], X-ray binaries [@gb03], or the first quasars [@madau04b; @kmm06]. A strong absorption signal might originate from cold, neutral gas if the Ly$\alpha$ background effectively couples the spin temperature to the gas temperature [@pf07]. In addition, there is the signal produced by stars (primordial or already metal-enriched) formed in the first dwarf galaxies [e.g. @nb08]. All of these compete with each other, and more work is required to understand their relative importance. One important task is to extend the simulations to larger cosmological volumes, to measure the aggregate signal from many sources in a more robust way.
Minihaloes may not have been the dominant formation sites for primordial stars, in terms of producing the bulk of the radiation that drove reionization, or of being the source for the majority of the heavy elements present at high redshifts [@gb06; @sbk08]. Nevertheless, they are the ideal laboratory to test our current standard model of the first stars, by providing an exceedingly simple environment for the star formation process [@bromm09]. The next step in the hierachical build-up of structure is already highly complex, due to the presence of metals, turbulent velocity fields, and possibly dynamically significant magnetic fields [@wa07b; @wa08b; @greif08; @schleicher09]. It is therefore crucial to empirically probe the minihalo environment, and the signature left in the radio background might provide us with one of the few avenues to accomplish this in the foreseeable future.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank the referee Naoki Yoshida for his valuable comments and suggestions that greatly improved this work. TG thanks Matthias Bartelmann and Simon Glover for many stimulating discussions. TG acknowledges financial support by the Heidelberg Graduate School of Fundamental Physics (HGSFP). The HGSFP is funded by the Excellence Initiative of the German government (grant number GSC 129/1). RSK acknowledge subsidies from the [*Deutsche Forschungsgemeinschaft*]{} via the priority program SFB 439 “Galaxies in the Early Universe” as well as via grants KL1358/1, KL1358/4 and KL1358/5. In addition, RSK also acknowledges partial support from a Frontier grant of Heidelberg University funded by the German Excellence Initiative. VB acknowledges support from NSF grant AST-0708795 and NASA ATFP grant NNX08AL43G. The simulations presented here were carried out at the Texas Advanced Computing Center (TACC).
T., [Bryan]{} G. L., [Norman]{} M. L., 2002, Sci, 295, 93
T., [Wise]{} J. H., [Bryan]{} G. L., 2007, [ApJ]{}, 659, L87
M. A., [Bromm]{} V., [Shapiro]{} P. R., 2006, [ApJ]{}, 639, 621
M. A., [Shapiro]{} P. R., [Ahn]{} K., [Iliev]{} I. T., 2006, [ApJ]{}, 644, L101
R., [Loeb]{} A., 2001, [Phys. Rep.]{}, 349, 125
V., [Coppi]{} P. S., [Larson]{} R. B., 2002, [ApJ]{}, 564, 23
V., [Kudritzki]{} R. P., [Loeb]{} A., 2001, [ApJ]{}, 552, 464
V., [Larson]{} R. B., 2004, [ARA&A]{}, 42, 79
V., [Loeb]{} A., 2003, [Nat]{}, 425, 812
V., [Loeb]{} A., 2004, New Astron., 9, 353
V., [Yoshida]{} N., [Hernquist]{} L., 2003, [ApJ]{}, 596, L135
V., [Yoshida]{} N., [Hernquist]{} L., [McKee]{} C. F., 2009, [Nat]{}, 459, 49
B., [Ferrara]{} A., 2005, Space Sci. Rev., 116, 625
P. C., [Glover]{} S. C. O., [Klessen]{} R. S., 2008, [ApJ]{}, 672, 757
M., [Haiman]{} Z., [Rees]{} M. J., [Weinberg]{} D. H., 2004, [ApJ]{}, 601, 666
B. T., [Bertoldi]{} F., 1996, [ApJ]{}, 468, 269
X., [Carilli]{} C. L., [Keating]{} B., 2006, [ARA&A]{}, 44, 415
G. B., 1959, [ApJ]{}, 129, 536
S. R., 2006, [MNRAS]{}, 371, 867
S. R., [Oh]{} S. P., 2006, [ApJ]{}, 652, 849
S. R., [Oh]{} S. P., [Briggs]{} F. H., 2006, [Phys. Rep.]{}, 433, 181
L., [White]{} S. D. M., [Jenkins]{} A., [Frenk]{} C. S., [Springel]{} V., 2005, [MNRAS]{}, 363, 379
J. P., et al., 2006, Space Sci. Rev., 123, 485
S. C. O., [Abel]{} T., 2008, [MNRAS]{}, 388, 1627
S. C. O., [Brand]{} P. W. J. L., 2001, [MNRAS]{}, 321, 385
S. C. O., [Brand]{} P. W. J. L., 2003, [MNRAS]{}, 340, 210
S. C. O., [Jappsen]{} A.-K., 2007, [ApJ]{}, 666, 1
K. M., [Hivon]{} E., [Banday]{} A. J., [Wandelt]{} B. D., [Hansen]{} F. K., [Reinecke]{} M., [Bartelmann]{} M., 2005, [ApJ]{}, 622, 759
T. H., [Bromm]{} V., 2006, [MNRAS]{}, 373, 128
T. H., [Johnson]{} J. L., [Bromm]{} V., [Klessen]{} R. S., 2007, [ApJ]{}, 670, 1
T. H., [Johnson]{} J. L., [Klessen]{} R. S., [Bromm]{} V., 2008, [MNRAS]{}, 387, 1021
Z., [Abel]{} T., [Rees]{} M. J., 2000, [ApJ]{}, 534, 11
Z., [Loeb]{} A., 1997, [ApJ]{}, 483, 21
A., [Woosley]{} S. E., 2002, [ApJ]{}, 567, 532
I. T., [Scannapieco]{} E., [Martel]{} H., [Shapiro]{} P. R., 2003, [MNRAS]{}, 341, 81
I. T., [Shapiro]{} P. R., [Ferrara]{} A., [Martel]{} H., 2002, [ApJ]{}, 572, L123
A.-K., [Glover]{} S. C. O., [Klessen]{} R. S., [Mac Low]{} M.-M., 2007, [ApJ]{}, 660, 1332
A.-K., [Klessen]{} R. S., [Glover]{} S. C. O., [MacLow]{} M.-M., 2009, [ApJ]{}, 696, 1065
J. L., [Greif]{} T. H., [Bromm]{} V., 2007, [ApJ]{}, 665, 85
J. L., [Greif]{} T. H., [Bromm]{} V., 2008, [MNRAS]{}, 388, 26
J. L., [Greif]{} T. H., [Bromm]{} V., [Klessen]{} R. S., [Ippolito]{} J., 2009, [MNRAS]{}, in press (arXiv:0902.3263)
A., [Arendt]{} R. G., [Mather]{} J., [Moseley]{} S. H., 2005, [Nat]{}, 438, 45
T., [Yoshida]{} N., [Susa]{} H., [Umemura]{} M., 2004, [ApJ]{}, 613, 631
A., et al., 2006, New Astron. Rev., 50, 925
E., et al., 2009, [ApJS]{}, 180, 330
M., [Madau]{} P., [Montgomery]{} R., 2006, [ApJ]{}, 637, L1
J., 2008, in [Minchin]{} R., [Momjian]{} E., eds, AIP Conf. Ser. Vol. 1035, The Evolution of Galaxies Through the Neutral Hydrogen Window. Am. Inst. Phys., New York, p. 303
A., [Rybicki]{} G. B., 1999, [ApJ]{}, 524, 527
P., [Meiksin]{} A., [Rees]{} M. J., 1997, [ApJ]{}, 475, 429
P., [Rees]{} M. J., [Volonteri]{} M., [Haardt]{} F., [Oh]{} S. P., 2004, [ApJ]{}, 604, 484
P., [Girardi]{} L., [Chiosi]{} C., [Wood]{} P. R., 2001, [A&A]{}, 371, 152
C. F., [Tan]{} J. C., 2008, [ApJ]{}, 681, 771
H. J., [White]{} S. D. M., 1996, [MNRAS]{}, 282, 347
S., [Barkana]{} R., 2008, [MNRAS]{}, 385, L63
S. P., 2001, [ApJ]{}, 553, 499
S. P., [Haiman]{} Z., 2002, [ApJ]{}, 569, 558
K., [Tsuribe]{} T., [Schneider]{} R., [Ferrara]{} A., 2005, [ApJ]{}, 626, 627
D., [Ferland]{} G., 2006, Astrophysics of Gaseous Nebulae and Active Galactic Nuclei. University Science Books, Sausalito
N., 2005, in [Corbelli]{} E., [Palla]{} F., [Zinnecker]{} H., eds, The Initial Mass Function 50 Years. Springer, Berlin, p. 479
W. H., [Schechter]{} P., 1974, [ApJ]{}, 187, 425
J. R., [Furlanetto]{} S. R., 2007, [MNRAS]{}, 376, 1680
D. S., [Bower]{} R., [Frenk]{} C. S., [Gao]{} L., [Jenkins]{} A., [Theuns]{} T., [White]{} S. D. M., 2005, [MNRAS]{}, 363, 393
G. B., [Lightman]{} A. P., 1979, Radiative Processes in Astrophysics. Wiley-Interscience, New York
M. R., [Bromm]{} V., [Kamionkowski]{} M., 2002, [MNRAS]{}, 336, 1082
E., [Madau]{} P., [Woosley]{} S., [Heger]{} A., [Ferrara]{} A., 2005, [ApJ]{}, 633, 1031
D., 2002, [A&A]{}, 382, 28
D., 2003, [A&A]{}, 397, 527
D. R. G., [Banerjee]{} R., [Klessen]{} R. S., 2008, [Phys. Rev. D]{}, 78, 083005
D. R. G., [Galli]{} D., [Glover]{} S. C. O., [Banerjee]{} R., [Palla]{} F., [Schneider]{} R., [Klessen]{} R. S., 2009, submitted (arXiv:0904.3970)
M., et al., 2009, submitted (arXiv:0901.0559)
F. H., [Lizano]{} S., [Galli]{} D., [Cant[ó]{}]{} J., [Laughlin]{} G., 2002, [ApJ]{}, 580, 969
V., 2005, [MNRAS]{}, 364, 1105
J. C., [McKee]{} C. F., 2004, [ApJ]{}, 603, 383
M., [Yoshida]{} N., [Oh]{} S. P., [Sugiyama]{} N., 2009, [MNRAS]{}, 395, 777
L., [Ferrara]{} A., [Schneider]{} R., 2007, [MNRAS]{}, 382, 945
M., [Stiavelli]{} M., 2009, [ApJ]{}, 694, 879
J., [Giroux]{} M. L., [Shull]{} J. M., 2001, [ApJ]{}, 550, L1
G. A., [Leitherer]{} C., [Schaerer]{} D., [Meynet]{} G., [Maeder]{} A., 2007, [ApJ]{}, 663, 995
D., [Abel]{} T., [Norman]{} M. L., 2004, [ApJ]{}, 610, 14
J. H., [Abel]{} T., 2007, [ApJ]{}, 665, 899
J. H., [Abel]{} T., 2008, [ApJ]{}, 685, 40
S. E., [Heger]{} A., 2006, [ApJ]{}, 637, 914
S. A., 1952, [AJ]{}, 57, 31
S.-C., [Langer]{} N., 2005, [A&A]{}, 443, 643
N., [Abel]{} T., [Hernquist]{} L., [Sugiyama]{} N., 2003, [ApJ]{}, 592, 645
N., [Oh]{} S. P., [Kitayama]{} T., [Hernquist]{} L., 2007, [ApJ]{}, 663, 687
N., [Omukai]{} K., [Hernquist]{} L., 2008, Sci, 321, 669
[^1]: E-mail: [email protected]
[^2]: http://www.skatelescope.org
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.